# Institute for Automated Research — Wiki (full corpus)
# Source: https://instituteforautomatedresearch.org/wiki  ·  Generated at build.
# License: open for reuse with attribution to the Institute for Automated Research.



==============================================================================
# A knowledge base for autonomous research
# https://instituteforautomatedresearch.org/wiki/
# Open knowledge base on autonomous research methodology: the data, tools, and protocols behind autonomous, adversarially-verified research.
==============================================================================

A public, machine-readable knowledge base from the
[Institute for Automated Research](https://instituteforautomatedresearch.org).

It documents the practical substrate of autonomous research: the datasets a
pipeline can actually reach, the distilled findings of the literature it
builds on, and honest, recorded provenance for all of it. Every
page is plain Markdown in
[a public Git repository](https://github.com/institute-for-automated-research/website),
served as both human pages and raw `.md`, and explicitly open to LLM crawlers.

## Start here

- **[Distilled literature](/wiki/papers/)**: papers reduced to their core
  results, datasets used, and theory tested, with source locators and honest
  provenance; read the full paper to replicate or extend it.
  Openly-licensed sources are also mirrored, machine-accessible, in the
  [Open Library](/library).
- **[Openly available datasets](/wiki/datasets/)**: free, downloadable data
  sources for finance and economics research, with working access recipes and
  gotchas, distilled from what the ZeroPaper pipeline actually runs.
- **[Commercial datasets](/wiki/commercial/)**: the paywalled core
  (WRDS/CRSP/Compustat) and the vendor feeds you can buy, plus what the open
  sources can and cannot substitute.
- **[Confidential datasets](/wiki/confidential/)**: supervisory and
  administrative microdata that is not purchasable at all, documented with its
  access conditions and gotchas.
- **[Browse by tag](/wiki/tags/)**: every page cross-indexed by topic,
  method, access, data shape, source, and status.

## Contributing

Found an error or want a topic covered? Use the **Edit** link on any page, open
an issue, or email `contact@instituteforautomatedresearch.org`. Content is
reviewed before publishing; provenance and accuracy are the point.


==============================================================================
# Browse by tag
# https://instituteforautomatedresearch.org/wiki/tags/
# Every wiki page grouped by tag: by topic (asset pricing, macro, fundamentals), access (free, no-API-key, licensed), data shape (panel, time-series), source, dataset, and verification status. The fast way to find the dataset you need.
==============================================================================

import TagIndex from '../../components/TagIndex.astro';

Pages are tagged on several axes: **topic**, **access**, **data shape**,
**source**, **dataset** (`data:<slug>`, one per dataset page), and
**verification** (how far the access recipe was actually run: fetched,
reachable, licensed, or unverified). Pick a tag to jump to everything under it.
A dataset tag marked "no wiki page yet" is one a paper cites but we have not
documented; it doubles as our to-write list. Tip: in-page search
(<kbd>Cmd/Ctrl</kbd> + <kbd>K</kbd>, top of any wiki page) searches the full
text of every page below.

<TagIndex />


==============================================================================
# Commercial datasets for finance & economics research
# https://instituteforautomatedresearch.org/wiki/commercial/
# Commercial data sources for finance and economics research: purchasable from a vendor or reachable through a licence such as WRDS. Access path and the gotchas that bite automated pipelines, with no provenance badge since they were not exercised here.
==============================================================================

Not free. Commercial data is obtainable by anyone who pays a vendor or holds a
licence (including WRDS-gated feeds), which is what separates it from the
[openly available set](/wiki/datasets/) and from the
[confidential tier](/wiki/confidential/), which cannot be purchased at all.

**WRDS / CRSP / Compustat / IBES / OptionMetrics** are subscription-gated, but
most universities license WRDS for affiliated researchers. See
[WRDS / CRSP / Compustat](/wiki/commercial/wrds/) for the academic-access path
and exactly which [openly available sources](/wiki/datasets/) substitute for
which paid ones.

These pages document the access path and the gotchas, but carry **no provenance
badge**: they were not exercised here (no vendor credentials), so they read as
unverified per the Verified discipline.

| Dataset | What it is | Page |
|---|---|---|
| **WRDS / CRSP / Compustat** | The paywalled core; academic access | [WRDS](/wiki/commercial/wrds/) |
| **RavenPack** | News and event analytics (sentiment, relevance) | [RavenPack](/wiki/commercial/ravenpack/) |
| **FactSet LionShares** | Global institutional and fund holdings | [FactSet LionShares](/wiki/commercial/factset-lionshares/) |
| **Revelio Labs** | Workforce / human-capital panel | [Revelio Labs](/wiki/commercial/revelio/) |
| **I/B/E/S** | Sell-side analyst estimates, consensus and actuals | [I/B/E/S](/wiki/commercial/ibes/) |
| **SDC Platinum** | M&A and new-issues deal-level data | [SDC Platinum](/wiki/commercial/sdc-platinum/) |
| **CoreLogic** | US property, deeds, repeat-sales, foreclosure | [CoreLogic](/wiki/commercial/corelogic/) |
| **DealScan** | Syndicated and large corporate loan deals | [DealScan](/wiki/commercial/dealscan/) |
| **PitchBook** | Venture, private-equity, and M&A deal data | [PitchBook](/wiki/commercial/pitchbook/) |
| **Preqin** | Private-capital and hedge-fund data | [Preqin](/wiki/commercial/preqin/) |
| **RepRisk** | ESG risk-incident feed (28 issue categories) | [RepRisk](/wiki/commercial/reprisk/) |
| **Trucost** | Firm-level carbon and environmental data | [Trucost](/wiki/commercial/trucost/) |
| **TRACE** | Corporate bond secondary-market transactions (Enhanced, via WRDS) | [TRACE](/wiki/commercial/trace/) |
| **Thomson 13F (s34)** | Institutional equity holdings from 13F filings | [Thomson 13F (s34)](/wiki/commercial/thomson-13f/) |
| **CRSP Mutual Funds** | Survivor-bias-free fund returns, fees, flows, holdings (+ Thomson s12, MFLINKS) | [CRSP Mutual Funds](/wiki/commercial/crsp-mutual-funds/) |
| **TAQ** | NYSE intraday trade and quote microstructure data | [TAQ](/wiki/commercial/taq/) |
| **Orbis (BvD)** | Global public and private firm financials and ownership | [Orbis (BvD)](/wiki/commercial/orbis-bvd/) |
| **Markit bond pricing** | Composite dealer quotes for individual bonds | [Markit bond pricing](/wiki/commercial/markit/) |
| **Markit CDS** | Single-name credit default swap spreads | [Markit CDS](/wiki/commercial/markit-cds/) |
| **CoStar** | Commercial real estate transactions and characteristics | [CoStar](/wiki/commercial/costar/) |
| **Bloomberg** | Cross-asset terminal market data | [Bloomberg](/wiki/commercial/bloomberg/) |
| **NETS** | Establishment-level employment and location panel | [NETS](/wiki/commercial/nets/) |
| **Crane Data** | Money market fund holdings and assets | [Crane Data](/wiki/commercial/crane-mmf/) |
| **OptionMetrics IvyDB** | Daily option prices, implied vols, and Greeks | [OptionMetrics](/wiki/commercial/optionmetrics/) |
| **Compustat Global** | Non-US company fundamentals | [Compustat Global](/wiki/commercial/compustat-global/) |
| **Compustat Segments** | Business- and geographic-segment financials | [Compustat Segments](/wiki/commercial/compustat-segments/) |
| **Datastream** | Global multi-asset and macro time series | [Datastream](/wiki/commercial/datastream/) |
| **Worldscope** | Global company fundamentals | [Worldscope](/wiki/commercial/worldscope/) |
| **Capital IQ** | Company, capital-structure, and transactions data | [Capital IQ](/wiki/commercial/capital-iq/) |
| **Audit Analytics** | Auditor, disclosure, and restatement data | [Audit Analytics](/wiki/commercial/audit-analytics/) |
| **NielsenIQ** | Retail scanner and Homescan consumer-panel data (via Kilts) | [NielsenIQ](/wiki/commercial/nielseniq/) |
| **Morningstar** | Fund, ETF, and fund-sustainability data | [Morningstar](/wiki/commercial/morningstar/) |
| **KLD / MSCI ESG** | Firm-level ESG strength/concern and ratings | [KLD / MSCI ESG](/wiki/commercial/kld/) |
| **Cboe options** | Index/equity option quotes, trades, and VIX | [Cboe options](/wiki/commercial/cboe-options/) |
| **FactSet Revere** | Inter-firm supplier/customer/competitor links | [FactSet Revere](/wiki/commercial/factset-revere/) |
| **Global Financial Data** | Long-run cross-country stock, bond, and macro series | [Global Financial Data](/wiki/commercial/global-financial-data/) |
| **Crunchbase** | Startup characteristics, funding rounds, investors | [Crunchbase](/wiki/commercial/crunchbase/) |
| **CSMAR** | Chinese listed-firm prices, financials, ownership | [CSMAR](/wiki/commercial/csmar/) |
| **MSCI Real Estate** | Property total-return indices and rental yields (IPD) | [MSCI Real Estate](/wiki/commercial/msci-real-estate/) |
| **Refinitiv transcripts** | Earnings-call transcripts for textual analysis | [Refinitiv transcripts](/wiki/commercial/refinitiv-transcripts/) |
| **Markit Securities Finance** | Stock borrow fees, utilization, lendable supply | [Markit Securities Finance](/wiki/commercial/markit-securities-finance/) |
| **Markit quanto** | Cross-currency quanto quotes (FX-equity covariance) | [Markit quanto](/wiki/commercial/markit-quanto/) |
| **Blue Chip Forecasts** | Professional forecaster rate/GDP survey | [Blue Chip Forecasts](/wiki/commercial/blue-chip-forecasts/) |
| **Consensus Economics** | Cross-country macro & FX forecast surveys | [Consensus Economics](/wiki/commercial/consensus-economics/) |
| **Cerved** | Italian company financials (incl. private firms) | [Cerved](/wiki/commercial/cerved/) |
| **Clarity Services** | Subprime/alternative-credit bureau (payday, installment) | [Clarity Services](/wiki/commercial/clarity-services/) |
| **Equifax** | Traditional consumer credit-bureau records | [Equifax](/wiki/commercial/equifax/) |
| **InfoUSA / Data Axle** | Establishment listings: location, industry, employment | [InfoUSA / Data Axle](/wiki/commercial/infousa/) |
| **Lipper eMAXX** | CUSIP-level bond holdings by insurer/fund/ETF | [Lipper eMAXX](/wiki/commercial/emaxx/) |
| **Eurodollar futures** | Intraday CME prices for policy-surprise windows | [Eurodollar futures](/wiki/commercial/eurodollar-futures/) |
| **FTSE All-Share** | UK equity index membership, market cap, returns | [FTSE All-Share](/wiki/commercial/ftse-all-share/) |
| **Siblis Research** | S&P/Nasdaq index addition-deletion dates and values | [Siblis Research](/wiki/commercial/siblis-research/) |
| **Optimal Blue** | Mortgage rate-locks and real-time lender offers | [Optimal Blue](/wiki/commercial/optimal-blue/) |
| **RateWatch** | Branch-level posted deposit and CD rates (weekly) | [RateWatch](/wiki/commercial/ratewatch/) |
| **BvD Zephyr** | Global M&A, IPO, and PE/VC deals, linked to Orbis | [BvD Zephyr](/wiki/commercial/bureau-van-dijk-zephyr/) |
| **VentureSource** | VC funds, financing rounds, startup locations | [VentureSource](/wiki/commercial/venturesource/) |
| **NYT archive** | Full-text New York Times corpus (text-as-data) | [NYT archive](/wiki/commercial/nyt-news/) |
| **LexisNexis court** | U.S. court filings and public records | [LexisNexis court](/wiki/commercial/lexisnexis-court/) |
| **Ancestry records** | Death indexes and genealogical records | [Ancestry records](/wiki/commercial/ancestry-death-records/) |
| **Getty CEO photos** | Dated executive press photos for facial measures | [Getty CEO photos](/wiki/commercial/gettyimages-ceo-photos/) |
| **AHA Annual Survey** | U.S. hospital services, beds, operations, system affiliation | [AHA Annual Survey](/wiki/commercial/aha-annual-survey/) |
| **Moody's URD** | Creditor recovery rates at resolution of corporate defaults | [Moody's URD](/wiki/commercial/moodys-urd/) |
| **Rystad Energy** | Asset-level global oil and gas production, costs, reserves | [Rystad Energy](/wiki/commercial/rystad/) |
| **SteelBenchmarker** | Biweekly reference prices for steel products | [SteelBenchmarker](/wiki/commercial/steelbenchmarker/) |
| **StockTwits** | Ticker-tagged retail social-media messages and sentiment | [StockTwits](/wiki/commercial/stocktwits/) |
| **Titlon (OSE)** | Oslo Stock Exchange prices, returns, accounting (Nordic academic) | [Titlon (OSE)](/wiki/commercial/titlon-ose/) |
| **ZTRAX** | Zillow property deeds and assessor records (discontinued 2023) | [ZTRAX](/wiki/commercial/ztrax/) |
</content>
</invoke>


==============================================================================
# Confidential datasets in finance & economics research
# https://instituteforautomatedresearch.org/wiki/confidential/
# Confidential supervisory, administrative, and single-counterparty microdata cited in the papers we distill: what the collection is and the gotchas, with no provenance badge since there is no access path that can be run here.
==============================================================================

Neither [free](/wiki/datasets/) nor [commercially licensed](/wiki/commercial/).
Confidential supervisory, administrative-registry, individual-level, or
single private-counterparty microdata, reachable only through a qualifying
affiliation or an approved restricted-data arrangement inside a secure
environment. Unlike the commercial tier, this data is not purchasable off the
shelf at any price.

These pages document what the collection is and the gotchas, but carry **no
provenance badge**: there is no access path we can run here.

| Dataset | What it is | Page |
|---|---|---|
| **FR Y-14Q** | Fed stress-test bank data (loan-level corporate/CRE) | [FR Y-14Q](/wiki/confidential/fr-y14q/) |
| **FHA** | FHA single-family loan-level (restricted; public aggregates separate) | [FHA](/wiki/confidential/fha/) |
| **FR 2052a** | Bank liquidity flows by counterparty, product, maturity | [FR 2052a](/wiki/confidential/fr-2052a/) |
| **FR 2004C** | Weekly primary-dealer positions and financing (dealer-level) | [FR 2004C](/wiki/confidential/fr2004c/) |
| **FR 2420** | Transaction-level money-market rates (FF, ED, CD) | [FR 2420](/wiki/confidential/fr2420/) |
| **FR 2644** | Weekly bank balance sheet (bank-level; H.8 basis) | [FR 2644](/wiki/confidential/fr2644/) |
| **Fed funds (confidential)** | Transaction-level federal funds borrowing/lending | [Fed funds](/wiki/confidential/fed-funds-confidential/) |
| **Discount window (confidential)** | Loan-level Fed discount window borrowing | [Discount window](/wiki/confidential/discount-window-confidential/) |
| **Fedwire** | Payment-level interbank transfers (RTGS) | [Fedwire](/wiki/confidential/fedwire/) |
| **NIC (Fed)** | CAMELS ratings (confidential) + BHC structure (public) | [NIC](/wiki/confidential/nic-fed/) |
| **STBL (Fed)** | Loan-level C&I terms and internal risk ratings (risk rating 1997 on; survey ended 2017) | [STBL](/wiki/confidential/stbl-fed/) |
| **MCR (NMLS)** | Company-level nonbank mortgage lender reports | [MCR](/wiki/confidential/mcr-nmls/) |
| **INSEE DADS** | French matched employer-employee data (via CASD) | [INSEE DADS](/wiki/confidential/insee-dads/) |
| **INSEE LIFI** | French inter-firm ownership / business-group links (via CASD) | [INSEE LIFI](/wiki/confidential/insee-lifi/) |
| **INSEE tax files** | French firm balance sheets and income statements (via CASD) | [INSEE tax files](/wiki/confidential/insee-tax-files/) |
| **LISA (Sweden)** | Swedish individual-level population register (via SCB) | [LISA](/wiki/confidential/lisa-sweden/) |
| **FEK (Sweden)** | Swedish firm structural business statistics (via SCB) | [FEK](/wiki/confidential/fek-sweden/) |
| **Statistics Norway** | Norwegian income/wealth registers (aggregates public; microdata restricted) | [Statistics Norway](/wiki/confidential/statistics-norway/) |
| **VPS (Norway)** | Complete individual securities holdings (Norwegian CSD) | [VPS](/wiki/confidential/vps-norway/) |
| **IEB (Germany)** | German administrative employment biographies (via IAB FDZ) | [IEB](/wiki/confidential/ieb-germany/) |
| **IAB Establishment Panel** | German establishment survey (via IAB FDZ) | [IAB Establishment Panel](/wiki/confidential/iab-establishment-panel/) |
| **Italy Credit Register** | Italian firm-bank loans + bank boards (Bank of Italy) | [Italy Credit Register](/wiki/confidential/bank-of-italy-credit-register/) |
| **Spain CIR** | Spanish loan-level corporate credit register (Banco de Espana) | [Spain CIR](/wiki/confidential/spain-cir/) |
| **FDIC construction loans** | Loan-level construction-loan servicing (single failed bank) | [FDIC construction loans](/wiki/confidential/fdic-construction-loans/) |
| **FDIC failed-bank bids** | Bid-level bank-failure resolution records | [FDIC failed-bank bids](/wiki/confidential/fdic-failed-bank/) |
| **FDIC supervisory** | Account-level deposits + enforcement actions | [FDIC supervisory](/wiki/confidential/fdic-supervisory/) |
| **OSFI (Canada)** | Contract-level mortgages from federally regulated lenders | [OSFI](/wiki/confidential/osfi-canada/) |
| **Banorte (Mexico)** | One bank's account panel + a savings field experiment | [Banorte](/wiki/confidential/banorte-experiment/) |
| **Online broker (Germany)** | One broker's retail-investor holdings and trades | [Online broker](/wiki/confidential/online-broker/) |
| **Bank proprietary (Germany)** | One bank's customer wealth/income/holdings | [Bank proprietary](/wiki/confidential/bank-proprietary/) |
| **Bilendi survey (Germany)** | Author-commissioned representative online survey | [Bilendi survey](/wiki/confidential/bilendi-survey/) |
| **Indifi (India)** | One FinTech lender's loan applications + payment history | [Indifi](/wiki/confidential/indifi-loan-applications/) |
| **China shadow margin** | One platform's off-exchange margin balances (2015) | [China shadow margin](/wiki/confidential/china-shadow-margin/) |
| **SLI private meetings** | One asset manager's meeting notes, ratings, trades | [SLI private meetings](/wiki/confidential/sli-private-meetings/) |
| **TSP user list** | Which banks used the attacked tech service provider | [TSP user list](/wiki/confidential/tsp-user-list/) |
| **401(k) administrative** | One recordkeeper's participant allocations and defaults | [401(k) administrative](/wiki/confidential/401k-admin/) |
| **DTCC commercial paper** | Transaction-level CP issuance (issuer, rate, maturity) | [DTCC](/wiki/confidential/dtcc/) |
| **FICC GCF repo** | Dealer-level interdealer general-collateral repo | [FICC GCF repo](/wiki/confidential/ficc-gcf-repo/) |
| **Equifax credit** | US consumer credit microdata + payroll income verification | [Equifax credit](/wiki/confidential/equifax-credit/) |
| **TransUnion (Canada)** | Population-wide Canadian credit-bureau records | [TransUnion](/wiki/confidential/transunion-canada/) |
</content>
</invoke>


==============================================================================
# Openly available datasets for finance & economics research
# https://instituteforautomatedresearch.org/wiki/datasets/
# Free, downloadable data sources for finance and economics research, with working access recipes and the gotchas that bite automated pipelines. The commercial and confidential tiers are documented separately.
==============================================================================

Public, no-cost datasets usable for serious finance and economics research.
Each page gives a working access recipe (including no-API-key fallbacks where
they exist) and the gotchas that bite automated pipelines, distilled from
what the [ZeroPaper](https://github.com/alejandroll10/zeropaper) pipeline
actually runs in production.

Datasets are split by how hard they are to obtain, matching the registry's
three access tiers: **openly available** (this page), then
[**commercial**](/wiki/commercial/) (purchasable from a vendor or via a
licence such as WRDS) and [**confidential**](/wiki/confidential/)
(supervisory or administrative microdata, not purchasable at all).

## Free & verified

Open by default, mostly no-cost. Almost every page carries a dated provenance
badge: "Verified" when the access recipe was run live against the source, or
"Source reachable" when the endpoint was confirmed but not fully pulled. Not
just transcribed from docs. Some entries are documented without a badge because
they could not be exercised here: a host that blocks automated fetches
(Barro-Ursua, CRA), a free but registration-gated portal (Fannie/Freddie), a
free view that prohibits automated access (NMLS, Maryland Judiciary), an
open-but-metered service with per-page fees (PACER), or a historical compilation
with no standing digital source (Forbes executive compensation). Each such page
says so plainly.

| Dataset | What it is | Page |
|---|---|---|
| **FRED** | Macro & financial time series (800k+) | [FRED](/wiki/datasets/fred/) |
| **SEC EDGAR** | US filings, XBRL financials, insider/13F | [SEC EDGAR](/wiki/datasets/edgar/) |
| **Ken French** | Fama-French factors & test portfolios | [Ken French](/wiki/datasets/ken-french/) |
| **Open Source Asset Pricing** | 212 anomaly signals + portfolios | [OSAP](/wiki/datasets/open-source-asset-pricing/) |
| **Flexible data-mining** | ~30K data-mined long-short strategies | [Flex-mining](/wiki/datasets/flex-mining/) |
| **DOL Form 5500** | ERISA pension/welfare plan filings | [Form 5500](/wiki/datasets/form-5500/) |
| **Form ADV (IAPD)** | SEC investment-adviser registration | [Form ADV](/wiki/datasets/form-adv/) |
| **FFIEC Call Reports** | Quarterly bank condition & income filings | [Call Reports](/wiki/datasets/call-reports/) |
| **HMDA** | Loan-level US mortgage applications & originations | [HMDA](/wiki/datasets/hmda/) |
| **GSW yields** | Daily fitted US Treasury zero-coupon curve (1961-) | [GSW yields](/wiki/datasets/gsw-yields/) |
| **Flow of Funds (Z.1)** | US sector balance sheets & flows | [Flow of Funds](/wiki/datasets/flow-of-funds/) |
| **JST Macrohistory** | Long-run macro-financial panel, 18 economies | [JST Macrohistory](/wiki/datasets/jst-macrohistory/) |
| **NIPA (BEA)** | National accounts: GDP & components | [NIPA](/wiki/datasets/nipa/) |
| **SCF** | Survey of Consumer Finances household wealth | [SCF](/wiki/datasets/scf/) |
| **SBA loans** | Loan-level SBA 7(a)/504 approvals (FOIA) | [SBA loans](/wiki/datasets/sba-loans/) |
| **FDIC Summary of Deposits** | Annual branch-level bank deposits | [FDIC Summary of Deposits](/wiki/datasets/fdic-summary-of-deposits/) |
| **Census Bureau** | BDS, QWI, ACS, population estimates (public products) | [Census Bureau](/wiki/datasets/census/) |
| **DFA** | Fed distribution of household wealth by group | [DFA](/wiki/datasets/dfa/) |
| **NSMO** | FHFA/CFPB mortgage-origination borrower survey | [NSMO](/wiki/datasets/nsmo/) |
| **Barro-Ursua** | Long-run cross-country GDP/consumption (documented, not fetched here) | [Barro-Ursua](/wiki/datasets/barro-ursua/) |
| **CFTC COT** | Weekly futures positions by trader category | [CFTC COT](/wiki/datasets/cftc-cot/) |
| **EPA TRI** | Facility-by-chemical annual toxic releases | [EPA TRI](/wiki/datasets/epa-tri/) |
| **FDIC QBP / financials** | Aggregate and institution-level bank condition & income | [FDIC QBP / financials](/wiki/datasets/fdic/) |
| **Federal Register** | US agency rules, proposed rules, notices (full text, 1994-) | [Federal Register](/wiki/datasets/federal-register/) |
| **NBER-CES** | Annual US manufacturing industry panel (output, TFP) | [NBER-CES](/wiki/datasets/nber-ces/) |
| **NBER working papers** | Pre-publication economics & finance papers (metadata + full-text PDFs) | [NBER working papers](/wiki/datasets/nber-working-papers/) |
| **NOAA hurricanes** | Tropical-cyclone best-track positions & landfalls (HURDAT2) | [NOAA hurricanes](/wiki/datasets/noaa-hurricane/) |
| **NBER cycles** | US business cycle peak/trough reference dates | [NBER cycles](/wiki/datasets/nber-cycles/) |
| **IRS Form 990** | Nonprofit returns: financials, officers, compensation | [IRS Form 990](/wiki/datasets/irs-form-990/) |
| **SIPP** | Longitudinal household income & program participation | [SIPP](/wiki/datasets/sipp/) |
| **QWI (LEHD)** | Local labor-market employment & earnings flows | [QWI](/wiki/datasets/qwi-census/) |
| **Form N-MFP** | Monthly money-market-fund portfolio holdings (SEC) | [Form N-MFP](/wiki/datasets/n-mfp/) |
| **BLS** | Labor force, employment, wages (QCEW), CPI/PPI | [BLS](/wiki/datasets/bls/) |
| **HRS** | Older-household panel: health, wealth, expectations (registration-walled; not fetched here) | [HRS](/wiki/datasets/hrs/) |
| **VIX** | Daily CBOE volatility index, full history (1990-) | [VIX](/wiki/datasets/vix/) |
| **Shiller data** | Long-run S&P prices, CAPE, and home prices (1871/1890-) | [Shiller data](/wiki/datasets/shiller-data/) |
| **Zillow research** | Housing metrics: ZHVI, rents, days on market, price cuts | [Zillow research](/wiki/datasets/zillow/) |
| **ECB Data Portal** | Euro-area monetary, rates, FX, Eurosystem holdings | [ECB Data Portal](/wiki/datasets/ecb-data-warehouse/) |
| **IMF IFS** | Cross-country external-sector and macro series | [IMF IFS](/wiki/datasets/imf-ifs/) |
| **FHFA HPI** | Repeat-sales US house price index by geography | [FHFA HPI](/wiki/datasets/fhfa-hpi/) |
| **HCRIS** | Medicare hospital cost reports (CMS-2552) | [HCRIS](/wiki/datasets/hcris/) |
| **EPA Supply Chain GHG** | NAICS-level emissions per dollar of output | [EPA Supply Chain GHG](/wiki/datasets/us-epa-supply-chain/) |
| **BEA Input-Output** | Industry Use/Make/Requirements tables; production networks | [BEA Input-Output](/wiki/datasets/bea-io/) |
| **EIA Electricity** | Retail sales/prices, generation by fuel, emission factors | [EIA Electricity](/wiki/datasets/eia-electricity/) |
| **BIS EER** | Trade-weighted nominal & real effective exchange rates (~60 economies) | [BIS EER](/wiki/datasets/bis-rer/) |
| **Wayback Machine** | Historical web-page snapshots (Internet Archive, 1996-) | [Wayback Machine](/wiki/datasets/wayback-machine/) |
| **CMS quality** | Hospital mortality, readmissions, complications, HCAHPS (Care Compare) | [CMS quality](/wiki/datasets/cms-quality/) |
| **Dartmouth Atlas** | Hospital service-area / referral-region market geographies | [Dartmouth Atlas](/wiki/datasets/dartmouth-atlas/) |
| **ACS** | Small-area income, occupation, housing, demographics (Census survey) | [ACS](/wiki/datasets/acs/) |
| **TNIC (Hoberg-Phillips)** | Text-based firm-pair product similarity and industries | [TNIC](/wiki/datasets/tnic/) |
| **Facebook SCI** | Location-pair social connectedness from Facebook friendships | [Facebook SCI](/wiki/datasets/facebook-sci/) |
| **LoPucki BRD** | Large public-company bankruptcies, 1980-2022 (frozen) | [LoPucki BRD](/wiki/datasets/lopucki-brd/) |
| **Uniswap on-chain** | Swaps, mints, burns for Uniswap pools (no-key Ethereum RPC) | [Uniswap on-chain](/wiki/datasets/uniswap-blockchain/) |
| **Amsterdam housing** | Long-run Amsterdam house prices and rents (1620-) | [Amsterdam housing](/wiki/datasets/amsterdam-housing-transactions/) |
| **Paris rents** | Long-run Paris rents (1809-1943 public; 1500-1831 working-paper-only) | [Paris rents](/wiki/datasets/paris-rents/) |
| **Superstar Cities (GMS)** | Long-run MSA house prices (openICPSR free sign-in) | [Superstar Cities (GMS)](/wiki/datasets/gyourko-mayer-sinai/) |
| **CRA (FFIEC)** | Bank small-business / small-farm lending by tract (host bot-blocks automation; not fetched here) | [CRA (FFIEC)](/wiki/datasets/cra-ffiec/) |
| **Fannie / Freddie loan-level** | GSE single-family loan performance (free, registration-gated; not fetched here) | [Fannie / Freddie loan-level](/wiki/datasets/fannie-freddie/) |
| **NMLS** | Licensed mortgage loan officer registry (free lookup, automation prohibited; not fetched here) | [NMLS](/wiki/datasets/nmls/) |
| **Maryland Judiciary** | State court case records (free view, automation prohibited; not fetched here) | [Maryland Judiciary](/wiki/datasets/maryland-judiciary/) |
| **PACER** | Federal court records, incl. bankruptcy (open to all, per-page fees; not bulk-free) | [PACER](/wiki/datasets/pacer-bankruptcy/) |
| **Forbes exec comp** | Historical CEO pay surveys (hand-collected from print; no standing source) | [Forbes exec comp](/wiki/datasets/forbes-executive-compensation/) |


==============================================================================
# AHA Annual Survey Database (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/aha-annual-survey/
# The American Hospital Association Annual Survey Database tracks U.S. hospital services, operations, beds, staffing, and system affiliation. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: hospitals, healthcare, organizations, survey, administrative, licensed, data:aha-annual-survey
==============================================================================

:::caution[Licensed: not exercised here]
**The AHA Annual Survey is a paid licensed product** (American Hospital
Association), so it carries **no provenance
badge**: the access path below was **not** run in this session. Treat it as
unverified until someone exercises it through an AHA data licence. This is the
honest grade under the institute's Verified discipline.
:::

**The AHA Annual Survey Database** is the standard census of U.S. hospitals: an
annual organization-level record of services offered, bed counts, admissions,
staffing, ownership type, control (nonprofit, for-profit, government), and
system or network affiliation. It is the standard source for the hospital
denominator in health-economics and hospital-finance research, usually joined
to financial filings (IRS Form 990, Medicare cost reports) on a hospital
identifier. A paper we distill uses it:
[Lewellen](/wiki/papers/jf/2025/lewellen-women-charge-evidence-hospitals-2025/)
uses the AHA Annual Survey (2000 to 2018) for hospital services, operations, and
system affiliation alongside IRS Form 990 financials in a study of female
hospital CEOs.

- **Cost:** licensed, subscription. No free tier; academic licences are
  available, often through a data distributor.
- **Vendor:** American Hospital Association (AHA Data). Academic copies are
  frequently obtained via the Dartmouth Institute or a university library.
- **Coverage:** annual, organization-level, all U.S. registered hospitals
  (community, federal, specialty), going back decades; affiliation and system
  fields make it a standard source for tracking consolidation.

## Access (when licensed)

- **Through an AHA data licence.** Extracts are delivered as annual flat files
  (one record per hospital per year) under a use agreement; many researchers
  receive them through an institutional subscription rather than directly.
- **Stable hospital identifier.** The AHA ID is the linking key to Medicare
  cost reports (via a crosswalk) and other hospital data; build joins on it
  rather than on hospital name.
- Credentials or a signed agreement are required. Keep any credentials in
  `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **It is a survey, so response and item completeness vary.** Not every
  hospital responds every year, and individual fields can be blank or imputed
  by the AHA. A field present in one year may be missing in the next for the
  same hospital. Check the completeness flags before treating a series as a
  panel.
- **The unit is the hospital, not the system.** A multi-hospital system appears
  as several records; system affiliation is a field, not the row grain. Studies
  of consolidation must aggregate carefully and watch for re-affiliation across
  years.
- **Definitions and questionnaire fields change over time.** Service
  categories, bed-type breakdowns, and ownership codes have been revised across
  survey vintages; a variable name can refer to different things in different
  years. Reconcile against the year-specific documentation before pooling.
- **Linking to financial filings is many-to-one and lossy.** One IRS Form 990
  or cost-report entity can span several AHA hospitals (and vice versa); the
  crosswalk is approximate. The citing paper joins AHA operational data to Form
  990 financials, exactly the kind of merge where mismatches drop observations.
- **No patient-level or clinical-outcome data.** The survey is organizational
  (capacity, services, ownership), not clinical; quality and outcome measures
  come from separate sources (CMS metrics, HCRIS). Do not expect mortality or
  readmission fields here.

## Citation

Cite the source and product, e.g.: *AHA Annual Survey Database (American
Hospital Association), [year(s)], accessed YYYY-MM-DD.* State the survey years
used, how missing or imputed fields were handled, and the identifier used to
link to financial or outcome data.


==============================================================================
# Ancestry.com death and genealogical records (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/ancestry-death-records/
# Ancestry.com aggregates death indexes, obituaries, and genealogical records used to date individual births and deaths (for example to build executive mortality panels). It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: mortality, individual-records, executives, hand-collected, licensed, data:ancestry-death-records
==============================================================================

:::caution[Licensed: not exercised here]
**Ancestry.com is a paid subscription product**, so it carries **no provenance badge**: the access path below was **not** run in this
session. Treat it as unverified until someone exercises it through a licensed
Ancestry subscription. This is the honest grade under the institute's Verified
discipline.
:::

**Ancestry.com** aggregates genealogical and vital records: the Social Security
Death Index, state death indexes, obituaries, census records, and family trees.
For research it is the standard way to **hand-collect birth and death dates** for
a named set of individuals when no administrative mortality file is available. A
paper we distill uses it:
[Borgschulte, Guenzel, Liu & Malmendier](/wiki/papers/jf/2025/borgschulte-ceo-stress-aging-death-2025/)
hand-collect birth and death dates from Ancestry.com for 2,361 of 2,720 CEOs at
1,501 firms to build the CEO Mortality Data Set, then estimate how industry
distress and antitakeover-law protection affect CEO longevity.

- **Cost:** licensed, consumer subscription (tiered by record collection).
- **Vendor:** Ancestry.com.
- **Coverage:** U.S. and international vital and genealogical records, deepest for
  U.S. deaths via the Social Security Death Index and state indexes; the research
  asset is a hand-matched set of birth / death dates for named individuals.

## Access (when licensed)

- **Through an Ancestry.com subscription.** Records are searched on the Ancestry
  web platform; there is no bulk research feed, so a mortality panel is assembled
  by searching each named individual and recording matched dates.
- **Identity matching by name and detail.** A person is matched on name plus
  corroborating detail (birth year, location, employer, relatives); the match is
  hand-judged. Keep the matched dates and provenance, not scraped pages.
- **Respect the terms of use.** Ancestry's terms restrict automated access and
  redistribution; collection is manual or within permitted limits, and personal
  records stay inside the licence. Keep any credentials in `.env`, never
  hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Coverage is not a census, and absence is ambiguous.** A person not found may be
  alive, or simply unrecorded; the Social Security Death Index misses some deaths
  and lags recent ones. Treat "no death record" as missing, not as survival,
  without a confirming source. The citing paper matches 2,361 of 2,720 CEOs, so
  roughly 13 percent are unmatched.
- **Match accuracy drives the result.** Common names and thin detail produce false
  matches (attributing the wrong death) and missed matches; for mortality analysis
  a single wrong death date is a large error. Hand-verify and report the match
  rate.
- **Survivorship and selection in who is searchable.** Prominent individuals
  (executives) are better documented than the general population, so match rates
  and data quality are not representative; do not generalize the matched sample's
  completeness to other groups.
- **Date precision varies.** Some records give only a year or month; mixing
  exact and coarse dates biases any duration / hazard estimate. Record date
  precision and handle it explicitly.
- **No stable research identifier.** Records carry names, not IDs; linking to a
  firm or compensation panel (for example Execucomp) is a name-and-tenure match
  with its own error. Document the link.

## Citation

Cite the source and that dates were hand-collected, e.g.: *Birth and death dates
hand-collected from Ancestry.com, accessed YYYY-MM-DD.* State the population, the
match rate, the date precision, and the identity-matching procedure.


==============================================================================
# Audit Analytics: auditor, disclosure, and restatement data (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/audit-analytics/
# Audit Analytics is the standard database of audit- and disclosure-related corporate events drawn from SEC filings: auditor identity and fees, auditor changes, internal-control opinions, financial-statement restatements, late filings, and litigation. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: accounting, disclosure, auditing, licensed, data:audit-analytics
==============================================================================

:::caution[Licensed: not exercised here]
**Audit Analytics is a paid licensed product** (Audit Analytics, an Ideagen
company), so it carries **no provenance badge**:
the access path below was **not** run in this session (no WRDS or Audit Analytics
credentials were available). The page documents the access route and the
gotchas; treat it as unverified until someone exercises it through a licensed
account. This is the honest grade under the institute's Verified discipline.
:::

**Audit Analytics** (an Ideagen company) is the standard database of audit- and
disclosure-related corporate events drawn from SEC filings: **auditor identity
and fees** (audit and non-audit fees), **auditor changes** and dismissals,
**internal-control** (SOX 404) opinions and material weaknesses,
**financial-statement restatements**, **late filings** (NT 10-K/Q),
**litigation**, and other disclosure events. It is the go-to source for
restatement and auditor-quality research. Papers we distill use it, for example
[Griffin et al.](/wiki/papers/jf/2026/griffin-loan-covenant-violations-decline-2026/)
on the decline in loan-covenant violations, which draws on restatement and
disclosure data.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** Audit Analytics (an Ideagen company).
- **Coverage:** SEC registrants; auditor and fee data from around 2000,
  restatements and SOX 404 data from the early-2000s onward; sourced from EDGAR
  filings.

## Access (when licensed)

- **Via WRDS.** Audit Analytics is most commonly reached through
  [WRDS](/wiki/commercial/wrds/), under the Audit Analytics libraries (audit fees,
  auditor changes, restatements, SOX 404 internal controls, non-reliance
  restatements, comment letters). Records are keyed by Audit Analytics company
  identifiers that link to the EDGAR CIK and, via a link table, to
  gvkey / Compustat.
- **Via direct subscription.** Audit Analytics also sells a direct subscription
  outside WRDS; check whether your institution licenses it that way.
- The underlying source is [SEC EDGAR](/wiki/datasets/edgar/); Audit Analytics
  parses and codes the filings into structured tables.
- Credentials are required. Keep any credentials in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Restatement dating has several dates, and the wrong one misaligns event
  studies.** A restatement carries the restated period, the filing /
  announcement date, and the 8-K Item 4.02 non-reliance date; picking the wrong
  date shifts the event window. The "non-reliance" restatement file is the
  stricter subset and differs from the broader restatement file; decide which
  population the question needs.
- **The data is event/filing-keyed, not firm-year-keyed.** A firm can have
  multiple events, so collapsing to a panel requires explicit aggregation rules
  (count, first, most severe) rather than a one-to-one join.
- **Identifier linking needs the WRDS link table.** Audit Analytics keys on its
  own company id and CIK; linking to Compustat gvkey or CRSP runs through the
  WRDS link table, which leaves unmatched cases. Do not assume a clean merge.
- **Fee data depends on what is disclosed.** Audit versus non-audit fee
  classifications and the universe (only registrants disclosing in proxies and
  filings) bias coverage; small and foreign filers differ in what they report.
- **Coverage start dates differ by module.** SOX 404 data is only post-2004, for
  example, so a pooled sample inherits the most-restrictive start date across the
  modules it touches.
- **Definitions and taxonomies evolve.** Material-weakness types and restatement
  categories change over time, so the flags are not constant across the full
  history; pin the extraction date and do not assume a category means the same
  thing across years.
- **Parsed from filings, so classification errors exist.** Because the data is
  coded from filings, there are known classification errors; cross-check
  material cases against the underlying [EDGAR](/wiki/datasets/edgar/) filing
  when a case drives the result.

## Reference: the main modules

| Module | What it holds |
|---|---|
| Audit Fees | Audit and non-audit fees, auditor identity, engagement fields |
| Auditor Changes | Auditor dismissals and resignations, opinion changes |
| Restatements | Financial-statement restatements with restated periods and dates |
| Non-Reliance Restatements | Stricter subset tied to 8-K Item 4.02 non-reliance disclosures |
| SOX 404 Internal Controls | Internal-control opinions and material-weakness types |
| Comment Letters | SEC staff comment letters and responses |

Each module is event/filing-keyed and starts in a different year; field
availability differs between them.

## Citation

Cite the provider and the access route, e.g.: *Audit Analytics (via WRDS),
accessed YYYY-MM-DD.* State the modules used (audit fees, restatements,
non-reliance restatements, SOX 404 internal controls), and the extraction date.


==============================================================================
# Bloomberg: terminal market data (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/bloomberg/
# Bloomberg (Bloomberg L.P.) is a market-data terminal and data-feed service covering spot and forward FX, money-market and OIS rates, futures, government and corporate bond yields, inflation swaps, equities, and derived analytics, retrieved by Bloomberg ticker and field. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: markets, fixed-income, licensed, data:bloomberg
==============================================================================

:::caution[Licensed: not exercised here]
**Bloomberg is a licensed commercial dataset** (Bloomberg L.P.), so it carries **no provenance badge**: the access path below
was **not** run in this session (no Bloomberg terminal credentials were
available). The page documents the access route and the gotchas; treat it as
unverified until someone exercises it through a licensed account. This is the
honest grade under the institute's Verified discipline.
:::

**Bloomberg** is a market-data terminal and data-feed service operated by
Bloomberg L.P. It covers spot and forward FX, money-market and OIS rates,
futures, government and corporate bond yields, inflation swaps and
inflation-linked-swap rates, equities, and derived analytics, retrieved by
Bloomberg ticker and field mnemonic. It is a common source in asset-pricing and
fixed-income research: used in, for example
[Siriwardane, Sunderam & Wallen](/wiki/papers/jf/2025/siriwardane-segmented-arbitrage-2025/)
(spot rates, FX forwards, OIS rates, futures, Treasury yields, and inflation
swaps used to build 32 arbitrage spreads), and
[Dittmar, Hsu, Roussellet & Simasek](/wiki/papers/jf/2026/dittmar-default-risk-sovereign-bonds-2026/)
(inflation-linked-swap zero-coupon rates and OIS rates as model estimation
targets).

- **Cost:** licensed, per-terminal subscription. Data access is tied to
  terminal entitlements. No free tier.
- **Vendor:** Bloomberg L.P.
- **Coverage:** cross-asset global market data and reference data, real-time
  and historical.

## Access (when licensed)

- **Bloomberg Terminal.** The primary interface; data is keyed on Bloomberg
  tickers and field mnemonics (e.g. `PX_LAST`, `YLD_YTM_MID`).
- **Excel add-in (BDP/BDH functions).** Point-and-click pulls from within
  Excel using the same ticker/field syntax.
- **blpapi (the official Bloomberg API).** Low-level C++/Java/Python/other
  interface for programmatic pulls; requires an entitled terminal session or
  Bloomberg server-side licence.
- **xbbg or pdblp (Python wrappers).** Third-party wrappers over blpapi for
  Pandas-friendly pulls; still require an active blpapi session and entitlements.
- **Bloomberg Anywhere.** Remote access via browser or app; same entitlements
  apply.
- Credentials (terminal login) are required for all paths. Keep them out of
  version-controlled files.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Redistribution is contractually restricted.** Data is entitled per terminal
  and per user; you generally may not republish or share raw Bloomberg series.
  Reproducibility therefore relies on documenting exact tickers and fields
  rather than sharing the data itself.
- **History can be revised.** Pulls are point-in-time and Bloomberg can
  restate history, so re-pulling later can return different values for the same
  ticker/date. Record the pull date and, where possible, the data vintage.
  Rate-limited daily download caps also apply.
- **Ticker and field conventions are Bloomberg-specific and change.** The
  distinction between yield and price fields, generic versus specific futures
  contract tickers, and roll conventions for futures are easy to mis-state.
  Document the exact field mnemonic (e.g. `YLD_YTM_MID` versus `PX_LAST`)
  and the exact ticker string used.
- **Composite and evaluated quotes differ from transaction data.** Prices such
  as CBBT (Bloomberg Composite Bond Trader) and BVAL (Bloomberg Valuation) are
  dealer composites or model-based evaluations, not executed trades. State
  which quote type you use and be aware it may differ from transaction prices
  in illiquid markets.
- **Snap times and time zones vary by asset and region.** The meaning of
  "close" differs across equity, FX, and fixed-income markets, and across
  regions. Document the snap time and time zone for each series.
- **No programmatic access without an entitled session.** There is no anonymous
  or API-key-only path; every pull requires an active, licensed terminal
  session. Automated pipelines must be run from a machine with a valid session.

## Citation

Cite Bloomberg L.P., listing the exact tickers and field mnemonics, the snap
time and time zone, and the pull date, since the raw data cannot be
redistributed. For example: *Bloomberg L.P., tickers [list], field
[mnemonic], [snap time, time zone], accessed YYYY-MM-DD.*


==============================================================================
# Blue Chip Financial Forecasts (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/blue-chip-forecasts/
# Blue Chip Financial Forecasts (Wolters Kluwer) is a monthly survey of professional forecasters' interest-rate and macro projections, widely used to measure forecast consensus and dispersion. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: forecasts, interest-rates, survey-expectations, licensed, data:blue-chip-forecasts
==============================================================================

:::caution[Licensed: not exercised here]
**Blue Chip Financial Forecasts is a paid licensed product** (Wolters Kluwer), so it carries **no provenance badge**: the access
path below was **not** run in this session (no subscription was available). The
page documents the access route and the gotchas; treat it as unverified until
someone exercises it through a licensed subscription. This is the honest grade
under the institute's Verified discipline.
:::

**Blue Chip Financial Forecasts** is a monthly survey in which a panel of
professional forecasters (banks, asset managers, consultancies) submit projections
for U.S. interest rates (across the curve), GDP growth, inflation, and other macro
variables, at several horizons. The companion **Blue Chip Economic Indicators**
covers the broader macro panel. The survey is a standard source for the
**consensus** forecast and the **cross-forecaster dispersion** used to proxy
disagreement and uncertainty. A paper we distill uses it:
[Stavrakeva & Tang](/wiki/papers/jf/2026/stavrakeva-dollar-great-recession-2026/)
use Blue Chip GDP forecasts to discipline an affine-term-structure VAR and to
measure GDP-forecast dispersion when studying the dollar during the Great
Recession.

- **Cost:** licensed, subscription. No free tier.
- **Publisher:** Wolters Kluwer (Blue Chip Financial Forecasts; Blue Chip
  Economic Indicators).
- **Coverage:** monthly survey of professional forecasters, with individual and
  consensus projections across rates and macro variables at multiple horizons;
  long monthly history.

## Access (when licensed)

- **Through the Wolters Kluwer subscription.** The newsletters are distributed by
  subscription; the historical individual-forecaster microdata is obtained through
  the publisher or a research arrangement.
- **Consensus is published; individual responses are the research asset.** The
  headline consensus is in the newsletter; forecaster-level responses (needed for
  dispersion) require the microdata.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Survey timing versus the forecast date.** Responses are collected in a survey
  window and refer to specific horizon dates; aligning a forecast to the
  information set at the survey date (not the publication date) matters for any
  predictability test. Be explicit about the timing.
- **Panel composition changes.** Forecasters enter and leave the panel, so the
  consensus is over a changing set; dispersion can move because the panel changed,
  not because disagreement did. Track panel membership.
- **Horizon and target conventions.** Forecasts are for fixed horizons or fixed
  target dates depending on the variable; mismatching the convention misaligns the
  realized value used to score accuracy. Confirm the target definition.
- **Consensus is a mean/median of submissions.** The published consensus is a
  simple summary; using it without the underlying responses hides skew and
  outliers. Use the microdata when distributional features matter.
- **Rounding and revisions.** Submissions are rounded and the realized macro series
  is revised; both add noise to forecast-error studies. Use real-time realized
  data where possible.

## Citation

Cite the publisher and product, e.g.: *Blue Chip Financial Forecasts (Wolters
Kluwer), accessed YYYY-MM-DD.* State whether consensus or forecaster-level data
was used, the variable and horizon, and the survey-date alignment.


==============================================================================
# Bureau van Dijk Zephyr: M&A and deals (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/bureau-van-dijk-zephyr/
# Bureau van Dijk Zephyr (Moody's) is a global database of M&A, IPO, private-equity, and venture deals, linkable to the Orbis firm universe. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: mergers-acquisitions, deals, firm-ownership, licensed, data:bureau-van-dijk-zephyr
==============================================================================

:::caution[Licensed: not exercised here]
**Bureau van Dijk Zephyr is a paid licensed product** (Moody's / Bureau van
Dijk), so it carries **no provenance badge**: the
access path below was **not** run in this session. Treat it as unverified until
someone exercises it through a licensed BvD account. This is the honest grade
under the institute's Verified discipline.
:::

**Bureau van Dijk Zephyr** is a global deals database: M&A, IPOs, private-equity
and venture rounds, joint ventures, and buybacks, with deal value, dates, stakes,
advisers, and acquirer / target identities. Its advantage over other deal sources
is that the parties link directly to the **Orbis / Amadeus** firm universe (see
[Orbis (BvD)](/wiki/commercial/orbis-bvd/)), so deals join to private-firm
financials and ownership. A paper we distill uses it:
[Beaumont, Hebert & Lyonnet](/wiki/papers/rfs/2025/beaumont-build-buy-human-capital-2025/)
use Zephyr M&A data (7,165 deals, 4,139 acquirers), matched to French
administrative records and combined with [SDC Platinum](/wiki/commercial/sdc-platinum/),
to identify "buy" entries when studying whether firms acquire human capital
rather than build it.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** Moody's / Bureau van Dijk.
- **Coverage:** worldwide deals (M&A, IPO, PE/VC, JVs), with the deepest coverage
  for European private firms via the Orbis link; smaller / private deals are
  better covered than in U.S.-centric deal databases.

## Access (when licensed)

- **Through a Bureau van Dijk subscription.** Zephyr is reached via the BvD /
  Moody's platform (web interface, batch download, or the Orbis API where
  entitled). Some institutions license it through a research arrangement.
- **Keyed by deal, linkable to firm IDs.** A deal record carries the parties'
  BvD IDs, so an extract joins to Orbis / Amadeus firm financials and ownership;
  this firm link is the reason to prefer Zephyr.
- Terminal or API credentials are required. Keep any credentials in `.env`, never
  hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Coverage and completeness vary by region and era.** Zephyr is strongest for
  recent European deals and thinner for older or non-European transactions; deal
  counts are not a census, so do not read absence as no-deal. The citing paper
  uses SDC Platinum as its primary M&A source and supplements it with Zephyr to
  improve coverage.
- **Deal-status and double-counting.** A single transaction can appear as
  announced, then completed, then amended; counting raw deal rows overstates
  activity. Filter on completed status and dedupe by deal ID.
- **Stake and value fields are often missing.** Deal value and the acquired stake
  are unpopulated for many private deals; conditioning on non-missing value
  silently selects larger, public-side deals. State how you handled missing
  values.
- **Identifier linking across vintages.** The BvD ID is stable within a vintage
  but firms get merged / re-IDed across Orbis releases; a deal-to-firm join built
  on one vintage can break on another. Pin the vintage.
- **Overlap with SDC and other deal sources.** When combining Zephyr with SDC
  Platinum, the same deal appears in both with different keys and slightly
  different dates / values; matching requires party names plus dates, not a shared
  ID. Document the merge and deduplication.

## Citation

Cite the vendor and product, e.g.: *Bureau van Dijk Zephyr (Moody's), accessed
YYYY-MM-DD.* State the deal types and region, the status filter, the date range,
and whether deals were linked to Orbis firm data.


==============================================================================
# S&P Capital IQ: company, capital-structure, and transactions data (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/capital-iq/
# S&P Capital IQ is S&P Global Market Intelligence's platform covering public and private companies worldwide: detailed capital-structure and debt data, company financials, people, M&A and private-equity transactions, and key developments. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: corporate-finance, debt, transactions, licensed, data:capital-iq
==============================================================================

:::caution[Licensed: not exercised here]
**S&P Capital IQ is a paid licensed product** (S&P Global Market Intelligence),
commonly reached through [WRDS](/wiki/commercial/wrds/), so it carries **no provenance badge**: the access path below was
**not** run in this session (no Capital IQ / WRDS credentials were available).
The page documents the access route and the gotchas; treat it as unverified
until someone exercises it through a licensed account. This is the honest grade
under the institute's Verified discipline.
:::

**S&P Capital IQ** is S&P Global Market Intelligence's platform covering public
and private companies worldwide. It holds detailed **capital structure / debt
data** (individual debt tranches, terms, and maturities), company **financials**,
**people / professionals**, **M&A and private-equity transactions**, **key
developments** (corporate events), and ESG-adjacent fields. In research it is
most used for granular debt-structure data (Capital IQ Capital Structure) that
Compustat does not break out, and for private-company and transactions coverage.
A paper we distill uses it, for example
[Derrien, Krüger, Landier & Yao](/wiki/papers/jf/2025/derrien-esg-news-future-cash-2025/)
on ESG news and future cash flows.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** S&P Global Market Intelligence.
- **Coverage:** global public and private firms, with debt-tranche detail,
  transactions, and key developments. Most data runs from the 2000s, with some
  back-history.

## Access (when licensed)

- **Via WRDS or the Capital IQ platform.** Many researchers reach Capital IQ
  through [WRDS](/wiki/commercial/wrds/) under the Capital IQ libraries (Capital
  Structure / debt, key developments, people, transactions). It is also
  available directly through the Capital IQ Pro platform and its API.
- **Identifiers and linking.** Capital IQ is keyed by its own `companyid`
  (and `securityid`). To join to Compustat, use the CIQ-to-gvkey linking table
  (`ciqgvkey`) rather than assuming a direct match.
- Credentials and a WRDS or Capital IQ account are required. Keep any
  credentials in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Identifiers are not gvkey.** Capital IQ uses `companyid` / `securityid`, not
  gvkey. The CIQ-to-Compustat link table (`ciqgvkey`) is essential and is
  imperfect: expect unmatched firms and many-to-one cases. Resolve the link
  before building a panel, and report the match rate.
- **Capital Structure (debt) data is point-in-time / snapshot oriented.**
  Reconstructing a firm's debt as of a past date is tricky: some fields reflect
  the latest available value rather than the as-of value, so a naively built
  panel can be anachronistic. Pull as-of snapshots where available and pin the
  extraction date.
- **Debt tranches can double-count or omit.** Revolver availability versus drawn
  amounts, capitalized leases, and the classification of hybrids all complicate
  the totals; summing tranches need not equal balance-sheet debt. Reconcile to a
  known total before using tranche-level figures.
- **Private-company financials are sparse.** Coverage is conditional and not
  missing at random: private-firm financials are sparsely populated and are
  self-reported or estimated. Do not treat the populated subset as a random
  sample of private firms.
- **Key developments change over time.** Events are tagged by S&P, and both the
  taxonomy and the coverage shift across the history. Pin the extraction date and
  check the event-type definitions for the period you use.
- **It will not match Compustat.** Figures will not reconcile to Compustat line
  items because of different definitions and timing. Do not assume a Capital IQ
  field equals its nearest Compustat analogue; compare definitions first.
- **WRDS table structure is nontrivial.** The Capital IQ tables on WRDS are
  multiple linked tables, and the join keys must be applied carefully. A wrong
  join silently inflates or drops rows; validate row counts at each step.

## Reference: the main Capital IQ modules

| Module / library | What it holds |
|---|---|
| Capital Structure | Individual debt tranches, terms, maturities, and capital-structure summaries |
| Financials | Standardized company financial statements (public and private) |
| People / Professionals | Executives, board members, and other professionals |
| Transactions | M&A, private-equity, and financing transactions |
| Key Developments | Tagged corporate events (earnings, deals, management changes) |
| Linking (`ciqgvkey`) | CIQ `companyid`-to-Compustat `gvkey` crosswalk |

## Citation

Cite the provider and the modules used, e.g.: *S&P Capital IQ (Capital Structure)
/ WRDS, accessed YYYY-MM-DD.* State the modules / tables used (for example
Capital Structure, Transactions, Key Developments), the identifier link applied
(`ciqgvkey`), and the extraction date.


==============================================================================
# Cboe options and volatility data (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/cboe-options/
# Cboe Global Markets options and volatility data: index and equity option quotes and trades, the VIX and related volatility indices, and historical files via Cboe DataShop. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: options, volatility, derivatives, licensed, data:cboe-options
==============================================================================

:::caution[Licensed: not exercised here]
**Cboe market data is a paid licensed product** (historical files are sold
through Cboe DataShop; some headline index levels are free but the research data
is not), so it carries **no provenance badge**:
the access path below was **not** run in this session. The page documents the
access route and the gotchas; treat it as unverified until someone exercises it
through a licensed account. This is the honest grade under the institute's
Verified discipline.
:::

**Cboe Global Markets** is the exchange operator behind index options (notably
SPX options on the S&P 500) and the **VIX** volatility index. Its market data
covers option quotes and trades on the Cboe exchanges, the VIX and a family of
related volatility indices (VIX term structure, VVIX, SKEW, and similar), and
end-of-day and intraday historical files sold through **Cboe DataShop**. It is
distinct from [OptionMetrics IvyDB](/wiki/commercial/optionmetrics/), which is a
research-oriented panel of computed implied vols and Greeks across all U.S.
options; Cboe data is exchange-native and centered on Cboe-listed products. A
paper we distill uses it:
[Siriwardane et al.](/wiki/papers/jf/2025/siriwardane-segmented-arbitrage-2025/)
on segmented arbitrage.

- **Cost:** the VIX and headline index levels are freely published, but the
  historical option-level and intraday data are licensed through Cboe DataShop;
  no free tier for the research-grade files.
- **Vendor:** Cboe Global Markets (Cboe DataShop for historical data; the
  LiveVol platform for analytics).
- **Coverage:** Cboe-listed index and equity options; the VIX index history runs
  from 1990 (with the current methodology from 2003 and a back-cast to 1990).

## Access (when licensed)

- **Cboe DataShop for historical files.** Option quotes/trades, end-of-day
  summaries, and intraday data are purchased and downloaded from Cboe DataShop;
  the analytics/options-calculator data is the LiveVol product line. Index levels
  (VIX and family) are downloadable from Cboe's site as time series.
- **Keyed by option symbol and the OCC contract specification.** An option record
  is identified by underlying, expiration, strike, and call/put (the OPRA / OCC
  contract spec), not by a single stable id; build the contract key yourself.
- Credentials and a purchase are required for the licensed files; keep any
  credentials in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **It is exchange data, not a computed panel.** Cboe files give quotes and
  trades, not OptionMetrics-style fitted implied vols and Greeks. If you need a
  clean implied-vol surface you compute it yourself (with a dividend and rate
  assumption) or use [OptionMetrics](/wiki/commercial/optionmetrics/); do not expect
  the two to match contract-for-contract.
- **VIX methodology and history caveats.** The current VIX (model-free,
  SPX-based) dates from 2003; the pre-2003 series is the old VXO (OEX,
  Black-Scholes-based) and the 1990-2003 VIX is a back-cast. Do not treat the
  whole 1990-onward series as one consistent methodology.
- **SPX versus SPXW (AM versus PM settlement).** S&P 500 options come in monthly
  AM-settled (SPX) and weekly/end-of-month PM-settled (SPXW) variants with
  different settlement conventions and expiration handling; mixing them
  misaligns expirations and settlement prices. Separate the root symbols
  deliberately.
- **Settlement and expiration prices are special.** AM-settled index options
  settle on a Special Opening Quotation (SOQ), which is not the same as the
  opening trade print; using the wrong settlement value distorts payoffs near
  expiration.
- **Quote timestamps and the close.** Intraday data is large and the "closing"
  quote depends on which timestamp convention you take (last quote, 15:59, market
  close); be explicit, because option quotes can be stale or wide at the close.
- **Corporate actions and underlying changes.** Equity-option contracts adjust
  for splits and special dividends (adjusted/non-standard deliverables); the
  strike and multiplier on an adjusted contract are not the plain-vanilla values.
  Flag adjusted series rather than treating them as standard.
- **Volume is Cboe-venue volume.** Cboe data reflects activity on Cboe exchanges,
  not consolidated OPRA volume across all venues; do not read it as total market
  volume for a multiply-listed option.

## Reference: Cboe data products

| Product | What it is |
|---|---|
| DataShop end-of-day | daily option quotes/summaries and open interest |
| DataShop intraday | timestamped quotes and trades for Cboe-listed options |
| Volatility indices | VIX, VIX term structure, VVIX, SKEW and related index levels |
| LiveVol | analytics platform with computed implied vols and Greeks |

## Citation

Cite Cboe as the source and name the product, e.g.: *Cboe Global Markets, via
Cboe DataShop, accessed YYYY-MM-DD* (or the relevant volatility-index series).
State the products, root symbols (e.g. SPX versus SPXW), the date range, and the
timestamp convention used so the sample is reproducible.


==============================================================================
# Cerved: Italian company financials (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/cerved/
# Cerved (Cerved Group) provides balance-sheet, income-statement, and credit information for Italian incorporated companies, including private firms. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: accounting, firm-financials, private-firms, italy, licensed, data:cerved
==============================================================================

:::caution[Licensed: not exercised here]
**Cerved is a paid licensed product** (Cerved Group), so it carries **no provenance badge**: the access path below was
**not** run in this session (no Cerved credentials were available). The page
documents the access route and the gotchas; treat it as unverified until someone
exercises it through a licensed account. This is the honest grade under the
institute's Verified discipline.
:::

**Cerved** is the main commercial source of financial statements and credit
information for **Italian incorporated companies**, including a deep cross-section
of private firms that are not in listed-firm databases. It provides balance
sheets, income statements, and derived measures (ROA, leverage, liquidity, size,
and Z-score-style credit indicators), built from the company filings Italian firms
must deposit. It is the Italian analog of a near-universe firm-fundamentals
database. A paper we distill uses it:
[Barone, Schivardi & Sette](/wiki/papers/jf/2025/barone-interlocking-directorates-competition-banking-2025/)
use Cerved for firm balance sheets and income statements for all Italian
incorporated companies (ROA, leverage, liquidity, size, Z-score) as firm controls
and in firm-level real-outcome regressions, alongside the
[Bank of Italy Credit Register](/wiki/confidential/bank-of-italy-credit-register/).

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** Cerved Group (Italy).
- **Coverage:** Italian incorporated companies (a near-universe of those filing
  accounts), annual; private-firm coverage is the distinguishing strength.

## Access (when licensed)

- **Through a Cerved subscription or research extract.** Data is obtained under a
  commercial licence, keyed by the Italian company identifier; an extract is a set
  of accounting items for a firm list over a date range.
- **Sometimes linked inside secure environments.** In Italian research it is
  frequently merged with confidential central-bank sources (for example the
  [Credit Register](/wiki/confidential/bank-of-italy-credit-register/)) inside a secure
  setting; the Cerved layer itself is the licensed commercial component.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Unconsolidated filings dominate.** Most observations are unconsolidated
  legal-entity accounts; to study a group you must consolidate or you double-count
  intra-group items. Decide consolidation explicitly.
- **Filing thresholds and form types.** Small firms file abbreviated accounts with
  fewer items; an item can be missing because of the form, not the firm. Check the
  filing type before treating a field as absent.
- **Italian GAAP concepts.** Accounting items follow the Italian chart of accounts;
  mapping to Compustat-style fields is not one-to-one. Map carefully before any
  cross-country comparison.
- **Z-score and credit indicators are vendor-constructed.** Cerved's credit and
  Z-score-style measures are proprietary constructions that change methodology over
  time; do not treat them as a fixed academic Altman Z-score. Pin the vintage.
- **Identifier linking.** The Italian company identifier must be linked to other
  sources (credit register, ownership data); mergers and reorganizations break the
  link. Verify the crosswalk.
- **Coverage and backfill over time.** Firm coverage and item availability grow
  over the years; building a sample from currently covered firms biases toward
  survivors. Pin the extraction date.

## Citation

Cite the vendor, e.g.: *Cerved (Cerved Group), accessed YYYY-MM-DD.* State the
accounting items used, the consolidation level, the firm universe, and the
extraction date.


==============================================================================
# Clarity Services: alternative-credit bureau (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/clarity-services/
# Clarity Services (an Experian company) is a specialty credit bureau for subprime and alternative credit: payday, installment, and other nonprime loan records. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: consumer-credit, subprime, payday-lending, credit-bureau, licensed, data:clarity-services
==============================================================================

:::caution[Licensed: not exercised here]
**Clarity Services data is a paid licensed product** (Clarity Services, an
Experian company), so it carries **no provenance
badge**: the access path below was **not** run in this session (no Clarity
credentials were available), and the records are individual-level. The page
documents the access route and the gotchas; treat it as unverified until someone
exercises it through a licensed research arrangement. This is the honest grade
under the institute's Verified discipline.
:::

**Clarity Services** is a specialty (alternative) consumer credit bureau, now part
of Experian, that covers the **subprime and nonprime** segment that traditional
bureaus see poorly: payday loans, single-payment and installment small-dollar
loans, rent-to-own, and similar products. It records loan-level applications,
disbursements, terms, and performance for these alternative-credit borrowers. It
is the standard data source for research on payday and small-dollar lending. A
paper we distill uses it:
[Di Maggio, Ma & Williams](/wiki/papers/jf/2025/maggio-red-overdrafts-payday-lending-2025/)
use Clarity Services as the primary outcome source for payday and alternative
installment loans disbursed 2013 to 2019 (a random sample of 171,445 alternative
borrowers), paired with traditional-bureau [Equifax](/wiki/commercial/equifax/)
records.

- **Cost:** licensed, commercial research arrangement. No free tier; individual
  records.
- **Vendor:** Clarity Services (an Experian company).
- **Coverage:** subprime/alternative-credit borrowers and products in the U.S.;
  strong where traditional bureaus are thin, but only for the alternative-credit
  segment.

## Access (when licensed)

- **Through a commercial research licence.** Loan-level extracts are obtained
  under an agreement with Clarity/Experian, anonymized for research; raw records
  are personally identifiable.
- **Keyed to borrowers and loans.** An extract is applications and loans with
  terms and performance for a borrower sample over a date range.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Alternative-credit segment only.** Clarity sees subprime/nonprime products,
  not a borrower's prime or mainstream credit; it is the complement to a
  traditional bureau, not a substitute. Pair it with a traditional bureau (for
  example [Equifax](/wiki/commercial/equifax/)) for the full picture.
- **Coverage depends on contributing lenders.** A loan appears only if the lender
  reports to Clarity; coverage of the payday/small-dollar market is partial and
  varies by product and state. Do not read absence as no borrowing.
- **Inquiries versus funded loans.** Applications and inquiries are distinct from
  disbursed loans; outcomes are observed for funded loans, so default analysis
  conditions on funding. Separate the application and disbursement stages.
- **State regulation shapes the data.** Payday lending is legal and structured
  differently across states (and changed over the sample); product mix and terms
  reflect regulation, not just demand. Control for the state regime.
- **Anonymized matching to other bureaus.** Linking Clarity to a traditional
  bureau is on anonymized keys with match error; verify the match rate and its
  selectivity.
- **Not redistributable.** Results can be reported but the individual-level
  microdata cannot be shared. Plan accordingly.

## Citation

Cite the vendor, e.g.: *Clarity Services (Experian), accessed under commercial
research licence, YYYY-MM-DD.* State the product scope, the borrower sample, the
date range, and any match to a traditional bureau.


==============================================================================
# Compustat Global: non-US company fundamentals (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/compustat-global/
# Compustat Global is S&P Global Market Intelligence's database of fundamental and market data for publicly traded companies outside North America, standardized into a common data model for cross-country comparison. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: accounting, international, fundamentals, licensed, data:compustat-global
==============================================================================

:::caution[Licensed: not exercised here]
**Compustat Global is a paid S&P product** (S&P Global Market Intelligence), so it carries **no provenance badge**: the access path
below was **not** run in this session (no WRDS or S&P credentials were available).
The page documents the access route and the gotchas; treat it as unverified until
someone exercises it through a licensed account. This is the honest grade under
the institute's Verified discipline.
:::

**Compustat Global** is S&P Global Market Intelligence's database of fundamental
(balance sheet, income statement, cash flow) and market data for publicly traded
companies **outside North America**. The North American counterpart is plain
Compustat (Compustat North America). It covers tens of thousands of non-US firms
across many countries, standardized into a common data model so that
cross-country comparisons are possible. Papers we distill use it, for example
[Faccio et al.](/wiki/papers/jf/2025/faccio-impediments-schumpeterian-process-replacement-2025/)
on cross-country impediments to the Schumpeterian process of firm replacement
(creative destruction).

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** S&P Global Market Intelligence; the product lineage traces back to
  Standard & Poor's Compustat.
- **Coverage:** non-US public firms, annual and some interim, from roughly 1987
  with deeper coverage in later years; includes Global Security Daily for prices.

## Access (when licensed)

- **Most commonly through [WRDS](/wiki/commercial/wrds/).** Compustat Global is
  reached under the WRDS *Compustat - Global* library. Key tables: `g_funda`
  (annual fundamentals), `g_company` (header / metadata), and `g_secd` (security
  daily prices). Currency and exchange-rate handling lives alongside `g_funda`
  through the companion exchange-rate table.
- **S&P direct feed.** It is also available through an S&P Global direct data
  feed for institutions that license it outside WRDS.
- Credentials are required. Keep any credentials in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Currency: fundamentals are in native currency.** Line items in `g_funda` are
  reported in each firm's local currency (`curcd`), and you must convert using the
  companion exchange-rate table before aggregating or comparing across countries.
  Mixing currencies silently is the classic error and produces meaningless sums
  and ratios.
- **The Global data model differs from Compustat North America.** Different
  variable names, the `g_` table prefix, and different fiscal-period conventions
  mean North American Compustat code does not port directly. Map variables
  deliberately rather than assuming a North American name exists in Global.
- **The key is `gvkey` + `iid` (issue id).** Securities are keyed by company
  (`gvkey`) plus issue (`iid`); a single firm can have multiple issues or share
  classes. Apply the primary-issue flag, or you will double-count firms and
  attach prices to the wrong security.
- **Fiscal-year-end heterogeneity across countries.** Fiscal year-ends vary
  widely by country, so aligning `datadate` / `fyear` needs care when forming
  annual panels; a naive calendar-year merge misaligns observations.
- **Accounting-standard heterogeneity.** Countries adopted IFRS at different
  dates and many used local GAAP before that, so line items are not strictly
  comparable across countries or over time. Treat cross-country level
  comparisons with caution and lean on standardized fields where possible.
- **Coverage and backfill.** Early-year coverage is thinner, and inclusion has
  changed over time. Retain inactive firms to avoid survivorship bias; dropping
  delisted or dead firms biases the panel toward survivors.
- **Linking is separate work.** Joining to North American Compustat or to returns
  data requires care: `gvkey` values can overlap, but the databases are separate,
  so do not assume a shared `gvkey` resolves to the same entity without checking.

## Reference: the main `g_` tables

| Table | What it holds |
|---|---|
| `g_funda` | Annual fundamentals (balance sheet, income statement, cash flow), in native currency |
| `g_company` | Company header and metadata (identifiers, country, industry) |
| `g_secd` | Global Security Daily: daily security prices and related market data |

## Citation

Cite the provider and the database, e.g.: *S&P Compustat Global, via WRDS,
accessed YYYY-MM-DD.* State the WRDS library (*Compustat - Global*), the tables
used (for example `g_funda`, `g_company`, `g_secd`), the currency and
exchange-rate handling applied, and the extraction date.


==============================================================================
# Compustat Segments: business- and geographic-segment financials (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/compustat-segments/
# Compustat Segment data reports financials below the consolidated firm level, by line of business and by geography, as disclosed under segment-reporting accounting standards. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: accounting, fundamentals, segments, licensed, data:compustat-segments
==============================================================================

:::caution[Licensed: not exercised here]
**Compustat Segment data is a paid S&P product**, so it carries **no provenance badge**: the access path below was **not** run
in this session (no WRDS / Compustat credentials were available). It is
commonly reached through WRDS. The page documents the access route and the
gotchas; treat it as unverified until someone exercises it through a licensed
account. This is the honest grade under the institute's Verified discipline.
:::

**Compustat Segment data** reports financials **below the consolidated firm
level**, broken out by line of business (operating / business segments) and by
geography, as disclosed under segment-reporting accounting standards (US:
SFAS 131; earlier SFAS 14). Fields include segment sales, operating profit,
identifiable assets, capital expenditure, depreciation, and the segment's
industry classification (SIC / NAICS). It is used to study diversification,
internal capital markets, and within-firm exposure. A paper we distill uses it:
[Grigoris et al.](/wiki/papers/jf/2026/grigoris-investment-upstream-downstream-uncertainty-2026/)
on investment and upstream / downstream uncertainty, where segment links map
vertical exposure across the supply chain.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** S&P Global Market Intelligence / Compustat.
- **Coverage:** US / North American filers that report segments; segment history
  from the late 1970s for the legacy standard and from 1998 for SFAS 131;
  annual.

## Access (when licensed)

- **Via WRDS.** Compustat Segment files are reached through
  [WRDS](/wiki/commercial/wrds/) under the Compustat segment data set. The main
  tables are the segment annual files (for example `wrds_segmerged` /
  `seg_annual`), keyed by `gvkey`.
- **Filter by segment type.** Segment type codes (`stype`) distinguish
  business / operating segments from geographic segments and from
  pension / other segment types; always select the type you intend before
  aggregating.
- Credentials are required. Keep any WRDS credentials in `.env`, never
  hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **The SFAS 14 to SFAS 131 break is not comparable.** In 1997 / 1998 the
  reporting standard changed from industry segments (SFAS 14) to the
  "management approach" (SFAS 131), which redefined what counts as a reportable
  segment. Segment counts and definitions are **not** comparable across that
  break; do not pool a panel straight through it without accounting for the
  change.
- **Segment type codes matter; do not double-count.** Business / operating,
  geographic, and operating segment types overlap. Segment sales summed across
  all types can exceed firm sales. Filter by `stype` to one type before summing,
  or you will double-count.
- **Segments are not a stable panel unit.** Firms self-define their segments and
  re-define them from year to year, so a segment is not a fixed entity over
  time and segment identifiers are not persistent. Do not assume a segment in
  year *t* is the same unit in year *t+1*.
- **Segment items do not reconcile to the consolidated total.** The sum of
  segment-level items does not equal the firm-level total because of
  eliminations, "other / corporate" segments, and unallocated items. Expect a
  residual; do not treat the segment sum as the firm number.
- **Industry codes attached to segments are coarse and move.** The SIC / NAICS
  code on a segment is at a high level and can change across years, so an
  industry-based segment classification is approximate.
- **Coverage is conditional on disclosure.** Small or single-segment firms
  report a single segment, and non-disclosure is not missing-at-random. A
  segment sample is selected toward larger, more diversified firms.
- **Linking needs gvkey plus care.** Joining segments over time, or to
  firm-level Compustat, uses `gvkey` plus careful handling of the segment IDs;
  do not rely on segment identifiers alone to track a segment across years.

## Reference: segment type codes and main fields

| Segment type (`stype`) | What it holds |
|---|---|
| Business / operating | Line-of-business segments (the SFAS 131 reportable segments) |
| Geographic | Sales and assets broken out by region or country |
| Pension / other | Pension and residual segment types reported alongside the above |

| Field | What it holds |
|---|---|
| `gvkey` | Firm identifier (links to firm-level Compustat) |
| Segment sales | Revenue attributed to the segment |
| Operating profit | Segment operating income |
| Identifiable assets | Assets attributed to the segment |
| Capital expenditure | Segment capex |
| Depreciation | Segment depreciation |
| SIC / NAICS | Segment industry classification |

## Citation

Cite the provider and the files, e.g.: *S&P Compustat Segment files (via WRDS),
accessed YYYY-MM-DD.* State the tables used (for example `wrds_segmerged` /
`seg_annual`), the segment types retained, and the extraction date.


==============================================================================
# Consensus Economics forecast surveys (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/consensus-economics/
# Consensus Economics surveys a panel of professional forecasters for cross-country macro and exchange-rate projections at several horizons. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: forecasts, exchange-rates, macro, survey-expectations, licensed, data:consensus-economics
==============================================================================

:::caution[Licensed: not exercised here]
**Consensus Economics surveys are a paid licensed product** (Consensus Economics
Inc.), so they carry **no provenance badge**: the
access path below was **not** run in this session (no subscription was
available). The page documents the access route and the gotchas; treat it as
unverified until someone exercises it through a licensed subscription. This is
the honest grade under the institute's Verified discipline.
:::

**Consensus Economics** runs monthly surveys of professional forecasters across
many countries, collecting projections for GDP growth, inflation, interest rates,
and **exchange rates** at fixed horizons (commonly 1, 3, 12, and 24 months, plus
fixed-year targets). It is the standard cross-country source for survey-based
expectations, widely used to test exchange-rate predictability and to measure
disagreement. A paper we distill uses it:
[Kremens, Martin & Varela](/wiki/papers/jf/2025/kremens-long-horizon-exchange-rate-2025/)
use Consensus Economics monthly consensus exchange-rate forecasts at the 1-, 3-,
12-, and 24-month horizons as the primary survey expectations series in their
long-horizon exchange-rate analysis.

- **Cost:** licensed, subscription. No free tier.
- **Publisher:** Consensus Economics Inc. (Consensus Forecasts; Foreign Exchange
  Consensus Forecasts).
- **Coverage:** professional-forecaster panels across major and many emerging
  economies, monthly, with consensus and individual responses at several horizons.

## Access (when licensed)

- **Through the Consensus Economics subscription.** The forecasts are distributed
  by subscription; historical individual-forecaster microdata is obtained through
  the publisher or a research arrangement.
- **Consensus is published; individual responses are the research asset.** The
  headline consensus is in the publications; forecaster-level responses (for
  dispersion) require the microdata.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Fixed-horizon versus fixed-event targets.** Some series are fixed horizons
  (for example 12 months ahead) and some are fixed calendar-year targets, which
  have a shrinking horizon through the year; mixing them misaligns the realized
  value. Confirm the target type per series.
- **Survey date and the information set.** Forecasts reflect the information at the
  survey collection date, not the publication date; predictability tests must use
  the collection-date information set. Align timing carefully.
- **Exchange-rate quote direction.** FX forecasts can be quoted as units of
  foreign per domestic or the reverse, and differ by country; a wrong convention
  flips the predicted change. Verify the quote direction per currency.
- **Panel composition changes.** Forecasters enter and leave; consensus and
  dispersion shift with the panel, not only with beliefs. Track membership.
- **Country and variable coverage is uneven.** Emerging-market and longer-horizon
  coverage is thinner and starts later; treat the panel as unbalanced. Pin the
  available window per country and horizon.

## Citation

Cite the publisher and product, e.g.: *Consensus Economics (Consensus Forecasts /
Foreign Exchange Consensus Forecasts), accessed YYYY-MM-DD.* State the country,
variable, horizon, target type (fixed-horizon vs fixed-year), and survey-date
alignment.


==============================================================================
# CoreLogic: property and housing microdata (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/corelogic/
# CoreLogic (Cotality) is a US property database: deed transactions, tax and assessor records, repeat-sales house-price indices, and foreclosure data at the property and zip-code level. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: real-estate, housing, microdata, licensed, panel-data, data:corelogic
==============================================================================

:::caution[Licensed: not exercised here]
**CoreLogic is a licensed commercial dataset** (CoreLogic, rebranded Cotality), so it carries **no provenance badge**: the
access path below was **not** run in this session (no CoreLogic credentials were
available). The page documents the access route and the gotchas; treat it as
unverified until someone exercises it through a licensed account. This is the
honest grade under the institute's Verified discipline.
:::

**CoreLogic** is a US **property and housing microdata** provider. Its
research-relevant content includes property-level **deed** records (arms-length
sale transactions with prices and dates), **tax and assessor** files (property
characteristics and assessed values), **repeat-sales** house-price indices, and
**foreclosure** data at the property and zip-code level. It appears across
housing-finance research: used in, for example
[Amaral, Dohmen, Kohl & Schularick](/wiki/papers/jf/2025/amaral-superstar-returns-spatial-heterogeneity-2025/)
for deeds and repeat-sales transactions across 248 MSAs to estimate
idiosyncratic housing-price risk,
[Piazzesi](/wiki/papers/jf/2025/piazzesi-presidential-address-housing-betas-2025/)
for the cross-section of idiosyncratic capital gains on individual houses, and
[Heitz, Martin & Ufier](/wiki/papers/jf/2026/heitz-bank-monitoring-onsite-inspections-2026/)
for zip-code monthly foreclosure rates as a local-stress measure.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** CoreLogic (rebranded Cotality in 2025).
- **Coverage:** US residential and commercial property, with deed, assessor,
  HPI, and foreclosure modules; county-level depth and history vary.

## Access (when licensed)

- **Direct from CoreLogic / Cotality.** Through their data files, platform, or a
  cloud data share, keyed on the CoreLogic property identifier (CLIP).
- **Possibly via a research data provider.** Some institutions reach CoreLogic
  microdata through licensed academic arrangements; availability is not
  confirmed here, so check with your library before assuming a given path.
- Credentials are required either way. Keep them in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **County coverage is uneven across time and geography.** Recording practices
  differ by county, and several non-disclosure states (for example Texas) do not
  release transaction prices, so price coverage has structural holes. Check
  coverage for your counties and period before reading gaps as no activity.
- **Deed and assessor files are separate and must be joined.** Transactions
  (deeds) and property characteristics (assessor) live in different files; join
  on the CoreLogic property identifier (CLIP) and confirm the match before
  treating a sale as fully attributed.
- **Repeat-sales needs at least two arms-length sales.** Repeat-sales indices and
  price-risk estimates require the same property selling twice at arms length;
  properties that transact once, or only through non-arms-length transfers, drop
  out, which selects toward more-traded properties.
- **Distinguish arms-length sales from transfers.** The deed file includes
  intra-family transfers, refinancings, and foreclosure-related transfers that
  are not market sales; filter on transaction type before computing returns.
- **Foreclosure data is aggregated geographically.** Foreclosure measures are
  often used at the zip-code or county level; state the geography and the
  monthly definition so the stress measure is reproducible.
- **Files are large and get refreshed.** Property microdata is voluminous and
  periodically restated; pin the data vintage and re-pull deliberately rather
  than mixing refreshes.

## Citation

Cite the product and vendor, e.g.: *CoreLogic property data, CoreLogic
(Cotality); data licensed and accessed YYYY-MM-DD.* State the data vintage, the
module (deed, assessor, HPI, or foreclosure), the geography, and the
transaction-type filter used.


==============================================================================
# CoStar: commercial real estate transactions (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/costar/
# CoStar (CoStar Group) is a commercial-real-estate database: property-level records of completed sales, listings, leases, assessments, and physical/location characteristics across US markets. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: real-estate, commercial-real-estate, licensed, transaction-data, data:costar
==============================================================================

:::caution[Licensed: not exercised here]
**CoStar is a licensed commercial dataset** (CoStar Group), so it carries **no provenance badge**: the access path below was
**not** run in this session (no CoStar credentials were available). The page
documents the access route and the gotchas; treat it as unverified until someone
exercises it through a licensed account. This is the honest grade under the
institute's Verified discipline.
:::

**CoStar** is a **commercial real estate** database published by CoStar Group.
It records property-level data on completed sales, active and historical
listings, leases, assessments, and physical and location characteristics across
US markets. It is a common source in commercial real estate research: used in,
for example
[Nozawa & Tsoy](/wiki/papers/jf/2025/nozawa-counter-markets-nonstandardized-assets-2025/)
for completed CRE sales across 15 US cities, 1998Q1 to 2022Q3, including sale
price, assessment value, and property characteristics, with days-to-sell as a
liquidity proxy.

- **Cost:** licensed, subscription (broker/investor product; academic access
  via a data agreement or negotiated extract). No free tier.
- **Vendor:** CoStar Group.
- **Coverage:** US commercial properties (office, retail, industrial,
  multifamily) by metro and submarket, with sale transactions, listing dates,
  and property attributes; depth varies by market and era.

## Access (when licensed)

- **Directly from CoStar.** The common path is the CoStar platform, with
  exports or a negotiated data extract arranged through CoStar Group. There is
  no standard WRDS path for CoStar.
- Credentials and a valid license are required. Keep them in `.env`, never
  hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Coverage is a selected sample, not a census.** CoStar records marketed or
  observed properties and skews toward larger metros and institutional-grade
  assets. Absence of a property in the data is not evidence that the property
  does not exist or that no transaction occurred.
- **Listing and delisting dates can be noisy or missing.** Any days-to-sell
  measure depends on these dates, which are entered and maintained by brokers.
  Verify the completeness of listing-date fields for your markets before
  constructing liquidity measures.
- **Prices mix multiple sources.** The price field can reflect a reported sale
  price, a broker estimate, or an assessment value. Keep the price type
  explicit and consistent; do not blend types without flagging them.
- **The live platform is continuously updated and overwritten.** Reproducing a
  historical pull requires a dated extract: a fresh pull from the platform will
  not match an older one. Archive the extract date and its scope.
- **Property and submarket definitions are CoStar-specific and not stable
  across vintages.** The CoStar property identifier and submarket boundaries
  can shift between platform versions; pin the vintage before linking across
  files or time periods.
- **Duplicate and revised listings need de-duplication.** A single property can
  appear under multiple listing records (original, revised, relisted); decide
  on a de-duplication rule before constructing the analysis sample and document
  it.

## Citation

Cite the product and vendor, stating the extract date and vintage, the markets
and time period covered, the price field used, and how listing and delisting
dates were cleaned, e.g.: *CoStar commercial real estate data, CoStar Group,
extract dated YYYY-MM-DD; markets: [list]; period: [start] to [end]; price
field: [field name]; listing-date cleaning: [description].*


==============================================================================
# Crane Data: money market fund holdings and assets (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/crane-mmf/
# Crane Data LLC is a money-market-fund (MMF) data service covering monthly fund-level total net assets, yields, and portfolio holdings (instrument type, issuer, maturity) for US money-market funds. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: money-market-funds, licensed, fund-data, data:crane-mmf
==============================================================================

:::caution[Licensed: not exercised here]
**Crane Data is a licensed commercial dataset** (Crane Data LLC), so it carries **no provenance badge**: the access path
below was **not** run in this session (no Crane Data subscription was
available). The page documents the access route and the gotchas; treat it as
unverified until someone exercises it through a licensed account. This is the
honest grade under the institute's Verified discipline.
:::

**Crane Data** is a **money-market-fund (MMF)** data service, operated by
Crane Data LLC. It provides monthly fund-level total net assets (AUM), yields,
and portfolio holdings (instrument type, issuer, maturity), with the
portfolio-weight detail needed to study fund composition (e.g. Treasury-bill
and Fed reverse-repo / RRP weights). It is a source in MMF and short-term
markets research: used in, for example
[Siriwardane, Sunderam & Wallen](/wiki/papers/jf/2025/siriwardane-segmented-arbitrage-2025/)
(MMF total net assets and holdings for the money-fund-reform analysis,
cross-checked with SEC Form N-MFP), and
[Stein & Wallen](/wiki/papers/jf/2025/stein-imperfect-intermediation-money-like-2025/)
(monthly money-fund holdings and AUM, T-bill and RRP portfolio weights,
fund-level elasticity estimation, and an AUM decomposition).

- **Cost:** licensed, subscription.
- **Vendor:** Crane Data LLC.
- **Coverage:** US money-market funds, monthly, fund and holding level. The
  free public counterpart is SEC Form N-MFP (the monthly MMF portfolio filing
  on EDGAR), which carries much of the same holdings detail but needs assembly.

## Access (when licensed)

- **Directly from Crane Data LLC.** Data are delivered by subscription via file
  delivery or downloads. License required; contact Crane Data LLC for terms.
- **Free alternative.** SEC Form N-MFP filings are available on EDGAR and cover
  similar holdings detail. The trade-off is more processing: filings must be
  fetched, parsed, and assembled into a panel before use.
- Credentials and subscription details should be stored in `.env`, never
  hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Commercial repackaging of regulatory filings.** Crane Data cleans and
  repackages fund filings; convenient, but the underlying source for holdings
  is the same N-MFP filings on EDGAR. Cross-check against EDGAR when accuracy
  matters for a specific holding or vintage.
- **Fund vs. share-class structure.** AUM and yields are reported at the
  share-class and fund levels. Decide the unit of analysis before aggregating
  and avoid double-counting share classes when computing fund-level totals.
- **Monthly snapshots only.** Holdings are month-end snapshots; intra-month
  dynamics and same-day repo turnover are not captured.
- **Proprietary instrument and issuer taxonomy.** Crane's instrument and issuer
  classifications are their own taxonomy. Map them explicitly before merging
  with other sources; do not assume they align with EDGAR N-MFP field codes or
  other data vendors.
- **Fund identity changes over time.** Fund-name and identifier changes across
  mergers and liquidations break a naive panel. Build a crosswalk before
  treating the fund identifier as a stable entity key.
- **Coverage is not a regulatory census.** Crane tracks a broad universe of US
  MMFs, but it is not identical to the full SEC-registered population the way
  N-MFP is. Funds near launch or wind-down may appear or disappear differently
  than in EDGAR.

## Citation

Cite Crane Data LLC money market fund data, stating the vintage (date range
pulled), the unit of observation (fund level or share-class level), and whether
holdings were cross-checked against SEC Form N-MFP filings on EDGAR.


==============================================================================
# CRSP Mutual Funds + Thomson holdings: the survivor-bias-free fund panel (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/crsp-mutual-funds/
# The CRSP Survivor-Bias-Free Mutual Fund Database (returns, TNA, fees, CRSP holdings) and the Thomson Reuters Mutual Fund Holdings (s12), linked by MFLINKS, are the standard US open-end mutual-fund panel. Licensed via WRDS: this page documents the access path and the gotchas; the path was exercised through a licensed WRDS session.
# Access confirmed (licensed) 2026-06-24 · via live crsp_q_mutualfunds (monthly_tna_ret_nav, holdings, fund_fees) + tr_mutualfunds.s12 + mfl.mflink1 queries through a licensed WRDS session
# Tags: equities, mutual-funds, institutional-investors, panel-data, licensed, wrds, data:crsp-mutual-funds
==============================================================================

:::note[Access confirmed via a licensed WRDS session]
The keystone queries against `crsp_q_mutualfunds.monthly_tna_ret_nav`,
`crsp_q_mutualfunds.holdings`, `crsp_q_mutualfunds.fund_fees`,
`tr_mutualfunds.s12`, and `mfl.mflink1` were run through a licensed WRDS
session on 2026-06-24 and returned real rows, so this page carries the amber
"Access confirmed (licensed)" badge. Reproducing it still requires the reader's
own WRDS account and the institution's CRSP and Thomson/Refinitiv entitlements
(licensed, not open).
:::

**The CRSP Survivor-Bias-Free Mutual Fund Database** (`crsp_q_mutualfunds` on
WRDS) is the standard US open-end mutual-fund panel: monthly and daily returns,
NAV, total net assets (TNA), expense ratios, style codes, flows, and
portfolio-level equity holdings. The **Thomson Reuters Mutual Fund Holdings**
database (`tr_mutualfunds`, the `s12` family) is the older holdings source.
**MFLINKS** (`mfl`) is the crosswalk that ties the two together. Together they
are the standard source for the mutual-fund performance, flow, and
active-management literature.

- **Cost:** licensed (via WRDS). No free cleaned tier.
- **Vendors:** CRSP (Center for Research in Security Prices) and Thomson
  Reuters / Refinitiv (LSEG); MFLINKS maintained by WRDS.
- **Coverage:** US open-end mutual funds; survivor-bias-free (dead funds
  retained). CRSP returns run from the 1960s; CRSP holdings begin in the early
  2000s; Thomson `s12` holdings cover the earlier period back to 1980.
- **Unit:** share class (`crsp_fundno`) for returns, fees, and TNA; portfolio
  (`crsp_portno`) for holdings.

## Access (when licensed)

- **Through WRDS.** Query the `crsp_q_mutualfunds`, `tr_mutualfunds`, and
  `mfl` libraries; see [WRDS / CRSP / Compustat](/wiki/commercial/wrds/) for the
  connection recipe and the persistent-server pattern that fires Duo 2FA once.

```python
from utils.wrds_client import wrds_query, wrds_start
wrds_start()                                   # no-op if already running
df = wrds_query("""
    SELECT crsp_fundno, caldt, mret, mtna, mnav
    FROM crsp_q_mutualfunds.monthly_tna_ret_nav
    WHERE caldt >= '2010-01-01'
    LIMIT 5
""")
```

- Credentials are required. Keep them in `.env` (`WRDS_USER`, `WRDS_PASS`),
  never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented.

- **Don't eager-load large local parquets.** The `crsp_q_mutualfunds.holdings`
  table is roughly 438M rows and the Thomson `s12` fund-holdings panel runs to
  ~150M+ rows; daily NAV/return is ~182M rows. After you cache a pull to local
  parquet, never reload it with a whole-file `pd.read_parquet(<file>)` -- it
  will OOM-kill the process. Stream it instead:
  `polars.scan_parquet(path).select([...]).filter(...).collect()`
  (column projection + predicate pushdown) so you filter before materializing
  and never hold the full table in RAM. Requires `polars` + `pyarrow`. The same
  rule applies at query time: always filter `holdings` by `crsp_portno` and a
  date range, never `SELECT *`.
- **Share class vs portfolio: aggregate or you double-count.** A fund's Class A
  and Class I are separate `crsp_fundno` rows with identical underlying assets
  but different fees, TNA, and investor types. Returns, fees, and TNA live at
  the share-class level; holdings live at the portfolio level (`crsp_portno`).
  For most analyses aggregate share classes to the portfolio (TNA-weighted
  returns, summed TNA and flows) using `crsp_q_mutualfunds.portnomap`. Forgetting
  this overweights funds with many share classes.
- **Linking CRSP to Thomson needs MFLINKS, and the keys differ.** CRSP uses
  `crsp_fundno`; Thomson uses its own `fundno`. The bridge is `wficn` (the
  Thomson/Wharton fund identifier) via `mfl.mflink1` (`crsp_fundno` -> `wficn`)
  and `mfl.mflink2` (`wficn` -> Thomson `fundno`). Do not try to match on names
  or tickers. MFLINKS coverage is incomplete and stops being updated at its last
  vintage, so newer funds can be unlinkable.
- **Two holdings sources with different vintages.** CRSP holdings
  (`crsp_q_mutualfunds.holdings`) are already matched to CRSP `permno` and are
  the more recent series; Thomson `s12` is legacy, CUSIP-keyed, and needed for
  the earlier sample. Mixing them without reconciling identifiers and report
  dates produces gaps and duplicates.
- **TNA is missing in spots and flows are implied, not reported.** `mtna` has
  gaps (it is NaN for some months); standard fund flows are computed, not given:
  `flow_t = (TNA_t - TNA_{t-1} * (1 + ret_t)) / TNA_{t-1}` (Sirri and Tufano,
  1998). Decide how to treat missing TNA before computing flows or returns
  divide by a missing base.
- **Filter for fund type and survivorship explicitly.** The database mixes
  equity, bond, balanced, index, and exchange-traded products. Use `crsp_obj_cd`
  or `lipper_class` to isolate the fund type, `et_flag = 'N'` to drop ETFs, and
  the index-fund flag to drop index funds. Because the panel is
  survivor-bias-free, performance studies must **include** dead funds
  (`dead_flag = 'Y'`); excluding them reintroduces the very bias the database
  removes.

## Key tables (reference)

### Returns, NAV, TNA (share-class level)

| Table | Description | Frequency | Key columns |
|---|---|---|---|
| `crsp_q_mutualfunds.monthly_tna_ret_nav` | Monthly returns, TNA, NAV | Monthly | `crsp_fundno`, `caldt`, `mret`, `mtna`, `mnav` |
| `crsp_q_mutualfunds.daily_nav_ret` | Daily returns and NAV | Daily | `crsp_fundno`, `caldt`, `dret`, `dnav` |
| `crsp_q_mutualfunds.monthly_returns` | Monthly returns only | Monthly | `crsp_fundno`, `caldt`, `mret` |

### Characteristics and fees

| Table | Description | Key columns |
|---|---|---|
| `crsp_q_mutualfunds.fund_hdr` | Fund header (name, ticker, CUSIP, dates, flags) | `crsp_fundno`, `fund_name`, `ticker`, `first_offer_dt`, `dead_flag`, `et_flag`, `index_fund_flag` |
| `crsp_q_mutualfunds.fund_style` | Style/objective codes over time | `crsp_fundno`, `begdt`, `enddt`, `crsp_obj_cd`, `lipper_class` |
| `crsp_q_mutualfunds.fund_fees` | Expense ratios, management fees, turnover | `crsp_fundno`, `begdt`, `enddt`, `exp_ratio`, `mgmt_fee`, `turn_ratio`, `actual_12b1` |
| `crsp_q_mutualfunds.fund_flows` | Sales and redemptions (portfolio level) | `crsp_portno`, `report_dt`, `new_sls`, `rein_sls`, `redemp` |

### Holdings and links

| Table | Description | Key columns |
|---|---|---|
| `crsp_q_mutualfunds.holdings` | Portfolio equity holdings (CRSP, ~438M rows) | `crsp_portno`, `report_dt`, `permno`, `percent_tna`, `nbr_shares`, `market_val` |
| `tr_mutualfunds.s12` | Thomson quarterly mutual-fund holdings (~150M+ rows) | `fundno`, `fdate`, `rdate`, `cusip`, `shares`, `change` |
| `tr_mutualfunds.s12type1` | Thomson Type 1 holdings (complete reports) | `fundno`, `fdate`, `cusip`, `shares` |
| `mfl.mflink1` | `crsp_fundno` -> `wficn` | `crsp_fundno`, `wficn` |
| `mfl.mflink2` | `wficn` -> Thomson `fundno` | `wficn`, `fundno` |
| `crsp_q_mutualfunds.portnomap` | `crsp_fundno` -> `crsp_portno` | `crsp_fundno`, `crsp_portno` |

### Common CRSP objective codes

| Code | Description |
|---|---|
| `EDCI` | Equity domestic, capital appreciation |
| `EDYB` | Equity domestic, equity income |
| `EDYM` | Equity domestic, income and growth |
| `IC` / `I` | International (core / general) |
| `M` | Mixed / balanced |
| `OB` | Bond, general |

Use `crsp_obj_cd LIKE 'ED%'` for all domestic-equity funds.

## Standard operations

### Equity-fund monthly returns with style

```python
funds = wrds_query("""
    SELECT a.crsp_fundno, a.caldt, a.mret, a.mtna,
           b.crsp_obj_cd, b.lipper_class
    FROM crsp_q_mutualfunds.monthly_tna_ret_nav AS a
    LEFT JOIN crsp_q_mutualfunds.fund_style AS b
      ON a.crsp_fundno = b.crsp_fundno
      AND a.caldt BETWEEN b.begdt AND b.enddt
    WHERE a.caldt >= '1990-01-01'
      AND b.crsp_obj_cd LIKE 'ED%'      -- domestic equity
""")
```

### Aggregate share classes to the portfolio

```python
portmap = wrds_query("SELECT crsp_fundno, crsp_portno FROM crsp_q_mutualfunds.portnomap")
import pandas as pd
merged = funds.merge(portmap, on='crsp_fundno')
portfolio = (merged.groupby(['crsp_portno', 'caldt'])
    .apply(lambda g: pd.Series({
        'wret': (g['mret'] * g['mtna']).sum() / g['mtna'].sum() if g['mtna'].sum() > 0 else None,
        'tna': g['mtna'].sum()})).reset_index())
```

### Link CRSP funds to Thomson holdings

```python
link = wrds_query("""
    SELECT l1.crsp_fundno, l1.wficn, l2.fundno
    FROM mfl.mflink1 AS l1
    JOIN mfl.mflink2 AS l2 ON l1.wficn = l2.wficn
    WHERE l1.wficn IS NOT NULL
""")
holdings = wrds_query("""
    SELECT fundno, fdate, cusip, shares
    FROM tr_mutualfunds.s12type1
    WHERE fdate >= '2000-01-01'
""")
```

### Stream a cached holdings parquet (don't eager-load)

```python
import polars as pl
sub = (pl.scan_parquet("data/mutual_funds/holdings.parquet")
         .select(["crsp_portno", "report_dt", "permno", "percent_tna"])
         .filter((pl.col("crsp_portno") == 1000001) &
                 (pl.col("report_dt") >= pl.date(2015, 1, 1)))
         .collect())
```

## Citation

Cite each underlying provider, not WRDS itself, e.g.: *CRSP Survivor-Bias-Free
US Mutual Fund Database, Center for Research in Security Prices, LLC, accessed
via WRDS, YYYY-MM-DD*; *Thomson Reuters Mutual Fund Holdings (s12), accessed via
WRDS, YYYY-MM-DD*; and MFLINKS where used. State the date range, the fund-type
filter, whether dead funds are included, whether figures are at the share-class
or portfolio level, and how missing TNA and implied flows were handled.


==============================================================================
# Crunchbase: startup and funding data (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/crunchbase/
# Crunchbase is a commercial database of startups, funding rounds, investors, and company characteristics. A limited free tier exists, but research-grade bulk access is licensed. This page documents the access path and the gotchas, but the data was not exercised here.
# Tags: startups, venture-capital, private-firms, funding, licensed, data:crunchbase
==============================================================================

:::caution[Licensed: not exercised here]
**Crunchbase research access is a paid licensed product** (Crunchbase, Inc.):
the public site exposes a limited free view, but bulk/API access for research is licensed, so this page carries **no provenance badge**. The
access path below was **not** run in this session (no Crunchbase API key or Pro
licence was available). The page documents the access route and the gotchas;
treat it as unverified until someone exercises it through a licensed account.
This is the honest grade under the institute's Verified discipline.
:::

**Crunchbase** is a commercial database of private and public companies focused
on the startup ecosystem: company profiles (founding year, industry, location),
funding rounds (amount, date, stage), investors, acquisitions, and key people.
It is one of the two standard vendor sources for startup and venture-funding
characteristics, the other being [PitchBook](/wiki/commercial/pitchbook/); the two
are frequently used together in the same study. A paper we distill uses it:
[Hu & Ma](/wiki/papers/jf/2025/hu-persuading-investors-video-based-2025/) draw
startup characteristics (founding year, industry, location, funding rounds and
amounts, investor count) from Crunchbase, and separately use
[PitchBook](/wiki/commercial/pitchbook/) business descriptions for idea-novelty
scoring.

- **Cost:** a limited free tier exists; research-grade bulk export and the
  full API require a paid Crunchbase Pro / Enterprise licence.
- **Vendor:** Crunchbase, Inc.
- **Coverage:** global, strongest for U.S. technology startups and venture
  funding; coverage thins for non-U.S. and non-tech firms and for earlier years.

## Access (when licensed)

- **REST API or bulk data export.** Licensed users pull profiles, funding
  rounds, and investor records through the Crunchbase API (entity and search
  endpoints) or a bulk CSV/data export, keyed by Crunchbase's permalink/UUID
  identifiers.
- **Free tier is for lookup, not bulk research.** The public website and the
  free tier allow individual lookups but rate-limit and cap bulk pulls; do not
  build a research panel off the free tier.
- API credentials are required for programmatic access. Keep any credentials in
  `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Crowd-sourced, self-reported entries.** Much of Crunchbase originates from
  user and company submissions, so fields are unevenly populated and can be
  promotional or stale. Treat missing funding amounts as missing, not zero, and
  expect duplicate or merged company records.
- **Selection toward funded, successful, U.S. tech startups.** Companies that
  raise visible rounds and generate press are over-represented; failed or
  bootstrapped firms are under-covered. A sample built from Crunchbase presence
  is a survivorship/visibility-biased sample, not a census of startups.
- **Undisclosed and rounded funding amounts.** Round sizes are often undisclosed
  or reported as ranges/round numbers; the "total funding" field aggregates only
  what was disclosed. Do not treat it as audited capital raised.
- **Date ambiguity.** Founding dates and round announcement dates can differ from
  the actual close date and are sometimes only year-precise; be explicit about
  which date you use for timing.
- **Merging with PitchBook double-counts.** The two vendors overlap heavily but
  use different identifiers, round definitions, and amounts; a naive union
  double-counts rounds and conflicts on amounts. Dedupe on a deliberate match key
  and pick a precedence rule for conflicting fields.
- **Schema and history changes.** Crunchbase has changed its data model and API
  versions over time; a panel assembled across vintages can have inconsistent
  fields. Pin the API version and extraction date.

## Reference: representative entities and fields

| Entity | Representative fields |
|---|---|
| Organization | name, founded-on, category/industry, country/region, status |
| Funding round | announced-on, investment type/stage, money raised, currency |
| Investor | investor name, type, number of investments |
| Acquisition | acquirer, acquiree, announced-on, price (if disclosed) |
| Person | name, role, founder/executive affiliations |

Pull the specific entities and fields you need and record the API version and
extraction date; field availability is uneven across firms and years.

## Citation

Cite the provider and database, e.g.: *Crunchbase / Crunchbase, Inc., accessed
YYYY-MM-DD.* State the entity universe, the fields used, the API version or
export vintage, and the extraction date so the sample is reproducible.


==============================================================================
# CSMAR: China Stock Market & Accounting Research (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/csmar/
# CSMAR is the standard vendor database of Chinese listed-firm prices, financials, ownership, and market microstructure. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: china, equities, accounting, ownership, licensed, data:csmar
==============================================================================

:::caution[Licensed: not exercised here]
**CSMAR is a paid licensed product** (GTA Information Technology), so it carries **no provenance badge**: the access path below
was **not** run in this session (no CSMAR / GTA institutional credentials were
available). The page documents the access route and the gotchas; treat it as
unverified until someone exercises it through a licensed account. This is the
honest grade under the institute's Verified discipline.
:::

**CSMAR** (China Stock Market & Accounting Research, from GTA Information
Technology) is the most widely used vendor database for Chinese capital markets:
daily and intraday stock prices and returns for the Shanghai and Shenzhen
exchanges, listed-firm financial statements, ownership and shareholder data,
fund holdings, corporate governance, and a range of market-microstructure and
specialty modules. It is the China-market analog to the CRSP/Compustat core and
is the standard source for studies of Chinese listed firms. A paper we distill
uses it:
[Hansman, Hong, Jiang, Liu & Meng](/wiki/papers/rfs/2025/hansman-effects-credit-expansions-stock-2025/)
draw stock prices and trading data, book-to-market, semiannual fund holdings,
and top-ten-shareholder ownership from CSMAR.

- **Cost:** licensed, institutional subscription. No general free tier.
- **Vendor:** GTA Information Technology (Shenzhen).
- **Coverage:** A-share, B-share, and (in places) H-share listed firms on the
  Shanghai and Shenzhen exchanges; prices from the early 1990s, with financials,
  ownership, fund, and governance modules of varying start dates.

## Access (when licensed)

- **Through the CSMAR/GTA platform or WRDS.** Tables are downloaded from the
  CSMAR web platform (module-by-module query and CSV export) or, where the
  institution licenses it, through [WRDS](/wiki/commercial/wrds/). Check which
  route your institution provides.
- **Modular schema.** CSMAR is organized into databases/modules (trading,
  financial statements, ownership, funds, governance, etc.); an extract is a set
  of fields from one module for a firm list over a date range, keyed by the
  six-digit stock code.
- Platform credentials are required. Keep any credentials in `.env`, never
  hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Stock-code reuse and exchange prefixes.** The six-digit ticker is not a
  stable permanent identifier: codes can be reused after delisting, and the same
  number can exist on different exchanges. Use CSMAR's internal firm identifier
  and an exchange prefix, not the raw six-digit code, as the join key.
- **Share classes split the same firm.** A and B (and offshore H) shares of one
  company trade under different codes, currencies, and investor bases; aggregating
  them naively double-counts the firm or mixes currencies. Decide your share-class
  treatment explicitly.
- **Module boundaries and frequency mismatches.** Holdings and ownership are
  often semiannual or quarterly while prices are daily; merging across modules
  requires aligning report dates and disclosure lags. Do not assume a common
  frequency across modules.
- **Accounting-standard and regime breaks.** Chinese listed-firm financials span
  the pre-2007 PRC GAAP regime and the post-2007 IFRS-converged CAS; statement
  items are not directly comparable across the break. Watch the standard-change
  date when building long accounting panels.
- **Trading halts, limits, and special treatment.** Daily price limits, frequent
  trading suspensions, and ST/*ST special-treatment status produce stale or
  missing prices and truncated returns; treat suspended days carefully rather
  than carrying prices forward.
- **English/Chinese field labels and code tables.** Field names and categorical
  codes are documented in CSMAR's manuals (often Chinese-first); misreading a
  code table silently mislabels a variable. Keep the module codebook for the
  exact vintage you pulled.

## Reference: representative modules

| Module | Contents |
|---|---|
| Trading | Daily/intraday prices, returns, volume, market cap |
| Financial statements | Balance sheet, income, cash flow for listed firms |
| Ownership | Top-ten shareholders, ownership structure, controllers |
| Funds | Mutual-fund holdings and characteristics |
| Governance | Board, executive, and related-party data |

Pull the specific module and fields you need, record the module/version, and
keep the codebook; coverage and start dates differ by module.

## Citation

Cite the provider and database, e.g.: *China Stock Market & Accounting Research
(CSMAR) / GTA Information Technology, accessed YYYY-MM-DD.* State the modules and
fields used, the firm/share-class universe, and the extraction date so the sample
is reproducible.


==============================================================================
# Refinitiv Datastream: global time-series of prices and macro series (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/datastream/
# Datastream is Refinitiv's (LSEG) historical financial time-series database covering equities, bonds, commodities, indices, exchange rates, interest rates, options/futures, and a large library of macroeconomic series across many countries. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: equities, international, macro, time-series, licensed, data:datastream
==============================================================================

:::caution[Licensed: not exercised here]
**Datastream is a paid licensed product** (Refinitiv / LSEG), so it carries **no provenance badge**: it is not on
[WRDS](/wiki/commercial/wrds/) for most institutions, and the access path below was
**not** run in this session (no Refinitiv / LSEG credentials were available). The
page documents the access route and the gotchas; treat it as unverified until
someone exercises it through a licensed account. This is the honest grade under
the institute's Verified discipline.
:::

**Datastream** is Refinitiv's (LSEG) historical financial time-series database. It
covers **equities, bonds, commodities, indices, exchange rates, interest rates,
and options/futures**, alongside a large library of **macroeconomic series**
across many countries, with long histories (some series run back to the 1960s
and 1970s). Data is accessed by **series mnemonic** and **datatype code**. It is
a standard source for international asset prices and macro series. A paper we
distill uses it:
[Costain et al.](/wiki/papers/jf/2025/costain-term-structure-interest-rates-2025/)
on the term structure of interest rates.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** Refinitiv / LSEG (lineage: Thomson Reuters Datastream).
- **Coverage:** global, multi-asset and macro; daily, weekly, and monthly
  frequencies; deep history but uneven by series.

## Access (when licensed)

- **Refinitiv / LSEG, not WRDS.** Datastream is usually reached through a
  Refinitiv terminal, the **Datastream for Office** Excel add-in, or the
  **Datastream API / DSWS** (DataStream Web Service); it is not generally
  reached through [WRDS](/wiki/commercial/wrds/). Check whether your institution
  licenses it separately.
- **Pulled by mnemonic plus datatype.** Each request names a series mnemonic and
  a datatype code, for example `RI` (total-return index), `P` (price), or `MV`
  (market value), plus a frequency and date range.
- A terminal seat or API credentials are required. The **DSWS** service has
  Python and R clients for programmatic extraction. Keep any credentials in
  `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Static / padded data.** Datastream carries forward the last value on
  non-trading days and after a stock delists, producing spurious zero returns
  and flat stretches. Screen these out: a run of repeated identical values, or
  zero returns after the last live date, is padding, not data.
- **Survivorship and dead stocks.** Research lists and constituent lists differ,
  and using current index constituents introduces survivorship bias. Use the
  dead / research lists so delisted securities stay in the sample.
- **Data errors at the extremes.** There are well-documented spikes and
  reversals (a price that jumps and then reverts) that inflate small-stock
  returns. The standard Ince and Porter screens trim returns above set
  thresholds and delete suspicious reversals before any analysis.
- **RI versus P.** Use `RI` (return index) for total returns; `P` (price)
  excludes dividends and has its own adjustment quirks. Mixing them understates
  returns by the dividend component.
- **Currency.** Series come in local currency or a chosen currency, and the
  conversion / datatype must be set explicitly. Do not assume the default
  currency; record which one was requested.
- **Mnemonics are not stable.** The same economic series can have multiple codes,
  and codes change over time. Document the exact mnemonic, datatype, frequency,
  and extraction date so the extract is reproducible.
- **Market-value and shares fields lag.** `MV` and number-of-shares fields lag
  and are periodically restated; do not treat a recent value as final or assume
  point-in-time accuracy.

## Reference: common datatypes

| Datatype | What it holds |
|---|---|
| `RI` | Total-return index (price plus reinvested dividends); use for returns |
| `P` | Price (unadjusted for dividends) |
| `MV` | Market value (market capitalisation) |
| `NOSH` | Number of shares |
| `UP` | Unadjusted price |
| `DY` | Dividend yield |
| `PE` | Price-earnings ratio |

## Citation

Cite the provider and the database, e.g.: *Refinitiv (LSEG) Datastream, accessed
YYYY-MM-DD.* State the exact mnemonics, datatypes, and frequency requested, and
the extraction date. When you apply the standard data screens, cite Ince and
Porter (2006), "Individual Equity Return Data from Thomson Datastream: Handle
with Care!", *Journal of Financial Research*.


==============================================================================
# LPC DealScan: syndicated-loan data (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/dealscan/
# DealScan (LSEG / LPC) is a deal-level database of syndicated and large corporate loans: facility pricing, amounts, maturities, covenants, and lender shares, reached by most researchers through WRDS. It is licensed: this page documents the access path and the gotchas; the access path was exercised through a licensed WRDS session.
# Access confirmed (licensed) 2026-06-09 · via live DealScan query (dealscan.currfacpricing) through a licensed WRDS session
# Tags: credit, syndicated-loans, banking, licensed, deal-data, data:dealscan
==============================================================================

:::note[Access confirmed via a licensed WRDS session]
The keystone query `SELECT * FROM dealscan.currfacpricing LIMIT 1` was run
against the live dataset through a licensed WRDS session on 2026-06-09,
so this page carries the amber "Access confirmed (licensed)" badge. Reproducing
the access path still requires the reader's own WRDS account and the
institution's DealScan entitlement (licensed, not open).
:::

**DealScan** is a **syndicated-loan** database, built by the Loan Pricing
Corporation (now part of LSEG / Refinitiv). It records corporate loan
**deals** (packages) and the **facilities** (tranches) within them: amount,
all-in spread and pricing, maturity, loan type and purpose, covenants, and the
lender syndicate with each lender's share. It is a common source in
bank-lending and loan-pricing research: used in, for example
[Flanagan](/wiki/papers/jf/2025/flanagan-value-bank-lending-2025/) for the value
of bank lending,
[Anderson, Du & Schlusche](/wiki/papers/jf/2025/anderson-arbitrage-capital-global-banks-2025/)
for global-bank loan exposures, and
[Griffin, Nini & Smith](/wiki/papers/jf/2026/griffin-loan-covenant-violations-decline-2026/)
for loan covenant structure.

- **Cost:** licensed, subscription. No free tier. Most academics reach it
  through [WRDS](/wiki/commercial/wrds/).
- **Vendor:** LSEG / Refinitiv (Loan Pricing Corporation).
- **Coverage:** syndicated and large bilateral corporate loans globally, with
  the deepest coverage of the US syndicated market; history back to the late
  1980s, thinner early on.

## Access (when licensed)

- **Through WRDS.** The common academic path is the WRDS DealScan tables; see
  [WRDS / CRSP / Compustat](/wiki/commercial/wrds/) for the connection recipe.
- **Direct from LSEG.** Through LSEG's loan-data products and the LoanConnector
  lineage, keyed on DealScan facility and package identifiers.
- Credentials are required either way. Keep them in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented.

- **It is an origination snapshot, not loan performance.** DealScan records the
  deal as arranged; it does not track drawn balances, repayments, or defaults
  over the life of the loan. Do not read it as a performance panel.
- **The DealScan-to-Compustat link is a separate, imperfect step.** Matching
  borrowers to public-firm identifiers uses the legacy Chava-Roberts link, now
  the WRDS DealScan-Compustat linking table, which has coverage gaps and
  vintage breaks. Re-merge deliberately and report match rates.
- **Coverage skews to large syndicated deals.** Small and bilateral loans are
  under-covered, and early-1990s coverage is thin; absence of a loan is not
  absence of borrowing.
- **Amendments create multiple records.** A single facility can appear several
  times as it is amended; decide whether you want origination terms or the
  amended terms and deduplicate accordingly.
- **Package versus facility level.** Many fields live at the facility (tranche)
  level and others at the package (deal) level; mixing the two double-counts or
  mis-weights. State the level of analysis.
- **Pricing-field definitions matter.** The headline "all-in-drawn spread"
  bundles fees in a specific way and is missing for many facilities; state the
  pricing field and how you handle missing spreads.
- **Identifier migration.** Legacy LoanConnector identifiers and the newer LSEG
  identifiers coexist across vintages; pin which identifier scheme your pull
  uses before linking across files.

## Citation

Cite the product and vendor, e.g.: *LPC DealScan syndicated-loan data, LSEG /
Refinitiv, accessed via WRDS; data licensed and accessed YYYY-MM-DD.* State the
data vintage, the level of analysis (package or facility), the pricing field,
and the borrower-link table used.


==============================================================================
# Lipper eMAXX fixed-income holdings (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/emaxx/
# Lipper eMAXX (LSEG / Refinitiv) is the standard CUSIP-level database of fixed-income holdings for insurers, mutual funds, ETFs, and annuities. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: fixed-income, bond-holdings, institutional-ownership, corporate-bonds, licensed, data:emaxx
==============================================================================

:::caution[Licensed: not exercised here]
**Lipper eMAXX is a paid licensed product** (LSEG / Refinitiv, Lipper), so it carries **no provenance badge**: the access path
below was **not** run in this session (no eMAXX credentials were available). The
page documents the access route and the gotchas; treat it as unverified until
someone exercises it through a licensed account. This is the honest grade under
the institute's Verified discipline.
:::

**Lipper eMAXX** is a quarterly, CUSIP-level database of fixed-income holdings:
who holds which corporate bond, by par amount, across insurance companies,
mutual funds, ETFs, pension funds, and annuities. It is the standard source for
**bond-level investor composition** (the demand side of the corporate bond
market), built from regulatory filings and fund disclosures and keyed to the
bond's CUSIP. A paper we distill uses it:
[Li & Yu](/wiki/papers/jf/2026/li-investor-composition-corporate-bond-liquidity-2026/)
use Lipper eMAXX quarterly holdings to build a bond-level investor-composition
measure (each holder weighted by how actively it trades) when showing that the
growing presence of short-term investors such as mutual funds and ETFs drives the
rising liquidity component of credit spreads.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** LSEG / Refinitiv (Lipper); lineage through Thomson Reuters.
- **Coverage:** corporate (and other) bond holdings at the CUSIP level for
  insurers, mutual funds, ETFs, pensions, and annuities, quarterly. The citing
  paper reports eMAXX captures roughly 40 to 50 percent of corporate bonds
  outstanding by par.

## Access (when licensed)

- **Through an LSEG / Refinitiv (Lipper) entitlement.** The holdings extracts are
  reached via a Lipper / eMAXX subscription; some institutions license it through
  [WRDS](/wiki/commercial/wrds/). Check which route your institution provides.
- **Keyed by CUSIP and quarter.** An extract is holder identity and par holding
  by bond CUSIP over a date range; aggregate to bond-by-quarter to get
  investor-type shares.
- Terminal or API credentials are required. Keep any credentials in `.env`, never
  hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Partial coverage, not the full holder base.** eMAXX captures roughly half of
  corporate bonds outstanding, not all holders; the unobserved tail (households,
  some foreign and private holders) is missing. A bond with low eMAXX coverage has
  a noisily measured investor composition. The citing paper restricts its
  cross-section to bonds with eMAXX coverage above 20 percent for this reason.
- **Holder-type classification drives the result.** Whether a holder is "short
  term" (mutual fund, ETF) versus "long term" (insurer, pension) is the analysis
  variable; the eMAXX entity classification has edge cases (separately managed
  accounts, fund-of-funds, insurer general versus separate accounts). Audit the
  classification before trusting composition shares.
- **Quarterly snapshots miss intra-quarter trading.** Holdings are point-in-time
  at quarter end; turnover and rebalancing inside the quarter are invisible. Do
  not read a quarter-end snapshot as a flow.
- **CUSIP linking and reorganizations.** Bonds must be linked to Mergent FISD /
  CRSP / Compustat identifiers; the join has known mismatches across calls,
  exchanges, and issuer reorganizations. Verify the link.
- **Benchmark coverage against an aggregate.** Because coverage is partial, check
  eMAXX totals against an external aggregate (for example the Federal Reserve
  Flow of Funds corporate-bond holdings by sector) before treating shares as the
  full market.

## Citation

Cite the vendor and product, e.g.: *Lipper eMAXX / LSEG (Refinitiv), accessed
YYYY-MM-DD.* State the holder universe, the coverage threshold applied, and the
quarter range.


==============================================================================
# Equifax traditional credit-bureau data (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/equifax/
# Equifax traditional consumer credit-bureau records (installment and revolving balances, limits, credit scores) obtained under a commercial research licence. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: consumer-credit, household-finance, credit-bureau, licensed, data:equifax
==============================================================================

:::caution[Licensed: not exercised here]
**Equifax credit-bureau data is a paid licensed product** (Equifax), so it carries **no provenance badge**: the access path below
was **not** run in this session (no Equifax research licence was available), and
the records are individual-level. The page documents the access route and the
gotchas; treat it as unverified until someone exercises it through a licensed
arrangement. This is the honest grade under the institute's Verified discipline.
:::

**Equifax traditional credit-bureau data** is the mainstream consumer-credit
record: installment and revolving balances, credit limits, credit scores, and
balances in good standing across a consumer's tradelines, obtained under a
**commercial research licence**. This page covers the licensed-commercial access
route; the closely related confidential route (the FRBNY Consumer Credit Panel
and other PII extracts) is documented separately at
[Equifax consumer-credit records](/wiki/confidential/equifax-credit/). A paper we
distill uses the licensed traditional bureau:
[Di Maggio, Ma & Williams](/wiki/papers/jf/2025/maggio-red-overdrafts-payday-lending-2025/)
use Equifax for installment borrowing, credit-card balances and limits, credit
scores, and total balances in good standing (a representative 10% sample of
680,856 borrowers, 2005 to 2018), paired with alternative-bureau
[Clarity Services](/wiki/commercial/clarity-services/) records.

- **Cost:** licensed, commercial research arrangement. No free tier; individual
  records.
- **Vendor:** Equifax.
- **Coverage:** U.S. consumers with a traditional credit file (most adults with
  credit history), with tradeline-level balances, limits, and scores; the
  credit-visible population only.

## Access (when licensed)

- **Through a commercial research licence.** Anonymized borrower-level extracts
  are obtained under an Equifax research agreement; raw records are personally
  identifiable.
- **Related confidential route.** The FRBNY Consumer Credit Panel (a random sample
  of Equifax files) and project-specific PII extracts are a separate, more
  restricted path, documented at
  [Equifax consumer-credit records](/wiki/confidential/equifax-credit/).

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Credit-visible population only.** Consumers with thin or no file are
  under-represented or absent; the sample is not the whole adult population. Do not
  generalize to the credit-invisible.
- **Bureau-specific coverage.** A tradeline appears only if the lender reports to
  Equifax; balances and accounts can differ across the three bureaus. Do not treat
  one bureau as the complete liability picture.
- **Traditional credit only.** Equifax sees mainstream installment and revolving
  credit, not payday/alternative products; pair it with an alternative bureau (for
  example [Clarity Services](/wiki/commercial/clarity-services/)) for the nonprime
  segment.
- **Scores and fields are vendor constructs.** Credit scores and account-type codes
  are Equifax constructions that change over time; a score is not a fixed
  cross-vintage or cross-bureau concept. Pin the score model and codebook.
- **Snapshot timing and reporting lags.** Tradelines update on lender reporting
  cycles, so a period's balance can lag actual events. Be explicit about the
  reporting-date basis.
- **Not redistributable.** Results can be reported but the individual-level
  microdata cannot be shared. Plan accordingly.

## Citation

Cite the vendor and route, e.g.: *Equifax consumer-credit records, accessed under
commercial research licence, YYYY-MM-DD.* State the borrower sample, the date
range, the fields used, and the score/codebook vintage.


==============================================================================
# Eurodollar futures intraday prices (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/eurodollar-futures/
# Intraday (tick) Eurodollar futures prices from CME Group, the standard instrument for high-frequency monetary-policy-surprise identification around FOMC announcements. Daily settlements are public; the intraday windows are licensed. This page documents the access path and the gotchas, but the data was not exercised here.
# Tags: monetary-policy, interest-rate-futures, high-frequency, identification, licensed, data:eurodollar-futures
==============================================================================

:::caution[Licensed: not exercised here]
**Intraday Eurodollar futures prices are a paid licensed product** (CME Group via
commercial intraday / tick feeds), so this page carries **no provenance badge**: the access path below was **not** run in this
session. Daily settlement prices are public, but the high-frequency windows that
make this dataset useful for policy-surprise identification are licensed. Treat
it as unverified until someone exercises the intraday feed. This is the honest
grade under the institute's Verified discipline.
:::

**Eurodollar futures** (CME Group) were exchange-traded contracts on the
three-month USD LIBOR rate, with quarterly expiries stretching years out. The
deferred contracts price the expected path of short rates, so the change in a
Eurodollar futures price in a **tight window around an FOMC announcement** is a
standard high-frequency **monetary-policy surprise** that strips out information
known before the announcement. A paper we distill uses it:
[Stavrakeva & Tang](/wiki/papers/jf/2026/stavrakeva-dollar-great-recession-2026/)
build their policy-surprise instrument from the change in the fourth Eurodollar
contract (ED4, expiring three quarters hence) over a one-hour window (15 minutes
before to 45 minutes after) around FOMC and out-of-meeting QE announcements,
using ED4 because it captures unconventional policy at the zero lower bound.

- **Cost:** daily settlement prices are public (CME); the intraday / tick history
  needed for event-window surprises is licensed via a data vendor.
- **Vendor:** CME Group (and commercial intraday redistributors).
- **Coverage:** quarterly Eurodollar contracts out several years, with the
  high-frequency series most useful from the 1990s onward.

## Access (when licensed)

- **Daily settlements are free; intraday is the licensed asset.** Settlement
  prices are published by CME; the tick or minute-bar history around specific
  announcement timestamps is obtained through a CME data subscription or an
  intraday-data vendor.
- **Keyed by contract and timestamp.** An event-window surprise needs the contract
  price immediately before and after each announcement, so you need exact
  announcement timestamps aligned to the futures clock.
- Vendor credentials are required for the intraday feed. Keep any credentials in
  `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **The contract is retired; mind the SOFR transition.** With LIBOR's wind-down,
  CME converted Eurodollar futures to SOFR futures (the conversion completed in
  2023), and liquidity migrated well before that. Use the historical Eurodollar
  series for the in-sample period; for recent or forward analysis the comparable
  instrument is SOFR (or fed funds) futures, and the two are not interchangeable
  tick-for-tick.
- **Which contract you pick is the identification choice.** The near contract,
  ED4, or a bundle measure different things; at the zero lower bound the front
  contract barely moves while deferred contracts carry the unconventional-policy
  signal. State the contract and justify it (the citing paper uses ED4).
- **Window width trades noise against contamination.** Too narrow a window misses
  the price adjustment; too wide a window lets other news in. The 15-before /
  45-after FOMC convention is common but a choice. Report robustness to the
  window.
- **Timestamp alignment is fragile.** Surprises are only as clean as the alignment
  between the announcement timestamp and the futures-feed clock (time zone,
  exchange versus wall clock, release-time revisions). A few seconds of
  misalignment around a release corrupts the surprise.
- **Off-cycle announcements need hand-collected timestamps.** Crisis-era QE and
  forward-guidance announcements made outside scheduled FOMC meetings are part of
  the signal but require event timestamps assembled by hand. Document the event
  list.

## Citation

Cite the exchange and the intraday source, e.g.: *Eurodollar futures (CME Group),
intraday prices via [vendor], accessed YYYY-MM-DD.* State the contract used (for
example ED4), the event-window convention, and the announcement-timestamp source.


==============================================================================
# FactSet LionShares: institutional ownership (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/factset-lionshares/
# FactSet LionShares (FactSet Ownership) is a commercial source for global institutional and fund holdings, with institution classification. It is a paid subscription: this page documents the access path and the gotchas that bite ownership pipelines, but the data was not exercised here.
# Tags: equities, institutional-investors, licensed, panel-data, data:factset-lionshares
==============================================================================

:::caution[Licensed: not exercised here]
**FactSet LionShares is a paid subscription**, so it carries **no provenance badge**: the access path below was **not**
run in this session (no FactSet credentials were available). The page
documents the access route and the gotchas; treat it as unverified until
someone exercises it through a licensed account. This is the honest grade
under the institute's Verified discipline.
:::

**FactSet LionShares** (now delivered as **FactSet Ownership**) is a
commercial source for **institutional and fund equity holdings**.
It assembles holdings from regulatory filings (US 13-F and equivalents) and
fund disclosures into a position-level panel, and **classifies each
holder** by type (mutual fund, hedge fund, pension, bank, insurance company,
and so on). The
[Kwan, Liu & Matthies](/wiki/papers/jf/2026/kwan-liu-matthies-2026/) attention paper
uses LionShares for institutional holdings and institution classification,
which it needs to separate hedge funds from other investor types.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** FactSet.
- **Coverage:** global institutional and fund holdings (broader than US-only
  13-F sources, which is part of the point), with holder-type classification.

## Access (when licensed)

- **Direct from FactSet.** Through the FactSet workstation, the holdings/
  ownership data feed, or the API, using FactSet entity identifiers.
- **Via WRDS.** Subscribing institutions can reach the FactSet ownership
  data through [WRDS](/wiki/commercial/wrds/); query it like any other WRDS
  library and filter on date and security before pulling.
- Credentials are required either way. Keep them in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Holdings are stale between report dates.** Positions are observed at
  disclosure dates (quarterly for US 13-F, less often for many non-US
  holders). Between dates a position is carried forward, not live. Date your
  panel to actual report dates and do not infer intra-quarter trading.
- **The 13-F lag and omissions.** US 13-F positions are reported up to 45 days
  after quarter-end, and confidential-treatment requests let filers omit
  positions. The holdings you see lag reality and can be incomplete.
- **Two levels, institution and fund: do not double count.** The same assets
  can appear at the managing-institution level and the individual-fund level.
  Pick the level your analysis needs and aggregate consistently.
- **Holder classification is a mapping, with edge cases.** The institution
  type (hedge fund, mutual fund, pension, etc.) is FactSet's classification;
  multi-strategy and changing managers blur the lines. Inspect the type field
  rather than assuming a clean partition.
- **Identifiers are FactSet's own.** Holders and securities use FactSet entity
  IDs; securities also carry CUSIP/ISIN/SEDOL. Build an explicit crosswalk to
  PERMNO/CIK rather than assuming a shared key.
- **International coverage is less frequent and less complete** than US 13-F,
  because non-US disclosure regimes differ; do not treat a global panel as
  uniformly dense.

## Citation

Cite the product and vendor, e.g.: *FactSet Ownership (LionShares), FactSet
Research Systems; data licensed and accessed YYYY-MM-DD.* State the holding
level (institution vs fund) and the report dates used.


==============================================================================
# FactSet Revere: supply-chain relationships (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/factset-revere/
# FactSet Revere is FactSet's database of inter-firm business relationships (supplier, customer, competitor, partner) compiled from company filings, presentations, and disclosures. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: supply-chains, production-networks, relationships, licensed, data:factset-revere
==============================================================================

:::caution[Licensed: not exercised here]
**FactSet Revere is a paid licensed product** (FactSet Research Systems),
sometimes reached through [WRDS](/wiki/commercial/wrds/), so it carries **no provenance badge**: the access path below was
**not** run in this session (no FactSet / WRDS credentials were available). The
page documents the access route and the gotchas; treat it as unverified until
someone exercises it through a licensed account. This is the honest grade under
the institute's Verified discipline.
:::

**FactSet Revere** (the FactSet Supply Chain Relationships, formerly the Revere
Data relationship database) maps directed business relationships between firms:
supplier, customer, competitor, and strategic-partner links, each with a
relationship type and date range. Relationships are sourced from public
disclosures (annual reports, investor presentations, press releases, regulatory
filings) rather than from a transaction ledger, so the graph reflects what firms
*disclose* about their counterparties. It is one of the two standard sources for
firm-level production-network links, the other being
[Compustat Segments](/wiki/commercial/compustat-segments/) (whose principal-customer
disclosures cover an earlier period and fewer links). A paper we distill uses
both together:
[Grigoris & Segal](/wiki/papers/jf/2026/grigoris-investment-upstream-downstream-uncertainty-2026/)
splice Compustat Segments (1976-2002) with FactSet Revere (2003-2019) to build
upstream and downstream supply-chain exposures.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** FactSet Research Systems (the relationship product originated at
  Revere Data, which FactSet acquired).
- **Coverage:** broad global coverage of public companies, denser from the
  mid-2000s onward; relationships are time-stamped with start and (where
  disclosed) end dates.

## Access (when licensed)

- **Through FactSet or WRDS.** The relationship tables are reached through a
  FactSet workstation/API entitlement, or through [WRDS](/wiki/commercial/wrds/)
  where the institution licenses the FactSet Supply Chain Relationships product.
  Check which route your institution provides.
- **Keyed by FactSet entity identifiers.** Links are between FactSet's own
  entity IDs (and security IDs); an extract is a list of directed edges
  (source entity, target entity, relationship type, start/end date) that you map
  to your firm universe.
- Terminal or API credentials are required. Keep any credentials in `.env`,
  never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Disclosure-driven, not transaction-driven.** A link exists because a firm
  disclosed the relationship, not because a flow was observed. Coverage is biased
  toward larger, more transparent firms and toward relationships material enough
  to disclose; absence of a link is not evidence of no relationship.
- **No weights or flow magnitudes.** Revere records that a relationship exists
  and its type, not the dollar volume of trade between the two firms. If you need
  edge weights you must impute them (e.g. from Compustat Segments customer sales
  or from input-output shares); do not treat the unweighted graph as a flow
  matrix.
- **Directionality and roles.** Supplier-versus-customer direction matters for
  upstream/downstream exposure measures; confirm the role convention in the
  field you read, and remember the same pair can carry multiple link types.
- **Splicing with Compustat Segments.** The two sources overlap in the mid-2000s
  and use different identifiers and disclosure rules; a naive concatenation
  double-counts or drops links at the seam. Pin the splice year and reconcile
  identifiers across the boundary.
- **Stale and lingering links.** End dates are only as good as the disclosure
  that retired a relationship; some links persist in the data after they have
  ended in reality. Use the as-of date when building a point-in-time network.
- **Identifier linking.** FactSet entity IDs must be linked to CRSP/Compustat
  (e.g. via CUSIP/permno crosswalks); the join has known mismatches, especially
  for subsidiaries and reorganized entities. Verify the link rather than assuming
  a clean one-to-one map.

## Reference: relationship link types

| Link type | Direction | Typical use |
|---|---|---|
| Supplier | source supplies target | Upstream exposure of the target |
| Customer | source sells to target | Downstream exposure of the source |
| Competitor | undirected | Peer/rival sets |
| Partner | undirected | Joint ventures, alliances |

Each edge carries a relationship type and a start/end date; build the
point-in-time network from the as-of date you need.

## Citation

Cite the provider and database, e.g.: *FactSet Supply Chain Relationships
(Revere) / FactSet Research Systems, accessed YYYY-MM-DD.* State the entity
universe, the relationship types used, the as-of date, and the access route
(FactSet or WRDS) so the network is reproducible.


==============================================================================
# FTSE All-Share index constituents and returns (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/ftse-all-share/
# FTSE All-Share (FTSE Russell) is the standard investable-universe index for UK equities: membership, market capitalisation, and returns. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: equity-indices, uk-equities, index-constituents, licensed, data:ftse-all-share
==============================================================================

:::caution[Licensed: not exercised here]
**FTSE All-Share index data is a paid licensed product** (FTSE Russell), so it carries **no provenance badge**: the access path
below was **not** run in this session. Treat it as unverified until someone
exercises it through a licensed FTSE Russell feed or a redistributor. This is the
honest grade under the institute's Verified discipline.
:::

**The FTSE All-Share** is the headline broad UK equity index, the union of the
FTSE 100, FTSE 250, and FTSE SmallCap, and it serves as the **investable
universe** for most UK equity research: index membership, free-float market
capitalisation, and daily total returns. A paper we distill uses it:
[Becht, Franks & Wagner](/wiki/papers/jf/2026/becht-private-meetings-portfolio-firms-2026/)
use the FTSE All-Share as the eligible-stock universe (the index held between
roughly 353 and 703 firms over their 2007 to 2015 sample), taking market
capitalisation and daily returns as controls when studying private meetings
between an active asset manager and its UK portfolio firms.

- **Cost:** licensed, subscription. No free download of historical constituents.
- **Vendor:** FTSE Russell (part of LSEG); often reached through redistributors
  such as [Datastream](/wiki/commercial/datastream/) or
  [WRDS](/wiki/commercial/wrds/).
- **Coverage:** UK-listed equities meeting the FTSE size and liquidity rules, with
  membership, market cap, and returns; long daily history.

## Access (when licensed)

- **Through FTSE Russell or a redistributor.** Membership and index returns come
  from a FTSE Russell licence directly, or via Datastream / Refinitiv / WRDS where
  the institution carries the feed. Check which route is available.
- **Keyed by security and date.** A point-in-time constituent extract gives the
  membership set and free-float market cap on each date; pair with the security's
  return series.
- Terminal or API credentials are required. Keep any credentials in `.env`, never
  hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Use point-in-time membership, not today's list.** Constituents change at
  quarterly reviews and via fast-entry / corporate-action rules; reconstructing
  the universe from the current membership induces survivorship and look-ahead
  bias. The citing paper's universe swings from roughly 353 to 703 firms over the
  sample for exactly this reason. Pull the historical constituent set.
- **Free-float weighting, not full market cap.** FTSE weights by investable free
  float, so the index market cap differs from full shares-outstanding market cap;
  using the wrong one misweights any value-weighted measure. Match the cap concept
  to the question.
- **Price versus total return.** The price index and the total-return index differ
  by reinvested dividends; performance and event-study work need the total-return
  series. State which you used.
- **Identifier mapping for cross-data joins.** Linking FTSE constituents to
  accounting or holdings data requires mapping SEDOL / ISIN to the other source's
  identifiers across listing changes and dual lines; the join has known
  mismatches. Verify it.
- **Overlap with the sub-indices.** The All-Share is the union of FTSE 100 / 250 /
  SmallCap; double-counting or gaps appear if you splice the sub-indices instead
  of taking the All-Share membership directly. Use the All-Share list.

## Citation

Cite the index provider, e.g.: *FTSE All-Share Index (FTSE Russell / LSEG),
accessed YYYY-MM-DD.* State whether membership, market cap, or returns were used,
the cap concept (free-float versus full), and the point-in-time date range.


==============================================================================
# Getty Images executive photographs (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/gettyimages-ceo-photos/
# Getty Images licenses dated press photographs of executives, the raw material for machine-learning apparent-age and facial measures of CEOs. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: images, executives, machine-learning, apparent-age, licensed, data:gettyimages-ceo-photos
==============================================================================

:::caution[Licensed: not exercised here]
**Getty Images is a paid licensed product**, so it carries **no provenance badge**: the access path below was **not** run in this
session. Treat it as unverified until someone exercises it through a Getty Images
licence. This is the honest grade under the institute's Verified discipline.
:::

**Getty Images** is a stock and editorial photo agency whose **dated editorial
photographs of public figures** let researchers assemble time-stamped image sets
for named executives. Paired with a facial-analysis model, the photos yield
apparent-age and other face-based measures over time. A paper we distill uses it:
[Borgschulte, Guenzel, Liu & Malmendier](/wiki/papers/jf/2025/borgschulte-ceo-stress-aging-death-2025/)
build a CEO Apparent Aging Data Set of 3,002 dated photographs of 453 Fortune 1000
CEOs (the 2006 cohort), spanning 2000 to 2016, sourced from Getty Images
(supplemented with Google Images), and run a deep convolutional apparent-age model
to show that industry distress visibly ages CEOs.

- **Cost:** licensed, per-image or subscription editorial licence.
- **Vendor:** Getty Images.
- **Coverage:** editorial / press photographs of public figures with capture dates
  and captions; depth depends on how often a person was photographed, so prominent
  executives are well covered and others sparsely.

## Access (when licensed)

- **Through a Getty Images licence.** Images are searched and licensed on the Getty
  platform (editorial collection); the date and caption metadata, needed to place
  each photo in time, come with the image.
- **Build a per-person, per-date image set.** A usable dataset is the set of dated
  photos for each named individual; the research asset is the derived measure
  (apparent age), not the images themselves.
- **Respect the editorial licence.** Getty terms restrict redistribution and some
  automated use; store derived features and licence the images properly. Keep any
  credentials in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Capture date versus publication date.** The research date must be when the
  photo was taken, not when it was uploaded or republished; a wrong date corrupts
  any aging trajectory. Verify the capture date, not just the metadata timestamp.
- **Photo selection is endogenous.** Which photos exist and get licensed depends on
  events (a CEO is photographed more during a crisis), and editorial photos are
  retouched, lit, and posed differently; both confound a face-based measure. The
  citing paper uses a difference-in-differences design over many images per person
  to net out fixed appearance. Account for selection.
- **Model bias and validation.** Apparent-age and facial models carry systematic
  error by lighting, pose, expression, gender, and race; the measure is only as
  good as the model's validation on this population. State the model and its
  validation.
- **Coverage is uneven across people and time.** Some CEOs have many dated photos,
  others very few; sparse coverage makes per-person trajectories noisy and biases
  toward the most-photographed. Report images per person.
- **Mixing sources changes the distribution.** Supplementing Getty with Google
  Images (as the citing paper does) mixes editorial and web images with different
  quality and date reliability; treat source as a covariate, not as
  interchangeable.

## Citation

Cite the source and that a measure was derived, e.g.: *Editorial photographs
licensed from Getty Images, accessed YYYY-MM-DD.* State the population, the number
of images and dates, any supplementary image source, and the facial model used to
derive the measure.


==============================================================================
# Global Financial Data (GFD): long-run cross-country series (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/global-financial-data/
# Global Financial Data is a commercial vendor of long-run historical stock, bond, commodity, and macroeconomic series spanning many countries and centuries. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: macro, international, historical, returns, licensed, data:global-financial-data
==============================================================================

:::caution[Licensed: not exercised here]
**Global Financial Data is a paid licensed product** (Global Financial Data,
Inc.), so it carries **no provenance badge**:
the access path below was **not** run in this session (no GFD credentials were
available). The page documents the access route and the gotchas; treat it as
unverified until someone exercises it through a licensed account. This is the
honest grade under the institute's Verified discipline.
:::

**Global Financial Data (GFD)** is a commercial database of long-run historical
financial and macroeconomic series: equity price and total-return indices, bond
yields, exchange rates, commodity prices, interest rates, and macro aggregates,
assembled for many countries with histories reaching back centuries in some
series. Its distinguishing feature is *long-run cross-country* coverage: it
stitches together historical sources to extend series far before the start of
modern machine-readable feeds, which makes it a standard input for
survivorship-bias and very-long-horizon return studies. A paper we distill uses
it:
[Van Binsbergen, Hua, Peeters & Wachter](/wiki/papers/jf/2025/binsbergen-united-states-lucky-survivor-2025/)
draw on annual total returns for 55 countries from 1920 to 2020 to estimate
survivorship-corrected crash risk.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** Global Financial Data, Inc.
- **Coverage:** thousands of series across dozens of countries; equity, fixed
  income, FX, commodities, and macro, with deep historical extensions assembled
  from archival sources.

## Access (when licensed)

- **Through the GFD web platform or API.** Series are pulled from the GFD
  Finaeon platform (web interface and an API), keyed by GFD's own ticker/series
  codes. An extract is a set of series codes over a date range.
- **Some series reach researchers via institutional libraries.** Universities
  sometimes provide GFD access through a library subscription rather than a
  direct lab licence; check which route your institution provides.
- API credentials are required. Keep any credentials in `.env`, never
  hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Spliced series carry a methodology break at every join.** A "long-run"
  series is a concatenation of underlying historical sources with different
  construction methods; the further back you go, the more the index reflects
  GFD's splicing choices rather than a single consistent methodology. Read the
  series documentation before treating a multi-century series as homogeneous.
- **Total-return versus price indices.** GFD provides both; mixing a price index
  for one country with a total-return index for another silently biases
  cross-country return comparisons. Confirm the index type for every series in a
  panel.
- **Currency and real-versus-nominal.** Series are in native currency and mostly
  nominal; cross-country aggregation requires explicit FX conversion and a
  consistent deflator choice. Do not mix currencies or nominal-with-real series.
- **Sparse and interpolated early history.** Early observations can be annual,
  interpolated, or carried from a single archival source; apparent low volatility
  in the deep past may be an artifact of sparse sampling rather than calm markets.
- **Country composition changes over time.** Borders, market existence, and
  index membership shift across a century-long sample; a "country" series can
  reflect different underlying markets at different dates. Treat the country
  panel as unbalanced and document entry/exit.
- **Provenance is the value and the risk.** Because the long histories come from
  archival reconstruction, cross-check GFD against an independent source
  (e.g. official statistics or another historical compilation) when a single
  series drives a result.

## Reference: representative series families

| Family | Examples |
|---|---|
| Equity indices | Country total-return and price indices, long-run composites |
| Fixed income | Government bond yields, bill rates, long-term interest rates |
| FX | Bilateral exchange rates, long-run currency series |
| Commodities | Gold, oil, agricultural and metal price histories |
| Macro | CPI, GDP, population, and other long-run aggregates |

Series are keyed by GFD codes; pull the specific codes you need, record the
index type (price vs. total return), and note the historical splice points.

## Citation

Cite the provider and database, e.g.: *Global Financial Data (GFD), accessed
YYYY-MM-DD.* State the series codes used, the index type, the currency, and the
date range so the sample is reproducible.


==============================================================================
# I/B/E/S: analyst estimates and actuals (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/ibes/
# I/B/E/S is the standard panel of sell-side analyst forecasts (EPS and other measures), consensus summaries, and matched "street" actuals, reached by most researchers through WRDS. It is licensed: this page documents the access path and the gotchas; the access path was exercised through a licensed WRDS session.
# Access confirmed (licensed) 2026-06-09 · via live I/B/E/S query (ibes.act_epsint) through a licensed WRDS session
# Tags: equities, analyst-forecasts, licensed, panel-data, data:ibes
==============================================================================

:::note[Access confirmed via a licensed WRDS session]
**I/B/E/S is a licensed commercial dataset** (LSEG / Refinitiv, reached for most
academics through [WRDS](/wiki/commercial/wrds/)). The keystone query
`SELECT * FROM ibes.act_epsint LIMIT 1` was run against the live dataset through
a licensed WRDS session on 2026-06-09, so this page carries the amber "Access
confirmed (licensed)" badge. Reproducing it requires your own WRDS account and
your institution's I/B/E/S entitlement (licensed, not open).
:::

**I/B/E/S** (Institutional Brokers' Estimate System) is the long-running panel
of **sell-side analyst forecasts**: per-analyst (Detail) and consensus (Summary)
estimates of earnings per share and other measures, plus recommendations, price
targets, and matched "street" **actuals**. It is the standard source for analyst
expectations, forecast dispersion, and earnings surprises in empirical finance.
Papers we distill use it for analyst coverage and forecast measures, for example
[Ben-Rephael, Carlin, Da & Israelsen](/wiki/papers/jf/2025/ben-rephael-uncovering-hidden-effort-problem-2025/)
and [Cookson, Niessner & Schiller](/wiki/papers/jf/2026/cookson-social-media-merger-withdrawals-2026/).

- **Cost:** licensed, subscription. No free tier. Most academics reach it
  through an institutional WRDS license.
- **Vendor:** LSEG / Refinitiv (the dataset passed through Thomson Reuters and
  Refinitiv ownership; older citations say Thomson Financial / Thomson Reuters).
- **Coverage:** US (`ibes` US files) and international (`epsint`) equities, from
  the mid-1970s for US EPS; per-analyst detail, consensus summary, actuals,
  recommendations, and price targets.

## Access (when licensed)

- **Through WRDS (the usual path).** The `ibes` libraries on
  [WRDS](/wiki/commercial/wrds/): Summary statistics (`ibes.statsum_epsus`),
  Detail history (`ibes.det_epsus`), actuals (`ibes.actu_epsus`),
  recommendations (`ibes.recddet`), and the broker/analyst translation files.
  Query through a licensed WRDS session exactly as for CRSP and Compustat; see
  the [WRDS page](/wiki/commercial/wrds/) for the connection pattern.
- **Direct from LSEG / Refinitiv.** Through Refinitiv Workspace / a data feed,
  for institutions that license I/B/E/S directly rather than via WRDS.
- Credentials are required either way. Keep them in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented from observed pipeline behavior.

- **Summary vs Detail are different objects.** The Summary file is the
  consensus snapshot computed on each month's "statistical period" cutoff; the
  Detail file is the per-analyst estimate history. Do not rebuild a consensus
  from Detail and expect it to equal Summary: the cutoffs, the included
  estimates, and the staleness rules differ.
- **Split adjustment and rounding corrupt small-cap EPS.** I/B/E/S historically
  stored estimates on a split-adjusted basis rounded to the cent, so for
  high-split or low-price stocks the adjusted EPS can round to zero or flip
  sign. Where possible work from unadjusted values and apply the CRSP
  cumulative adjustment factor yourself, and be wary of tiny denominators.
- **"Street" actuals are not GAAP and not Compustat.** I/B/E/S actuals are the
  analyst-basis ("street") earnings that exclude items analysts deem
  non-recurring; they do not equal Compustat GAAP EPS. Use the I/B/E/S actual,
  not a Compustat figure, when computing a surprise against an I/B/E/S forecast,
  or the surprise is mismatched at the numerator.
- **Forecast period indicator (FPI) and measure must be filtered.** A single
  permno-month carries many rows: FY1/FY2 annual, quarterly, long-term growth,
  and non-EPS measures (sales, CFPS, EBITDA). Always filter `fpi` and `measure`
  explicitly; an unfiltered pull mixes horizons and metrics.
- **Look-ahead bias from review and activation dates.** Use the estimate's
  announce/review and activation dates, not just the statistical period, to know
  what was actually observable as-of a date. The Summary file's "as-of" is the
  cutoff date, not the date the underlying estimates became public.
- **Historical revisions and the anonymity recoding.** I/B/E/S has restated
  parts of its history (the well-documented late-1990s/2000s changes), and
  analyst identities are masked codes that have been recoded over time; a broker
  or analyst code is not a stable cross-vintage key. Pin the data vintage.
- **Identifiers are CUSIP-based and need linking.** I/B/E/S keys on its own
  `ticker` plus historical CUSIP, not PERMNO or GVKEY; join through the
  WRDS-supplied I/B/E/S-CRSP link (or `iclink`) rather than a naive CUSIP merge,
  and respect that CUSIPs are reused over time.
- **US vs international files are separate.** The US (`epsus`) and international
  (`epsint`) files have different conventions and currencies; do not pool them
  without aligning currency and measure definitions.

## Reference: key WRDS I/B/E/S tables

| Table | What it holds |
|---|---|
| `ibes.statsum_epsus` | Consensus Summary statistics (mean/median/stdev, count) |
| `ibes.det_epsus` | Detail: individual analyst EPS estimates |
| `ibes.actu_epsus` | Reported "street" actuals matched to estimates |
| `ibes.recddet` | Analyst recommendations (detail) |
| `ibes.id` / link files | Identifier and broker/analyst translation tables |

## Citation

Cite the provider and the access route, e.g.: *I/B/E/S, LSEG / Refinitiv,
accessed via WRDS, YYYY-MM-DD.* State whether estimates are split-adjusted or
unadjusted, the FPI and measure used, and the data vintage.


==============================================================================
# InfoUSA / Data Axle business and consumer files (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/infousa/
# Data Axle (formerly InfoUSA) compiles business and consumer reference files: establishment listings with location, industry, and employment, plus consumer household files. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: establishments, employment, marketing-data, licensed, data:infousa
==============================================================================

:::caution[Licensed: not exercised here]
**InfoUSA / Data Axle files are a paid licensed product** (Data Axle, formerly
InfoUSA), so they carry **no provenance badge**:
the access path below was **not** run in this session (no Data Axle credentials
were available). The page documents the access route and the gotchas; treat it as
unverified until someone exercises it through a licensed account. This is the
honest grade under the institute's Verified discipline.
:::

**Data Axle** (formerly **InfoUSA**, also known historically as infoGROUP)
compiles commercial **business** and **consumer** reference files: establishment
listings with name, address, geocode, industry (SIC/NAICS), and employment/sales
estimates, plus consumer household files. The business file is used in research to
get establishment-level location and employment counts, including for small and
private firms not in other databases. A paper we distill uses it:
[Allcott, Montanari, Ozaltun & Tan](/wiki/papers/jf/2026/allcott-corporate-social-impact-2026/)
use InfoUSA for firm-level county employment counts in their analysis of corporate
social impact.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** Data Axle (formerly InfoUSA / infoGROUP).
- **Coverage:** U.S. establishments and consumer households, broad but
  compiled-and-estimated rather than administrative; annual snapshots with
  geocodes and industry codes.

## Access (when licensed)

- **Through a Data Axle licence or a research archive.** Files are obtained under a
  commercial licence; some historical InfoUSA business-file vintages are available
  to researchers through data archives. Check which route your institution
  provides.
- **Keyed by establishment.** An extract is a set of establishments with location,
  industry, and employment for a geography and year.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Compiled and estimated, not administrative.** Listings are assembled from
  directories, phone records, and modeling; employment and sales are often
  estimated, not reported. Treat employment counts as estimates, not a census like
  the [QWI](/wiki/datasets/qwi-census/) or BLS.
- **Coverage and accuracy vary by firm size and industry.** Small and new
  establishments are missed or stale; the file lags openings and closings. Do not
  read absence as nonexistence.
- **Establishment, not firm.** Records are establishments (locations); rolling up
  to the firm requires the corporate-linkage fields, which are imperfect. Decide
  the unit deliberately.
- **Vintage snapshots and definition changes.** The file is a periodic snapshot and
  the compilation methodology has changed across the InfoUSA-to-Data-Axle history;
  a panel across vintages is not perfectly consistent. Pin the vintage.
- **Industry coding.** SIC/NAICS assignment is vendor-applied and can be coarse or
  wrong for diversified establishments. Verify industry codes before conditioning
  on them.
- **Duplicate and merged records.** The same establishment can appear under
  variant names or be merged across files; dedupe before counting.

## Citation

Cite the vendor, e.g.: *Data Axle (formerly InfoUSA) business file, accessed
YYYY-MM-DD.* State the vintage/year, the geography, the unit (establishment vs
firm), and that employment figures are compiled estimates.


==============================================================================
# KLD / MSCI ESG ratings (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/kld/
# Firm-level environmental, social, and governance ratings: the historical KLD STATS strength/concern indicators and the successor MSCI ESG (KLD STATS and IVA) ratings, commonly reached through WRDS. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: esg, ratings, firm-level, licensed, data:kld, data:msci-esg
==============================================================================

:::caution[Licensed: not exercised here]
**KLD / MSCI ESG data is a paid licensed product**, often reached through
[WRDS](/wiki/commercial/wrds/), so it carries
**no provenance badge**: the access path below was **not** run in this session
(no MSCI / WRDS credentials were available). The page documents the access route
and the gotchas; treat it as unverified until someone exercises it through a
licensed account. This is the honest grade under the institute's Verified
discipline.
:::

This page covers one dataset lineage under two slugs (`data:kld` and
`data:msci-esg`). **KLD STATS** is the original firm-level ESG ratings dataset
built by KLD Research & Analytics: for each firm-year it records binary
**strength** and **concern** indicators across categories (environment,
community, human rights, employee relations, diversity, product, governance) plus
"controversial business" involvement screens. KLD was acquired and the product
became **MSCI ESG** (MSCI ESG KLD STATS, and the forward-looking MSCI ESG IVA
ratings). It is the most widely used ESG dataset in the asset-pricing and
corporate-finance literature. Papers we distill use it:
[Duchin et al.](/wiki/papers/jf/2025/duchin-sustainability-greenwashing-evidence-asset-2025/)
on sustainability versus greenwashing, and
[Starks et al.](/wiki/papers/jf/2026/starks-esg-profiles-investor-horizons-2026/)
on corporate ESG profiles and investor horizons.

- **Cost:** licensed; commonly accessed through [WRDS](/wiki/commercial/wrds/)
  where the institution subscribes, or directly from MSCI. No free tier.
- **Vendor:** MSCI ESG Research (formerly KLD Research & Analytics).
- **Coverage:** U.S. public firms; KLD STATS begins in 1991 with a small
  universe that expands over time (to roughly the largest 1,000 U.S. firms in
  2001 and the largest 3,000 in 2003), continuing under MSCI.

## Access (when licensed)

- **Through WRDS or MSCI.** Most academic users pull KLD STATS from the WRDS
  MSCI/KLD library; the forward-looking IVA ratings and the current MSCI ESG
  feed are licensed directly from MSCI. Check which your institution provides.
- **Annual firm-year panel of indicators.** A KLD STATS extract is one row per
  firm per year with the strength and concern indicator columns; you join it to
  returns and fundamentals on a firm identifier (CUSIP/ticker in the file).
- Credentials are required; keep any credentials in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **The universe expands over time, so an unbalanced panel is mostly coverage,
  not change.** KLD covers a few hundred firms in the early 1990s and expands in
  steps (notably 2001 and 2003). A firm "becoming rated" is usually the universe
  growing, not a new ESG event. Control for the coverage expansion or restrict to
  a consistently covered set.
- **Indicators are added and dropped across years.** The set of strength and
  concern indicators changes over time (categories are introduced, split, or
  retired), so the raw count of strengths or concerns is not comparable across
  years without rescaling. Normalize by the number of indicators available that
  year before summing or netting.
- **Do not net strengths against concerns naively.** A common net score
  (strengths minus concerns) mixes two distinct constructs and is sensitive to
  the indicator-count changes above; the literature documents that this measure
  is fragile. Be explicit about how you aggregate and test robustness to the
  aggregation choice.
- **The KLD-to-MSCI-IVA methodology break is real.** After the 2010 MSCI
  acquisition the legacy KLD STATS strength/concern format continued for a time
  (through roughly 2013-2014), but the flagship MSCI ESG product moved to the IVA
  industry-relative, letter-graded
  (AAA to CCC) methodology, which is not the same measure and is not directly
  comparable to the old binary indicators. Do not splice KLD STATS and IVA into
  one continuous series.
- **Ratings get revised and restated.** Vintages differ; a rating for a past year
  can change in a later release. Pin the data vintage you used and avoid
  look-ahead from a restated rating.
- **Identifier linking.** Files carry CUSIP and ticker, which are not stable panel
  keys; link to a permanent identifier (e.g. PERMNO via the CRSP/Compustat link)
  and verify the match rather than joining on ticker.

## Reference: KLD STATS category structure

| Category | Indicator type |
|---|---|
| Environment | strengths and concerns |
| Community | strengths and concerns |
| Human rights | strengths and concerns |
| Employee relations | strengths and concerns |
| Diversity | strengths and concerns |
| Product | strengths and concerns |
| Corporate governance | strengths and concerns |
| Controversial business involvement (alcohol, tobacco, gambling, firearms, military, nuclear) | exclusionary screens |

Each indicator is binary (0/1) per firm-year; the count of available indicators
in a category changes across years.

## Citation

Cite the data owner and product explicitly, distinguishing the legacy from the
successor, e.g.: *MSCI ESG KLD STATS (formerly KLD Research & Analytics),
accessed via WRDS YYYY-MM-DD* or *MSCI ESG Ratings (IVA), accessed YYYY-MM-DD.*
State which product and vintage, the years, and the firm identifier used so the
sample is reproducible.


==============================================================================
# LexisNexis court records (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/lexisnexis-court/
# LexisNexis aggregates U.S. court filings and public records (civil lawsuits, judgments, dockets), the raw material for hand-collected litigation datasets. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: court-records, litigation, debt-collection, household-finance, licensed, data:lexisnexis-court
==============================================================================

:::caution[Licensed: not exercised here]
**LexisNexis is a paid licensed product** (LexisNexis / RELX), so it carries **no provenance badge**: the access path below was
**not** run in this session. Treat it as unverified until someone exercises it
through a licensed LexisNexis account. This is the honest grade under the
institute's Verified discipline.
:::

**LexisNexis** aggregates U.S. legal and public records: civil and criminal court
filings, dockets, judgments, liens, and related public records, searchable across
jurisdictions. For research it is the source from which **litigation datasets are
hand-collected**, because no single open registry spans all U.S. civil courts. A
paper we distill uses it:
[Di Maggio, Kalda & Yao](/wiki/papers/jf/2026/maggio-student-debt-second-chance-2026/)
hand-collect, from LexisNexis court filings across U.S. civil courts (2010 to
2017), the National Collegiate Student Loan Trusts debt-collection lawsuits
(borrower identity, court, filing date, outcome) and match them to credit-bureau
panels to study what happens when collection suits are dismissed and the debt
discharged.

- **Cost:** licensed, subscription. No free tier for bulk extraction.
- **Vendor:** LexisNexis (RELX).
- **Coverage:** U.S. court filings and public records across federal, state, and
  local jurisdictions, with uneven depth by court; the research asset is typically
  a hand-collected subset matching a named docket, party, or case type.

## Access (when licensed)

- **Through a LexisNexis subscription.** Records are reached via the LexisNexis
  research platform (or CourtLink / public-records products); bulk or programmatic
  extraction needs the appropriate product entitlement and may require a separate
  data agreement.
- **Hand-collection is the norm.** Because coverage and formats vary by court, a
  usable dataset is usually assembled by querying for a specific party, case type,
  or docket and extracting fields by hand or with scripts the licence permits;
  the citing paper's dataset is hand-collected this way.
- **Identity matching is the join.** Court records carry names and addresses, not
  research IDs; linking to credit or administrative data is a name / address /
  date match. Keep any credentials in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Coverage is uneven across courts.** Jurisdictions digitize and expose filings
  to LexisNexis at different rates; a "national" extract is really the union of
  courts that report, so absence of a case is not absence of litigation. Map which
  courts are covered before treating counts as complete.
- **Records are filings, not outcomes.** A lawsuit record is the filing; the
  disposition (dismissed, settled, judgment) may be in a separate later record or
  missing. Linking filing to outcome is itself work and a source of attrition.
- **Name matching is error-prone.** Matching court parties to a credit panel or
  firm registry on names and addresses produces false positives and missed
  matches, especially for common names; the match rate and rule materially affect
  the sample. Report the matching procedure and rate.
- **Selection into being sued.** Who appears in collection lawsuits is not random
  (it reflects creditor strategy, balance, geography); a sample of sued borrowers
  is selected, so external validity must be argued, not assumed.
- **Redistribution and privacy limits.** Court records contain personal
  identifiers under a restrictive licence; store de-identified derived data, keep
  raw records inside the licence, and do not redistribute. Respect the terms.

## Citation

Cite the vendor and the records used, e.g.: *LexisNexis court records, accessed
YYYY-MM-DD.* State the courts and case type, the date range, that the dataset was
hand-collected, and the identity-matching procedure.


==============================================================================
# IHS Markit CDS: single-name credit default swap spreads (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/markit-cds/
# Markit CDS (IHS Markit, now S&P Global Market Intelligence) provides daily composite single-name and index credit-default-swap spreads contributed by dealers, across maturities, currencies, seniority, and restructuring clauses. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: credit, cds, licensed, data:markit-cds
==============================================================================

:::caution[Licensed: not exercised here]
**Markit CDS is a licensed commercial dataset** (IHS Markit, now S&P Global
Market Intelligence), so it carries **no
provenance badge**: the access path below was **not** run in this session (no
Markit credentials were available). The page documents the access route and the
gotchas; treat it as unverified until someone exercises it through a licensed
account. This is the honest grade under the institute's Verified discipline.
:::

**Markit CDS** is a daily composite single-name credit-default-swap spread
product, built from marks contributed by dealer banks and aggregated by IHS
Markit (now S&P Global Market Intelligence). It covers individual corporate and
sovereign reference entities as well as CDS indices (CDX, iTraxx), across
maturities (the five-year tenor is the common reference point), currencies,
seniority tiers, and restructuring clauses (for example, modified restructuring
"MR" and no-restructuring "XR"). Note: this is the Markit single-name CDS
spread product; Markit bond pricing and Markit Securities Finance are separate
products under their own slugs. It is a common source in credit-risk and
fixed-income research: used in, for example
[Siriwardane, Sunderam & Wallen](/wiki/papers/jf/2025/siriwardane-segmented-arbitrage-2025/)
(CDS pricing in the CDS-bond basis construction), and
[Copeland & Martin](/wiki/papers/jf/2025/copeland-repo-financial-crisis-2025/)
(five-year modified-restructuring USD CDS spreads as a proxy for dealer
counterparty credit risk, matched to 13 of 18 dealers).

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** IHS Markit / S&P Global Market Intelligence.
- **Coverage:** single-name corporate and sovereign reference entities plus CDS
  indices (CDX, iTraxx), daily composite spreads, from the mid-2000s.

## Access (when licensed)

- **Direct from IHS Markit / S&P Global.** Data is delivered as a feed keyed
  on the Markit RED code (Reference Entity Database) per reference entity and
  seniority tier. Credentials and a subscription are required.
- Pulls are parameterized by: RED code, maturity tenor, currency, seniority
  (e.g. SNRFOR), and document clause / restructuring convention.
- Keep credentials in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **You must pin all contract dimensions.** Maturity tenor (5y), currency
  (USD vs EUR), seniority (SNRFOR vs SUBLT2), and restructuring clause (MR vs
  XR vs CR) are distinct series; mixing them silently produces non-comparable
  spreads. State all four dimensions in every pull.
- **Spreads are composite dealer marks, not executed trades.** Thin or
  infrequently-traded names have stale or interpolated marks; treat them with
  caution and check the number of contributing dealers per observation.
- **The RED code is the linking key and is not an equity or bond identifier.**
  Matching to CUSIP, ISIN, or ticker requires a separate crosswalk. Entity
  successions from mergers, spin-offs, or defaults can reassign RED codes;
  verify entity continuity when working across long panels.
- **The 2009 CDS Big Bang changed contract conventions.** Standardized fixed
  coupons (100 bps or 500 bps), upfront points-in-lieu, and restructuring
  standardization took effect in April 2009 for North American names. Pre- and
  post-2009 quoted spreads are not directly comparable; document how your
  pipeline handles this break (par spread conversion, separate pre/post
  samples, or exclusion of the transition window).
- **Index and single-name data are separate files.** CDX/iTraxx index spreads
  live in a different feed from single-name spreads; they require separate
  pulls and should not be mixed without explicit justification.
- **Coverage thins for high-yield and sovereign names.** Investment-grade
  US corporates have the densest coverage; high-yield names and many sovereign
  entities have fewer contributing dealers and more gaps.

## Citation

Cite the product, vendor, and the full contract specification: *IHS Markit CDS
single-name spread data, S&P Global Market Intelligence, accessed YYYY-MM-DD;
maturity [e.g. 5y], currency [USD], seniority [SNRFOR], restructuring clause
[MR/XR/CR], RED-code vintage [date].* Also state how the April 2009 Big Bang
convention change was handled in the sample construction.


==============================================================================
# Markit quanto and cross-currency quotes (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/markit-quanto/
# Markit quanto (cross-currency) derivative quotes from S&P Global (IHS Markit), used to extract the quanto-implied covariance between exchange rates and equity returns. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: derivatives, exchange-rates, risk-premia, licensed, data:markit-quanto
==============================================================================

:::caution[Licensed: not exercised here]
**Markit quanto quotes are a paid licensed product** (S&P Global / IHS Markit), so they carry **no provenance badge**: the access
path below was **not** run in this session (no Markit credentials were
available). The page documents the access route and the gotchas; treat it as
unverified until someone exercises it through a licensed account. This is the
honest grade under the institute's Verified discipline.
:::

**Markit quanto quotes** are dealer-derived prices for **quanto** (cross-currency)
derivatives, where the payoff references a foreign asset but settles in the
domestic currency. Quanto-versus-vanilla price differences pin down the
risk-neutral **covariance between the exchange rate and the underlying asset's
return**, which is otherwise hard to observe. They are part of the Markit
composite-quote franchise alongside [Markit bond pricing](/wiki/commercial/markit/)
and [Markit CDS](/wiki/commercial/markit-cds/). A paper we distill uses them:
[Kremens, Martin & Varela](/wiki/papers/jf/2025/kremens-long-horizon-exchange-rate-2025/)
use Markit quanto forwards on the S&P 500 (24-month quotes from December 2009
onward) to construct a quanto-implied risk premium (the risk-neutral covariance
between FX and equity).

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** S&P Global (IHS Markit).
- **Coverage:** quanto/cross-currency quotes for major currency-equity pairs,
  with history that thickens after the late 2000s; availability depends on dealer
  contributions.

## Access (when licensed)

- **Through a Markit / S&P Global feed.** Quotes are delivered as part of a Markit
  pricing entitlement, keyed by the underlying, currency pair, and tenor.
- **Composite dealer quotes, not exchange trades.** Like other Markit pricing
  products, the values are composites built from contributing dealers, not
  executed-trade prints. Confirm the methodology for the series you use.
- Terminal or API credentials are required. Keep any credentials in `.env`, never
  hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Composite quotes carry a methodology, not a trade.** The quote is a modeled
  composite from dealer contributions; thin contribution makes a tenor unreliable.
  Check the contributor count and the staleness of each quote.
- **The signal is a difference, so it is noise-sensitive.** The quanto-implied
  covariance comes from differencing quanto and vanilla prices; small quote errors
  are amplified in the difference. Treat short samples and illiquid tenors with
  care.
- **Tenor and history availability.** Long tenors (for example 24-month) exist
  only for some pairs and only from the late 2000s; do not assume a balanced panel
  across pairs and tenors. Pin the available window per pair.
- **Currency-pair conventions.** Quote direction (domestic vs foreign) and the
  settlement currency must be read carefully, or the sign of the implied
  covariance flips. Confirm the convention before interpreting.
- **Reconciliation with vanilla quotes.** The vanilla leg used to difference
  against the quanto leg must come from a consistent source and tenor; mismatched
  legs contaminate the estimate. Source both legs consistently.

## Citation

Cite the vendor and product, e.g.: *Markit quanto quotes / S&P Global (IHS
Markit), accessed YYYY-MM-DD.* State the underlying, currency pair, tenor, and the
vanilla source used for the differencing.


==============================================================================
# Markit Securities Finance: securities-lending data (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/markit-securities-finance/
# Markit Securities Finance (S&P Global / IHS Markit) is the standard securities-lending dataset: stock borrow fees, utilization, and lendable supply from a broad contributor base. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: short-selling, securities-lending, borrow-fees, licensed, data:markit-securities-finance
==============================================================================

:::caution[Licensed: not exercised here]
**Markit Securities Finance is a paid licensed product** (S&P Global / IHS
Markit), so it carries **no provenance badge**:
the access path below was **not** run in this session (no Markit credentials were
available). The page documents the access route and the gotchas; treat it as
unverified until someone exercises it through a licensed account. This is the
honest grade under the institute's Verified discipline.
:::

**Markit Securities Finance** (the former Data Explorers securities-lending
database) is the standard source for **short-selling cost and supply**: stock
borrow fees, utilization, lendable supply, and on-loan quantities, aggregated from
a large contributor base of lenders and prime brokers. It is the data behind most
empirical work on borrow costs, the short-sale-cost wedge in anomaly returns, and
crowded shorts. A paper we distill uses it:
[Muravyev, Pearson & Pollet](/wiki/papers/jf/2025/muravyev-anomalies-short-sale-costs-2025/)
use the Markit Securities Finance Buy Side Analytics feed (daily from June 28,
2006) for the indicative borrow fee (the buy-side expected borrow cost) and
utilization when measuring how short-sale costs erode anomaly returns.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** S&P Global (IHS Markit); lineage through Data Explorers.
- **Coverage:** global equities (and some other securities) with daily borrow
  fees, utilization, and supply; deeper and broader from the mid-2000s onward.

## Access (when licensed)

- **Through a Markit / S&P Global feed or WRDS.** The data is reached via a Markit
  entitlement (for example the Buy Side Analytics data feed) or, where the
  institution licenses it, through [WRDS](/wiki/commercial/wrds/). Check which route
  your institution provides.
- **Keyed by security and date.** An extract is borrow fee, utilization, and
  supply by security over a date range. Distinguish indicative (buy-side) fees
  from other fee concepts in the product.
- Terminal or API credentials are required. Keep any credentials in `.env`, never
  hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Several fee concepts, not one.** The product reports multiple fee measures
  (indicative buy-side fee, value-weighted average fee, and others); they differ in
  level and timing. Use the fee concept that matches your question and state it.
- **Contributor-based, so coverage is uneven.** Fees and supply come from
  contributing lenders/brokers; a hard-to-borrow name with thin contribution is
  noisily measured. Check contributor depth before trusting a fee.
- **Fees are stale and sticky for some names.** Indicative fees update on
  contribution cycles and can lag the true marginal borrow cost, especially for
  specials around events. Treat event-window fees cautiously.
- **Utilization needs the supply denominator.** Utilization depends on the
  lendable-supply estimate, which is itself contributor-based; a low-supply name
  inflates utilization mechanically. Read utilization with supply.
- **History start and the proxy fallback.** The daily series starts in the
  mid-2000s; for earlier periods or without access, researchers proxy borrow cost
  with the short-interest-to-institutional-ownership ratio, which is a coarse
  substitute. Do not splice the proxy and the Markit fee without care.
- **Identifier linking.** Securities must be linked to CRSP/Compustat
  identifiers; the join has known mismatches across reorganizations. Verify the
  link.

## Citation

Cite the vendor and product, e.g.: *Markit Securities Finance / S&P Global (IHS
Markit), accessed YYYY-MM-DD.* State the fee concept used (for example indicative
buy-side fee), the security universe, and the date range.


==============================================================================
# IHS Markit bond pricing: composite quotes for corporate bonds (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/markit/
# The Markit Bond Pricing Database (IHS Markit / S&P Global) provides daily evaluated composite price quotes for individual corporate and other bonds, aggregated from contributing dealers, together with the dealer-count per bond. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: fixed-income, corporate-bonds, bond-pricing, licensed, data:markit
==============================================================================

:::caution[Licensed: not exercised here]
**The Markit Bond Pricing Database is a licensed commercial dataset** (IHS
Markit, now part of S&P Global Market Intelligence), so it carries **no provenance badge**: the access path below was
**not** run in this session (no IHS Markit / S&P Global credentials were
available). The page documents the access route and the gotchas; treat it as
unverified until someone exercises it through a licensed account. This is the
honest grade under the institute's Verified discipline.
:::

**Markit Bond Pricing** (the Markit Bond Pricing Database, IHS Markit / S&P
Global Market Intelligence) provides daily evaluated composite price quotes for
individual corporate and other bonds, aggregated from contributing dealers,
together with the count of distinct dealers contributing a quote per bond per
day. Note: Markit's single-name CDS spreads, Markit Securities Finance
(securities lending), and Markit quanto forwards are separate products
documented under their own slugs and are not covered here. Used in, for example
[Huang, Nozawa & Shi](/wiki/papers/jf/2025/huang-global-credit-spread-puzzle-2025/)
(daily trader quotes and the number of dealers quoting each bond, used as a
dealer-meeting-intensity proxy), and
[Siriwardane, Sunderam & Wallen](/wiki/papers/jf/2025/siriwardane-segmented-arbitrage-2025/)
(Markit cash-bond and CDS pricing for the CDS-bond basis).

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** IHS Markit / S&P Global Market Intelligence.
- **Coverage:** cross-market corporate and credit bonds; daily composite quotes;
  history depending on the feed, with deeper history for investment-grade names.

## Access (when licensed)

- **Directly from IHS Markit / S&P Global.** The common path is a data-feed
  or file-delivery agreement with IHS Markit / S&P Global, keyed on bond
  identifiers (ISIN or CUSIP).
- **Through an institutional data-services agreement.** Some universities and
  asset managers reach Markit products under a broader data-services arrangement
  with S&P Global; confirm whether your institution's agreement covers the Bond
  Pricing feed specifically.
- Credentials and a signed license are required in all cases. Keep them in
  `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Composite quotes, not executed trades.** These are dealer-contributed
  evaluated prices, not transaction prices from TRACE. Quotes can be stale or
  thin where few dealers contribute; for trade-level data you need TRACE (FINRA).
- **Dealer-count is contribution, not depth.** The field reflects how many
  dealers submitted a quote that day; it is not a direct measure of market depth
  or trading volume, and it is bursty around month-ends and holidays.
- **Bond-identifier linking is a separate, imperfect step.** Joining ISIN or
  CUSIP to FISD (Fixed Income Securities Database) issuer, rating, or maturity
  fields requires a separate linking table; match rates are imperfect, and
  idiosyncratic re-issuances or identifier changes can drop bonds silently.
- **Coverage and quote frequency vary by bond and era.** Off-the-run, high-yield,
  or small-issue bonds receive fewer contributing quotes; absence of a quote on a
  given day is not absence of a bond from the universe.
- **Do not conflate with other Markit products.** Markit CDS (single-name
  spreads), Markit Securities Finance (securities lending), and Markit iBoxx
  indices are separate products with separate licenses; mixing feeds or assuming
  a single license covers all Markit data is a common error.
- **Evaluated-price methodology can change across vintages.** The aggregation
  method (trimming, weighting of dealer quotes) can differ across feed versions;
  pin the feed version in your data dictionary and flag any vintage breaks in the
  sample period.

## Citation

Cite the product and vendor, stating the feed version and the bond-identifier
link used, e.g.: *Markit Bond Pricing Database, IHS Markit / S&P Global Market
Intelligence, data licensed and accessed YYYY-MM-DD; bonds linked via CUSIP/ISIN
to FISD.* State the sample period, quote frequency, and how you handle
bond-days with no contributing quote.


==============================================================================
# Moody's Ultimate Recovery Database (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/moodys-urd/
# Moody's Ultimate Recovery Database (URD) records firm- and instrument-level creditor recovery rates at the resolution of U.S. corporate defaults. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: credit-markets, default, recovery, bankruptcy, corporate-debt, licensed, data:moodys-urd
==============================================================================

:::caution[Licensed: not exercised here]
**Moody's URD is a paid licensed product** (Moody's Analytics), so it carries **no provenance badge**: the access path below
was **not** run in this session. Treat it as unverified until someone exercises
it through a Moody's Analytics licence. This is the honest grade under the
institute's Verified discipline.
:::

**Moody's Ultimate Recovery Database (URD)** records the recovery actually
realized by creditors when a U.S. corporate default is resolved: the value of
the cash, new debt, or equity received per dollar of claim, by instrument and
seniority, measured at emergence from bankruptcy rather than from secondary
market prices at default. It is the reference source for loss-given-default and
recovery-rate research because it captures the full resolution rather than a
trading-price proxy. A paper we distill uses it:
[Griffin, Nini & Smith](/wiki/papers/jf/2026/griffin-loan-covenant-violations-decline-2026/)
use the URD for the par-value-weighted firm-level recovery rates of 403
corporate defaults (1997 to 2020) when showing that loan covenant violations,
and the creditor control they trigger, protect recovery at bankruptcy.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** Moody's Analytics (part of the CreditView / Default & Recovery
  data family).
- **Coverage:** U.S. nonfinancial corporate defaulters with large rated debt;
  instrument-level recoveries observed at the resolution of the default event.
  Coverage starts in the late 1980s and is deepest for bond and loan issuers
  that Moody's rated.

## Access (when licensed)

- **Through a Moody's Analytics data licence.** Extracts are delivered at two
  grains: the default event (firm, default date, type) and the individual debt
  instrument (seniority, collateral, recovery value). Join on the deal or
  family identifier.
- **Recovery is the realized payout at emergence.** The headline "ultimate
  recovery" is nominal or discounted value received at the resolution of the
  case, not the 30-day-post-default trading price; the database also stores the
  trading-based measure for comparison.
- Credentials are required. Keep any credentials in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **The sample is selected toward large, rated defaulters.** The URD covers
  firms with Moody's-rated debt that went through a formal default and
  resolution; small private-debt defaults and out-of-court workouts are
  underrepresented. Recovery statistics from it do not describe the universe of
  corporate distress. The citing paper's 403-default sample is exactly this
  selected population.
- **Two recovery concepts, easy to mix up.** "Ultimate" (settlement-value)
  recovery and trading-price recovery are different numbers for the same
  instrument and answer different questions. State which you use; comparisons
  across papers break when one uses each.
- **Firm-level numbers are a weighted aggregation of instrument-level rows.**
  The par-value-weighted firm recovery used by the citing paper is built up from
  per-instrument recoveries; the weighting scheme (par, market, by seniority)
  changes the headline. Document the aggregation.
- **Resolution lag truncates recent cohorts.** "Ultimate" recovery is only
  observed once the case closes, so the most recent default years are
  incomplete until their cases resolve. A panel that ends near the data vintage
  systematically misses slow resolutions.
- **Seniority and collateral coding requires care.** Recovery is highly
  sensitive to lien position and instrument type; mislabeling a second-lien or
  unsecured tranche distorts the loss-given-default. Reconcile the instrument
  fields against the credit agreement where it matters.

## Citation

Cite the vendor and product, e.g.: *Moody's Ultimate Recovery Database (Moody's
Analytics), accessed YYYY-MM-DD.* State the default years and sample, which
recovery concept (ultimate versus trading) was used, and the weighting used to
move from instrument-level to firm-level recovery.


==============================================================================
# Morningstar fund and sustainability data (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/morningstar/
# Morningstar mutual-fund and ETF data: returns, holdings, categories, star ratings, and the Sustainability Rating (globes) and carbon metrics, reached through Morningstar Direct or a data licence. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: funds, esg, holdings, licensed, data:morningstar
==============================================================================

:::caution[Licensed: not exercised here]
**Morningstar data is a paid licensed product**, so it carries **no provenance badge**: the access path below was **not** run
in this session (no Morningstar Direct or data-licence credentials were
available). The page documents the access route and the gotchas; treat it as
unverified until someone exercises it through a licensed account. This is the
honest grade under the institute's Verified discipline.
:::

**Morningstar** is the standard commercial source for mutual-fund and ETF data:
monthly and daily returns, portfolio holdings, the Morningstar Category
classification, the star rating, fund flows, and fee/share-class detail. For
sustainability research it also publishes the **Morningstar Sustainability
Rating** (the "globes"), portfolio ESG and carbon-risk scores (built on
Sustainalytics inputs), and fund-level sustainable-investing flags. A paper we
distill uses it:
[Parise & Rubin](/wiki/papers/jf/2025/parise-green-window-dressing-2025/) on
green window dressing.

- **Cost:** licensed, subscription. No free tier for the research data (the
  consumer Morningstar.com site is a separate, limited product).
- **Vendor:** Morningstar, Inc. (sustainability inputs from Morningstar
  Sustainalytics).
- **Coverage:** open-end funds, ETFs, and other vehicles across many domiciles,
  with long return histories; sustainability ratings begin in the mid-2010s.

## Access (when licensed)

- **Morningstar Direct or a data feed.** Most academic users query through the
  **Morningstar Direct** desktop/web platform (interactive screening and export)
  or receive a bulk **data feed / Direct Web Services API** under a licence.
  Check which your institution provides.
- **Keyed by SecId and FundId.** Morningstar uses its own identifiers
  (`SecId` for a share class / security, `FundId` for the fund) as the join keys;
  CUSIP and ticker are available but are share-class level and change over time.
- Credentials are required; keep any credentials in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Share class versus fund.** Morningstar data is organized by share class, but
  most analysis wants the fund. Returns, fees, and flows differ across share
  classes of the same fund; aggregate to the fund (value-weight by share-class
  assets) deliberately rather than picking one class or double-counting.
- **Survivorship and the dead-fund question.** The default screen in Direct can
  return only surviving funds; a sample built that way is survivorship-biased.
  Explicitly include obsolete/merged/liquidated funds and use the full universe
  when computing performance.
- **Category reclassification.** The Morningstar Category is reassigned over time
  as a fund's holdings drift, so a fund's category is not constant. Pin the
  category as of the date you need rather than using the current category
  retroactively.
- **The Sustainability Rating is rank-based, recent, and revised.** The globe
  rating is a percentile rank within category, exists only from the mid-2010s
  forward, and its methodology was revised (the 2019 shift to the Sustainalytics
  ESG Risk Rating framework for the underlying portfolio score). Do not treat
  early-vintage and post-revision globes as the same measure, and do not extend
  the series before it existed.
- **Carbon and ESG scores need portfolio coverage.** Portfolio-level
  sustainability metrics are computed only over the rated portion of holdings;
  a fund with low coverage gets a metric off a thin base. Check the coverage
  percentage before conditioning on the score.
- **Backfill and incubation bias.** Funds can enter the database with backfilled
  history, and incubated funds appear only after a successful track record;
  both bias performance upward. Use the date a fund was added when it matters.
- **Identifiers are not stable through corporate actions.** Share-class mergers,
  ticker reuse, and CUSIP changes mean a ticker is not a panel key; join on
  `SecId`/`FundId` and verify links across mergers.

## Reference: common Morningstar data items

| Item | Level | Notes |
|---|---|---|
| Monthly/daily total return | share class | load-adjusted and unadjusted variants exist |
| Net assets / TNA | share class | aggregate to fund for fund-level totals |
| Morningstar Category | fund | reassigned over time |
| Star rating | share class | risk-adjusted, category-relative, trailing |
| Sustainability Rating (globes) | fund | percentile rank within category, mid-2010s on |
| Portfolio ESG / carbon-risk score | fund | depends on holdings coverage |
| Estimated net flow | share class | derived from assets and returns |

## Citation

Cite Morningstar as the source, e.g.: *Morningstar Direct, Morningstar, Inc.,
accessed YYYY-MM-DD* (note Sustainalytics as the sustainability-data source where
relevant). State the access route (Direct or data feed), the identifier used
(`SecId`/`FundId`), and the share-class-versus-fund aggregation so the sample is
reproducible.


==============================================================================
# MSCI Real Estate (IPD): property indices and yields (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/msci-real-estate/
# MSCI Real Estate (formerly IPD) provides property total-return indices and rental-yield benchmarks across countries and sectors, built from appraised institutional portfolios. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: real-estate, indices, international, yields, licensed, data:msci-real-estate
==============================================================================

:::caution[Licensed: not exercised here]
**MSCI Real Estate is a paid licensed product** (MSCI, formerly IPD), so it carries **no provenance badge**: the access path below
was **not** run in this session (no MSCI Real Estate credentials were available).
The page documents the access route and the gotchas; treat it as unverified until
someone exercises it through a licensed account. This is the honest grade under
the institute's Verified discipline.
:::

**MSCI Real Estate** (the IPD indices, after MSCI acquired Investment Property
Databank) measures the performance of directly held institutional real estate:
total returns, capital growth, income return, and rental-yield benchmarks, by
country and by property sector (office, retail, industrial, residential). The
indices are built bottom-up from the appraised portfolios that institutional
investors contribute, so they reflect *valuation-based* performance rather than
transaction prices. It is a standard benchmark source for real-estate yield and
return levels. A paper we distill uses it:
[Amaral, Dohmen, Kohl & Schularick](/wiki/papers/jf/2025/amaral-superstar-returns-spatial-heterogeneity-2025/)
anchor their long-run city housing-return database to 2018 MSCI rental-yield
benchmarks.

- **Cost:** licensed, subscription. No free tier (headline aggregates are
  sometimes published, but the underlying series are licensed).
- **Vendor:** MSCI (the IPD databank lineage).
- **Coverage:** many countries and property sectors, with annual and (in some
  markets) quarterly frequency; history is deeper in markets with long-standing
  IPD contribution (UK, several European and developed markets).

## Access (when licensed)

- **Through MSCI Real Estate (IPD) products.** Series are obtained via MSCI's
  real-estate index and analytics products under a licence, keyed by country,
  sector, and index variant (total return, capital growth, income return, yield).
- **Aggregates versus contributed micro-data.** The published indices are
  available to licensees; the underlying contributed portfolio data is a separate,
  more restricted arrangement. Confirm which you have access to.
- Credentials are required. Keep any credentials in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Appraisal-based, so returns are smoothed.** Index values come from periodic
  property appraisals, not market transactions; this *smooths* and lags returns
  relative to a transaction-price index, understating volatility and the speed of
  repricing. Do not treat MSCI real-estate volatility as comparable to equity or
  transaction-based indices without an unsmoothing adjustment.
- **Survivorship and contributor composition.** The index reflects the
  portfolios institutions choose to contribute; the contributor set changes over
  time and skews toward prime institutional-grade assets. Levels and trends can
  shift as composition changes, not only as the market moves.
- **Yield definitions differ.** "Net initial yield," "equivalent yield," and
  "income return" are distinct concepts; using a yield series for the wrong
  purpose (e.g. a benchmark anchor) silently mismatches the level. Confirm the
  exact yield variant.
- **Frequency and geography are uneven.** Many markets are annual-only, some are
  quarterly; sector breakdowns and history depth vary by country. Treat the panel
  as unbalanced and do not assume a common frequency across markets.
- **Currency.** Returns and yields are reported per local market; cross-country
  aggregation needs explicit currency handling. Do not mix currencies in a panel.
- **Cross-source anchoring.** When using an MSCI yield as a benchmark anchor for
  a separately built series (as in the distilled paper), the anchor inherits
  MSCI's appraisal and composition properties; document the anchor year and
  variant so the extrapolation is reproducible.

## Reference: representative index variants

| Variant | What it measures |
|---|---|
| Total return | Income return plus capital growth |
| Capital growth | Change in appraised capital value |
| Income return | Net income relative to capital value |
| Rental / initial yield | Income as a share of value (benchmark anchor) |

Series are keyed by country, sector, and variant; pull the specific variant you
need and record the country, sector, frequency, and as-of year.

## Citation

Cite the provider and database, e.g.: *MSCI Real Estate (IPD) / MSCI, accessed
YYYY-MM-DD.* State the country, sector, index variant, and as-of year so the
series is reproducible.


==============================================================================
# NETS: National Establishment Time Series (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/nets/
# NETS (Walls & Associates, from Dun & Bradstreet source data) is an establishment-level panel tracking US establishments annually from the early 1990s: location, industry, employment, sales, and ownership links. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: firms, establishments, employment, licensed, panel-data, data:nets
==============================================================================

:::caution[Licensed: not exercised here]
**NETS is a licensed commercial dataset** (Walls & Associates, built from Dun
& Bradstreet source data), so it carries **no
provenance badge**: the access path below was **not** run in this session (no
Walls & Associates credentials were available). The page documents the access
route and the gotchas; treat it as unverified until someone exercises it
through a licensed account. This is the honest grade under the institute's
Verified discipline.
:::

**NETS** (the National Establishment Time Series) is an establishment-level
panel built by Walls & Associates from annual snapshots of the Dun & Bradstreet
(D&B) business database. Each establishment carries a location (geocoded to
county and census tract), industry code, employment, sales, and ownership and
headquarters links tracked over time, which lets researchers follow
establishment births, deaths, relocations, and ownership changes. It is a
common source in establishment-level research: used in, for example
[Barkai & Panageas](/wiki/papers/jf/2025/barkai-value-employment-2025/)
(establishment-level employment and ownership/acquirer-age changes, 1998 to
2014, 213,792 acquisitions), and
[Kruttli, Roth Tran & Watugala](/wiki/papers/jf/2025/kruttli-pricing-poseidon-extreme-weather-2025/)
(firm establishment locations by county, annual, to construct a hurricane
landfall-region exposure measure).

- **Cost:** licensed, purchased extract. No free tier or standard academic
  subscription path.
- **Vendor:** Walls & Associates (from Dun & Bradstreet source data).
- **Coverage:** US establishments annually from the early 1990s; tens of
  millions of establishments per vintage.

## Access (when licensed)

- **Directly from Walls & Associates.** NETS is purchased as a dataset extract,
  often as a one-time vintage. Contact Walls & Associates for current pricing
  and extraction options.
- **No standard WRDS path.** Unlike many licensed academic datasets, NETS does
  not flow through [WRDS](/wiki/commercial/wrds/); the purchase is a direct
  arrangement with the vendor.
- A license agreement is required. Keep any credentials or delivery tokens in
  `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Sales and employment are often imputed, not reported.** Walls & Associates
  derives establishment-level sales and employment from D&B industry-level
  ratios when the underlying D&B record lacks a reported figure. Establishment-
  level sales growth cannot be measured cleanly and level comparisons are noisy;
  Barkai & Panageas flag exactly this limitation.
- **D&B updates lag and carry stale records.** Establishment births and deaths
  are measured with delay and error because D&B does not instantly remove
  defunct entities or add new ones; treat event timing as approximate.
- **Periodic snapshot stitched into a time series.** Within-establishment
  changes over time are more reliable than cross-section levels in any given
  year; design tests accordingly.
- **DUNS and ownership links change with corporate restructurings.** The
  establishment identifier (DUNS) and headquarters/parent ownership links are
  point-in-time; treat them as such rather than as stable longitudinal keys.
- **Geocoding precision varies by vintage.** County-level assignment is
  generally stable; tract-level precision degrades in older vintages and for
  less-populous areas.
- **Coverage and accuracy are debated.** The literature has noted discrepancies
  relative to County Business Patterns (CBP) and the Quarterly Census of
  Employment and Wages (QCEW), particularly for aggregate employment counts.
  Benchmark NETS figures against CBP or QCEW before drawing aggregate
  conclusions.

## Citation

Cite NETS by vendor and source, stating the vintage, whether employment and
sales figures are reported or imputed, and how establishment births and deaths
were defined in your sample: e.g., *NETS (National Establishment Time Series),
Walls & Associates, from Dun & Bradstreet source data, vintage YYYY; employment
figures are largely imputed from industry ratios.* Report the DUNS-based
identifier scheme and any ownership-link vintages used.


==============================================================================
# NielsenIQ retail scanner and consumer panel (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/nielseniq/
# NielsenIQ retail scanner and Homescan consumer-panel data, distributed for academic research through the Kilts Center at Chicago Booth. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: consumption, retail, micro, licensed, data:nielseniq
==============================================================================

:::caution[Licensed: not exercised here]
**NielsenIQ data is a paid licensed product**, distributed to researchers under
a Kilts Center data agreement, so it carries
**no provenance badge**: the access path below was **not** run in this session
(no Kilts / NielsenIQ subscription was available). The page documents the access
route and the gotchas; treat it as unverified until someone exercises it through
a licensed account. This is the honest grade under the institute's Verified
discipline.
:::

**NielsenIQ** (formerly Nielsen) sells two distinct consumer datasets that
academic users reach through the **Kilts Center for Marketing at the University
of Chicago Booth School of Business**: the **Retail Scanner** data (weekly
store-level sales, prices, and quantities aggregated from point-of-sale systems)
and the **Consumer Panel** (Homescan, a longitudinal panel of households that
record their purchases). They are the standard source for high-frequency
household consumption and retail-price research. A paper we distill uses it:
[Allcott et al.](/wiki/papers/jf/2026/allcott-corporate-social-impact-2026/)
on an economic view of corporate social impact.

- **Cost:** licensed; access for academics runs through a Kilts Center data
  agreement and an annual subscription paid by the institution. No free tier.
- **Vendor:** NielsenIQ, distributed for research by the Kilts Center
  (Chicago Booth).
- **Coverage:** Retail Scanner covers a large set of participating U.S. retail
  chains at the store-week-UPC level; the Consumer Panel tracks tens of
  thousands of U.S. households at the trip-UPC level. Both run from the mid-2000s
  forward.

## Access (when licensed)

- **Through the Kilts Center.** An institution signs the Kilts data agreement;
  named researchers are then granted access and download the data from the Kilts
  Center marketing-data server. Each product (Retail Scanner, Consumer Panel) is
  licensed and downloaded separately.
- **Annual files, not a live feed.** The data is delivered as annual flat-file
  releases (tab-delimited), organized by year and product module, plus master
  files for products, stores, and households. You build a panel by stitching the
  annual releases together yourself.
- Credentials and the signed agreement are required; keep any credentials in
  `.env`, never hard-coded, and respect the agreement's redistribution limits.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Retail Scanner is store-level, the Consumer Panel is household-level: they
  are not the same data.** Scanner is what stores sold; the panel is what
  households reported buying. They cover different universes, use different
  weights, and do not reconcile to each other. Pick the one that matches the
  question and do not splice them naively.
- **Projection and panel weights are mandatory for any population total.** Both
  products ship sampling/projection weights; raw row counts and raw sums are
  sample artifacts, not population quantities. Always apply the provided weights
  before reporting national or market totals.
- **UPC versioning and product churn.** A UPC can be reused or revised over time,
  and the product master changes across annual releases, so a UPC is not a stable
  panel identifier on its own. Join through the version-stamped product master
  for the matching year rather than assuming a UPC means the same item across
  years.
- **Product modules and departments, not free-text categories.** Items are
  classified into NielsenIQ product modules, groups, and departments; analysis by
  "category" means choosing a level of that hierarchy. Misaligning the hierarchy
  level across years silently changes the category definition.
- **Store coverage changes as chains enter and leave.** The set of participating
  retailers is not constant, so a store-week panel has entry and exit that is a
  data-coverage artifact, not a market event. Do not read a store dropping out as
  a closure.
- **Magnet and special items.** Some non-UPC items (random-weight produce, meat,
  bulk) are captured through "magnet" data with different conventions and
  coverage; treat them separately from scanned UPC items.
- **It is large.** Retail Scanner is hundreds of gigabytes across years; plan for
  out-of-core processing and pull only the product modules and years you need.

## Reference: the two products

| Product | Unit of observation | Typical key fields |
|---|---|---|
| Retail Scanner | store x week x UPC | store code, week ending, UPC, units, dollar sales, price |
| Consumer Panel (Homescan) | household x trip x UPC | household code, trip date, retailer code, UPC, quantity, total price paid |

Master files (products, stores, households, retailers) carry the descriptive
fields and the weights; join the transaction files to them through the matching
annual release.

## Citation

Cite NielsenIQ as the data owner and the Kilts Center as the distributor, per the
required acknowledgment in the data agreement, e.g.: *Researcher(s) own analyses
calculated (or derived) based in part on data from NielsenIQ and marketing
databases provided through the Kilts Center for Marketing Data Center at the
University of Chicago Booth School of Business.* State the product (Retail Scanner
or Consumer Panel), the years, and the product modules used. The conclusions are
the researcher's own and not those of NielsenIQ.


==============================================================================
# New York Times article archive (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/nyt-news/
# The full-text New York Times archive (back to 1851) is a long, consistent news corpus used for text-as-data measures of sentiment, attention, and discourse. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: text-as-data, news, media-archive, narrative, licensed, data:nyt-news
==============================================================================

:::caution[Licensed: not exercised here]
**The New York Times full-text archive is a licensed source** (The New York
Times), so it carries **no provenance badge**: the
access path below was **not** run in this session. The free NYT API exposes
article metadata and abstracts, but the full body text used for corpus-scale
text analysis is licensed. Treat it as unverified until someone exercises the
full-text archive. This is the honest grade under the institute's Verified
discipline.
:::

**The New York Times archive** is a continuously published newspaper running since
1851, which makes its full text one of the few news corpora long and consistent
enough to build **century-scale** text-as-data measures. It is used for narrative,
attention, sentiment, and topic measures in finance and economics. A paper we
distill uses it:
[Hirshleifer, Mai & Pukthuanthong](/wiki/papers/jf/2025/hirshleifer-war-discourse-cross-section-2025/)
run semisupervised topic modelling (sLDA) over roughly 7 million New York Times
articles spanning January 1871 to October 2019 to construct a war-discourse index,
whose innovations form a factor (WarFac) that prices the cross section of stock
returns.

- **Cost:** article metadata and abstracts are free via the NYT Developer API;
  the full body-text archive for bulk text mining is licensed.
- **Publisher:** The New York Times.
- **Coverage:** the full run of the paper from the 19th century to the present;
  millions of articles, with section, date, and length metadata.

## Access (when licensed)

- **Metadata is open; full text is licensed.** The NYT Developer API (Article
  Search, Archive) returns headlines, abstracts, sections, and dates for free with
  a key, but not the full body text. Corpus-scale full text comes through a
  licensed archive (for example ProQuest Historical Newspapers / TDM Studio, or a
  direct NYT text-and-data-mining licence).
- **Keyed by article and date.** A corpus extract is article text plus date,
  section, and length; topic models weight by article length (the citing paper
  uses length-weighted topic shares).
- API keys or archive credentials are required. Keep any credentials in `.env`,
  never hard-coded; respect the licence's rate and redistribution limits.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **The free API is not the full text.** A pipeline that pulls the NYT API and
  treats abstracts as the article body silently truncates the corpus; topic and
  sentiment measures built on abstracts differ from full-text measures. Confirm
  which layer you have.
- **OCR and vocabulary drift over a long span.** A corpus spanning 1871 to today
  carries OCR errors in the older scanned material and large shifts in vocabulary,
  spelling, and section structure; a model trained naively across the whole span
  conflates language change with topic change. The citing paper uses a rolling
  120-month estimation window to handle this. Account for drift.
- **Composition changes (sections, length, supply).** Article counts, average
  length, and the mix of sections grow and shift over decades; a raw topic count
  rises with the volume of print, not necessarily with attention. Normalize
  (length-weight, share-of-coverage) rather than count.
- **Redistribution is restricted.** Licensed full text cannot be republished;
  store derived features (topic loadings, counts), not the raw articles, and keep
  the corpus inside the licence. Do not redistribute text.
- **Reproducibility needs the exact corpus vintage.** The archive and API results
  change as articles are added or corrected; a text measure is only reproducible
  if the corpus snapshot and query are pinned. Record the vintage.

## Citation

Cite the publisher and the access route, e.g.: *New York Times article archive
(full text via [archive provider]), accessed YYYY-MM-DD.* State the corpus span,
the article count, the access layer (metadata API versus licensed full text), and
any length weighting.


==============================================================================
# Optimal Blue mortgage rate-lock data (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/optimal-blue/
# Optimal Blue captures mortgage rate-lock agreements and real-time lender offer distributions from its pricing-engine platform, a near-real-time view of locked rates and the offers borrowers could have gotten. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: mortgage-markets, rate-locks, household-finance, price-dispersion, licensed, data:optimal-blue
==============================================================================

:::caution[Licensed: not exercised here]
**Optimal Blue data is a paid licensed product** (Optimal Blue, a Constellation
Software business), so it carries **no provenance
badge**: the access path below was **not** run in this session. Treat it as
unverified until someone exercises it through a licensed Optimal Blue research
agreement. This is the honest grade under the institute's Verified discipline.
:::

**Optimal Blue** runs a mortgage product-and-pricing-engine platform used by a
large share of U.S. originators, so its data captures **rate-lock agreements**
(the locked rate, loan and borrower characteristics, program, points) and, via
the **Pricing Insight** product, the **distribution of lender offers** available
for an identical loan at a point in time. That combination lets researchers
compare what a borrower locked against what they could have gotten. A paper we
distill uses it:
[Bhutta, Fuster & Hizmo](/wiki/papers/jf/2026/bhutta-mortgage-overpayment-borrower-sophistication-2026/)
use 3.6 million Optimal Blue rate-lock agreements (January 2015 to December 2019)
and, for the subset of locks in 20 MSAs where offer data is sold, link each lock
to the real-time distribution of competing lender offers, to measure how much
borrowers overpay relative to the offers available for the same loan.

- **Cost:** licensed, research agreement. No free tier.
- **Vendor:** Optimal Blue (Constellation Software).
- **Coverage:** locked mortgages flowing through Optimal Blue's pricing engine,
  with loan / borrower attributes and (in Pricing Insight) the contemporaneous
  lender offer surface; the citing paper covers 2015 to 2019, with offer-level
  data for 20 MSAs.

## Access (when licensed)

- **Through an Optimal Blue research / data agreement.** Extracts are obtained
  under a licence with Optimal Blue; the lock-level data and the Pricing Insight
  offer distributions are separate products that must both be licensed for the
  overpayment comparison.
- **Keyed by loan and lock date.** A lock record carries the locked rate, points,
  FICO, LTV, loan amount, program, purpose, and MSA; matching to offers requires
  the same-day, same-MSA offer surface for a near-identical loan.
- **No credential is run here.** Keep any access credentials in `.env`, never
  hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Platform coverage, not the whole market.** The data covers originators that
  use Optimal Blue's pricing engine, not every U.S. lender; the panel of lenders
  and channels (retail, wholesale, correspondent) is selected. Check
  representativeness before generalizing to all originations.
- **Points and the rate are a package.** The locked rate is meaningless without
  the points paid; comparing rates across loans requires adjusting for the
  points-rate trade-off (the citing paper adjusts locked rates for points using
  the empirical relationship). Never compare bare rates.
- **The offer match defines the benchmark.** "Overpayment" is only as clean as the
  match between a lock and the offer distribution for an *identical* loan (same
  day, MSA, FICO, LTV, program, amount, points). A loose match contaminates the
  benchmark; the offer data also exists only for the subset of MSAs Optimal Blue
  sells, which is why the matched sub-sample (roughly 67,500 loans) is far smaller
  than the full 3.6 million locks.
- **A lock is not a closed loan.** Rate locks can expire, be renegotiated, or fall
  through; a lock record is the lock, not a funded mortgage. Do not treat lock
  counts as origination counts.
- **Linking to other mortgage data.** Joining to [HMDA](/wiki/datasets/hmda/) or
  to lender financials such as the [Mortgage Call Report (NMLS)](/wiki/confidential/mcr-nmls/)
  requires careful mapping of lender identity and loan attributes; the join has no
  shared key and must be built on characteristics. Document the merge.

## Citation

Cite the vendor and product, e.g.: *Optimal Blue rate-lock and Pricing Insight
data, accessed YYYY-MM-DD.* State the lock-date range, whether the offer
distribution was used, the MSAs covered, and the points adjustment applied.


==============================================================================
# OptionMetrics IvyDB: option prices, implied vols, and Greeks (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/optionmetrics/
# OptionMetrics IvyDB is the standard database of end-of-day option prices, OptionMetrics-computed implied volatilities and Greeks, and the standardized volatility surface for US exchange-listed equity and index options from 1996. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: options, equities, licensed, data:optionmetrics
==============================================================================

:::caution[Licensed: not exercised here]
**OptionMetrics IvyDB is a paid licensed product** (from OptionMetrics, an
independent data provider), so it carries **no
provenance badge**: the access path below was **not** run in this session (no
OptionMetrics or WRDS credentials were available). The page documents the access
route and the gotchas; treat it as unverified until someone exercises it through
a licensed account. This is the honest grade under the institute's Verified
discipline.
:::

**OptionMetrics IvyDB US** is the standard source for historical US option
data: **daily closing bid/ask quotes**, volume, and open interest for
exchange-listed equity and index options, plus **OptionMetrics-computed implied
volatilities and Greeks** (delta, gamma, vega, theta) and a **volatility
surface** (standardized, interpolated implied vols indexed by delta and
maturity). Coverage begins in 1996. The option records link to the underlying
security prices and to distribution (dividend and split) data, so a pipeline can
join options to their underlying. Papers we distill use it for option-pricing
work, for example
[Kruttli, Roth Tran & Watugala](/wiki/papers/jf/2025/kruttli-pricing-poseidon-extreme-weather-2025/)
on the pricing of extreme-weather and hurricane risk in options.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** OptionMetrics, an independent data provider.
- **Coverage:** US equity and index options 1996-present (IvyDB US); also IvyDB
  Global, IvyDB Europe, IvyDB Canada, and futures options are sold as separate
  products.

## Access (when licensed)

- **Most commonly through WRDS.** Like CRSP, Compustat, and I/B/E/S, IvyDB US is
  usually reached through [WRDS](/wiki/commercial/wrds/) under the OptionMetrics
  library, with tables such as `opprcd` (option prices), `secprd` (security
  prices), `vsurfd` (the standardized volatility surface), and the
  `optionm` metadata tables. Check whether your institution's WRDS subscription
  includes OptionMetrics; it is licensed separately.
- **Direct subscription.** IvyDB is also available as a direct OptionMetrics
  subscription (file delivery or API), independent of WRDS.
- Credentials and a licensed seat are required. Keep any credentials in `.env`,
  never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Implied vols and Greeks are model output, not market-observed.**
  OptionMetrics computes implied volatility and the Greeks (delta, gamma, vega,
  theta) from its own pricing model (a binomial / Cox-Ross-Rubinstein tree with
  OptionMetrics' dividend and interest-rate assumptions). These fields inherit
  the model's assumptions; they are not quoted by the exchange. Do not treat a
  reported implied vol or delta as a primitive observation.
- **The volatility surface is interpolated and smoothed.** The `vsurfd`
  product gives constant-maturity, fixed-delta implied vols produced by a
  kernel-smoothed fit across the available options, not raw quotes. Surface
  points are standardized constructs; do not treat them as tradable prices or
  as observed quotes for a specific contract.
- **Prices are end-of-day snapshot bid/ask, with quality issues.** Option
  records are closing bid/ask quotes (and their midpoint) from a daily snapshot,
  so they include stale quotes, crossed quotes, and illiquid contracts with a
  zero bid or a very wide spread. Filter on bid > 0, on spread, on volume, and
  on open interest before using prices.
- **The identifier is OptionMetrics' `secid`, which must be linked.** Underlying
  securities are keyed by OptionMetrics' own `secid`, not by permno or CUSIP.
  Use the provided linking table to map `secid` to CRSP permno / CUSIP rather
  than assuming a match; tickers are reused and change over time.
- **Survivorship and changing coverage.** Delisted underlyings and expired
  contracts are handled by point-in-time records, and exchange coverage (which
  options classes are included) has changed over the sample. A naive
  current-universe filter will drop history; check coverage by period.
- **Early-sample data revisions and known errors.** OptionMetrics has revised
  historical records, and the early years (late 1990s) have documented
  data-quality issues and thinner liquidity. Pin the extraction date and treat
  early-sample illiquid options cautiously.

## Reference: the main IvyDB tables/files

| Table / file | What it holds |
|---|---|
| `opprcd` (option prices) | Daily closing bid/ask, volume, open interest, implied vol, and Greeks per contract |
| `secprd` (security prices) | Daily prices and returns for the underlying securities |
| `vsurfd` | Standardized, interpolated implied vols by constant maturity and fixed delta |
| `optionm` metadata | Security and contract reference data, distributions, and the `secid` linking table |

Each is keyed by `secid` and date; option records carry contract terms (strike,
expiration, call/put) and the `optionid`.

## Citation

Cite the provider and the database, e.g.: *OptionMetrics IvyDB US, via WRDS,
accessed YYYY-MM-DD.* State the library and the exact tables used (for example
`opprcd`, `secprd`, `vsurfd`), the date range, and the extraction date.


==============================================================================
# Orbis (Bureau van Dijk): global firm financials and ownership (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/orbis-bvd/
# Orbis (Bureau van Dijk / Moody's Analytics) is a global firm-level database covering financial statements, ownership and corporate-structure links, and firm identifiers for public and private companies across countries. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: firms, financial-statements, licensed, panel-data, data:orbis-bvd
==============================================================================

:::caution[Licensed: not exercised here]
**Orbis is a licensed commercial dataset** (Bureau van Dijk, a Moody's
Analytics company), so it carries **no
provenance badge**: the access path below was **not** run in this session
(no BvD credentials were available). The page documents the access route
and the gotchas; treat it as unverified until someone exercises it through
a licensed account. This is the honest grade under the institute's Verified
discipline.
:::

**Orbis** is a global firm-level database from Bureau van Dijk (a Moody's
Analytics company). It aggregates financial statements, ownership and
corporate-structure links, and firm identifiers sourced from national
business registers and filings for public and (especially) private companies
across countries. It is used in, for example
[Greenwald, Krainer & Paul](/wiki/papers/jf/2025/greenwald-credit-line-channel-2025/)
(firm financials in the credit-line-channel analysis), and
[Bias, Lochner, Obernberger & Sevilir](/wiki/papers/jf/2026/bias-going-public-internal-organization-2026/)
(firm structure around going public).

- **Cost:** licensed, subscription. Institution-specific modules and country
  coverage determine scope and price. Usually a direct BvD license, not via
  WRDS.
- **Vendor:** Bureau van Dijk / Moody's Analytics.
- **Coverage:** hundreds of millions of firms globally, public and private.
  Financial-statement depth is uneven by country: deeper in countries with
  mandatory account filing (much of Europe), thinner in the US.

## Access (when licensed)

- **BvD Orbis platform.** The common access route is the Orbis web interface
  for ad hoc queries, batch extract requests, or the BvD data feed or API,
  keyed on the BvD ID.
- **Institution-hosted vintage.** Some institutions maintain a historical Orbis
  vintage on a research server; access rules are institution-specific.
- Credentials and an active license are required. Keep them in `.env`, never
  hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Country coverage is deeply uneven.** Europe has detailed filings due to
  mandatory account-filing requirements; US coverage is thin for private
  firms. Cross-country comparisons are biased by differences in reporting
  regimes; document the country-specific coverage gaps in your sample.
- **The live platform does not retain prior vintages.** BvD historically
  overwrote records rather than preserving history, so building a true panel
  requires archived annual disks or institutional vintages. Back-filled and
  restated figures change history silently in the live product.
- **Survivorship bias.** Dead and dissolved firms can drop out of the live
  product, biasing samples toward survivors. Use historical vintages and
  document how you handle firm exit.
- **Consolidation codes must be chosen deliberately.** Orbis carries both
  consolidated and unconsolidated accounts for the same firm; mixing them
  double-counts economic activity. State the consolidation code and apply it
  consistently.
- **The BvD ID is not stable.** Identifiers can change across BvD products
  (Orbis, Amadeus, Zephyr) and across vintages. Linking across products or
  vintages requires a crosswalk; pin the identifier scheme and vintage before
  any merge.
- **Ownership links are point-in-time.** Ownership-structure data reflects
  a snapshot, not a panel. Building ownership histories requires multiple
  vintages and explicit handling of the observation date for each link.
- **Currency, accounting standard, and units fields must be normalized.**
  Orbis mixes currencies, IFRS and local GAAP, and sometimes inconsistent
  units across countries and sources; pooling without normalization produces
  garbage comparisons.

## Citation

Cite the product and vendor, e.g.: *Orbis firm-level data, Bureau van Dijk
(Moody's Analytics), vintage [disk/year], consolidation code [C1/C2/U1/U2],
country sample [list], accessed [date].* State the vintage or disk, the
consolidation code, the country sample, and how survivorship and restatements
were handled in your sample construction.


==============================================================================
# PitchBook: private-capital and deal data (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/pitchbook/
# PitchBook (Morningstar) is a deal-level database of venture capital, private equity, and M&A: startups and their funding rounds, investors, valuations, and exits. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: private-equity, venture-capital, licensed, deal-data, data:pitchbook
==============================================================================

:::caution[Licensed: not exercised here]
**PitchBook is a licensed commercial dataset** (Morningstar), so it carries **no provenance badge**: the access path below was
**not** run in this session (no PitchBook credentials were available). The page
documents the access route and the gotchas; treat it as unverified until someone
exercises it through a licensed account. This is the honest grade under the
institute's Verified discipline.
:::

**PitchBook** is a **deal-level private-capital** database. It tracks startups
and private companies with their founding year, industry, and location; venture
and private-equity **funding rounds** (date, amount, investors, post-money
valuation); funds and their limited partners; and **exits** through IPO and M&A.
It is used in, for example
[Chen & Ewens](/wiki/papers/jf/2025/chen-venture-capital-startup-agglomeration-2025/)
for VC fund and startup data and LP commitment information,
[Hu & Ma](/wiki/papers/jf/2025/hu-persuading-investors-video-based-2025/) for
startup characteristics and funding rounds (alongside Crunchbase), and
[Barkai & Panageas](/wiki/papers/jf/2025/barkai-value-employment-2025/) for IPO
and M&A exit valuations by founding-year cohort.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** PitchBook Data (a Morningstar company).
- **Coverage:** private companies, deals, funds, and investors globally, with
  the deepest coverage in venture capital and private equity.

## Access (when licensed)

- **Direct from PitchBook.** Through the PitchBook platform, data exports, the
  Excel plug-in, or an API, keyed on PitchBook company, deal, and fund
  identifiers.
- **Possibly via a library or vendor feed.** Some institutions reach PitchBook
  through bundled products; availability is not confirmed here, so check with
  your library before assuming a given path.
- Credentials are required either way. Keep them in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Deal coverage is built from public sources plus voluntary input, so it skews
  to visible deals.** Larger, announced, and venture-backed deals are covered
  better than small or undisclosed private transactions; absence of a deal is
  not absence of activity.
- **Valuations and round terms are often estimated or missing.** Post-money
  valuations and deal amounts can be modeled or undisclosed; treat them as
  noisy, and check whether a figure is reported or estimated before using it as
  an outcome.
- **History gets restated.** Deals, companies, and valuations are added and
  revised retroactively as information surfaces, so a current pull is not what
  was knowable at an earlier date. Pin the data vintage and re-pull deliberately
  for point-in-time designs.
- **It overlaps imperfectly with Crunchbase and Preqin.** The same company or
  fund can carry different rounds, amounts, or valuations across vendors. Pick a
  source of truth per field and document discrepancies; papers that use both
  (for example Hu & Ma with Crunchbase) reconcile explicitly.
- **Industry, location, and founding-year fields carry classification noise.**
  These are PitchBook's own tags; state how you bucket them so the categories
  are reproducible.
- **Entity mapping is its own step.** Join on PitchBook company, deal, and fund
  identifiers and confirm the parent/subsidiary and ticker mapping before merging
  with public-market data.

## Citation

Cite the product and vendor, e.g.: *PitchBook private-capital data, PitchBook
Data (Morningstar); data licensed and accessed YYYY-MM-DD.* State the data
vintage, the module (company, deal, fund, or exit), and whether valuations are
reported or estimated.


==============================================================================
# Preqin: private-capital and hedge-fund data (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/preqin/
# Preqin is a fund-level database of private capital (private equity, venture, private debt, real assets) and hedge funds: fund sizes, vintages, returns, cash flows, and limited-partner commitments. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: private-equity, venture-capital, hedge-funds, licensed, panel-data, data:preqin
==============================================================================

:::caution[Licensed: not exercised here]
**Preqin is a licensed commercial dataset** (Preqin, now part of BlackRock), so it carries **no provenance badge**: the access
path below was **not** run in this session (no Preqin credentials were
available). The page documents the access route and the gotchas; treat it as
unverified until someone exercises it through a licensed account. This is the
honest grade under the institute's Verified discipline.
:::

**Preqin** is a **private-capital and hedge-fund** database. It covers private
equity, venture capital, private debt, real estate, and infrastructure funds,
plus hedge funds, with fund sizes, vintages, strategies, net returns and cash
flows (for performance and PME work), and **limited-partner (LP) commitment**
records that map institutions to the funds they back. It is used in, for example
[Siriwardane, Sunderam & Wallen](/wiki/papers/jf/2025/siriwardane-segmented-arbitrage-2025/)
for hedge-fund returns and a fixed-income-arbitrage strategy flag,
[Chen & Ewens](/wiki/papers/jf/2025/chen-venture-capital-startup-agglomeration-2025/)
for LP commitment data in a home-bias robustness test, and
[Johnston-Ross, Ma & Puri](/wiki/papers/jf/2025/johnston-ross-private-equity-financial-stability-2025/)
for PE fund size, vintage, and first-time-fund indicators.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** Preqin (acquired by BlackRock).
- **Coverage:** private-capital and hedge funds globally, with fund, performance,
  LP, and deal modules; depth varies by strategy and region.

## Access (when licensed)

- **Direct from Preqin.** Through the Preqin Pro platform, data exports, or an
  API, keyed on Preqin fund and firm identifiers.
- **Possibly via a library or vendor feed.** Some institutions reach Preqin
  through bundled products; availability is not confirmed here, so check with
  your library before assuming a given path.
- Credentials are required either way. Keep them in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Reporting is voluntary and partly FOIA-sourced, so selection bias is
  structural.** Fund performance comes from manager submissions and from
  freedom-of-information disclosures by public pension LPs. Funds that report,
  and the LPs subject to disclosure, are not a random sample; survivorship and
  backfill bias both apply, especially to hedge-fund returns.
- **The fund universe is incomplete and overlaps imperfectly with peers.**
  Coverage differs from PitchBook, Burgiss, and Cambridge Associates; the same
  fund can carry different sizes or returns across vendors. Decide on one source
  of truth per field and document discrepancies.
- **LP commitment data is partial.** Commitment records are strongest where
  public LPs must disclose; private LP commitments are thinly covered. Do not
  treat the LP map as the full investor base of a fund.
- **Net versus gross, and PME conventions, must be stated.** Returns are usually
  net of fees, but fee treatment and the public-market-equivalent benchmark
  choice drive the numbers. State the return basis and PME convention explicitly.
- **History gets restated.** Funds and performance are added and revised
  retroactively as managers report; a current pull is not what was known at an
  earlier date. Pin the data vintage and re-pull deliberately.
- **Entity mapping is its own step.** Join on Preqin fund and firm identifiers
  and confirm the manager, fund-family, and portfolio-company mapping before
  merging with other sources.

## Citation

Cite the product and vendor, e.g.: *Preqin private-capital data, Preqin; data
licensed and accessed YYYY-MM-DD.* State the data vintage, the module (fund,
performance, LP, or deal), the return basis, and any PME convention used.


==============================================================================
# RateWatch deposit-rate surveys (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/ratewatch/
# RateWatch (S&P Global Market Intelligence) is the standard branch-level survey of U.S. deposit and CD rates, posted-rate data at weekly frequency. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: deposit-rates, banking, branch-level, deposit-competition, licensed, data:ratewatch
==============================================================================

:::caution[Licensed: not exercised here]
**RateWatch is a paid licensed product** (S&P Global Market Intelligence), so it carries **no provenance badge**: the access path
below was **not** run in this session. Treat it as unverified until someone
exercises it through a licensed S&P Global / RateWatch subscription. This is the
honest grade under the institute's Verified discipline.
:::

**RateWatch** surveys the **posted deposit and CD rates** of U.S. bank and credit
union branches, at the branch level and weekly frequency, across products
(savings, money market, certificates of deposit by term and balance tier). It is
the standard source for the **price of deposits** in the cross-section of
branches, used to study deposit competition, rate-setting, and the deposits
channel. A paper we distill uses it:
[Martin, Puri & Ufier](/wiki/papers/jf/2026/martin-deposit-flows-failing-banks-2026/)
use RateWatch 12-month CD rate spreads to show that large banks under regulatory
enforcement raise deposit rates (paying on average roughly 75 basis points above
the FDIC national average, near the regulatory rate cap) to attract insured term
deposits.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** S&P Global Market Intelligence (RateWatch).
- **Coverage:** branch-level posted deposit and CD rates for U.S. depository
  institutions, weekly, by product and balance tier; broad branch coverage with a
  long weekly history.

## Access (when licensed)

- **Through an S&P Global / RateWatch subscription.** Extracts are obtained under
  a RateWatch licence (directly or through S&P Global Market Intelligence); some
  institutions reach it via a market-data agreement.
- **Keyed by branch, product, and week.** A rate record is a posted rate for a
  product (for example a 12-month CD at a balance tier) at a branch in a survey
  week; aggregate to bank level with branch weights when you need an
  institution-level rate.
- **No credential is run here.** Keep any access credentials in `.env`, never
  hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Posted rates, not paid rates.** RateWatch records the rate a branch advertises,
  not the average rate actually paid on the existing deposit book; the posted rate
  leads and overstates marginal cost relative to the realized interest expense in
  Call Reports. Match the concept to your question and do not equate the two.
- **Product and balance tier must be fixed.** Deposit rates vary sharply by product
  (savings versus CD), CD term, and balance tier; comparing across branches
  requires holding the product and tier constant (the citing paper fixes on the
  12-month CD). Specify the product.
- **Branch versus institution aggregation.** A bank has many branches that may post
  different rates; rolling up to a bank-level rate needs a weighting choice
  (deposit-weighted, simple, rate-setter branch). State the aggregation.
- **Rate-setter and follower structure.** RateWatch identifies rate-setting
  branches whose rates other branches mirror; treating every branch as an
  independent observation overstates the effective sample size. Account for the
  rate-setter design.
- **Linking to bank financials.** Joining branch rates to bank-level Call Report
  data (see [Call Reports](/wiki/datasets/call-reports/)) requires mapping branch
  to RSSD / institution across mergers and branch sales; the join has known
  mismatches around acquisition events. Verify the link.

## Citation

Cite the vendor and product, e.g.: *RateWatch (S&P Global Market Intelligence),
accessed YYYY-MM-DD.* State the deposit product and balance tier, the branch-to-
bank aggregation, and the weekly date range.


==============================================================================
# RavenPack: news and event analytics (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/ravenpack/
# RavenPack turns text news into structured, timestamped entity-event records with sentiment, relevance, and novelty scores. It is a paid subscription: this page documents the access path and the gotchas that bite event-study pipelines, but the data was not exercised here.
# Tags: equities, licensed, event-data, data:ravenpack
==============================================================================

:::caution[Licensed: not exercised here]
**RavenPack is a paid subscription**, so it carries **no provenance badge**: the access path below was **not** run in this
session (no RavenPack credentials were available). The page documents the
access route and the gotchas; treat it as unverified until someone exercises
it through a licensed account. This is the honest grade under the institute's
Verified discipline, which forbids a badge for an unrun claim.
:::

**RavenPack** converts unstructured news and other text into structured,
machine-readable **entity-event records**: each detected company, person, or
place in a story gets a timestamped row carrying a sentiment score, a
relevance score, a novelty score, and a topic taxonomy. It is a
commercial feed for news-based signals in finance. The
[Kwan, Liu & Matthies](/wiki/papers/jf/2026/kwan-liu-matthies-2026/) attention paper
uses **RavenPack 1.0** to classify the topic, subject, and sentiment of the
news that institutional investors read, and to map stories to stock tickers.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** RavenPack (news and event analytics; the analytics product has
  versioned editions, e.g. a legacy **1.0** and later editions, with
  different fields and taxonomies; confirm the edition against current vendor
  documentation).
- **Coverage:** global news (newswires such as Dow Jones, plus web and
  press-release sources); millions of entity-event records per day.

## Access (when licensed)

- **Direct from the vendor.** Delivered as historical flat-file dumps plus an
  ongoing feed (SFTP / API / cloud share). You pull dated files and the
  point-in-time analytics for each story.
- **Via WRDS.** Subscribing institutions can reach RavenPack News Analytics
  through [WRDS](/wiki/commercial/wrds/), which is the easiest academic route if
  your university licenses it. Query it like any other WRDS library; filter on
  date and relevance before pulling.
- Either way, credentials are required. Keep them in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Filter on relevance first.** Each record has a `RELEVANCE` score (0 to 100).
  A story that merely mentions a company in passing scores low; keep
  `RELEVANCE = 100` (or a high threshold) when you want stories that are
  *about* the entity, or your signal is mostly noise.
- **Deduplicate with the novelty score.** Newswires re-run and syndicate the
  same story. The Event Novelty Score (`ENS`) flags the first report
  (`ENS = 100`) versus echoes; without it you double-count one event many
  times.
- **Timestamps are UTC; align to the trading calendar.** The production
  timestamp is when the record was created, in UTC. Convert to the market
  timezone and decide deliberately how to treat after-close and weekend news
  before forming a daily signal.
- **Use the point-in-time timestamp, not the story date, to avoid look-ahead.**
  Signals must be built from when the analytics were available, not when the
  event nominally happened.
- **Sentiment scales differ by version.** The Event Sentiment Score and
  Composite Sentiment Score conventions changed between RavenPack 1.0 and
  later editions (e.g. a 0 to 100 scale with 50 neutral). Confirm which edition
  and scale you are on before thresholding.
- **Entity mapping is its own step.** Records key on `RP_ENTITY_ID`, not a
  ticker or PERMNO. Use RavenPack's mapping files to join to security
  identifiers, and remember entities include private firms, people, and
  places, not just listed equities.
- **Volume.** The raw feed is large; filter on entity, relevance, and date at
  read time rather than loading everything.

## Citation

Cite the product and edition, e.g.: *RavenPack News Analytics (RavenPack 1.0),
RavenPack; data licensed and accessed YYYY-MM-DD.* State the relevance and
novelty filters used, since they materially define the sample.


==============================================================================
# Refinitiv (LSEG) earnings-call transcripts (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/refinitiv-transcripts/
# Refinitiv (now LSEG) distributes transcripts of analyst-management conference calls, a standard corpus for textual analysis of disclosure. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: text, disclosure, earnings-calls, nlp, licensed, data:refinitiv-transcripts
==============================================================================

:::caution[Licensed: not exercised here]
**Refinitiv (LSEG) transcripts are a paid licensed product**, so they carry **no provenance badge**: the access path below was
**not** run in this session (no Refinitiv / LSEG credentials were available). The
page documents the access route and the gotchas; treat it as unverified until
someone exercises it through a licensed account. This is the honest grade under
the institute's Verified discipline.
:::

**Refinitiv (LSEG) earnings-call transcripts** are textual transcripts of
public-company conference calls (earnings calls and other analyst-management
calls), with the call segmented into presentation and question-and-answer
sections and tagged by speaker (management, analyst, operator). They are one of
the standard corpora for textual analysis of corporate disclosure, alongside
S&P Global / Capital IQ transcripts; researchers use them to measure tone,
topic attention, and the timing of management discussion of specific issues. A
paper we distill uses them:
[Kruttli, Roth Tran & Watugala](/wiki/papers/jf/2025/kruttli-pricing-poseidon-extreme-weather-2025/)
run textual analysis of hurricane-channel discussion across the 120 trading days
after landfall over 2002-2019.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** Refinitiv / LSEG (lineage through Thomson Reuters; the transcript
  product traces to StreetEvents).
- **Coverage:** global public companies, denser for large U.S. and developed-market
  firms; usable history reaches back to the early 2000s, with broader coverage in
  later years.

## Access (when licensed)

- **Through a Refinitiv/LSEG feed or a research data licence.** Transcripts are
  delivered as structured documents (often per-call files with speaker and
  section tags) through a Refinitiv/LSEG entitlement or a negotiated bulk research
  licence. Check which route your institution provides.
- **Keyed by company identifier and event date.** Each transcript is tied to a
  company (Refinitiv identifiers, mappable to RIC/PERMNO) and a call date/event;
  an extract is a set of calls for a firm list over a date range.
- Credentials are required. Keep any credentials in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Coverage is selective and grows over time.** Not every call is transcribed;
  coverage skews to larger, more-followed firms and thickens in later years.
  Absence of a transcript is not evidence a call did not happen, and a firm panel
  conditioned on transcript presence is size-biased.
- **Speaker and section tags are the analysis unit and they vary.** Tone and
  attention measures differ sharply between the scripted presentation and the
  unscripted Q&A; tagging quality and the operator/analyst/management labels are
  not perfectly consistent across vintages. Parse and validate the segmentation
  rather than treating the whole transcript as one block.
- **Transcription is approximate.** Transcripts are produced from audio and
  contain misspellings, mis-attributions, and "[inaudible]" gaps; named-entity
  and keyword counts inherit these errors. Build dictionaries and matching to be
  robust to transcription noise.
- **Event timing versus call timing.** The call date is not the period the call
  discusses; aligning text to an event window (as in the distilled paper's
  post-landfall windows) requires the actual call date, not the fiscal period.
  Be explicit about which date drives the window.
- **Licensing forbids redistribution.** The corpus cannot be re-posted; published
  work can report derived measures (counts, tone scores) but not the raw text.
  Keep derived features, not the source documents, in any shared artifact.
- **Identifier linking.** Refinitiv company identifiers must be linked to
  CRSP/Compustat; the join has known mismatches across reorganizations and
  ticker changes. Verify the link rather than assuming a clean one-to-one map.

## Reference: transcript structure

| Element | What it is |
|---|---|
| Header | Company, call date/time, event type, identifiers |
| Presentation | Scripted management remarks |
| Q&A | Unscripted analyst questions and management answers |
| Speaker tags | Management / analyst / operator attribution |

The presentation/Q&A split and speaker tags are the units most textual measures
are built on; validate the segmentation for the vintage you pulled.

## Citation

Cite the provider and database, e.g.: *Refinitiv / LSEG earnings-call
transcripts, accessed YYYY-MM-DD.* State the firm universe, the date range, the
sections analyzed (presentation, Q&A, or both), and the extraction date so the
text sample is reproducible.


==============================================================================
# RepRisk: ESG risk-incident data (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/reprisk/
# RepRisk is a daily firm-level feed of negative environmental, social, and governance incidents sourced from media and stakeholder reports, scored for severity, reach, and novelty across 28 issue categories. It is licensed: this page documents the access path and the gotchas; the access path was exercised through a licensed WRDS session.
# Access confirmed (licensed) 2026-06-09 · via live RepRisk query (reprisk.pm_company_identifiers) through a licensed WRDS session
# Tags: equities, esg, news, licensed, event-data, data:reprisk
==============================================================================

:::note[Access confirmed via a licensed WRDS session]
The keystone query `SELECT * FROM reprisk.pm_company_identifiers LIMIT 1` was run against the live dataset through a licensed WRDS session on 2026-06-09, so this page carries the amber "Access confirmed (licensed)" badge. Reproducing it still requires the reader's own WRDS account and the institution's RepRisk entitlement (licensed, not open).
:::

**RepRisk** is an **ESG risk-incident** database. Rather than a static rating, it
records discrete **negative incidents** at the firm level, captured daily from
media, NGO, government, and other stakeholder sources, classified into 28 ESG
issue categories and scored for **severity**, **reach** (source breadth), and
**novelty**. It also publishes the firm-level RepRisk Index (RRI), a running
measure built from the incident flow. Coverage starts in 2007. It is used in, for
example
[Derrien, Kruger, Landier & Yao](/wiki/papers/jf/2025/derrien-esg-news-future-cash-2025/)
as the main independent variable (daily negative ESG incidents, 2007 to 2019),
[Duchin, Gao & Xu](/wiki/papers/jf/2025/duchin-sustainability-greenwashing-evidence-asset-2025/)
for ESG risk incidents, and
[Starks, Venkat & Zhu](/wiki/papers/jf/2026/starks-esg-profiles-investor-horizons-2026/)
for negative environmental and social incident data.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** RepRisk AG.
- **Coverage:** firm-level incidents and RRI from 2007, global, across listed
  and unlisted companies, mapped to a RepRisk company identifier.

## Access (when licensed)

- **Direct from RepRisk.** Through their data feed, API, or a cloud data share,
  keyed on the RepRisk company identifier.
- **Via WRDS.** RepRisk is available on WRDS; the keystone query above runs
  against `reprisk.pm_company_identifiers` through a licensed WRDS session. Some
  institutions also reach it through a direct RepRisk feed or bundled products.
  Check your library's entitlement.
- Credentials are required either way. Keep them in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented.

- **It is an incident flow, not a rating, and only negative.** RepRisk records
  negative events; there are no positive offsets. A firm with no recorded
  incidents is not certified clean, only unmentioned. Do not read a zero as good
  ESG performance.
- **Media-attention bias confounds with size and visibility.** Incidents are
  captured from public sources, so larger, more-covered, English-language, and
  US/European firms accumulate more incidents mechanically. Control for media
  coverage and firm size before reading incident counts as risk.
- **Severity, reach, and novelty are RepRisk's own scores.** These are vendor
  classifications, not observed outcomes; state which fields you use and how you
  aggregate daily incidents to your event window.
- **The RRI is a derived, decaying index.** The RepRisk Index is constructed
  from the incident flow with its own decay and weighting; it is not the raw
  count. Decide deliberately whether your design wants raw incidents or the RRI.
- **History gets restated and methodology evolves.** Categories and scoring have
  been refined over time, and the back-catalogue can be reprocessed. Pin the
  data vintage and re-pull deliberately.
- **Entity mapping is its own step.** Join on the RepRisk company identifier and
  confirm the parent/subsidiary and ticker mapping before merging with returns
  or fundamentals.

## Citation

Cite the product and vendor, e.g.: *RepRisk ESG risk data, RepRisk AG; data
licensed and accessed YYYY-MM-DD.* State the data vintage, whether you use raw
incidents or the RRI, and the issue categories and scores included.


==============================================================================
# Revelio Labs: workforce and human-capital data (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/revelio/
# Revelio Labs builds a firm-level workforce panel from public professional profiles and job postings: headcount, hiring and attrition, role and seniority mix, and education. It is a paid subscription: this page documents the access path and the gotchas; the access path was exercised through a licensed WRDS session on 2026-06-09.
# Access confirmed (licensed) 2026-06-09 · via live Revelio Labs query (revelio.individual_positions) through a licensed WRDS session
# Tags: equities, licensed, panel-data, data:revelio
==============================================================================

:::note[Access confirmed via a licensed WRDS session]
The keystone query `SELECT * FROM revelio.individual_positions LIMIT 1` was run
against the live dataset through a licensed WRDS Postgres session on 2026-06-09
and returned a real row (24 columns). The page carries the amber "Access
confirmed (licensed)" badge. Reproducing this still requires your own WRDS
account and your institution's Revelio entitlement; the data is licensed, not
open.
:::

**Revelio Labs** builds a firm-level **workforce / human-capital panel** from
public professional profiles and online job postings: headcount, hiring and
attrition flows, role and seniority composition, skills, education (including
the **advanced-degree share** of staff), inferred compensation, and sentiment.
It maps this to companies and, where listed, to tickers. The
[Kwan, Liu & Matthies](/wiki/papers/jf/2026/kwan-liu-matthies-2026/) attention paper
uses Revelio / LinkedIn data for fund human capital (the advanced-degree share
of staff) in its result that more efficient funds employ more highly educated
analysts.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** Revelio Labs (workforce intelligence built from aggregated
  public professional profiles and job postings).
- **Coverage:** company-level workforce panels, mapped to a Revelio company
  identifier and, where applicable, to listed-equity tickers.

## Access (when licensed)

- **Direct from Revelio Labs.** Through their data feed, API, or a cloud data
  share (e.g. Snowflake), keyed on the Revelio company identifier.
- **Via WRDS.** Revelio Labs is available on [WRDS](/wiki/commercial/wrds/); the
  keystone query above runs against `revelio.individual_positions` through a
  licensed WRDS session. Some institutions also reach it through a direct
  Revelio feed or cloud share. Check your library's entitlement.
- Credentials are required either way. Keep them in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented below.

- **Coverage skews to white-collar, US, and large firms.** The data derives
  from public professional profiles, which over-represent office roles, the
  US, and big employers. Headcount levels are estimates, not a census; treat
  cross-firm level comparisons with care and prefer within-firm changes.
- **Profiles are self-reported and inferred.** Roles, seniority, and dates are
  inferred from self-authored profiles, so titles and start/end dates carry
  noise. Education fields (the advanced-degree share) inherit that noise.
- **Historical panels get restated.** As the underlying models and profile
  coverage are reworked, prior periods can be revised. Pin the data vintage
  and re-pull deliberately rather than mixing vintages.
- **Entity mapping is its own step.** Join on the Revelio company identifier
  and confirm the parent/subsidiary and ticker mapping; a single listed parent
  can span many subsidiary employers, and vice versa.
- **Timing is inferred, not point-in-time disclosure.** Hiring and attrition
  are reconstructed from profile changes, which surface with a lag and
  irregular timing; do not treat a monthly series as a clean as-of snapshot.
- **Define the metric explicitly.** "Advanced-degree share" and similar
  measures depend on how degrees and the staff denominator are defined; state
  the definition so the number is reproducible.

## Citation

Cite the product and vendor, e.g.: *Revelio Labs workforce data, Revelio Labs;
data licensed and accessed YYYY-MM-DD.* State the data vintage and the exact
metric definition used.


==============================================================================
# Rystad Energy database (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/rystad/
# Rystad Energy maintains asset-level oil and gas data: production, costs, reserves, and field economics for operators worldwide. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: energy, oil-gas, commodities, production, firm-level, licensed, data:rystad
==============================================================================

:::caution[Licensed: not exercised here]
**The Rystad Energy database is a paid licensed product** (Rystad Energy), so it carries **no provenance badge**: the access path
below was **not** run in this session. Treat it as unverified until someone
exercises it through a Rystad subscription. This is the honest grade under the
institute's Verified discipline.
:::

**The Rystad Energy database** is an asset-level model of the global oil and gas
industry: production volumes, operating and capital expenditures, reserves, and
field-level economics, built bottom-up from individual fields and tied to the
operating companies. Its products (UCube and related modules) are the reference
source for firm-by-field production and cost data when public filings are too
aggregated. A paper we distill uses it:
[Allcott, Montanari, Ozaltun & Tan](/wiki/papers/jf/2026/allcott-corporate-social-impact-2026/)
use Rystad for the oil production and operating expenses of the seven oil
supermajors across all of their oil fields worldwide (2018) when estimating the
consumer-surplus component of corporate social impact and the high impact per
dollar of revenue of oil producers under inelastic global demand.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** Rystad Energy (UCube and related upstream data products).
- **Coverage:** global upstream oil and gas at the field and company level, with
  production, cost, and reserve estimates; some figures are modeled rather than
  reported.

## Access (when licensed)

- **Through a Rystad Energy subscription.** Data are accessed via Rystad's
  platform (UCube and related modules) with extracts or API access where
  entitled; the grain runs from individual field to operating company to
  country aggregate.
- **Company-by-field is the useful join.** The citing paper aggregates
  field-level production and operating expense up to each supermajor; the
  field-to-operator mapping is the key Rystad provides that public filings do
  not.
- Credentials are required. Keep any credentials in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Many figures are modeled, not reported.** Rystad estimates field economics
  (costs, reserves, decline curves) from a proprietary model; they are
  estimates, not audited filings. Treat operating-cost and reserve numbers as
  vendor estimates and state that, as the citing paper's supply-elasticity
  argument rests on them.
- **Operator versus equity ownership of a field.** Production attributed to a
  company can be on an operated basis or a working-interest basis; the two
  differ for joint ventures and partial stakes. Confirm which attribution you
  pulled before summing a firm's output.
- **Vintage revisions.** Reserve and production estimates are revised as new
  information arrives, so the same field-year can change between data vintages.
  Pin the extract date; a re-pull will not reproduce an earlier number exactly.
- **Coverage and quality vary by region and operator.** Disclosure-poor
  jurisdictions and small private operators are modeled with less precision than
  listed majors. Country and company aggregates inherit that uneven precision.
- **Unit and currency conventions.** Volumes (barrels of oil equivalent versus
  oil only), cost units, and currency must be normalized before comparison; the
  oil-only restriction the citing paper applies is a deliberate scoping choice,
  not the database default.

## Citation

Cite the vendor and product, e.g.: *Rystad Energy (UCube), accessed
YYYY-MM-DD.* State which module and grain (field, company, country) were used,
the extract vintage, and that production or cost figures are vendor estimates
where applicable.


==============================================================================
# SDC Platinum: M&A and new-issues deal data (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/sdc-platinum/
# SDC Platinum is the standard deal-level database of mergers and acquisitions and new security issues (IPOs, SEOs, debt), assembled by LSEG / Refinitiv. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: equities, mergers-and-acquisitions, licensed, event-data, data:sdc-platinum
==============================================================================

:::caution[Licensed: not exercised here]
**SDC Platinum is a paid licensed product** (LSEG / Refinitiv), so it carries **no provenance badge**: the access path below was
**not** run in this session (no Refinitiv / SDC credentials were available). The
page documents the access route and the gotchas; treat it as unverified until
someone exercises it through a licensed account. This is the honest grade under
the institute's Verified discipline.
:::

**SDC Platinum** (Securities Data Company) is the long-standing deal-level
database used across corporate finance: **mergers and acquisitions** (the M&A
database) and **new issues** (the Global New Issues database: IPOs, seasoned
equity offerings, and debt issuance), with deal terms, dates, parties,
advisers, and deal values. It is the standard source for M&A and issuance
samples. Papers we distill use it for deal samples, for example
[Guenzel](/wiki/papers/jf/2025/guenzel-too-deep-effect-sunk-2025/) on sunk-cost
effects in acquisitions and
[Cookson, Niessner & Schiller](/wiki/papers/jf/2026/cookson-social-media-merger-withdrawals-2026/)
on merger withdrawals.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** LSEG / Refinitiv (formerly Thomson Reuters / Thomson Financial;
  the product traces back to Securities Data Company).
- **Coverage:** global M&A from the 1980s and new security issues, deal-level,
  with target/acquirer identifiers, deal status and dates, consideration, and
  league-table fields. Newer access is migrating into Refinitiv Workspace
  (Deals), with the legacy SDC Platinum desktop being retired.

## Access (when licensed)

- **Refinitiv / LSEG, not WRDS.** Unlike CRSP, Compustat, and I/B/E/S, SDC is
  usually reached through the Refinitiv terminal (legacy SDC Platinum desktop or
  the newer Refinitiv Workspace "Deals" application), not through
  [WRDS](/wiki/commercial/wrds/). Check whether your institution licenses it
  separately from its WRDS subscription.
- **Custom reports.** Data is pulled by building a query (deal type, date range,
  nation, status) and exporting a custom report; there is no single canonical
  table. Save the request definition so the extract is reproducible.
- Credentials and a terminal seat are required. Keep any credentials in `.env`,
  never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **The sample is the query, so the query is the method.** Because every extract
  is defined by a point-and-click request (deal flags, date window, nation,
  public/private status, minimum deal value), two researchers asking "the same"
  question can get different samples. Record the exact request definition and
  every filter; the extract is not reproducible without it.
- **Deal value is missing for a large share of deals.** Many transactions,
  especially private and small deals, have no disclosed value. Filtering on
  deal value silently selects toward large, public, US deals; treat a
  value-conditioned sample as non-random.
- **Coverage and flags change over time and get backfilled.** SDC has revised
  historical records and added deals retroactively, and definitions of fields
  (deal attitude, completion status, percent sought/acquired) are not uniform
  across the full history. Pin the extraction date and do not assume a field
  means the same thing across decades.
- **Status and double-counting need care.** A single transaction can appear as
  multiple records (amended terms, competing bids, tranches), and the
  completed/withdrawn/pending status drives the sample. Define inclusion on
  status explicitly and de-duplicate on the deal identifier, not on
  target-acquirer-date alone.
- **Documented data-quality errors.** The reliability of SDC M&A fields has been
  questioned in the accounting and finance literature; do not treat individual
  flags as ground truth. Cross-check material deals against filings
  ([SEC EDGAR](/wiki/datasets/edgar/)) or news when the deal drives the result.
- **Identifiers need linking to returns data.** Target and acquirer are keyed by
  SDC's own identifiers plus CUSIP/ticker; join to CRSP via CUSIP with the usual
  historical-CUSIP caveats rather than assuming a clean permno match.

## Reference: the two main databases

| Database | What it holds |
|---|---|
| M&A | Deal-level mergers, acquisitions, tender offers, buybacks, stake purchases |
| Global New Issues | IPOs, seasoned equity offerings, and debt issuance (terms, dates, underwriters) |

Each is queried by deal type, date range, nation, and status into a custom
exported report; field availability differs between the two.

## Citation

Cite the provider and the database, e.g.: *SDC Platinum (M&A) / Refinitiv (LSEG),
accessed YYYY-MM-DD.* State the database (M&A or New Issues), the exact query
filters (deal type, dates, nation, status, deal-value cutoff), and the
extraction date.


==============================================================================
# Siblis Research index-constituent data (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/siblis-research/
# Siblis Research sells historical index addition and deletion dates and constituent market values for the S&P 500, MidCap 400, SmallCap 600, and Nasdaq 100. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: equity-indices, index-constituents, index-reconstitution, licensed, data:siblis-research
==============================================================================

:::caution[Licensed: not exercised here]
**Siblis Research data is a paid subscription product**, so it carries **no provenance badge**: the access path below was
**not** run in this session. Treat it as unverified until someone exercises it
through a Siblis subscription. This is the honest grade under the institute's
Verified discipline.
:::

**Siblis Research** is a subscription data provider whose research-relevant
product is **historical index membership**: the dates on which a stock was added
to or deleted from the S&P 500, S&P MidCap 400, S&P SmallCap 600, and Nasdaq 100,
along with historical constituent market values and weights. Reconstructed index
event histories are the raw material for index-effect and demand-curve studies. A
paper we distill uses it:
[Greenwood & Sammon](/wiki/papers/jf/2025/greenwood-disappearing-index-effect-2025/)
take 736 S&P 500 additions and 731 deletions over 1980 to 2020 from Siblis
Research (plus MidCap 400, SmallCap, and Nasdaq 100 changes from 1995) when
documenting the disappearance of the index-inclusion effect.

- **Cost:** licensed, subscription. No free historical-constituent download.
- **Vendor:** Siblis Research.
- **Coverage:** addition / deletion dates and constituent market values for the
  major S&P and Nasdaq indices; S&P 500 history back to 1980 in the citing paper,
  MidCap / SmallCap / Nasdaq from 1995.

## Access (when licensed)

- **Through a Siblis Research subscription.** The historical constituent and
  index-change files are bought from Siblis directly (spreadsheet / download
  products by index).
- **Two distinct products.** The change history (dates of additions and
  deletions) and the point-in-time constituent market values are different
  extracts; an index-event study needs the dated change list, while a weighting or
  demand-shock measure needs the constituent values.
- **No credential is run here.** Keep any account credentials in `.env`, never
  hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Announcement date versus effective date.** Index changes have an announcement
  date and a (later) effective date; the abnormal return accrues between them, so
  using the wrong date misplaces the event window. The citing paper anchors its
  baseline window to the trading day before announcement through the day after the
  effective date. Confirm which date each field is.
- **Coverage starts later for the smaller indices.** S&P 500 history reaches back
  to 1980, but MidCap / SmallCap / Nasdaq histories begin around 1995, and
  announcement dates are not available for some sub-indices (events are dated off
  the effective date). Do not assume uniform history across indices.
- **Migrations are not plain additions.** A stock moving from the MidCap 400 to
  the S&P 500 is simultaneously a deletion and an addition; treating it as a clean
  addition double-counts the demand shock (the migration mechanism is central to
  the citing paper). Flag migrations explicitly.
- **Pre-1990 gaps need an outside source.** Early S&P 500 announcement dates are
  incomplete in the historical record; the citing paper fills pre-1990
  announcements from Barberis, Shleifer & Wurgler (2005). Plan for a supplemental
  source if your sample reaches back that far.
- **Identifier matching to CRSP / Compustat.** Constituents must be matched to
  CRSP permnos for returns; the join has known mismatches across ticker reuse and
  reorganizations. Verify the link.

## Citation

Cite the vendor, e.g.: *Siblis Research index-constituent data, accessed
YYYY-MM-DD.* State the index, the date field used (announcement versus effective),
and the constituent-value vintage.


==============================================================================
# SteelBenchmarker steel price index (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/steelbenchmarker/
# SteelBenchmarker publishes biweekly reference prices for hot-rolled band, cold-rolled coil, scrap, and other steel products. Current spot reports are free, but the full historical product-level series is a subscription product. This page documents the access path and the gotchas; the series was not exercised here.
# Tags: commodities, steel, prices, reference-price, market-data, licensed, data:steelbenchmarker
==============================================================================

:::caution[Licensed: not exercised here]
**SteelBenchmarker's full historical series is a paid product** (World Steel
Dynamics / American Metal Market), so it carries
**no provenance badge**: the access path below was **not** run in this session.
The latest spot prices are posted free, but the product-level history used for
research is a subscription. Treat it as unverified until someone exercises the
full series. This is the honest grade under the institute's Verified discipline.
:::

**SteelBenchmarker** publishes biweekly reference transaction prices for U.S.
and world steel products: hot-rolled band (coil), cold-rolled coil, hot-dipped
galvanized, standard plate, and steel scrap grades. Prices are built from a
survey of buyers and sellers and serve as the public benchmark that physical
steel contracts and the NYMEX steel futures settle against. A paper we distill
uses it:
[Martin](/wiki/papers/rfs/2025/martin-real-effects-centralized-markets-2025/)
uses SteelBenchmarker product-level biweekly prices (January 2007 to December
2017, 19,653 product-publication-date observations across six products) as the
primary source for measuring price dispersion and price levels when testing
whether the introduction of steel futures reduced dispersion in the physical
product market.

- **Cost:** the latest spot reference prices are posted free on the
  SteelBenchmarker site; the full historical, product-level series is a paid
  subscription. No open bulk download of the history.
- **Vendor:** World Steel Dynamics with American Metal Market (the citing paper
  calls it a proprietary price database).
- **Coverage:** biweekly, by product and region (USA, China, World Export),
  going back to 2006; the cross-product structure is what enables
  dispersion-across-products tests.

## Access (when licensed)

- **Spot prices free; history licensed.** The current biweekly release is
  public; the multi-year, product-level back series used for panel work comes
  through a SteelBenchmarker / American Metal Market subscription.
- **The grain is product by publication date.** Each release carries one price
  per product per region; the panel is assembled by stacking releases, as the
  citing paper does around the futures-introduction events.
- Credentials are required for the historical series. Keep any credentials in
  `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Biweekly, not daily.** Prices update roughly every two weeks, so event
  windows are measured in publication dates, not trading days; the citing paper
  counts 40 publication dates before and after each event. Do not assume a daily
  series exists.
- **Survey-based reference prices, not transaction tape.** Each print is a
  reference level derived from a buyer-and-seller survey, not a clearing price
  from a tape. It is a benchmark, with the smoothing and judgment that implies;
  dispersion measures must be built across reporting firms or products, not
  within a single tick.
- **Free spot page is not the research series.** The public site shows only the
  latest values; scraping it does not reconstruct the licensed history, and the
  free numbers can be revised relative to the archived series. Use the licensed
  back series for any panel.
- **Product definitions and regional baskets matter.** Hot-rolled band,
  cold-rolled coil, and the scrap grades are distinct series with different
  liquidity and different futures linkage; treating them as one steel price
  conflates the treated and control products the design relies on.
- **Units and currency.** Prices are quoted per ton in USD for U.S. products;
  reconcile metric versus short ton and region before comparing across baskets.

## Citation

Cite the source and product, e.g.: *SteelBenchmarker (World Steel Dynamics /
American Metal Market), [dates], accessed YYYY-MM-DD.* State the products and
regions used, the date range, and whether the free spot page or the licensed
historical series was the source.


==============================================================================
# StockTwits social-media messages (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/stocktwits/
# StockTwits is a finance-focused social platform whose ticker-tagged messages, often self-labeled bullish or bearish, are used as a retail-sentiment signal. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: social-media, sentiment, text-as-data, retail-investors, fintech, licensed, data:stocktwits
==============================================================================

:::caution[Licensed: not exercised here]
**StockTwits message data is a paid licensed product** (StockTwits Inc., often
via Social Market Analytics), so it carries **no
provenance badge**: the access path below was **not** run in this session. Treat
it as unverified until someone exercises it through a StockTwits or SMA data
licence. This is the honest grade under the institute's Verified discipline.
:::

**StockTwits** is a finance-focused social network on which users post short
messages tagged with stock tickers (`$AAPL`), frequently attaching a
self-declared **Bullish** or **Bearish** label. That explicit sentiment tag,
plus the ticker structure, makes it a widely used measure of retail investor
sentiment at the firm-day level for text-as-data research. A paper we distill
uses it:
[Cookson, Niessner & Schiller](/wiki/papers/jf/2026/cookson-social-media-merger-withdrawals-2026/)
use roughly 260 million StockTwits posts (January 2010 to December 2021) matched
to 6,438 U.S. M&A deals to build firm-specific abnormal sentiment after a merger
announcement, showing that negative social-media sentiment predicts deal
withdrawal.

- **Cost:** licensed. Some current data is reachable through StockTwits' own
  API under terms, but the historical research corpus is typically obtained
  through Social Market Analytics or a direct StockTwits data agreement.
- **Vendor:** StockTwits Inc.; redistribution and history often via Social
  Market Analytics (SMA).
- **Coverage:** message-level, ticker-tagged, from the platform's launch around
  2009 to present; deepest for liquid U.S. equities that attract retail
  attention.

## Access (when licensed)

- **Through a StockTwits or SMA data licence.** The research corpus (full
  message history with timestamps, tickers, and the Bullish/Bearish label)
  comes as bulk extracts under agreement; the public API is rate-limited and
  governed by terms that restrict bulk historical collection.
- **The grain is the message.** Each row is a post with author, timestamp,
  ticker(s), text, and the optional sentiment tag; firm-day sentiment measures
  are aggregated up from messages, as the citing paper's abnormal-sentiment
  measure is.
- Credentials are required. Keep any credentials in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **The self-labeled sentiment tag is optional and sparse.** Only a fraction of
  messages carry an explicit Bullish/Bearish flag; the rest need a text
  classifier, and the labeled subset is not a random sample of opinion. State
  how unlabeled messages are handled.
- **Population is self-selected retail, not the market.** Posters skew toward
  active retail traders and toward high-attention tickers; coverage is thin for
  small or neglected stocks and absent for most of the cross section. Sentiment
  is informative about the attentive crowd, not the marginal investor.
- **Volume confounds sentiment.** Message counts spike around news, so an
  abnormal-sentiment measure must net out the attention spike; the citing paper
  builds an *abnormal* sentiment for exactly this reason. Raw sentiment tracks
  volume.
- **Bots, promotion, and manipulation.** Ticker streams contain spam, paid
  promotion, and coordinated posting; naive aggregation imports that noise.
  Filter duplicates and low-quality accounts.
- **Terms restrict bulk reuse.** Both the API and the licensed corpus carry
  redistribution limits; cache derived features (firm-day sentiment), not raw
  messages, and do not republish the text. The same firm-day measure built from
  different filters will not match another paper's; document the pipeline.

## Citation

Cite the source and provider, e.g.: *StockTwits messages (StockTwits Inc., via
Social Market Analytics), [dates], accessed YYYY-MM-DD.* State the message
count and date range, how sentiment was assigned (self-label versus classifier),
and how attention/volume was netted out.


==============================================================================
# NYSE TAQ: trade and quote microstructure data (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/taq/
# NYSE TAQ (Trade and Quote) is a tick-level database of intraday trades and quotes for all US-listed equities on the consolidated tape, covering two lineages: Monthly TAQ (1993 onward) and Daily TAQ (millisecond to nanosecond stamps, 2003/2014+), reached by most academics through WRDS. It is licensed: the access path was exercised through a licensed WRDS session.
# Access confirmed (licensed) 2026-06-09 · via live TAQ NBBO query (taqm_2020.complete_nbbo_2020) through a licensed WRDS session
# Tags: equities, market-microstructure, licensed, event-data, wrds, data:taq
==============================================================================

:::note[Access confirmed via a licensed WRDS session]
**NYSE TAQ is a licensed commercial dataset** (NYSE / ICE Data Services,
reached for most academics through
[WRDS](/wiki/commercial/wrds/)). The keystone query `SELECT * FROM taqm_2020.complete_nbbo_2020 LIMIT 1` was run against a live WRDS Postgres session on 2026-06-09 and returned a real row, so this page carries the amber "Access confirmed (licensed)" badge. Reproducing it requires the reader's own WRDS account and the institution's TAQ entitlement.
:::

**NYSE TAQ** is a **tick-level trade and quote** database covering all US-listed
equities on the consolidated tape. It records every trade (price, size,
timestamp, exchange code, condition flags) and every quote (bid, ask, size,
exchange) reported to the consolidated tape. It is a common data source for
market-microstructure research: used in, for example
[Brogaard, Ringgenberg & Roesch](/wiki/papers/jf/2025/brogaard-floor-trading-matter-2025/)
for a study of whether floor trading matters in market microstructure, and
[Schwarz, Barber, Huang, Jorion & Odean](/wiki/papers/jf/2025/schwarz-actual-retail-price-equity-2025/)
for execution-quality analysis in the actual retail price of equity.

- **Cost:** licensed, subscription. No free tier. Most academics reach it
  through [WRDS](/wiki/commercial/wrds/).
- **Vendor:** NYSE (ICE Data Services).
- **Coverage:** all US-listed equities, consolidated tape, intraday (tick level);
  Monthly TAQ from 1993 onward; Daily TAQ from 2003 onward (millisecond stamps),
  with nanosecond resolution introduced in 2014.

## Access (when licensed)

- **Through WRDS.** The common academic path: WRDS exposes both the Monthly TAQ
  and Daily TAQ tables. WRDS also provides the Intraday Indicators product
  (WRDS-computed microstructure measures such as effective spreads, price
  impact, and realized spreads), so many research designs do not require
  processing raw ticks at all; use the derived measures where they suffice.
- **Directly from NYSE / ICE Data Services.** Available through NYSE's data
  licensing programs.
- Credentials are required either way. Keep them in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect when working with the data.

- **Raw Daily TAQ is enormous.** Daily TAQ runs to terabytes per year; do not
  attempt a naive full pull. Use WRDS server-side queries (SAS or Python with
  WRDS Cloud filters), push filtering and aggregation to the server, or use the
  WRDS Intraday Indicators for pre-aggregated measures.
- **Trade-signing rules must match the era.** Timestamp alignment between trades
  and quotes changed over the sample period. The original Lee-Ready rule used a
  5-second lookback; later literature moved to a contemporaneous-quote convention
  once timestamps tightened. The sign-of-trade algorithm must be chosen to match
  the period; mixing conventions introduces systematic errors in effective spread
  and order-flow estimates.
- **The data requires cleaning before use.** Key steps: drop cancelled and
  corrected trades (condition codes), exclude opening and closing auction prints,
  filter out-of-sequence and late-reported trades, and apply standard quote
  filters (zero or locked/crossed quotes). Failing to clean produces biased
  spread and impact estimates.
- **Monthly TAQ and Daily TAQ are not the same product.** They differ in field
  names, timestamp resolution, and symbology conventions. Code written for one
  does not transfer directly to the other; state which product your pull uses.
- **Consolidated tape mixes venues.** TAQ records trades and quotes from all
  reporting venues (NYSE, Nasdaq, regional exchanges, dark pools). Venue-level
  inference requires the exchange code; studies of a specific venue must filter
  on it, not take the full tape.
- **Symbology and the CRSP link require care.** TAQ identifies securities by
  symbol (which can be reused across firms over time), not CUSIP or PERMNO.
  Linking to CRSP or Compustat requires the TAQ-CRSP linking table and attention
  to date ranges; symbol reuse and symbol changes cause silent mismatches if the
  link is applied without effective dates.

## Citation

Cite the specific product and vendor, e.g.: *NYSE TAQ [Monthly TAQ / Daily TAQ],
NYSE / ICE Data Services, accessed via WRDS; data licensed and accessed
YYYY-MM-DD.* State which product (Monthly or Daily), the timestamp resolution
used, the trade-signing rule applied, and the cleaning criteria (condition codes
excluded, auction prints, quote filters).


==============================================================================
# Thomson Reuters institutional (13F) holdings: the s34 database (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/thomson-13f/
# The Thomson Reuters / Refinitiv (now LSEG) Institutional (13F) Holdings database, known by its WRDS table family "s34", is a manager-by-quarter panel of US institutional equity holdings built from SEC Form 13F filings. It is licensed: this page documents the access path and the gotchas; the access path was exercised through a licensed WRDS session.
# Access confirmed (licensed) 2026-06-09 · via live Thomson Reuters institutional 13F query (tfn.s34) through a licensed WRDS session
# Tags: equities, institutional-investors, licensed, panel-data, wrds, data:thomson-13f
==============================================================================

:::note[Access confirmed via a licensed WRDS session]
The keystone query `SELECT fdate, mgrno FROM tfn.s34 LIMIT 1` was run against
the Thomson Reuters s34 institutional holdings table through a licensed WRDS
session on 2026-06-09 and returned a real row, so this page carries the amber
"Access confirmed (licensed)" badge. Reproducing it still requires the reader's
own WRDS account and the institution's Thomson/Refinitiv entitlement (licensed,
not open).
:::

**The Thomson Reuters / Refinitiv Institutional (13F) Holdings database**
(WRDS table family: `s34`) is a **manager-by-quarter panel** of US
institutional equity holdings, built from mandatory SEC Form 13F disclosures.
Each record captures a reporting manager's long US equity positions (shares
held and market value) as of each quarter-end, along with a manager-type
classification. It is a common source in institutional-ownership and
corporate-finance research: used in, for example
[Falato & Scharfstein](/wiki/papers/jf/2025/falato-stock-market-bank-risk-2025/)
for the stock-market and bank-risk analysis (institutional ownership channel),
and
[Guernsey, Guo, Liu & Serfling](/wiki/papers/jf/2025/guernsey-thirty-years-change-evolution-2025/)
for the evolution of institutional ownership over thirty years.

- **Cost:** licensed (via WRDS). No free cleaned tier.
- **Vendor:** Thomson Reuters / Refinitiv (LSEG).
- **Coverage:** 13F-reporting institutions (managers above the 13F AUM
  threshold) quarterly from 1980; US long equity positions only.

## Access (when licensed)

- **Through WRDS.** The common academic path uses the `s34` table family
  on WRDS; see [WRDS / CRSP / Compustat](/wiki/commercial/wrds/) for the
  connection recipe.
- **Raw from SEC EDGAR.** The underlying Form 13F filings are freely available
  from EDGAR, but they are un-cleaned, delivered in raw XML or text with
  uncleaned CUSIPs, and are not a drop-in substitute for the Thomson/Refinitiv
  tables.
- Credentials are required for the Thomson/Refinitiv tables on WRDS. Keep
  them in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented.

- **Coverage gaps and stale holdings, worse in the later sample.** A
  data-quality literature (Ben-David, Franzoni, Moussawi, and others) documents
  that the Thomson/Refinitiv 13F series has coverage gaps and stale or
  carried-forward holdings, and that its coverage of filings degraded relative
  to the raw SEC 13F filings in the later part of the sample (Ben-David,
  Franzoni, Moussawi, and Sedunov, 2021, date a sharp increase in stale and
  dropped holdings to around 2013, with roughly 10% underreporting by 2015).
  Holdings can be carried forward
  across quarters, and the report date (rdate) versus file date (fdate) labels
  can be misaligned. Validate against the raw SEC EDGAR 13F filings, especially
  for recent years, before relying on the series.
- **13F covers only long US equity above the threshold.** Short positions,
  positions below the per-security reporting threshold, non-equity holdings,
  and managers below the $100 million AUM threshold are all absent. The
  database does not represent total ownership; aggregate holdings from it will
  fall short of float.
- **The manager-type (typecode) classification is unreliable and frozen.**
  The typecode field (bank, insurance, mutual fund, etc.) is known to be
  inaccurate for many managers and was not updated as institutions changed
  structure. Do not use typecode splits without re-classifying from an
  independent source.
- **The manager identifier (mgrno) is not stable across mergers.**
  When institutions merge or rename, mgrno assignments can break continuity.
  The mgrno is also separate from security identifiers; linking to CRSP
  requires the s34-to-CRSP link table and careful CUSIP handling (including
  CUSIP changes over time).
- **Confidential-treatment filings and amendments create gaps.**
  Managers may receive confidential treatment on individual positions, delaying
  or omitting those holdings from the public (and therefore the database)
  record. Amendments also mean the same quarter can have multiple filing
  vintages.
- **Raw EDGAR 13F filings differ structurally from the WRDS tables.**
  Raw filings use uncleaned CUSIPs, inconsistent formatting, and lack the
  manager-type and numeric identifiers the Thomson tables supply. They are a
  source for independent verification, not a cleaned substitute.

## Citation

Cite the Thomson Reuters / Refinitiv Institutional (13F) Holdings database
(s34), accessed via WRDS, stating the data vintage, the date of access, how
coverage gaps and stale holdings were handled (exclusion, supplementation with
the raw SEC EDGAR 13F filings, or restriction to a period less affected), and
how the typecode classification was treated
(whether typecodes were used as-is or re-classified from an independent
source).


==============================================================================
# Titlon Oslo Stock Exchange data (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/titlon-ose/
# Titlon is the University of Tromso's financial database for the Oslo Stock Exchange: prices, returns, shares outstanding, and accounting data for Nordic listed firms. It is free to Nordic academic users but credential-gated, not openly public; this page documents the access path and the gotchas, and the data was not exercised here.
# Tags: equities, asset-pricing, nordic, market-data, academic, licensed, data:titlon-ose
==============================================================================

:::caution[Credential-gated: not exercised here]
**Titlon is free to Nordic academic users but credential-gated**, so it carries **no provenance badge**: the
access path below was **not** run in this session, and access requires an
eligible Nordic academic affiliation. Treat it as unverified until someone
exercises it through a Titlon account. This is the honest grade under the
institute's Verified discipline.
:::

**Titlon** is the financial-data service maintained by UiT The Arctic University
of Norway (formerly the University of Tromso) covering the Oslo Stock Exchange
(OSE) and other Nordic markets: daily and monthly stock prices, returns, shares
outstanding, dividends, and linked accounting data for listed firms. It is the
standard source for Norwegian equity returns in academic work, the OSE
counterpart to CRSP. A paper we distill uses it:
[Betermier, Calvet, Knupfer & Kvaerner](/wiki/papers/jf/2025/betermier-investor-factors-2025/)
use Titlon for prices, returns, and shares outstanding of 535 OSE stocks (1997
to 2017), joined to complete individual investor holdings from the Norwegian
central securities depository, when building investor-based factors that price
the Norwegian cross section.

- **Cost:** free to eligible Nordic academic users, but access is
  credential-gated through a participating institution; it is not an open
  public download.
- **Vendor:** UiT The Arctic University of Norway (Titlon), with data sourced
  from the Oslo Stock Exchange and Oslo Bors.
- **Coverage:** OSE-listed equities and related Nordic securities, daily and
  monthly, with prices, returns, corporate actions, and accounting links; the
  citing paper's window runs 1997 to 2017.

## Access (when licensed)

- **Through a Titlon academic account.** Eligible users at participating Nordic
  institutions log in and query the database (a SQL-style client / data
  portal); access is tied to the institution's agreement, so a non-Nordic
  researcher needs a collaborator or affiliation.
- **CRSP-style return building.** Prices, shares outstanding, and corporate
  actions support standard return and market-cap construction; corporate-action
  adjustment is the key step, as with any clean-return source.
- Credentials are required. Keep any credentials in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Access is affiliation-bound, not purchasable.** Titlon is free but gated to
  Nordic academic users; a pipeline cannot simply buy a key. Plan for a
  collaborator with eligible access, and do not assume the data can be exercised
  outside that channel.
- **A small, thin market.** The OSE cross section is a few hundred names (535 in
  the citing paper), with thin trading and energy/shipping concentration;
  illiquidity, non-synchronous prices, and a short cross section make
  cross-sectional asset-pricing tests noisier than on U.S. data. Weight and
  filter accordingly.
- **Corporate-action and delisting handling is on you.** As with CRSP, returns
  must be adjusted for splits, dividends, and delistings; survivorship creeps in
  if delisted names are dropped. Confirm the corporate-action coverage before
  building a return series.
- **Identifier mapping to other Nordic data.** Joining Titlon to holdings or tax
  registers (as the citing paper does) relies on consistent security and firm
  identifiers (ISIN, organization number) that can change across time and source;
  build joins on stable keys plus dates.
- **Redistribution is restricted despite being free.** Free-to-academics does
  not mean republishable; the underlying OSE data carries reuse limits. Cache
  derived series, not raw vendor tables, and do not redistribute.

## Citation

Cite the source, e.g.: *Titlon financial database (UiT The Arctic University of
Norway), Oslo Stock Exchange data, [dates], accessed YYYY-MM-DD.* State the
securities and date range, the return-construction and corporate-action handling,
and the identifier used to link to any holdings or accounting data.


==============================================================================
# FINRA TRACE: corporate bond transactions (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/trace/
# TRACE (Trade Reporting and Compliance Engine) is FINRA's facility for secondary-market transaction reporting in US fixed-income securities, primarily corporate bonds. The version used by most academic researchers is the historical Enhanced TRACE file, reached for most researchers through WRDS. It is licensed: this page documents the access path and the gotchas, and the keystone query was exercised through a licensed WRDS session.
# Access confirmed (licensed) 2026-06-09 · via live Enhanced TRACE query (trace.trace_enhanced) through a licensed WRDS session
# Tags: fixed-income, corporate-bonds, licensed, transaction-data, wrds, data:trace
==============================================================================

:::note[Access confirmed via a licensed WRDS session]
The keystone query `SELECT cusip_id, trd_exctn_dt FROM trace.trace_enhanced LIMIT 1` was run against `trace.trace_enhanced` through a licensed WRDS session on 2026-06-09, so this page carries the amber "Access confirmed (licensed)" badge. Reproducing it still requires your own WRDS account and your institution's TRACE entitlement; the data is licensed, not open.
:::

**TRACE** (Trade Reporting and Compliance Engine) is FINRA's facility into
which dealers report secondary-market transactions in eligible US fixed-income
securities: primarily corporate bonds, with agency and securitized products
added in later phases. The version used by most academic researchers is the historical
Enhanced TRACE file, which carries fuller fields and uncapped trade sizes
compared with the free public dissemination feed. It is a common source in
fixed-income research: used in, for example,
[Nozawa & Tsoy](/wiki/papers/jf/2025/nozawa-counter-markets-nonstandardized-assets-2025/)
for corporate bond transaction prices and volumes 2002-2020 (plus a TRACE
agency-MBS subset used as a liquidity placebo), and
[Li & Yu](/wiki/papers/jf/2026/li-investor-composition-corporate-bond-liquidity-2026/)
for corporate bond liquidity and investor composition.

- **Cost:** Enhanced TRACE is licensed (via WRDS); the real-time dissemination
  feed is free but volume-capped (not a substitute for the academic history).
- **Vendor / originator:** FINRA. Distributed to academics via WRDS.
- **Coverage:** secondary-market corporate bond transactions reported to FINRA;
  coverage phased in from July 2002 through 2005, so the early period is
  partial. Agency MBS and securitized-product TRACE are separate, later-phase
  files.

## Access (when licensed)

- **Through WRDS.** The common academic path is the Enhanced TRACE tables on
  WRDS; see [WRDS / CRSP / Compustat](/wiki/commercial/wrds/) for the connection
  recipe. Credentials are required.
- **Directly from FINRA.** FINRA publishes a real-time public dissemination
  feed at no cost, but it masks large trade sizes (e.g., as "5MM+" or "1MM+"
  indicators) and does not carry the full historical depth of the Enhanced
  file. Use it only if volume-capping is acceptable for your purpose.
- Keep credentials in `.env`; never hard-code them.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect when working with the data.

- **Do not mix the free dissemination feed with the Enhanced file.** The
  public feed caps reported trade sizes (large trades are flagged as "5MM+" or
  "1MM+" rather than showing the actual notional). Merging it with Enhanced
  TRACE records, or treating it as a substitute for the historical Enhanced
  file, produces silent volume errors.
- **TRACE requires cleaning before use.** The raw file contains cancellations,
  corrections, reversals, and double-counted interdealer and agency legs. These
  must be filtered out before computing prices or volumes. The Dick-Nielsen
  cleaning procedure is the common reference in the literature; document which
  version you apply.
- **Coverage phased in over 2002-2005.** Phase 1 (July 2002) covered only
  investment-grade bonds above a size threshold. Phases 2 and 3 expanded
  coverage through 2005. Research samples starting before 2005 have partial
  bond-market coverage; control or restrict accordingly.
- **Prices are dealer-reported transaction prices, not quotes.** They reflect
  executed trades rather than bid-ask midpoints. Small-trade prices are
  noisier; large trades are subject to time delays in the dissemination feed.
- **Linking to bond characteristics requires Mergent FISD.** TRACE records
  CUSIPs but does not carry issue-level characteristics (coupon, maturity,
  seniority, rating). Merging those fields requires a separate link to Mergent
  FISD at the CUSIP level; treat this as an explicit join step and report match
  rates.
- **Agency MBS and securitized-product TRACE are separate files.** They were
  added in later phases, have different field structures, and cover different
  instrument types. Do not assume the corporate bond cleaning procedure applies
  to them without checking.

## Citation

Cite FINRA TRACE (Enhanced TRACE via WRDS), stating: the version (Enhanced vs.
dissemination feed), the data vintage (access date or coverage window), the
cleaning procedure applied (e.g., Dick-Nielsen with version reference), and the
FISD link used for bond characteristics. Example: *FINRA TRACE Enhanced
file, accessed via WRDS, [coverage dates]; cleaned following Dick-Nielsen
([year]); bond characteristics linked from Mergent FISD via CUSIP.*


==============================================================================
# Trucost: firm-level environmental and carbon data (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/trucost/
# Trucost (S&P Global) is a firm-level environmental panel: scope 1, 2, and 3 greenhouse-gas emissions, intensities, and other environmental metrics, with much of it modeled rather than disclosed. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: equities, esg, climate, licensed, panel-data, data:trucost
==============================================================================

:::caution[Licensed: not exercised here]
**Trucost is a licensed commercial dataset** (S&P Global), so it carries **no provenance badge**: the access path below was
**not** run in this session (no S&P credentials were available). The page
documents the access route and the gotchas; treat it as unverified until
someone exercises it through a licensed account. This is the honest grade under
the institute's Verified discipline.
:::

**Trucost** is a firm-level **environmental dataset** maintained by S&P Global.
Its most-used content is **greenhouse-gas emissions**: scope 1 (direct), scope 2
(purchased electricity), and scope 3 (value chain), reported in tonnes of CO2
equivalent (tCO2e) by fiscal year, alongside emission intensities (per revenue),
other pollutants, natural-capital costs, and fossil-fuel reserve exposure. It is
a common carbon-emissions source in climate-finance papers: used in, for example
[Zhang](/wiki/papers/jf/2025/zhang-carbon-returns-globe-2025/) for firm-level
annual scope 1 and 2 emissions with their actual release dates, and
[Pedersen](/wiki/papers/jf/2026/pedersen-carbon-pricing-green-finance-2026/) for
fiscal-year 2021 scope 1 and 2 emissions in a sustainable-discount-rate
calibration.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** S&P Global (Trucost; the environmental data line of S&P Global
  Sustainable1).
- **Coverage:** thousands of listed companies globally, with emissions and other
  environmental metrics by fiscal year; history back to roughly 2005 for the
  broadest panels.

## Access (when licensed)

- **S&P Global / Sustainable1.** Through S&P's data feeds, the S&P Capital IQ
  platform, or the Trucost data files, keyed on the S&P company identifier.
- **Possibly via a data vendor or library feed.** Some institutions reach
  Trucost through bundled S&P Global products; availability is not confirmed
  here, so check with your library before assuming a given path.
- Credentials are required either way. Keep them in `.env`, never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Most emissions are modeled, not disclosed.** A large share of firm-year
  emissions are estimated by Trucost's environmental model from sector and
  activity data, not taken from a company report. Disclosed and modeled values
  are mixed in the same field; if your design needs disclosed numbers, filter on
  the disclosure flag rather than treating every value as reported.
- **Fiscal-year emissions are released with a long lag.** Emissions for a fiscal
  year are published well after that year ends. Using a value as if it were
  known at fiscal-year close builds in look-ahead. Match on the actual data
  release date, not the fiscal year, for any return or pricing test.
- **Scope 3 is the noisiest.** Value-chain emissions depend on the most
  estimation and on boundary choices; treat scope 3 levels and changes with more
  caution than scope 1 and 2.
- **History gets restated.** As the model and disclosures are reworked, prior
  firm-years can be revised. Pin the data vintage and re-pull deliberately
  rather than mixing vintages.
- **Intensity denominators matter.** Emission intensity uses revenue (or another
  denominator) that itself changes with currency and restatements; state the
  denominator and the units so the number is reproducible.
- **Entity and identifier mapping is its own step.** Join on the S&P company
  identifier and confirm the parent/subsidiary and ticker mapping before
  merging with returns or fundamentals.

## Citation

Cite the product and vendor, e.g.: *Trucost environmental data, S&P Global;
data licensed and accessed YYYY-MM-DD.* State the data vintage, the emission
scope, the units (tCO2e), and whether values are disclosed or modeled.


==============================================================================
# VentureSource venture-capital data (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/venturesource/
# VentureSource (Dow Jones / CB Insights) tracks venture-capital funds, financing rounds, valuations, and startup locations. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: venture-capital, startups, private-markets, financing-rounds, licensed, data:venturesource
==============================================================================

:::caution[Licensed: not exercised here]
**VentureSource is a paid licensed product** (Dow Jones / CB Insights), so it carries **no provenance badge**: the access path below
was **not** run in this session. Treat it as unverified until someone exercises it
through a licensed account. This is the honest grade under the institute's
Verified discipline.
:::

**VentureSource** is a venture-capital deals and fund database: VC fund
characteristics (size, vintage, location), startup financing rounds (amount,
date, investors), pre-money valuations, and the startup's headquarters location
over time. It is one of the two standard VC sources (alongside
[PitchBook](/wiki/commercial/pitchbook/) and the older Thomson VentureXpert) for
financing-round and fund-flow research. A paper we distill uses it:
[Chen & Ewens](/wiki/papers/jf/2025/chen-venture-capital-startup-agglomeration-2025/)
use VentureSource as their primary VC source (2010 to 2018) for fund
characteristics, startup financing rounds, and the startup's HQ state over time
when showing that local VC supply drives startup agglomeration in VC hubs, with
PitchBook used for robustness.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** Dow Jones (VentureSource); now part of CB Insights.
- **Coverage:** VC-backed startups, financing rounds, and funds, with the startup
  location and round-level detail that makes geographic and time-series analysis
  possible; deepest for U.S. VC.

## Access (when licensed)

- **Through a Dow Jones / CB Insights subscription.** Extracts are obtained under
  a VentureSource / CB Insights data licence (download or API where entitled).
- **Two linked grains.** Fund-level records (fund size, vintage, LPs where
  available) and round-level records (startup, amount, date, investors,
  valuation) are different extracts that join on fund and company identifiers.
- Terminal or API credentials are required. Keep any credentials in `.env`, never
  hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Coverage is built from disclosure, so it is selective.** Rounds enter the
  database when they are reported; small, early, or undisclosed rounds are
  undercounted, and coverage thickens over time. A rising round count can reflect
  better reporting rather than more activity. Check coverage stability.
- **Cross-check against a second source.** VentureSource and PitchBook disagree on
  round amounts, dates, and even whether a round happened; conclusions that turn
  on one source are fragile. The citing paper uses PitchBook for robustness for
  exactly this reason.
- **Valuations are sparse and noisy.** Pre-money valuations are missing for many
  rounds and are often imputed or self-reported; conditioning on non-missing
  valuation selects larger, later rounds. State the missingness handling.
- **Startup location changes over time.** A startup's HQ state is time-varying
  (relocations to hubs are themselves an outcome); using a single static location
  induces look-ahead bias in any migration analysis. Use the point-in-time
  location.
- **Entity linking and renames.** Startups and funds get renamed, merged, or
  re-IDed; joins to other data (Crunchbase, SDC, regulatory filings) must be built
  on names plus dates, not a shared key. Verify the link.

## Citation

Cite the vendor and product, e.g.: *VentureSource (Dow Jones / CB Insights),
accessed YYYY-MM-DD.* State whether fund-level or round-level data was used, the
date range, and how location and valuation missingness were handled.


==============================================================================
# Refinitiv Worldscope: global company fundamentals (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/worldscope/
# Worldscope is Refinitiv (LSEG) global database of standardized company fundamentals (balance sheet, income statement, cash flow, ratios, per-share data) and descriptive information for public companies across many countries. It is licensed: this page documents the access path and the gotchas, but the data was not exercised here.
# Tags: accounting, international, fundamentals, licensed, data:worldscope
==============================================================================

:::caution[Licensed: not exercised here]
**Worldscope is a paid licensed product** (Refinitiv / LSEG), sometimes reached
through [WRDS](/wiki/commercial/wrds/), so it carries **no provenance badge**: the access path below was **not** run in this
session (no Refinitiv / LSEG / WRDS credentials were available). The page
documents the access route and the gotchas; treat it as unverified until someone
exercises it through a licensed account. This is the honest grade under the
institute's Verified discipline.
:::

**Worldscope** is Refinitiv's (LSEG) global database of standardized company
fundamentals: balance sheet, income statement, and cash flow items, along with
ratios, per-share data, and descriptive information for public companies across
many countries, with a long international history. It is one of the two main
sources of international company fundamentals, the other being
[Compustat Global](/wiki/commercial/compustat-global/). Worldscope is the
fundamentals complement to [Datastream](/wiki/commercial/datastream/) (which
supplies prices, returns, and market data) and is frequently used together with
it. A paper we distill uses it:
[Faccio et al.](/wiki/papers/jf/2025/faccio-impediments-schumpeterian-process-replacement-2025/)
on impediments to the Schumpeterian process of creative destruction.

- **Cost:** licensed, subscription. No free tier.
- **Vendor:** Refinitiv / LSEG (lineage traces through Thomson Reuters /
  Thomson Financial, where the product was the Worldscope database).
- **Coverage:** tens of thousands of firms across developed and emerging
  markets, with annual and interim fundamentals; history reaches into the 1980s,
  with deeper and broader coverage in later years.

## Access (when licensed)

- **Through Datastream or WRDS.** Worldscope fundamentals are typically reached
  through [Datastream](/wiki/commercial/datastream/), where they are exposed as
  Datastream datatypes (the Worldscope "WC" item codes), or through
  [WRDS](/wiki/commercial/wrds/) where the institution licenses it. Check which
  route your institution provides.
- **Keyed by Worldscope identifiers and item codes.** Data is keyed by
  Worldscope's own company identifier and by item ("WC") codes; an extract is a
  set of WC items pulled for a list of firms over a date range.
- Terminal or API credentials are required. Keep any credentials in `.env`,
  never hard-coded.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **The "WC" item codes are the schema, and you must map them carefully.** A
  Worldscope extract is a list of WC item codes, and the same labeled item can
  be computed differently than its Compustat analog. As a result Worldscope and
  [Compustat Global](/wiki/commercial/compustat-global/) do **not** reconcile
  item-for-item; do not assume a WC item equals the same-named Compustat field.
- **Accounting-standard heterogeneity limits comparability.** Firms report under
  local GAAP or IFRS with different adoption dates by country, so even after
  Worldscope's standardization, cross-country comparisons of a given item remain
  imperfect. Treat cross-border comparisons with caution.
- **Currency.** Fundamentals are reported in native currency and need explicit
  conversion before any cross-country aggregation or comparison; do not mix
  currencies in a panel.
- **Backfill and survivorship.** Worldscope has added historical data and firms
  retroactively, and the active list excludes dead firms; building a sample from
  only currently covered firms biases it toward survivors. Pin the extraction
  date and source a delisted/dead-firm list when survivorship matters.
- **Identifier linking.** Worldscope codes must be linked to Datastream
  mnemonics and to returns and prices; the Worldscope-to-Datastream link is the
  usual join and has known mismatches. Verify the link rather than assuming a
  clean one-to-one map.
- **Coverage and item availability vary by country and year.** Item availability
  and firm coverage are uneven, with more missingness in emerging markets and in
  early years. Do not read missing as zero, and check item-level availability
  before conditioning a sample on it.
- **Documented data errors.** The comparative-database literature documents
  discrepancies between Worldscope and Compustat Global; do not treat individual
  items as ground truth. Cross-check material observations when a single firm or
  item drives the result.

## Reference: representative WC item codes

| WC code | Item (category) |
|---|---|
| WC02999 | Total assets (balance sheet) |
| WC03351 | Total liabilities (balance sheet) |
| WC03995 | Common equity (balance sheet) |
| WC01001 | Net sales / revenue (income statement) |
| WC01651 | Net income before preferred dividends (income statement) |
| WC04860 | Net cash flow, operating activities (cash flow) |
| WC05001 | Market price, year end (per-share / market) |
| WC05101 | Earnings per share (per-share) |

Item codes are the extraction schema; pull the specific WC items you need and
record them. Availability differs by country and year.

## Citation

Cite the provider and database, e.g.: *Worldscope / Refinitiv (LSEG), accessed
YYYY-MM-DD.* State the WC item codes used, the access route (Datastream or WRDS),
and the extraction date so the sample is reproducible.


==============================================================================
# WRDS / CRSP / Compustat: the paywalled core (academic access)
# https://instituteforautomatedresearch.org/wiki/commercial/wrds/
# WRDS (CRSP, Compustat, IBES, OptionMetrics…) is not free, but most universities license it. How affiliated researchers get access, and what the free sources here can and cannot substitute for it.
# Access confirmed (licensed) 2026-05-16 · via live CRSP (crsp.msf) + Compustat (comp.funda) query through a licensed WRDS session
# Tags: equities, fundamentals, panel-data, licensed, wrds, data:wrds
==============================================================================

:::caution[Licensed, not free]
**WRDS is subscription-gated.** The badge above reads **"Access confirmed (licensed)"**,
deliberately *not* the green "Verified" used on free pages: the access path
was exercised in-session against a real institutional WRDS account (live
`crsp.msf` and `comp.funda` rows returned), but the data itself is **not
free**. Without a subscription you cannot reproduce this; see the
free-substitute map below for what you can.
:::

**WRDS** (Wharton Research Data Services) is the standard delivery layer for
the paid datasets that most empirical finance still runs on: **CRSP** (the
canonical US return history), **Compustat** (fundamentals), **IBES**
(analyst forecasts), **OptionMetrics**, **Thomson/Refinitiv**, **ExecuComp**,
**BoardEx**, and more. The [ZeroPaper](https://github.com/alejandroll10/zeropaper)
pipeline uses it when a license is present; when one isn't, the free sources
in this section cover a large share of the same ground.

## Who can get it (academic access)

- **Most universities already license WRDS.** If you are faculty, a PhD
  student, or a research staff member at a subscribing institution, you can
  request a personal WRDS account through your library or finance department
  at no cost to you; the institution pays the subscription.
- Accounts use a username/password plus (typically) Duo 2FA. Keep credentials
  in `.env` (`WRDS_USER`, `WRDS_PASS`); never hard-coded.
- No institutional subscription → no WRDS. There is no individual free tier.
  Use the free substitutes below.

## How the pipeline uses it (when licensed)

A persistent local WRDS server is started once per session so Duo 2FA fires a
single time; scripts then call a thin client rather than reconnecting:

```python
import sys; sys.path.insert(0, "code")
from utils.wrds_client import wrds_query, wrds_start
wrds_start()                                  # no-op if already running
df = wrds_query("SELECT permno, date, ret FROM crsp.msf "
                "WHERE date >= '2000-01-01' AND shrcd IN (10,11) LIMIT 100")
```

Direct `wrds.Connection()` works too, but Duo fires on **every** connection;
open one per script and reuse it.

## Gotchas (the ones that bite pipelines)

- **Don't eager-load large local parquets.** After you cache a large WRDS pull
  to local parquet (CRSP daily ~100M rows, TAQ, 13F/`s34` institutional
  holdings), never reload it with a whole-file `pd.read_parquet(<file>)` -- it
  will OOM-kill the process. Stream it instead:
  `polars.scan_parquet(path).select([...]).filter(...).collect()`
  (column projection + predicate pushdown) so you filter before materializing
  and never hold the full table in RAM. Requires `polars` + `pyarrow`. The same
  filter-first discipline applies at query time (never `SELECT *` on the big
  tables). For the mutual-fund holdings tables, see
  [CRSP Mutual Funds](/wiki/commercial/crsp-mutual-funds/).

## What the free sources here substitute for

| Need | Paid (WRDS) | Free substitute on this wiki |
|---|---|---|
| Asset-pricing factors / test assets | CRSP + own sorts | [Ken French](/wiki/datasets/ken-french/) |
| Cross-sectional anomaly signals | CRSP/Compustat merge | [Open Source Asset Pricing](/wiki/datasets/open-source-asset-pricing/) |
| Data-mining benchmark | Compustat ratios | [Flexible data-mining](/wiki/datasets/flex-mining/) |
| Fundamentals / disclosure | Compustat | [SEC EDGAR](/wiki/datasets/edgar/) (XBRL) |
| Macro calibration | n/a | [FRED](/wiki/datasets/fred/) |
| Retirement / pension assets | n/a | [DOL Form 5500](/wiki/datasets/form-5500/) |

What the free set does **not** replace: survivorship-bias-free long CRSP daily
returns, full Compustat history with point-in-time discipline, IBES detail,
and OptionMetrics. For those, an institutional WRDS license remains necessary.

## Key WRDS libraries (reference)

`crsp` (`msf`, `dsf`, `msenames`, `ccmxpf_linktable`), `comp` (`funda`,
`fundq`, `company`), `ibes` (`statsum_epsus`, `det_epsus`), `optionm`,
`tfn`, `ff`, `execcomp`, `boardex`. Always filter on date and share/exchange
codes; never `SELECT *` on CRSP daily (~100M rows).

## Citation

Cite the underlying provider, not WRDS itself, e.g.: *CRSP, Center for
Research in Security Prices, LLC, accessed via WRDS, YYYY-MM-DD*; *S&P Global
Market Intelligence Compustat, accessed via WRDS, YYYY-MM-DD.*


==============================================================================
# ZTRAX: Zillow Transaction and Assessment Dataset (licensed)
# https://instituteforautomatedresearch.org/wiki/commercial/ztrax/
# ZTRAX was Zillow's national property-level dataset of deed transactions and assessor records, distributed free to academics under a data-use agreement until the program was discontinued in 2023. This page documents the access path and the gotchas; the data was not exercised here.
# Tags: real-estate, housing, property-records, transactions, administrative, licensed, data:ztrax
==============================================================================

:::caution[Licensed and discontinued: not exercised here]
**ZTRAX was distributed under a Zillow data-use agreement, not an open
download, and the academic program was discontinued in 2023.** It carries
**no provenance badge**: the access path below was
**not** run in this session, and Zillow no longer issues new ZTRAX agreements.
Treat it as unverified and effectively closed to new users. This is the honest
grade under the institute's Verified discipline.
:::

**ZTRAX (the Zillow Transaction and Assessment Dataset)** was a national,
property-level database covering hundreds of millions of records nationwide: deed
transfers (sales, prices, mortgages, foreclosures) and assessor / tax-roll data
(property characteristics, assessed values, owner information) across U.S.
counties. It was the standard research source for parcel-level housing
transactions before Zillow ended the academic program. A paper we distill uses
it:
[Slutzky & Xu](/wiki/papers/rfs/2025/slutzky-financial-consequences-pretrial-detention-2025/)
use ZTRAX real-estate transactions (1993 to 2020; about 9 million Maryland
transactions, with foreclosure events identifiable post-2007) for foreclosure
events and the property-matched sample behind the foreclosure and combined
insolvency tests when estimating the effect of pretrial detention on household
financial distress.

- **Cost:** was free to academic and nonprofit researchers under a signed Zillow
  data-use agreement; not an open public download. The program was discontinued
  in 2023, so new access is no longer granted.
- **Vendor:** Zillow (ZTRAX). Coverage and successor products differ; some
  researchers now turn to commercial deed/assessor vendors instead.
- **Coverage:** parcel-level deeds and assessor records for most U.S. counties,
  back to the early 1990s where county records allow; transaction history depth
  varies by county.

## Access (when licensed)

- **Was through a Zillow data-use agreement.** Approved researchers downloaded
  the full transaction and assessment files (large fixed-width tables with a
  layout document) and matched on property identifiers. New agreements are no
  longer issued.
- **Two linked grains.** A transaction table (deed events: sale, mortgage,
  foreclosure, dates, amounts) and an assessment table (property characteristics,
  assessed value, owner), joined on a Zillow property identifier (RowID / ImportParcelID).
- Where a researcher still holds a legacy copy under their prior agreement, keep
  any credentials or keys in `.env`, never hard-coded, and respect the
  agreement's redistribution limits.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **The program is discontinued.** ZTRAX is no longer obtainable from Zillow;
  results built on it are not reproducible by a new user without a legacy copy.
  A pipeline cannot acquire it today. Plan for a commercial deed/assessor vendor
  if you need fresh access.
- **County coverage and history depth are uneven.** Transaction back-history,
  field completeness, and whether prices are recorded depend on each county's
  recording practice; absence of a record is often missing data, not absence of
  a sale. The citing paper's coverage scopes to Maryland for this reason.
- **Deed records mix arms-length and non-arms-length transfers.** Quitclaims,
  intra-family transfers, refinancings, and foreclosure deeds all appear;
  filtering to genuine market sales (by document type and price) is required
  before treating a transfer as a sale. Foreclosure identification is reliable
  only from the post-2007 deeds in the citing paper.
- **Matching to people or addresses is fuzzy.** Owner names and addresses need
  standardization; linking ZTRAX parcels to individuals (as the citing paper
  does to court records) is a probabilistic match that drops and mismatches
  observations. Document the match rate.
- **Large, layout-heavy files.** The raw tables are big fixed-width extracts
  with a separate layout document and shifting schemas across vintages; a
  loader built for one delivery can silently misparse another. Validate against
  the layout file.

## Citation

Cite the source and product, e.g.: *Zillow Transaction and Assessment Dataset
(ZTRAX), Zillow, [dates], accessed YYYY-MM-DD (academic data-use agreement; program
discontinued 2023).* State the counties and date range, the transaction filters
applied, and the match procedure used to link parcels to other records.


==============================================================================
# 401(k) plan administrative records (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/401k-admin/
# Plan-administration microdata from a large U.S. retirement-plan recordkeeper: participant portfolio allocations, participation, contribution rates, and plan defaults across many plans. It is confidential single-counterparty data. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: retirement, household-finance, portfolios, united-states, data:401k-admin
==============================================================================

:::caution[Confidential single-counterparty data: not exercised here]
**This is confidential data from a single retirement-plan recordkeeper.** It is
**not for sale and not publicly downloadable**: it carries **no provenance badge**
because there is no access path we can run here. Researchers obtained it under a
confidential agreement with the recordkeeper, anonymized. The page documents what
the data is and the gotchas; treat it as unverified. This is the honest grade
under the institute's Verified discipline.
:::

**The 401(k) administrative records** are the plan-administration microdata held
by a large U.S. retirement-plan **recordkeeper**: for each participant, the
portfolio allocation, participation status, contribution rate, and the features of
the employer plan (notably the **default** investment and auto-enrollment
settings). Because a recordkeeper administers many employers' plans, the data
spans a large cross-section of plans and workers with the plan-design detail that
public surveys lack. A paper we distill uses it:
[Choukhmane & de Silva](/wiki/papers/jf/2026/choukhmane-portfolio-choices-risk-preferences-2026/)
use anonymized 401(k) administrative records from a large recordkeeper (December
2006 to December 2017, about 4 million employees across more than 600,000 plans)
for portfolio allocations, participation, contribution rates, and plan defaults,
to study what drives investors' portfolio choices.

- **Cost:** not for sale. Confidential single-counterparty data.
- **Source:** a large U.S. retirement-plan recordkeeper (anonymized).
- **Coverage:** participants in the plans this recordkeeper administers, not all
  U.S. workers; high-detail allocation and plan-design data over the sample window.

## Access (restricted)

- **No public download and no resale.** The data is the recordkeeper's
  confidential administrative records; it cannot be purchased or redistributed.
- **Through a confidential agreement with the recordkeeper.** Access required a
  research agreement, and the recordkeeper is anonymized. There is no standing way
  for a third party to reach the same data.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **One recordkeeper is a selected set of plans.** The plans administered by this
  recordkeeper are not a random sample of U.S. employers; plan size, industry, and
  design skew the data. Do not treat it as the universe of 401(k) plans.
- **Defaults drive observed choices.** A large share of allocations reflect the
  plan default and auto-enrollment, not active choice; treating observed allocation
  as revealed preference confounds default effects. Separate active from defaulted
  participants.
- **Within-plan, not whole-portfolio.** The data shows assets in this 401(k), not
  the household's other accounts (IRAs, taxable, a spouse's plan); apparent
  risk-taking can reflect partial coverage. Treat it as one account.
- **Plan menu constrains allocations.** Participants can only hold the funds the
  plan offers; the menu, not just preferences, shapes allocations. Condition on the
  plan menu.
- **Job and plan changes break panels.** Participants leave plans on job change and
  the recordkeeper may lose them; the panel is unbalanced for administrative
  reasons. Model entry/exit.
- **Not redistributable.** Results can be reported but the microdata cannot be
  shared and cannot be re-pulled by others. Plan for non-reproducibility of the raw
  inputs.

## Citation

Cite the source and arrangement, e.g.: *401(k) plan administrative records (large
U.S. recordkeeper, anonymized), confidential; used under agreement, YYYY-MM-DD.*
State the sample window, the number of plans/participants, and the treatment of
defaults.


==============================================================================
# Italian Credit Register (Centrale dei Rischi, restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/bank-of-italy-credit-register/
# The Centrale dei Rischi is the Bank of Italy's confidential credit register: firm-bank loan quantities and interest rates above a reporting threshold, alongside the Or.So. bank-board register and supervisory balance-sheet reports. It is confidential supervisory data. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: credit-register, banks, credit, italy, central-bank, data:bank-of-italy-credit-register
==============================================================================

:::caution[Confidential supervisory data: not exercised here]
**The Italian Credit Register is confidential supervisory data** held by the Bank
of Italy. It is **not for sale and not publicly downloadable**: it carries **no
provenance badge** because there is no access path we can run here. Researchers
reach it only through a Bank of Italy affiliation or an approved restricted-data
arrangement, inside a secure environment, with output subject to disclosure
review. The page documents what the collection is and the gotchas; treat it as
unverified. This is the honest grade under the institute's Verified discipline.
:::

**The Centrale dei Rischi** is the Bank of Italy's credit register: banks report
their exposures to each borrower above a reporting threshold, so the register
gives **firm-bank loan quantities and interest rates** across the banking system.
Two companion Bank of Italy sources are accessed under the same terms: the
**Or.So.** register of bank board members (who sits on which bank board, at
quarterly frequency) and the **supervisory reports** (bank and banking-group
balance sheets). Together they support credit-supply and bank-competition
research. A paper we distill uses them:
[Barone, Schivardi & Sette](/wiki/papers/jf/2025/barone-interlocking-directorates-competition-banking-2025/)
use the Credit Register for quarterly firm-bank loan quantities and gross
interest rates on overdraft lines, for relationships with total lending above
75,000 euros (2011Q1 to 2014Q4, 3.5 million firm-bank-quarters), Or.So. for
interlocking bank directorships, and the
supervisory reports for bank balance sheets.

- **Cost:** not for sale. Restricted-access confidential supervisory data.
- **Collector:** Bank of Italy (Centrale dei Rischi; Or.So.; supervisory reports).
- **Coverage:** exposures above the reporting threshold across Italian banks,
  monthly/quarterly; the threshold has changed over time, so small exposures move
  in and out of scope.

## Access (restricted)

- **No public download.** The microdata is confidential and is not posted.
- **Through a Bank of Italy affiliation or approved program.** Access is limited
  to researchers at the Bank of Italy or others granted entry to the restricted
  data, worked inside a secure environment with output subject to disclosure
  review.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **The reporting threshold censors small exposures.** Only exposures above the
  threshold are reported, and the threshold was lowered over time; a borrower can
  appear or disappear for reporting reasons, not economic ones. Do not read a
  missing exposure as no lending.
- **Drawn versus granted.** The register distinguishes credit granted (the
  commitment) from credit drawn (utilization), especially for revolving overdraft
  lines; conflating them mismeasures both supply and demand. Pick the concept that
  matches your question.
- **Interest rates cover specific products.** Rate data is reported for particular
  loan categories (the overdraft/revolving lines), not every exposure; do not
  treat a rate field as the firm's overall borrowing cost.
- **Identifier joins.** Banks (and groups) and firms must be linked across the
  Credit Register, Or.So., the supervisory reports, and any firm-financials source
  (such as the Cerved firm database); mergers and group restructurings break the
  links. Verify the crosswalk and consolidate groups deliberately.
- **Bank versus banking group.** Reporting is at the bank level but ownership is
  at the group level; intra-group exposures and shared boards require the group
  perimeter. Decide consolidation explicitly.
- **Output is disclosure-reviewed and cannot be redistributed.** Results leave the
  secure environment only after review, and the microdata itself cannot be shared.

## Citation

Cite the source, e.g.: *Centrale dei Rischi (Bank of Italy), confidential
supervisory data; accessed under restricted-data arrangement, YYYY-MM-DD*; cite
Or.So. and the supervisory reports likewise. State the sample window, the
threshold in force, and the drawn-versus-granted concept used.


==============================================================================
# German bank proprietary customer data (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/bank-proprietary/
# Individual-level customer records (product holdings, wealth, income, equity participation) from one anonymous German bank, used in household-finance research. It is confidential single-counterparty data. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: household-finance, banking, wealth, germany, data:bank-proprietary
==============================================================================

:::caution[Confidential single-counterparty data: not exercised here]
**This is confidential data from a single bank** (kept anonymous). It is **not
for sale, not publicly downloadable, and not re-obtainable without the same
agreement**: it carries **no provenance badge** because there is no access path
we can run here. Researchers obtained it under a confidential agreement with the
bank. The page documents what the data is and the gotchas; treat it as
unverified. This is the honest grade under the institute's Verified discipline.
:::

**The proprietary bank data** is the individual-level customer record from one
anonymous German bank: financial-product holdings, wealth, income, and
stock-market participation for a sample of customers. It complements survey and
broker data by observing balance-sheet and participation measures inside a single
institution. A paper we distill uses it:
[Laudenbach, Malmendier & Niessen-Ruenzi](/wiki/papers/jf/2026/laudenbach-communism-attitudes-2026/)
use proprietary bank data on 326,437 randomly selected customers (2019) for
financial-product holdings, wealth, income, and stock-market participation,
alongside the [Bilendi survey](/wiki/confidential/bilendi-survey/) and the
[online-broker](/wiki/confidential/online-broker/) data.

- **Cost:** not for sale. Confidential single-counterparty data.
- **Source:** an anonymous German bank.
- **Coverage:** the bank's own customers, not the German population; a
  cross-sectional sample of customer records.

## Access (restricted)

- **No public download and no resale.** The data is the bank's confidential
  customer records; it cannot be purchased or redistributed.
- **Through a confidential agreement with the bank.** Access required a research
  agreement, and the bank's identity is not disclosed. There is no standing way
  for a third party to reach the same data.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **One bank is a selected clientele.** The sample is this bank's customers, who
  differ from the population on who banks where; external validity is the central
  caveat. Do not generalize participation or wealth to all Germans.
- **Holdings at this bank, not total wealth.** Wealth and product holdings are what
  this bank sees; assets at other institutions are missing, so wealth and
  participation can be understated. Treat it as a partial balance sheet.
- **Cross-section, not panel.** The data is a snapshot (here 2019); it cannot speak
  to within-person dynamics. Match the question to the cross-sectional design.
- **Income is bank-observed.** Income is inferred from what the bank records
  (inflows, declared income), not a tax authority; it is noisier than
  administrative income. Treat it as a proxy.
- **Not redistributable and bank stays anonymous.** Results can be reported but the
  microdata cannot be shared and the bank is not named, limiting external
  replication. Plan for non-reproducibility of the raw inputs.

## Citation

Cite the source and arrangement, e.g.: *Proprietary retail-bank customer data
(anonymous German bank), confidential; used under agreement, YYYY-MM-DD.* State
the snapshot date, the sample size, and that holdings reflect one institution.


==============================================================================
# Banorte bank-account panel and savings experiment (Mexico, restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/banorte-experiment/
# Individual-level account and transaction records for millions of customers of one Mexican bank (Banorte), plus a randomized savings field experiment run with the bank. It is confidential single-counterparty data. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: household-finance, banking, field-experiment, mexico, data:banorte-experiment
==============================================================================

:::caution[Confidential single-counterparty data: not exercised here]
**This is confidential data from a single bank** (Banorte). It is **not for sale,
not publicly downloadable, and not re-obtainable without the same institutional
agreement**: it carries **no provenance badge** because there is no access path
we can run here. Researchers obtained it under a confidential agreement with the
bank, and the field experiment was run jointly with it. The page documents what
the data is and the gotchas; treat it as unverified. This is the honest grade
under the institute's Verified discipline.
:::

**The Banorte data** combines two things from one large Mexican bank: an
individual-level **account panel** (checking balances, credit-card balances,
limits and interest charges, deposits, payments, and transaction-level spending
across ATM, card, and transfers, for millions of customers) and a **randomized
savings field experiment** run with the bank. The experiment is what makes the
panel valuable for causal work: treatment is assigned by the researchers, while
the account panel measures outcomes at high frequency. A paper we distill uses
it:
[Medina & Pagel](/wiki/papers/jf/2025/medina-saving-cause-borrowing-implications-2025/)
use the Banorte account panel (161 pretreatment variables) and bimonthly credit
bureau pulls (for non-Banorte balances) to study whether nudging saving causes
households to borrow more.

- **Cost:** not for sale. Confidential single-counterparty data plus a joint
  field experiment.
- **Source:** Banorte (a Mexican bank); credit-bureau pulls obtained through the
  bank.
- **Coverage:** the bank's own customers, not the Mexican population; high-frequency
  account and transaction records over the experiment window.

## Access (restricted)

- **No public download and no resale.** The data is the bank's confidential
  customer records; it cannot be purchased or redistributed.
- **Through a confidential agreement with the bank.** Access required a research
  agreement with Banorte; the field experiment was designed and run jointly with
  the bank. There is no standing way for a third party to reach the same data.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **One bank is not the population.** The sample is Banorte customers, who are
  selected (banked, in this bank, eligible for the product); external validity to
  all Mexican households is a real limit. Do not generalize without argument.
- **Outside-bank activity is only partly observed.** Spending and balances at other
  institutions are seen only through the bimonthly credit-bureau pulls, which are
  coarser than the own-bank panel; cross-institution substitution is measured with
  error. Treat the credit-bureau view as lower resolution.
- **The experiment defines the causal sample.** Clean identification comes from the
  randomized treatment, not the full panel; observational comparisons in the panel
  are not the experiment. Keep the experimental and observational analyses
  separate.
- **Account, not person, and not household.** A customer can hold several products
  and a household several customers; aggregating requires care. Fix the unit before
  measuring saving or borrowing.
- **Not redistributable.** Results can be reported but the microdata cannot be
  shared, and the data cannot be re-pulled by others. Plan for non-reproducibility
  of the raw inputs.
- **Currency and PPP.** Balances are in pesos; cross-country reporting needs an
  explicit PPP conversion (the paper uses 2019 OECD PPP). State the conversion.

## Citation

Cite the source and arrangement, e.g.: *Banorte proprietary account panel and
savings field experiment, confidential; used under agreement with the bank,
YYYY-MM-DD.* State the experiment window, the unit (account, customer, household),
and any currency/PPP conversion.


==============================================================================
# Bilendi commissioned online survey (Germany, restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/bilendi-survey/
# An author-commissioned representative online survey of Germans, fielded through the panel provider Bilendi, with individual-level responses on attitudes and financial behavior. It is a bespoke confidential collection, not an off-the-shelf product. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: survey, household-finance, attitudes, germany, data:bilendi-survey
==============================================================================

:::caution[Bespoke confidential survey: not exercised here]
**This is an author-commissioned survey**, fielded through the panel provider
Bilendi. It is **not an off-the-shelf product and not publicly downloadable**: it
carries **no provenance badge** because there is no standing access path we can
run here. The individual-level responses belong to a specific research project
and were collected under that arrangement. The page documents what the survey is
and the gotchas; treat it as unverified. This is the honest grade under the
institute's Verified discipline.
:::

**The Bilendi survey** is a one-off, author-commissioned online survey of German
respondents, fielded through the survey-panel provider Bilendi. It collects
individual-level responses on attitudes, stock-market participation, demographics,
trust, and risk tolerance, designed to be representative of the target population.
Because it is bespoke, the questionnaire and sample are defined by the research
project rather than by a vendor catalog. A paper we distill uses it:
[Laudenbach, Malmendier & Niessen-Ruenzi](/wiki/papers/jf/2026/laudenbach-communism-attitudes-2026/)
use a 2023 Bilendi survey of 9,695 Germans (5,286 East, 4,409 West) on attitudes,
stock-market participation, trust, and risk tolerance, alongside the
[online-broker](/wiki/confidential/online-broker/) and
[proprietary bank](/wiki/confidential/bank-proprietary/) data.

- **Cost:** commissioned (the authors pay the panel provider to field it); not a
  resold product.
- **Provider:** Bilendi (the online-panel fieldwork); design by the authors.
- **Coverage:** the surveyed sample, weighted to be representative of the target
  German population; a single cross-section, not a panel.

## Access (restricted)

- **No off-the-shelf access.** The survey is specific to the project; there is no
  catalog product to buy and no public file.
- **Project-bound.** The microdata belongs to the commissioning researchers under
  the fielding arrangement; obtaining the same data means commissioning a new
  survey, not licensing this one.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Online panels are not random samples.** Respondents are drawn from an opt-in
  panel and quota-balanced, not a probability sample; representativeness depends on
  the weights and quotas, and some groups are reached poorly. Use the weights and
  report them.
- **Self-reported, single cross-section.** Attitudes, participation, and risk
  tolerance are self-reported at one point in time; there is no panel dimension and
  no administrative check. Treat the measures as stated preferences.
- **Recruitment and wording drive answers.** Question order, framing, and the
  recruitment script shape responses; results are not comparable to a differently
  worded survey. Pin the instrument.
- **Region splits rely on self-report or current location.** East/West (or similar)
  splits depend on where respondents say they grew up or now live; misclassification
  is possible. Check how the split is defined.
- **Not reproducible from a public source.** Others cannot re-pull the same
  responses; replication means a new survey with the same instrument. Plan for that.

## Citation

Cite the design and provider, e.g.: *Author-commissioned online survey fielded by
Bilendi, YYYY; individual responses used under the project arrangement.* State the
fielding year, the sample size and quotas, and the use of weights.


==============================================================================
# China shadow-margin lending data (single provider, restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/china-shadow-margin/
# Daily stock-level off-exchange ("shadow") margin balances from one large Chinese lending platform, used to study the 2015 boom and bust. It is confidential single-counterparty data. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: china, leverage, margin, equities, data:china-shadow-margin
==============================================================================

:::caution[Confidential single-counterparty data: not exercised here]
**This is confidential data from a single shadow-margin platform.** It is **not
for sale and not publicly downloadable**: it carries **no provenance badge**
because there is no access path we can run here. Off-exchange margin lending was
not officially reported, so this is one private platform's records, shared under
a confidential arrangement and not purchasable off the shelf. The page documents
what the data is and the gotchas; treat it as unverified. This is the honest
grade under the institute's Verified discipline.
:::

**The China shadow-margin data** is the daily, stock-level record of off-exchange
("shadow") margin balances from one large lending platform during the Chinese
equity boom and bust of 2014 to 2015. Shadow margin sat outside the regulated,
exchange-reported brokerage margin system and allowed much higher leverage, so it
is central to understanding the 2015 crash but is unobserved in official
statistics. This dataset is one platform's book, estimated at roughly 5% of the
shadow-margin market. A paper we distill uses it:
[Hansman, Hong, Jiang, Liu & Meng](/wiki/papers/rfs/2025/hansman-effects-credit-expansions-stock-2025/)
use the proprietary shadow-margin balances (daily, stock-level, about 5% of the
market) for the bust analysis, alongside exchange formal-margin data and
[CSMAR](/wiki/commercial/csmar/) prices and holdings.

- **Cost:** not for sale. Confidential single-counterparty data.
- **Source:** one large (undisclosed) shadow-margin lending platform.
- **Coverage:** daily stock-level balances on this platform, a partial slice
  (about 5%) of the total shadow-margin market, concentrated around the 2015
  episode.

## Access (restricted)

- **No public download and no resale.** Off-exchange margin was not officially
  collected; this is a private platform's book, shared confidentially.
- **Through a confidential arrangement.** Access required an agreement with the
  platform (or an intermediary); there is no standing way for a third party to
  reach the same data.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **It is about 5% of the market, from one provider.** The platform's book is a
  partial and possibly non-representative slice of total shadow margin; scaling to
  the market requires an assumption about representativeness. State it explicitly.
- **Shadow margin is, by nature, incompletely measured.** Because it was
  unregulated and unreported, there is no benchmark to validate coverage against;
  the true market size is itself uncertain. Treat market-share claims as estimates.
- **Episode-specific.** The data is concentrated around the 2014 to 2015 boom and
  bust; it is not a long panel and the platform's behavior in the crash is
  endogenous to the event. Do not extrapolate beyond the episode.
- **Stock-level balances, not borrower-level.** Balances are by stock on the
  platform; the borrower composition behind them is not the unit, so inferences
  about who delevered are indirect. Match the inference to the stock-level unit.
- **Not redistributable.** Results can be reported but the microdata cannot be
  shared and cannot be re-pulled by others. Plan for non-reproducibility of the raw
  inputs.

## Citation

Cite the source and arrangement, e.g.: *Proprietary shadow-margin balances (single
Chinese platform), confidential; used under agreement, YYYY-MM-DD.* State the
window, the estimated market share, and the stock-level unit.


==============================================================================
# Federal Reserve discount window lending (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/discount-window-confidential/
# Loan-level records of Federal Reserve discount window borrowing (primary credit and related facilities). Contemporaneous borrower-level data is confidential; transaction details are released only with a statutory lag. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: banks, central-bank-lending, liquidity, banking, federal-reserve, data:discount-window-confidential
==============================================================================

:::caution[Confidential supervisory data: not exercised here]
**Contemporaneous discount window borrower data is confidential** and held by the
Federal Reserve. It carries **no provenance badge** because there is no
contemporaneous access path we can run here. Under Dodd-Frank, transaction-level
details are released to the public only after a roughly two-year lag; the
contemporaneous, borrower-identified data is restricted to a Federal Reserve
affiliation or an approved restricted-data arrangement, inside a secure
environment, with output subject to disclosure review. The page documents what
the collection is and the gotchas; treat it as unverified. This is the honest
grade under the institute's Verified discipline.
:::

**Federal Reserve discount window lending** records are the loan-level details of
borrowing at the Fed's discount window (primary credit, secondary credit,
seasonal credit, and, in stress periods, related emergency facilities): which
institution borrowed, how much, at what rate, and for how long. Borrowing is
historically **stigmatized**, so the identity of contemporaneous borrowers is
sensitive and protected. A paper we distill uses it:
[Kotidis & Schreft](/wiki/papers/jf/2025/kotidis-propagation-cyberattacks-financial-system-2025/)
use discount window borrowing records to measure whether banks exposed to a
cyberattack increase their probability of borrowing from the window.

- **Cost:** not for sale contemporaneously. Lagged transaction details are
  released publicly under Dodd-Frank.
- **Source:** Federal Reserve (discount window lending).
- **Coverage:** institutions that borrow at the window; loan-level terms. Most
  banks do not borrow in normal times, so usage is sparse outside stress periods.

## Access (restricted)

- **Lagged release is public; contemporaneous borrower data is not.**
  Dodd-Frank requires the Fed to disclose discount window transaction details
  after about two years; the contemporaneous, borrower-identified data is
  confidential until then.
- **Through a Federal Reserve affiliation or approved program.** Contemporaneous
  loan-level access is limited to researchers at the Federal Reserve System or
  others granted entry to the restricted data, worked inside a secure environment
  with output subject to disclosure review.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Usage is sparse and stigma-driven.** Most banks never borrow in normal times;
  borrowing spikes in stress. The data is a selected, episodic signal, not a
  continuous panel, and the decision to borrow is itself informative (stigma).
- **The two-year lag shapes what is usable contemporaneously.** For recent
  periods only the restricted data exists; the public lagged release is unavailable
  for the most recent two years. Match your access route to your sample window.
- **Facility heterogeneity.** Primary, secondary, and seasonal credit, plus
  crisis facilities, have different terms, eligibility, and meaning; pooling them
  conflates very different borrowing. Separate by facility.
- **Borrowing is an outcome, not an exposure.** Whether a bank borrows is an
  endogenous response to its condition; treating window use as exogenous
  mismeasures causality. Be explicit about identification.
- **Mapping to bank identifiers.** Borrowers must be linked to RSSD IDs and other
  bank data; mergers break the link over time. Verify the crosswalk.
- **Output is disclosure-reviewed and cannot be redistributed.** Contemporaneous
  loan-level results leave the secure environment only after review.

## Citation

Cite the source, e.g.: *Federal Reserve discount window lending records,
confidential; accessed under restricted-data arrangement, YYYY-MM-DD* (or the
public Dodd-Frank lagged release for older periods). State the facility, the
sample window, and whether contemporaneous or lagged data was used.


==============================================================================
# DTCC commercial paper transaction data (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/dtcc/
# Transaction-level commercial paper issuance records (issuer, volume, rate, maturity) from DTCC, used in money-market and bank-funding research. It is confidential, not an off-the-shelf feed. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: money-markets, commercial-paper, funding, post-trade, data:dtcc
==============================================================================

:::caution[Confidential data: not exercised here]
**This is confidential DTCC transaction data.** It is **not an off-the-shelf
feed and not publicly downloadable**: it carries **no provenance badge** because
there is no access path we can run here. Researchers obtained it under a
restricted arrangement with DTCC (the post-trade infrastructure provider). The
page documents what the data is and the gotchas; treat it as unverified. This is
the honest grade under the institute's Verified discipline.
:::

**DTCC commercial paper (CP) transaction data** is the transaction-level record
of CP issuance cleared through DTCC: for each issue, the issuer, volume, rate,
and maturity. Because DTCC clears the bulk of U.S. CP, the data is a broad
cross-section of U.S. CP at the security level, more detailed than the published
aggregate CP outstanding series, but defined by what clears through DTCC. A paper
we distill uses it:
[Anderson, Du & Schlusche](/wiki/papers/jf/2025/anderson-arbitrage-capital-global-banks-2025/)
use DTCC CP transaction data (issuer, volume, rate, maturity) to measure banks'
unsecured funding and build their arbitrage-capital measures, alongside
[FR 2420](/wiki/confidential/fr2420/) money-market rates and
[FR 2644](/wiki/confidential/fr2644/) balance sheets.

- **Cost:** not for sale as a standard feed. Restricted-access confidential data.
- **Source:** DTCC Solutions LLC (the data-licensing subsidiary of the DTCC
  group; CP settles through DTC, the Depository Trust Company, also part of DTCC).
- **Coverage:** CP issuance cleared through DTCC at the transaction level; a
  large share of U.S. CP, but defined by what clears through DTCC, not literally
  every CP trade.

## Access (restricted)

- **No public download.** The transaction microdata is confidential; only
  aggregate CP statistics are public (for example the Fed's CP releases).
- **Through a restricted arrangement with DTCC.** Access is granted per project
  under confidential terms; there is no standing way for a third party to reach
  the same data.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Issuance flow, not outstanding stock.** The data is new issues with their
  terms; building an outstanding-amount series requires aggregating issues by
  maturity, not reading a stock field. Do not confuse issuance with outstanding.
- **CP programs and issuer identity.** An issuer can run multiple programs and
  appear under different names or conduits (especially ABCP); mapping issues to
  the ultimate parent requires care. Verify the issuer crosswalk.
- **Asset-backed versus unsecured CP.** ABCP and unsecured financial/nonfinancial
  CP behave differently; pooling them conflates very different funding. Separate by
  CP type.
- **Rate conventions.** CP is quoted on a discount basis with specific day-count
  conventions; comparing CP rates to other money-market rates needs conversion.
  Convert before comparing.
- **Coverage is defined by DTCC clearing.** Anything not cleared through DTCC is
  outside the data; do not treat it as literally every CP trade. State the
  coverage assumption.
- **Not redistributable.** Results can be reported but the microdata cannot be
  shared or re-pulled by others. Plan for non-reproducibility of the raw inputs.

## Citation

Cite the source and arrangement, e.g.: *DTCC commercial paper transaction data
(DTCC Solutions LLC), confidential; used under restricted arrangement,
YYYY-MM-DD.* State the sample window, the CP type (ABCP versus unsecured), and the
issuance-versus-outstanding treatment.


==============================================================================
# Equifax consumer-credit records (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/equifax-credit/
# Individual-level consumer credit microdata from Equifax (balances, delinquency, scores, account types), often reached as an anonymized matched panel, plus Equifax payroll-based employment and income verification. It is restricted PII, not an off-the-shelf purchase. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: consumer-credit, household-finance, credit-bureau, united-states, data:equifax-credit
==============================================================================

:::caution[Restricted individual-level data: not exercised here]
**This is restricted Equifax consumer-credit microdata** (personally identifiable
in raw form). It is **not an off-the-shelf research purchase and not publicly
downloadable**: it carries **no provenance badge** because there is no access
path we can run here. Researchers reach it only as an anonymized panel under a
restricted data agreement (for example through the New York Fed Consumer Credit
Panel, or a project-specific Equifax extract). The page documents what the data
is and the gotchas; treat it as unverified. This is the honest grade under the
institute's Verified discipline.
:::

**Equifax consumer-credit records** are individual-level credit-bureau data:
monthly balances, delinquency, credit scores, and account types across a
consumer's tradelines. For research it is usually accessed as an **anonymized
matched panel** (the FRBNY Consumer Credit Panel is one standing example). Equifax
also offers **employment and income verification** built on payroll data (its
"Work Number" product), which links credit outcomes to labor-market outcomes. A
paper we distill uses both:
[Di Maggio, Kalda & Yao](/wiki/papers/jf/2026/maggio-student-debt-second-chance-2026/)
use an anonymized Equifax credit-bureau panel for monthly balances, delinquency,
scores, and account types, and Equifax payroll-based employment/income
verification (5,000-plus U.S. firms) for monthly earnings, hours, and employer,
to study student-debt relief.

- **Cost:** not an off-the-shelf research purchase. Restricted-access individual
  microdata.
- **Source:** Equifax (consumer credit-bureau records; payroll-based income
  verification).
- **Coverage:** consumers with a credit file (effectively most U.S. adults with
  credit history), monthly; the income-verification product covers only employees
  of contributing employers.

## Access (restricted)

- **No public download; PII in raw form.** Raw credit records are personally
  identifiable and cannot be posted; research uses anonymized, matched extracts.
- **Through a restricted data agreement.** Access is via a standing anonymized
  panel (for example the FRBNY Consumer Credit Panel) or a project-specific
  Equifax agreement, under strict privacy terms.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Only the credit-visible population.** People with thin or no credit files are
  under-represented or absent; the panel is not the whole adult population. Do not
  generalize to the unbanked or credit-invisible.
- **Bureau-specific coverage.** A consumer's tradelines appear only if the lender
  reports to Equifax; balances and accounts can differ across bureaus. Do not treat
  one bureau as the complete liability picture.
- **Scores and account types follow vendor definitions.** Credit scores and
  account-type codes are Equifax constructs that change over time; a score is not a
  fixed cross-vintage concept. Pin the score model and codebook.
- **Income verification is selected and partial.** The payroll-based income data
  covers only employees of contributing employers, so it is a non-random slice of
  workers; absence of income is not unemployment. Treat coverage carefully.
- **Anonymized matching introduces error.** Linking the credit panel to income or
  to a treatment population is done on anonymized keys with match error; verify the
  match rate and its selectivity.
- **Not redistributable.** Results can be reported but the microdata cannot be
  shared or re-pulled by others. Plan for non-reproducibility of the raw inputs.

## Citation

Cite the source and access route, e.g.: *Equifax consumer-credit records
(anonymized panel; FRBNY Consumer Credit Panel or project-specific extract),
restricted; accessed under data agreement, YYYY-MM-DD*; cite the Equifax
income-verification product separately if used. State the panel, the sample
window, and the score/codebook vintage.


==============================================================================
# FDIC construction-loan servicing records (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/fdic-construction-loans/
# Loan-level construction-loan servicing data for a single failed bank held by the FDIC: terms, draw requests, on-site inspection reports, and outcomes. It is confidential FDIC data. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: banks, credit, real-estate, monitoring, fdic, data:fdic-construction-loans
==============================================================================

:::caution[Confidential FDIC data: not exercised here]
**This is confidential FDIC loan-servicing data** for a single failed bank. It is
**not for sale and not publicly downloadable**: it carries **no provenance badge**
because there is no access path we can run here. Researchers reach it only through
an FDIC affiliation or an approved restricted-data arrangement, inside a secure
environment, with output subject to disclosure review. The page documents what
the collection is and the gotchas; treat it as unverified. This is the honest
grade under the institute's Verified discipline.
:::

**The FDIC construction-loan servicing records** are the loan-level servicing
files for the construction-loan book of **one failed bank**, held by the FDIC as
receiver. Their distinguishing feature is the **draw-request and on-site
inspection** detail: each construction loan funds in stages against physical
inspections, so the data records inspection dates, inspection reports, draw
requests, loan terms, borrower identifiers, and default outcomes. It lets
researchers observe bank monitoring at the loan-action level. A paper we distill
uses it:
[Heitz, Martin & Ufier](/wiki/papers/jf/2026/heitz-bank-monitoring-onsite-inspections-2026/)
use it as the primary dataset (about 11.6 million loan-day observations, 28,939
loans, roughly ten years) to study how on-site inspections affect monitoring and
moral hazard.

- **Cost:** not for sale. Restricted-access confidential FDIC data.
- **Source:** the FDIC (records of a failed bank held in receivership).
- **Coverage:** the construction-loan portfolio of a single failed institution
  over its history; it is one bank, not a system-wide sample.

## Access (restricted)

- **No public download.** The servicing microdata is confidential and is not
  posted.
- **Through an FDIC affiliation or approved program.** Access is limited to
  researchers at the FDIC or others granted entry to the restricted data, worked
  inside a secure environment with output subject to disclosure review.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **It is one bank.** The data is the portfolio of a single failed institution, so
  external validity is the central caveat; patterns may reflect that bank's
  practices, region, and cycle. Do not generalize to the banking system without
  argument.
- **Construction loans are staged and unusual.** Funds disburse in draws against
  inspections, so balances rise over the loan's life and a "draw" is not a new
  loan; standard loan-level conventions (one balance, one origination) do not
  apply. Model the draw schedule explicitly.
- **Loan-day structure inflates observation counts.** The ~11.6M figure is
  loan-days, not loans; clustering and serial correlation within a loan are severe.
  Cluster at the loan (and borrower) level.
- **Inspection timing is partly endogenous.** Inspections are triggered by draw
  requests and by the bank's own concern; treating inspection timing as exogenous
  mismeasures the monitoring effect. Be explicit about identification.
- **A failed bank is a selected sample over time.** The portfolio is observed up
  to and through failure, so later vintages are conditioned on the bank's
  deterioration. Watch for survivorship within the book.
- **Output is disclosure-reviewed and cannot be redistributed.** Results leave the
  secure environment only after review, and the microdata itself cannot be shared.

## Citation

Cite the source, e.g.: *FDIC construction-loan servicing records (single failed
bank), confidential; accessed under restricted-data arrangement, YYYY-MM-DD.*
State the observation unit (loan-day versus loan), the clustering, and the caveat
that it is one institution.


==============================================================================
# FDIC failed-bank bidding and resolution records (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/fdic-failed-bank/
# Bid-level records from FDIC bank-failure resolutions: bidder identities, bid terms, the FDIC's least-cost estimates, and loss-share claims. It is confidential FDIC data. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: banks, bank-failures, resolution, fdic, data:fdic-failed-bank
==============================================================================

:::caution[Confidential FDIC data: not exercised here]
**This is confidential FDIC resolution data.** It is **not for sale and not
publicly downloadable**: it carries **no provenance badge** because there is no
access path we can run here. Researchers reach it only through an FDIC affiliation
or an approved restricted-data arrangement, inside a secure environment, with
output subject to disclosure review. The page documents what the collection is
and the gotchas; treat it as unverified. This is the honest grade under the
institute's Verified discipline.
:::

**The FDIC failed-bank bidding and resolution records** are the proprietary
bid-level files from how the FDIC resolves failed banks. When a bank fails, the
FDIC runs a Purchase and Assumption (P&A) auction; the records capture **bidder
identities, bid terms, the FDIC's internal least-cost estimates**, and subsequent
**loss-share** claims by the acquirer. Because some auctions produce near-tied
bids between acquirer types, the data supports close-bid quasi-random designs. A
paper we distill uses it:
[Johnston-Ross, Ma & Puri](/wiki/papers/jf/2025/johnston-ross-private-equity-financial-stability-2025/)
use the failed-bank bidding records (P&A bid values, acquirer identities, FDIC
least-cost estimates) for a close-bid design on private-equity acquirers, with the
loss-share records for claims by acquirer type, alongside public FDIC
[Call Reports](/wiki/datasets/call-reports/) and
[Summary of Deposits](/wiki/datasets/fdic-summary-of-deposits/).

- **Cost:** not for sale. Restricted-access confidential FDIC data.
- **Source:** the FDIC (resolution and receivership records).
- **Coverage:** resolved bank failures (concentrated in and after the 2008-2013
  wave), at the bid and transaction level.

## Access (restricted)

- **No public download.** The bid-level microdata is confidential and is not
  posted; the FDIC publishes only summary failure information openly.
- **Through an FDIC affiliation or approved program.** Access is limited to
  researchers at the FDIC or others granted entry to the restricted data, worked
  inside a secure environment with output subject to disclosure review.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Bids are conditioned on the failure happening.** The sample is failed banks
  that went to resolution; selection into failure (and into receiving qualified
  bids) precedes the auction. The close-bid design addresses who wins, not who
  fails. Keep the two selection stages distinct.
- **Bidder eligibility is screened.** Not everyone can bid; the FDIC pre-qualifies
  bidders, so the bidder pool is selected, especially for non-bank or PE acquirers.
  Do not treat the bidder set as the universe of potential buyers.
- **Bid structures are not scalar.** A P&A bid bundles deposit premium, asset
  discounts, and loss-share terms; reducing it to one number discards the
  structure that determines the least-cost ranking. Use the FDIC's own least-cost
  comparison.
- **Loss-share claims realize over years.** Loss-share payments accrue well after
  resolution and depend on later asset performance; a snapshot understates
  eventual claims. Pin the as-of date.
- **Identifier joins to public data.** Linking to
  [Call Reports](/wiki/datasets/call-reports/) and other bank data needs the
  failed-bank identifier crosswalk; acquirers also change over time. Verify the
  link.
- **Output is disclosure-reviewed and cannot be redistributed.** Results leave the
  secure environment only after review, and the microdata itself cannot be shared.

## Citation

Cite the source, e.g.: *FDIC failed-bank bidding and resolution records,
confidential; accessed under restricted-data arrangement, YYYY-MM-DD.* State the
resolution window, the bid concept used, and the loss-share as-of date.


==============================================================================
# FDIC confidential supervisory and account-level deposit data (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/fdic-supervisory/
# Confidential FDIC microdata: account-level deposit balances and transactions for a failed bank, plus supervisory enforcement actions and brokered-deposit waivers. It is confidential supervisory data. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: banks, deposits, supervision, fdic, data:fdic-supervisory
==============================================================================

:::caution[Confidential supervisory data: not exercised here]
**This is confidential FDIC supervisory data.** It is **not for sale and not
publicly downloadable**: it carries **no provenance badge** because there is no
access path we can run here. Researchers reach it only through an FDIC affiliation
or an approved restricted-data arrangement, inside a secure environment, with
output subject to disclosure review. The page documents what the collection is
and the gotchas; treat it as unverified. This is the honest grade under the
institute's Verified discipline.
:::

**FDIC confidential supervisory data** spans two complementary kinds of restricted
microdata. The first is **account-level deposit microdata**: daily account
balances and transactions for an individual bank, which lets researchers watch
deposit inflows and outflows account-by-account as a bank approaches failure. The
second is **supervisory information** on the wider bank population: enforcement
actions (cease-and-desist orders, less-than-well-capitalized status) and
brokered-deposit waivers. A paper we distill uses both:
[Martin, Puri & Ufier](/wiki/papers/jf/2026/martin-deposit-flows-failing-banks-2026/)
use FDIC account-level deposit microdata (daily balances and transactions for one
failed bank, from early 2006 to failure) for their core analysis, and the
confidential supervisory data (enforcement actions, capitalization status,
brokered-deposit waivers) to generalize across a panel of banks built on public
[Call Reports](/wiki/datasets/call-reports/).

- **Cost:** not for sale. Restricted-access confidential supervisory data.
- **Source:** the FDIC (account-level records of a failed bank; supervisory
  enforcement records).
- **Coverage:** account-level data is one bank over a window to failure; the
  supervisory enforcement data covers the broader supervised population.

## Access (restricted)

- **No public download.** The account-level and enforcement microdata is
  confidential and is not posted; enforcement actions are public only in summary
  form.
- **Through an FDIC affiliation or approved program.** Access is limited to
  researchers at the FDIC or others granted entry to the restricted data, worked
  inside a secure environment with output subject to disclosure review.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Account-level data is one bank.** The daily deposit microdata is a single
  failed institution; external validity is the central caveat, which is exactly
  why the supervisory panel is used to generalize. Keep the two layers distinct.
- **Account versus depositor versus insured balance.** One depositor can hold
  several accounts, and the insured portion differs from the balance; runs look
  different at the account, depositor, and insured-exposure level. Fix the unit
  before measuring outflows.
- **Enforcement actions are timed and lagged.** A cease-and-desist order or a
  capitalization downgrade is dated by the supervisory process, which lags the
  bank's actual deterioration; treating the action date as the onset of trouble
  mismeasures timing. Use the action date carefully.
- **Brokered-deposit flags are regulatory categories.** Brokered-deposit status
  and waivers follow definitions that have changed; the label is a regulatory
  construct, not a clean economic measure of funding source. Read the definition
  for your period.
- **Identifier joins.** Linking the supervisory panel to public
  [Call Reports](/wiki/datasets/call-reports/) needs the bank identifier crosswalk
  and care with mergers. Verify the link.
- **Output is disclosure-reviewed and cannot be redistributed.** Results leave the
  secure environment only after review, and the microdata itself cannot be shared.

## Citation

Cite the source, e.g.: *FDIC confidential supervisory and account-level deposit
data, confidential; accessed under restricted-data arrangement, YYYY-MM-DD.*
State which layer is used (account-level single bank versus supervisory panel),
the unit (account, depositor, or insured balance), and the sample window.


==============================================================================
# Confidential federal funds transaction data (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/fed-funds-confidential/
# Confidential, transaction-level federal funds borrowing and lending records held by the Federal Reserve, beyond what published benchmark rates reveal. It is restricted supervisory data: this page documents what it is and the gotchas, but it was not exercised here and is not publicly accessible.
# Tags: money-markets, fed-funds, interbank, banking, federal-reserve, data:fed-funds-confidential
==============================================================================

:::caution[Confidential supervisory data: not exercised here]
**Confidential federal funds transaction data is restricted** and held by the
Federal Reserve. It is **not for sale and not publicly downloadable**: it carries
**no provenance badge** because there is no access path we can run here. The
effective federal funds rate and other published benchmarks are public, but the
underlying transaction-level borrowing and lending records are confidential.
Researchers reach them only through a Federal Reserve affiliation or an approved
restricted-data arrangement, inside a secure environment, with output subject to
disclosure review. The page documents what the collection is and the gotchas;
treat it as unverified. This is the honest grade under the institute's Verified
discipline.
:::

**Confidential federal funds transaction data** is the transaction-level record
of overnight interbank borrowing and lending in the federal funds market, held
by the Federal Reserve (collected largely through [FR 2420](/wiki/confidential/fr2420/)
and related reporting). It lets researchers see who borrowed from whom, at what
rate and amount, beyond the published effective rate. A paper we distill uses
it:
[Kotidis & Schreft](/wiki/papers/jf/2025/kotidis-propagation-cyberattacks-financial-system-2025/)
use confidential federal funds data to measure how exposed banks turn to the fed
funds market for borrowing after a cyberattack disrupts their normal payment
flows.

- **Cost:** not for sale. Restricted-access confidential supervisory data.
- **Source:** Federal Reserve (transaction reporting, including FR 2420).
- **Coverage:** federal funds transactions of reporting institutions; the
  reporting universe is a size-thresholded set of banks and FBOs, not every
  market participant.

## Access (restricted)

- **Benchmark rates are public; the transactions are not.** The effective federal
  funds rate and related statistics are published; the transaction-level records
  are confidential.
- **Through a Federal Reserve affiliation or approved program.** Access is limited
  to researchers at the Federal Reserve System or others granted entry to the
  restricted data, worked inside a secure environment with output subject to
  disclosure review.
- See [FR 2420](/wiki/confidential/fr2420/) for the reporting form whose instructions
  are public even though the data cannot be pulled.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **The reporting universe is not the whole market.** Only institutions above a
  size threshold report, so the observed transactions are a selected subset.
  Inferring market-wide totals requires care about coverage.
- **Identifying fed funds versus look-alike trades.** Fed funds, Eurodollars, and
  some repo are economically similar overnight transactions; classification rules
  determine what counts as fed funds. Use the collection's definition, not your
  own.
- **Borrowing and lending sides.** A transaction has two reporters with possibly
  different reporting obligations; double-counting or one-sided coverage can bias
  volumes. Confirm which side is observed.
- **Post-crisis structural change.** The fed funds market changed markedly after
  2008 (large reserves, FHLB lending, IOER); volumes and participants are not
  comparable across that break. Do not pool pre- and post-crisis naively.
- **Mapping to bank identifiers.** Reporters must be linked to RSSD IDs and other
  bank data; mergers break the link over time. Verify the crosswalk.
- **Output is disclosure-reviewed and cannot be redistributed.** Transaction-level
  results leave the secure environment only after review, and the micro-data
  itself cannot be shared.

## Citation

Cite the source, e.g.: *Confidential federal funds transaction data (Federal
Reserve), accessed under restricted-data arrangement, YYYY-MM-DD.* For the public
series, cite the effective federal funds rate. State the reporting basis (for
example FR 2420), the sample window, and the reporting population.


==============================================================================
# Fedwire Funds Service: payment-level transaction data (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/fedwire/
# Fedwire transaction data is the Federal Reserve's confidential record of real-time gross-settlement interbank payments: sender, receiver, value, and timestamp. It is restricted supervisory data: this page documents what it is and the gotchas, but it was not exercised here and is not publicly accessible.
# Tags: payments, interbank, banking, federal-reserve, data:fedwire
==============================================================================

:::caution[Confidential supervisory data: not exercised here]
**Fedwire transaction data is confidential** and held by the Federal Reserve.
It is **not for sale and not publicly downloadable**: it carries **no provenance
badge** because there is no access path we can run here. Aggregate Fedwire
volume and value statistics are published, but the **payment-level** records are
restricted. Researchers reach them only through a Federal Reserve affiliation or
an approved restricted-data arrangement, inside a secure environment, with output
subject to disclosure review. The page documents what the collection is and the
gotchas; treat it as unverified. This is the honest grade under the institute's
Verified discipline.
:::

**Fedwire Funds Service** is the Federal Reserve's real-time gross-settlement
(RTGS) system for large-value interbank payments. The transaction data is the
**payment-level** record of each transfer: sending institution, receiving
institution, value, and timestamp. It is the canonical source for studying
interbank payment flows, liquidity, and contagion through the payment system. A
paper we distill uses it:
[Kotidis & Schreft](/wiki/papers/jf/2025/kotidis-propagation-cyberattacks-financial-system-2025/)
use Fedwire payment-level data (payments sent and received by banks, in number
and value) to trace how a cyberattack on a set of banks propagates through the
payment system.

- **Cost:** not for sale. Restricted-access confidential payment data.
- **Source:** Federal Reserve (the Fedwire Funds Service).
- **Coverage:** large-value interbank transfers settled over Fedwire; complete at
  the payment level within the system, aggregated to the institution (RSSD) level
  for most research.

## Access (restricted)

- **Aggregate statistics are public; payment-level data is not.** The Fed
  publishes Fedwire volume and value aggregates; the payment-level records are
  confidential and are not posted.
- **Through a Federal Reserve affiliation or approved program.** Payment-level
  access is limited to researchers at the Federal Reserve System or others
  granted entry to the restricted data, worked inside a secure environment with
  output subject to disclosure review.
- Records are typically aggregated to the **depository-institution (RSSD)** level
  for analysis; the raw transfer-level data does not leave the secure
  environment.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Fedwire is not all payments.** It settles large-value transfers; small-value
  and many retail flows go through other rails (ACH, CHIPS, card networks).
  Fedwire flows are not the firm's or system's entire payment activity.
- **Identity is at the institution, not the underlying customer.** A payment is
  between member institutions; the ultimate originator and beneficiary are not
  the Fedwire participants. Do not read a participant as the economic party.
- **Bilateral flows can net or gross differently.** Gross RTGS records every
  transfer, but research aggregations may net within a day or a pair; a netted
  series understates gross throughput. Be explicit about the aggregation.
- **Timing and intraday dynamics matter.** Payments carry timestamps, and
  liquidity and queuing dynamics are intraday; collapsing to daily totals discards
  the timing that often drives the question. Decide your time resolution
  deliberately.
- **Mapping to bank identifiers.** Participants must be linked to RSSD IDs and
  then to other bank data; mergers and reorganizations break the link over time.
  Verify the crosswalk for your window.
- **Output is disclosure-reviewed and cannot be redistributed.** Payment-level
  results leave the secure environment only after review, and the micro-data
  itself cannot be shared.

## Citation

Cite the source, e.g.: *Fedwire Funds Service transaction data (Federal
Reserve), confidential; accessed under restricted-data arrangement, YYYY-MM-DD.*
For the public series, cite the Fed's published Fedwire statistics. State the
aggregation level (for example RSSD), the time resolution, and the sample window.


==============================================================================
# FEK: Swedish Structural Business Statistics (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/fek-sweden/
# FEK is Statistics Sweden's firm-level structural business statistics: employment, payroll, productivity, and balance-sheet items for Swedish firms. It is restricted administrative microdata accessed in SCB's secure environment. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: firm-financials, productivity, sweden, administrative, data:fek-sweden
==============================================================================

:::caution[Restricted administrative microdata: not exercised here]
**FEK is confidential Swedish administrative microdata** (Statistics Sweden /
SCB). It is **not for sale and not publicly downloadable**: it carries **no
provenance badge** because there is no access path we can run here. Researchers
reach it only through an approved project, accessed remotely in SCB's secure
environment (MONA), with output subject to confidentiality review. The page
documents what the collection is and the gotchas; treat it as unverified. This is
the honest grade under the institute's Verified discipline.
:::

**FEK** (Foretagens ekonomi, the Structural Business Statistics) is Statistics
Sweden's firm-level economic register: employment, payroll, value added,
productivity, return on assets, investment, and balance-sheet items for Swedish
firms, compiled from accounts and administrative sources. It is the firm-side
complement to the individual-level [LISA](/wiki/confidential/lisa-sweden/) register
and is the standard source for Swedish firm fundamentals in register-based
research. A paper we distill uses it:
[Olsson & Tag](/wiki/papers/jf/2025/olsson-what-cost-privatization-workers-2025/)
use FEK for firm-level employees, payroll, productivity (value added per
employee), ROA, investment ratio, and leverage, 1997 to 2017, linked to LISA
worker outcomes.

- **Cost:** not for sale. Restricted administrative microdata.
- **Producer:** Statistics Sweden (SCB).
- **Coverage:** Swedish firms on an annual panel; firm identifiers link to LISA
  individuals and to establishment registers.

## Access (restricted)

- **Through SCB's secure remote environment (MONA).** Approved researchers
  access pseudonymized extracts remotely; the microdata does not leave SCB, and
  outputs are checked before release.
- **Project approval required.** Access requires an approved research project;
  data is delivered as a project-specific extract, not an open download.
- SCB publishes documentation of the FEK variables, so the schema can be read
  even though the data cannot be pulled.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Firm versus establishment.** FEK is firm-level; linking to worker data in
  [LISA](/wiki/confidential/lisa-sweden/) goes through establishment identifiers, and
  a multi-establishment firm must be handled deliberately. Do not equate firm and
  establishment.
- **Definitional and threshold changes.** The structural business statistics have
  revised variable definitions and reporting frames over time; a long FEK panel is
  not perfectly consistent. Read the documentation for your years.
- **Productivity is a constructed ratio.** Value added per employee depends on how
  value added and headcount are measured; small denominator firms produce
  unstable ratios. Trim or check small firms before using productivity.
- **Coverage of small firms.** Very small firms can be estimated or imputed rather
  than fully reported; do not treat every cell as a clean filing. Check the
  reporting basis by size class.
- **Pseudonymized, project-specific keys.** Firm keys are pseudonyms created per
  delivery; you cannot merge across separately delivered extracts. Plan all links
  inside one delivery.
- **Output is disclosure-reviewed and cannot be redistributed.** Results leave the
  secure environment only after review, and the microdata itself cannot be shared.

## Citation

Cite the producer and access route, e.g.: *Structural Business Statistics (FEK,
Statistics Sweden), accessed via MONA under approved project, YYYY-MM-DD.* State
the years, the firm population, and how firm-to-worker links were built.


==============================================================================
# FHA single-family mortgage data
# https://instituteforautomatedresearch.org/wiki/confidential/fha/
# HUD publishes public aggregate data on FHA-insured single-family mortgages, but the loan-level origination-and-performance microdata used in default research is restricted administrative data. This page documents both and the gotchas; the restricted file was not exercised here.
# Tags: mortgage, housing, microdata, data:fha
==============================================================================

:::caution[Restricted loan-level data: not exercised here]
The FHA **loan-level** origination-and-performance microdata used in academic
default research is **restricted administrative data**, often matched to
confidential HMDA, reached through HUD or a supervisory arrangement; it carries
**no provenance badge** because there is no access path we can run here. HUD
does publish **public aggregate** FHA products (described below), which are
free, but they are not the loan-level file. The
[Frame, Huang, Jiang, Lee, Liu, Mayer & Sunderam](/wiki/papers/jf/2025/frame-impact-minority-representation-mortgage-2025/)
paper states its underlying FHA microdata is proprietary and was not accessed.
This is the honest grade under the institute's Verified discipline.
:::

**FHA single-family data** describes mortgages insured by the Federal Housing
Administration (part of HUD). Two distinct things travel under this name:

- **Public aggregates (free).** HUD publishes the FHA Single Family Portfolio
  Snapshot (monthly endorsement and portfolio counts), the FHA Single Family
  Loan Performance Trends report, the Annual Report to Congress on the Mutual
  Mortgage Insurance Fund, and Neighborhood Watch (lender-level early-default
  and early-claim rates).
- **Loan-level microdata (restricted).** The population of FHA single-family
  originations with borrower, loan-term, and delinquency-outcome fields is
  administrative data. It is used in, for example
  [Frame, Huang, Jiang, Lee, Liu, Mayer & Sunderam](/wiki/papers/jf/2025/frame-impact-minority-representation-mortgage-2025/)
  for the population of FHA originations and a default analysis (90+ days
  delinquent), reached as confidential data and matched to confidential HMDA.

- **Cost:** public aggregates are free; the loan-level microdata is restricted,
  not for sale.
- **Source:** HUD / Federal Housing Administration.
- **Coverage:** FHA-insured single-family mortgages (a selected, higher-LTV,
  more first-time-buyer population), not the whole mortgage market.

## Access

- **Public aggregates.** Reach the FHA single-family data products from HUD's
  FHA data pages and Neighborhood Watch (`neighborhoodwatch.hud.gov`). These
  are aggregate or lender-level, downloadable without an account.
- **Restricted loan-level.** The loan-level origination-and-performance file is
  obtained through HUD or a supervisory arrangement, frequently matched to the
  confidential HMDA microdata at the Federal Reserve, inside a controlled
  environment with disclosure review. There is no public download.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; the restricted file was not run here.

- **Public aggregates are not the loan-level file.** The portfolio snapshot is
  monthly counts and the performance trends are aggregates; neither substitutes
  for loan-level outcomes. Do not infer borrower-level results from the public
  products.
- **FHA is a selected population.** FHA insures lower-down-payment, higher-LTV,
  more first-time-buyer loans, so FHA default rates do not generalise to the
  conventional or jumbo market. State the population.
- **"Default" is defined several ways.** Ninety-plus days delinquent, insurance
  claim, and foreclosure are different events; pick one and state it. Cure and
  reinstatement create spells, so a static default flag misses the dynamics.
- **Servicing transfers fragment loan histories.** A loan can change servicers
  over its life, so assembling a clean performance panel requires tracking the
  loan, not the servicer.
- **The research file is disclosure-controlled.** Loan-level FHA work matched
  to confidential HMDA leaves the secure environment only after review and
  cannot be redistributed, so results are not reproducible from public files.

## Citation

Cite the specific FHA product used (a named public aggregate, or the restricted
loan-level administrative data), with the vintage, the population, and the
default definition. For restricted loan-level work, state the access route and
that output was disclosure-reviewed.


==============================================================================
# FICC GCF Repo Service data (dealer-level, restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/ficc-gcf-repo/
# Dealer-level daily interdealer general-collateral repo and reverse-repo activity by asset class from the FICC GCF Repo Service, licensed through the New York Fed. It is confidential. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: repo, dealers, money-markets, post-trade, data:ficc-gcf-repo
==============================================================================

:::caution[Confidential data: not exercised here]
**This is confidential FICC repo data.** It is **not for sale and not publicly
downloadable**: it carries **no provenance badge** because there is no access
path we can run here. Researchers obtained dealer-level GCF data under a
restricted licence through the Federal Reserve Bank of New York. The page
documents what the data is and the gotchas; treat it as unverified. This is the
honest grade under the institute's Verified discipline.
:::

**FICC GCF Repo Service data** is the dealer-level record of the **General
Collateral Finance (GCF) repo** segment: an interdealer, blind-brokered,
general-collateral repo market cleared by the Fixed Income Clearing Corporation
(FICC, part of DTCC). The data gives daily interdealer repo and reverse-repo
amounts by asset class at the dealer level, a view of the interdealer GC segment
that the published aggregate series does not resolve. A paper we distill uses it:
[Copeland & Martin](/wiki/papers/jf/2025/copeland-repo-financial-crisis-2025/)
use the dealer-level FICC GCF data for interdealer GC repo and reverse repo by
asset class, alongside [FR 2004C](/wiki/confidential/fr2004c/) primary-dealer
positions, to study the repo market over the financial crisis.

- **Cost:** not for sale. Restricted-access confidential data, licensed via the
  FRBNY.
- **Source:** DTCC / FICC (the GCF Repo Service), reached through the Federal
  Reserve Bank of New York under restricted terms.
- **Coverage:** the interdealer GCF (general-collateral) repo segment by asset
  class, daily, at the dealer level; it is one segment of the broader repo market.

## Access (restricted)

- **No public download.** The dealer-level data is confidential; only aggregate
  GCF statistics are public.
- **Through a restricted FRBNY arrangement.** Access to the dealer-level data is
  licensed through the New York Fed under confidential terms; there is no standing
  way for a third party to reach it.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **GCF is one segment, not the whole repo market.** The repo market has tri-party,
  bilateral/DVP, and GCF segments; GCF is the interdealer general-collateral slice.
  Do not read GCF activity as total repo. Combine segments deliberately.
- **General collateral, not specific securities.** GCF trades general-collateral
  baskets by asset class, not individual CUSIPs; you cannot identify a specific
  security's repo rate from it. Match the question to the GC level.
- **Blind-brokered and netted through FICC.** Trades are intermediated and novated
  to FICC as central counterparty; dealer-level positions reflect that netting, so
  gross bilateral exposure is not directly visible. Account for central clearing.
- **Asset-class definitions and segment changes.** The GCF product and its
  asset-class buckets have changed over time (and interact with tri-party reform);
  a series across the change needs reconciliation. Read the definitions for your
  window.
- **Dealer identity and entry/exit.** The set of GCF participants changes,
  especially around 2008; dealer identifiers must be tracked across mergers and
  exits. Verify the dealer roster.
- **Not redistributable.** Results can be reported but the microdata cannot be
  shared or re-pulled by others. Plan for non-reproducibility of the raw inputs.

## Citation

Cite the source and arrangement, e.g.: *FICC GCF Repo Service data (DTCC/FICC),
confidential; licensed via the Federal Reserve Bank of New York, YYYY-MM-DD.*
State the asset class, the sample window, and that it is the interdealer GCF
segment only.


==============================================================================
# FR 2052a: Complex Institution Liquidity Monitoring Report (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/fr-2052a/
# FR 2052a is the Federal Reserve's confidential liquidity-monitoring collection from large banking organizations: daily and monthly cash inflows and outflows by counterparty, product, and maturity. It is confidential supervisory data: this page documents what it is and the gotchas, but it was not exercised here and is not publicly accessible.
# Tags: banks, liquidity, funding, banking, federal-reserve, data:fr-2052a
==============================================================================

:::caution[Confidential supervisory data: not exercised here]
**FR 2052a is confidential supervisory data** collected by the Federal Reserve.
It is **not for sale and not publicly downloadable**: it carries **no provenance
badge** because there is no access path we can run here. Researchers reach it
only through a Federal Reserve affiliation or an approved restricted-data
arrangement, inside a secure environment, with output subject to disclosure
review. The page documents what the collection is and the gotchas; treat it as
unverified. This is the honest grade under the institute's Verified discipline.
:::

**FR 2052a** (the Complex Institution Liquidity Monitoring Report) is the
granular liquidity-flow collection that supports the Federal Reserve's
liquidity supervision and the Liquidity Coverage Ratio (LCR) framework. Large
banking organizations report cash and securities inflows and outflows broken
down by counterparty, product type, and maturity bucket, at daily frequency for
the largest filers and monthly for smaller ones. It is the most detailed view of
a bank's funding structure available to supervisors. A paper we distill uses it:
[Cooperman, Duffie, Luck, Wang & Yang](/wiki/papers/jf/2025/cooperman-bank-funding-risk-reference-2025/)
use FR 2052a (with [FR Y-14Q](/wiki/confidential/fr-y14q/)) for bank balance sheet
and funding composition by counterparty and product, the main panel for their
COVID-period empirics (20 BHCs, July 2017 to April 2022).

- **Cost:** not for sale. Restricted-access confidential supervisory data.
- **Collector:** Federal Reserve Board (the FR 2052a collection).
- **Coverage:** large banking organizations subject to enhanced liquidity
  standards; daily for the largest, monthly for smaller filers. The reporting
  population and granularity have changed across collection vintages.

## Access (restricted)

- **No public download.** The microdata is confidential and is not posted on the
  Fed website or FRED.
- **Through a supervisory affiliation or approved program.** Access is limited to
  researchers at the Federal Reserve System or others granted entry to the
  restricted data, worked inside a secure environment with output subject to
  disclosure review.
- The published **form and instructions** are public (the Fed posts the FR 2052a
  reporting template), so the schedule and field definitions can be read even
  though the data cannot be pulled.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Coverage is only the largest banks, so it is not the banking system.** Only
  organizations subject to enhanced liquidity standards report, so the sample is
  a selected set of the largest filers. Do not generalize liquidity findings to
  smaller banks.
- **Mixed daily and monthly frequency.** The largest filers report daily while
  smaller ones report monthly; aligning the panel to a common frequency drops or
  coarsens part of the sample. Be explicit about which frequency a given filer
  contributes.
- **The product and counterparty taxonomy is the schema, and it changed.** Flows
  are classified into a detailed product-by-counterparty-by-maturity grid that
  has been revised across vintages; a category can change meaning between
  versions. Read the form instructions for your sample period.
- **Reporting population shifts for regulatory reasons.** Filers enter and exit as
  thresholds and tailoring rules change, breaking the panel for non-economic
  reasons. Control for the changing population.
- **Flows, not stocks.** The report is built around projected inflows and
  outflows under the LCR construct; do not read a flow cell as a balance-sheet
  stock without mapping it to the reporting definition.
- **Output is disclosure-reviewed and cannot be redistributed.** Results leave
  the secure environment only after review, and the microdata itself cannot be
  shared. Plan for aggregation and review when designing what you report.

## Citation

Cite the collection and collector, e.g.: *FR 2052a (Federal Reserve Board),
confidential supervisory data; accessed under restricted-data arrangement,
YYYY-MM-DD.* State the reporting frequency, the sample window, and the filer
population for your period.


==============================================================================
# FR Y-14Q: confidential bank supervisory data (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/fr-y14q/
# FR Y-14Q is the Federal Reserve's quarterly stress-test data collection from large bank holding companies: loan-level corporate (H.1) and commercial real estate (H.2) records and more. It is confidential supervisory data: this page documents what it is and the gotchas, but it was not exercised here and is not publicly accessible.
# Tags: banks, credit, banking, federal-reserve, panel-data, data:fr-y14q
==============================================================================

:::caution[Confidential supervisory data: not exercised here]
**FR Y-14Q is confidential supervisory data** collected by the Federal Reserve.
It is **not for sale and not publicly downloadable**: it carries **no provenance
badge** because there is no access path we can run here. Researchers reach it
only through a Federal Reserve, FDIC, or OCC affiliation or an approved
restricted-data arrangement, inside a secure environment, and results are
subject to disclosure review. The page documents what the collection is and the
gotchas; treat it as unverified. This is the honest grade under the institute's
Verified discipline.
:::

**FR Y-14Q** is the **quarterly** data collection that supports the Federal
Reserve's stress tests (CCAR / DFAST). Large bank holding companies report
detailed, often **loan-level**, records across many schedules. The two
schedules most used in credit research are **Schedule H.1 (corporate loans)**
and **Schedule H.2 (commercial real estate)**, which carry loan terms, interest
rates, the bank's own probability-of-default (PD) and loss-given-default (LGD)
estimates, borrower financials, and utilization. It is used in, for example
[Beyhaghi, Fracassi & Weitzner](/wiki/papers/jf/2026/beyhaghi-adverse-selection-corporate-loans-2026/)
for loan-level interest rates, PD, LGD, and firm financials (2014Q4 to 2019Q4),
[Cooperman, Duffie, Luck, Wang & Yang](/wiki/papers/jf/2025/cooperman-bank-funding-risk-reference-2025/)
for loan-level credit commitments and utilization, and
[Greenwald, Krainer & Paul](/wiki/papers/jf/2025/greenwald-credit-line-channel-2025/)
for the credit-line draw analysis.

- **Cost:** not for sale. Restricted-access confidential supervisory data.
- **Collector:** Federal Reserve Board (the FR Y-14Q collection).
- **Coverage:** bank holding companies above the stress-test asset threshold;
  loan-level and aggregate schedules, quarterly. The reporting population and
  threshold have changed over time, so the panel is not fixed.

## Access (restricted)

- **No public download.** Unlike the Fed's published statistical releases, the
  microdata is confidential and is not posted on the Fed website or FRED.
- **Through a supervisory affiliation or approved program.** Access is limited to
  researchers at the Federal Reserve System, the FDIC, the OCC, or others
  granted entry to the restricted data, worked inside a secure environment with
  output subject to disclosure review.
- The published **form and instructions** are public (the Fed posts the FR Y-14Q
  reporting templates), so the schedule and field definitions can be read even
  though the data cannot be pulled.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Coverage is only large BHCs, so it is not the banking system.** Only holding
  companies above the stress-test asset threshold report, so the sample is a
  selected set of the largest banks. Do not generalise loan-level findings to
  community or mid-size banks.
- **The reporting threshold and population changed over time.** The asset
  threshold and the set of filers shifted across the 2010s, which breaks the
  panel: firms enter and exit the collection for regulatory reasons, not
  economic ones. Control for the changing population.
- **Loan-level corporate detail starts mid-decade.** The collection began early
  in the 2010s, but the granular loan-level corporate (H.1) reporting builds up
  over time; check the field-by-field start date for the window you use.
- **PD and LGD are the banks' own model outputs.** Risk estimates are produced by
  each reporting bank's internal models, so they are heterogeneous across filers
  and not a common-methodology benchmark. Treat cross-bank comparisons of PD/LGD
  with care.
- **Definitions change across collection vintages.** Schedules and instructions
  are revised over the years; a field can change meaning or reporting basis
  between vintages. Read the form instructions for your sample period.
- **Output is disclosure-reviewed and cannot be redistributed.** Results leave
  the secure environment only after review, and the microdata itself cannot be
  shared. Plan for aggregation and review when designing what you report.

## Citation

Cite the collection and collector, e.g.: *FR Y-14Q (Federal Reserve Board),
confidential supervisory data; accessed under restricted-data arrangement,
YYYY-MM-DD.* State the schedule (for example H.1 or H.2), the sample window, and
the reporting population for your period.


==============================================================================
# FR 2004C: weekly primary-dealer positions (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/fr2004c/
# FR 2004C is the dealer-level detail behind the Federal Reserve Bank of New York's weekly primary-dealer statistics: positions, transactions, and financing in government and other securities. Only aggregates are published; the dealer-level data is confidential. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: dealers, repo, treasuries, fixed-income, federal-reserve, data:fr2004c
==============================================================================

:::caution[Confidential supervisory data: not exercised here]
**FR 2004C dealer-level data is confidential** and collected by the Federal
Reserve Bank of New York. The Fed publishes the **aggregate** primary-dealer
statistics, but the **dealer-level** records carry **no provenance badge**:
there is no access path we can run here. Researchers reach the dealer-level
data only through a Federal Reserve affiliation or an approved restricted-data
arrangement, inside a secure environment, with output subject to disclosure
review. The page documents what the collection is and the gotchas; treat it as
unverified. This is the honest grade under the institute's Verified discipline.
:::

**FR 2004C** is the weekly collection of **primary government securities dealer**
positions, transactions, settlements, and financing that the Federal Reserve
Bank of New York uses to produce its published primary-dealer statistics. The
dealer-level micro-records (and related FR 2004 special-purpose surveys) cover
all primary dealers at weekly frequency, with positions and financing broken out
by asset class. A paper we distill uses it:
[Copeland & Martin](/wiki/papers/jf/2025/copeland-repo-financial-crisis-2025/)
use FR 2004C as the primary source for repo and reverse-repo outstanding by
asset class across all primary dealers at weekly frequency, 2004 to 2013.

- **Cost:** not for sale at the dealer level. Aggregates are published free.
- **Collector:** Federal Reserve Bank of New York (the FR 2004 dealer report
  series).
- **Coverage:** the primary-dealer population, weekly; positions, transactions,
  and financing by asset class. The set of primary dealers changes over time.

## Access (restricted)

- **Aggregates are public; the dealer-level micro-data is not.** The New York
  Fed posts weekly aggregate primary-dealer statistics on its website; the
  underlying dealer-level FR 2004 records are confidential and are not posted.
- **Through a Federal Reserve affiliation or approved program.** Dealer-level
  access is limited to researchers at the Federal Reserve System or others
  granted entry to the restricted data, worked inside a secure environment with
  output subject to disclosure review.
- The published **form and instructions** are public, so the field definitions
  can be read even though the dealer-level data cannot be pulled.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Primary dealers are not the whole market.** The report covers only the
  primary-dealer population, a selected set of large dealers; positions and
  financing here are not the entire dealer or repo market. Do not generalize to
  non-primary dealers.
- **The dealer set changes over time.** Primary dealers are added and removed,
  and firms merge or exit (notably around 2008), breaking the panel for
  structural reasons. Track the dealer roster for your window.
- **Net versus gross positions.** Reported positions can be net of offsetting
  long and short exposures within an asset class; a net figure understates gross
  market footprint. Confirm the reporting basis before interpreting levels.
- **Asset-class definitions and form revisions.** The FR 2004 forms have been
  revised over the years, and asset-class buckets can change; a category can
  shift meaning between vintages. Read the instructions for your sample period.
- **Aggregates can be revised.** The published weekly aggregates are subject to
  revision as dealers resubmit; a real-time series differs from the final one.
  Pin the vintage if timing matters.
- **Output is disclosure-reviewed and cannot be redistributed.** Dealer-level
  results leave the secure environment only after review, and the micro-data
  itself cannot be shared.

## Citation

Cite the collection and collector, e.g.: *FR 2004C dealer-level data (Federal
Reserve Bank of New York), confidential; accessed under restricted-data
arrangement, YYYY-MM-DD.* For the public series, cite the New York Fed primary
dealer statistics. State the asset class, the sample window, and the dealer
population for your period.


==============================================================================
# FR 2420: Report of Selected Money Market Rates (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/fr2420/
# FR 2420 is the Federal Reserve's confidential transaction-level collection of money-market rates: federal funds, Eurodollars, and certificates of deposit from banks and FBOs. It underlies published benchmarks but the transaction data is confidential. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: money-markets, rates, fed-funds, banking, federal-reserve, data:fr2420
==============================================================================

:::caution[Confidential supervisory data: not exercised here]
**FR 2420 is confidential supervisory data** collected by the Federal Reserve.
It is **not for sale and not publicly downloadable**: it carries **no provenance
badge** because there is no access path we can run here. Published benchmark
rates derived from it (for example the effective federal funds rate) are public,
but the transaction-level micro-data is restricted. Researchers reach it only
through a Federal Reserve affiliation or an approved restricted-data
arrangement, inside a secure environment, with output subject to disclosure
review. The page documents what the collection is and the gotchas; treat it as
unverified. This is the honest grade under the institute's Verified discipline.
:::

**FR 2420** (the Report of Selected Money Market Rates) is the daily,
**transaction-level** collection of money-market borrowing rates that banks and
foreign banking organizations (FBOs) report to the Federal Reserve: federal
funds (FF), Eurodollars (ED), and negotiable certificates of deposit (CDs). It
is the data behind published benchmarks and a backbone for studies of
short-term funding and rate arbitrage. A paper we distill uses it:
[Anderson, Du & Schlusche](/wiki/papers/jf/2025/anderson-arbitrage-capital-global-banks-2025/)
use FR 2420 daily transaction-level FF, ED, and CD data for U.S. banks and FBOs
as the basis of their IOER and CIP arbitrage-capital measures;
[Cooperman, Duffie, Luck, Wang & Yang](/wiki/papers/jf/2025/cooperman-bank-funding-risk-reference-2025/)
use it for wholesale-funding rate sensitivity.

- **Cost:** not for sale. Restricted-access confidential supervisory data.
- **Collector:** Federal Reserve Board (the FR 2420 collection).
- **Coverage:** reporting banks and FBOs above a size threshold; daily,
  transaction-level FF, ED, and CD borrowing. The reporting population and the
  instrument scope have changed over time.

## Access (restricted)

- **Derived benchmarks are public; the transaction data is not.** Published rates
  computed from FR 2420 (such as the effective federal funds rate) are posted by
  the Fed, but the underlying transaction records are confidential.
- **Through a Federal Reserve affiliation or approved program.** Access to the
  micro-data is limited to researchers at the Federal Reserve System or others
  granted entry to the restricted data, worked inside a secure environment with
  output subject to disclosure review.
- The published **form and instructions** are public, so the field definitions
  can be read even though the data cannot be pulled.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Reporting threshold means it is not all banks.** Only institutions above a
  size threshold report, so the transaction universe is a selected set of larger
  banks and FBOs. Do not treat it as the entire money market.
- **Borrowing side only, by instrument.** The report captures the reporter's
  borrowing transactions in the named instruments; it is not a complete two-sided
  view of the market. Match the instrument definition to your question.
- **Instrument scope and definitions changed.** The collection has been revised
  (for example as benchmark-rate reform changed what is collected); a field can
  change basis between vintages. Read the instructions for your sample period.
- **Settlement and timing conventions.** Rates are tied to transaction and
  settlement dates with specific conventions; mis-dating a transaction shifts it
  across a rate-decision boundary. Be explicit about which date you use.
- **Population shifts for regulatory reasons.** Filers enter and exit as
  thresholds change, breaking the panel for non-economic reasons. Control for the
  changing population.
- **Output is disclosure-reviewed and cannot be redistributed.** Transaction-level
  results leave the secure environment only after review, and the micro-data
  itself cannot be shared.

## Citation

Cite the collection and collector, e.g.: *FR 2420 (Federal Reserve Board),
confidential supervisory data; accessed under restricted-data arrangement,
YYYY-MM-DD.* State the instrument (FF, ED, or CD), the sample window, and the
reporting population for your period.


==============================================================================
# FR 2644: Weekly Report of Selected Assets and Liabilities (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/fr2644/
# FR 2644 is the Federal Reserve's confidential weekly bank balance-sheet collection from a sample of domestic banks and FBO branches. It underlies the published H.8 aggregates but the bank-level data is confidential. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: banks, balance-sheet, banking, federal-reserve, data:fr2644
==============================================================================

:::caution[Confidential supervisory data: not exercised here]
**FR 2644 bank-level data is confidential** and collected by the Federal Reserve.
The published H.8 aggregates derived from it are public, but the **bank-level**
records carry **no provenance badge**: there is no access path we can run here.
Researchers reach the bank-level data only through a Federal Reserve affiliation
or an approved restricted-data arrangement, inside a secure environment, with
output subject to disclosure review. The page documents what the collection is
and the gotchas; treat it as unverified. This is the honest grade under the
institute's Verified discipline.
:::

**FR 2644** (the Weekly Report of Selected Assets and Liabilities of Domestically
Chartered Commercial Banks and U.S. Branches and Agencies of Foreign Banks) is
the **weekly** balance-sheet collection from a sample of banks that supports the
Federal Reserve's published H.8 release on commercial-bank assets and
liabilities. Key asset and liability categories are reported weekly for domestic
banks and FBO branches. A paper we distill uses it:
[Anderson, Du & Schlusche](/wiki/papers/jf/2025/anderson-arbitrage-capital-global-banks-2025/)
use FR 2644 weekly bank balance-sheet data for U.S.-based entities (domestic
banks and FBO branches) to study balance-sheet adjustments to funding shocks
(their balance-sheet regressions use 50 banks with FR 2644 data).

- **Cost:** not for sale at the bank level. The H.8 aggregates are published free.
- **Collector:** Federal Reserve Board (the FR 2644 collection, basis of H.8).
- **Coverage:** a weekly-reporting sample of domestic banks and FBO branches;
  selected balance-sheet categories. The sample is a panel of reporters, not the
  full banking system.

## Access (restricted)

- **The H.8 aggregates are public; the bank-level data is not.** The Fed posts
  the weekly H.8 aggregates; the underlying FR 2644 bank-level records are
  confidential and are not posted.
- **Through a Federal Reserve affiliation or approved program.** Bank-level
  access is limited to researchers at the Federal Reserve System or others
  granted entry to the restricted data, worked inside a secure environment with
  output subject to disclosure review.
- The published **form and instructions** are public, so the field definitions
  can be read even though the bank-level data cannot be pulled.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **It is a weekly-reporter sample, not a census.** Only sampled banks report
  weekly, and the published H.8 is estimated by benchmarking the sample to the
  full population (from the quarterly Call Reports). Bank-level FR 2644 is the
  sample, not every bank.
- **Selected categories only.** The report carries selected asset and liability
  lines, not a full balance sheet; do not expect every Call Report item. Map your
  variable to the FR 2644 line definition.
- **Weekly versus quarterly reconciliation.** FR 2644 is weekly while the
  population benchmark (Call Reports) is quarterly; the two do not line up
  observation-for-observation. Be explicit about which series you use and why.
- **Sample composition changes.** The set of weekly reporters is refreshed over
  time, breaking the panel for non-economic reasons. Track the reporter set for
  your window.
- **Definitions change across vintages.** Lines and instructions are revised; a
  category can change basis between versions. Read the instructions for your
  sample period.
- **Output is disclosure-reviewed and cannot be redistributed.** Bank-level
  results leave the secure environment only after review, and the micro-data
  itself cannot be shared.

## Citation

Cite the collection and collector, e.g.: *FR 2644 (Federal Reserve Board),
confidential supervisory data; accessed under restricted-data arrangement,
YYYY-MM-DD.* For the public series, cite the Federal Reserve H.8 release. State
the balance-sheet lines, the sample window, and the reporter population for your
period.


==============================================================================
# IAB Establishment Panel (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/iab-establishment-panel/
# The IAB Establishment Panel is an annual representative survey of German establishments covering employment, wages, investment, and business practices. It is restricted microdata accessed through the IAB Research Data Centre. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: establishments, survey, germany, administrative, data:iab-establishment-panel
==============================================================================

:::caution[Restricted survey microdata: not exercised here]
**The IAB Establishment Panel is confidential German survey microdata**
(Institute for Employment Research, IAB). It is **not for sale and not publicly
downloadable**: it carries **no provenance badge** because there is no access
path we can run here. Researchers reach it only through the **IAB Research Data
Centre (FDZ)**, on-site or via remote execution, with output subject to
disclosure review. The page documents what the collection is and the gotchas;
treat it as unverified. This is the honest grade under the institute's Verified
discipline.
:::

**The IAB Establishment Panel** is an annual, representative survey of German
**establishments** (workplaces), interviewing thousands of establishments each
year about employment, hiring and separations, wages, working time, training,
investment, business expectations, and internal practices. Because it samples
establishments drawn from the social-security universe, it complements the
administrative [IEB](/wiki/confidential/ieb-germany/) (which has the worker spells but
not the survey questions on practices and expectations). A paper we distill uses
it:
[Bias, Lochner, Obernberger & Sevilir](/wiki/papers/jf/2026/bias-going-public-internal-organization-2026/)
use the IAB Establishment Panel to test whether the hierarchical changes they
measure correlate with the formalization of internal processes, alongside the
worker-level [IEB](/wiki/confidential/ieb-germany/).

- **Cost:** not for sale. Restricted survey microdata.
- **Producer:** Institute for Employment Research (IAB); access via the FDZ.
- **Coverage:** a representative annual panel of German establishments (the survey
  began in the western states in 1993 and added the east in 1996), with sampling
  weights for population inference.

## Access (restricted)

- **Through the IAB Research Data Centre (FDZ).** Approved researchers work on-site
  at an FDZ guest workstation or via the FDZ remote-execution service (JoSuA);
  the microdata does not leave the secure environment, and outputs are checked.
- **Project approval required.** Access requires an approved project and a data-use
  agreement; data is a project-specific extract, not an open download.
- The FDZ publishes documentation and questionnaires, so the variables can be read
  even though the data cannot be pulled.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **It is a survey, so use the weights.** Unlike the administrative IEB, this is a
  sample; population statements require the provided sampling weights, and
  large establishments are over-sampled by design. Do not treat raw counts as
  population totals.
- **Establishment, not firm.** The unit is the establishment (workplace), not the
  legal firm; a multi-site firm appears as separate establishments or not at all.
  Do not roll up to the firm without external links.
- **Panel attrition and rotation.** Establishments enter, drop out, and are
  refreshed; the panel is unbalanced and attrition is non-random (closures). Model
  entry/exit rather than assuming a fixed panel.
- **Questionnaire changes year to year.** Topics rotate and question wording
  changes across waves; a variable may exist only in some years or shift meaning.
  Check the questionnaire for each wave you use.
- **Self-reported and recall-based.** Many items are self-reported by the
  respondent and refer to a reference date or prior year; they carry recall and
  reporting error. Cross-check against IEB administrative measures where possible.
- **Output is disclosure-reviewed and cannot be redistributed.** Results leave the
  FDZ only after review, and the microdata itself cannot be shared.

## Citation

Cite the producer and access route, e.g.: *IAB Establishment Panel, Institute for
Employment Research (IAB), accessed via the FDZ under approved project,
YYYY-MM-DD.* State the waves used, the application of sampling weights, and the
establishment population.


==============================================================================
# IEB: German Integrated Employment Biographies (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/ieb-germany/
# The Integrated Employment Biographies (IEB) are the German Institute for Employment Research's administrative day-level employment records for the universe of workers covered by social security. They are restricted microdata accessed through the IAB Research Data Centre. This page documents what they are and the gotchas, but they were not exercised here.
# Tags: labor, employment-biographies, germany, administrative, data:ieb-germany
==============================================================================

:::caution[Restricted administrative microdata: not exercised here]
**The IEB are confidential German administrative microdata** (Institute for
Employment Research, IAB). They are **not for sale and not publicly downloadable**:
they carry **no provenance badge** because there is no access path we can run
here. Researchers reach them only through the **IAB Research Data Centre (FDZ)**,
on-site or via remote execution, with output subject to disclosure review. The
page documents what the collection is and the gotchas; treat it as unverified.
This is the honest grade under the institute's Verified discipline.
:::

**The Integrated Employment Biographies (IEB)** are the IAB's integrated
administrative source covering the working life of essentially every person in
Germany subject to social security: employment spells (with daily start and end
dates), wages, occupation, establishment, benefit receipt, and job-search
registration, assembled from the underlying social-security and employment-agency
notifications. They are the backbone of German register-based labor research and
the source from which many IAB sample products are drawn. A paper we distill uses
them:
[Bias, Lochner, Obernberger & Sevilir](/wiki/papers/jf/2026/bias-going-public-internal-organization-2026/)
use the IEB as the main source for occupational codes, wages, hierarchy layering,
functions, and tenure when measuring how going public reshapes a firm's internal
organization, alongside the [IAB Establishment Panel](/wiki/confidential/iab-establishment-panel/).

- **Cost:** not for sale. Restricted administrative microdata.
- **Producer:** Institute for Employment Research (IAB); access via the FDZ.
- **Coverage:** the universe of workers in social-security-covered employment
  (employment, marginal employment, benefits, job search), with daily spell
  resolution; civil servants and the self-employed are not covered.

## Access (restricted)

- **Through the IAB Research Data Centre (FDZ).** Approved researchers work on-site
  at an FDZ guest workstation or via the FDZ remote-execution service (JoSuA);
  the microdata does not leave the secure environment, and outputs are checked.
- **Project approval required.** Access requires an approved project and a data-use
  agreement; data is provided as a project-specific extract, not an open download.
- The FDZ publishes documentation of the IEB variables and spell structure, so
  the schema can be read even though the data cannot be pulled.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **It is a spell file, not a person-year panel.** Records are employment spells
  with daily start/end dates that overlap and stack (parallel jobs); collapsing to
  a person-year requires deliberate rules about dominant job and overlaps. Do not
  treat raw spells as observations.
- **Wages are right-censored at the contribution ceiling.** Earnings are capped at
  the social-security contribution limit, which varies by year and region;
  high-wage observations are censored. Use an imputation or account for the ceiling.
- **No civil servants or self-employed.** Coverage is social-security employment
  only; whole occupational groups are absent. Do not read the IEB as the entire
  labor force.
- **Occupation and education coding has breaks.** The occupation classification
  (KldB) changed (notably the 2010 revision) and education is imputed by known
  procedures; a code or an education level is not stable across the break. Pin the
  classification and imputation vintage.
- **Establishment identifiers can split or merge.** Establishment IDs change with
  reorganizations and ID renumbering; an apparent open/close can be an
  administrative artifact. Use the FDZ establishment-history flags.
- **Output is disclosure-reviewed and cannot be redistributed.** Results leave the
  FDZ only after review, and the microdata itself cannot be shared.

## Citation

Cite the producer and access route, e.g.: *Integrated Employment Biographies
(IEB), Institute for Employment Research (IAB), accessed via the FDZ under
approved project, YYYY-MM-DD.* State the years, the spell-to-panel rules, and the
treatment of the wage ceiling.


==============================================================================
# Indifi FinTech loan application records (India, restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/indifi-loan-applications/
# Loan-level application records from one Indian FinTech lender (Indifi): applications with payment-transaction history, applicant characteristics, credit-bureau data, and outcomes. It is confidential single-counterparty data. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: fintech, lending, credit, india, data:indifi-loan-applications
==============================================================================

:::caution[Confidential single-counterparty data: not exercised here]
**This is confidential data from a single FinTech lender** (Indifi
Technologies). It is **not for sale, not publicly downloadable, and not
re-obtainable without the same agreement**: it carries **no provenance badge**
because there is no access path we can run here. Researchers obtained it under a
confidential agreement with the lender. The page documents what the data is and
the gotchas; treat it as unverified. This is the honest grade under the
institute's Verified discipline.
:::

**The Indifi data** is the loan-level application record of one Indian FinTech
lender: complete loan applications with applicant characteristics, the applicant's
**payment-transaction history** (digital payments flowing through the business),
merged credit-bureau scores, and loan outcomes. The payment history is the
distinguishing feature, because it lets the lender (and the researcher) see
cashless-payment flows as a credit signal. A paper we distill uses it:
[Ghosh, Vallee & Zeng](/wiki/papers/jf/2026/ghosh-fintech-lending-cashless-payments-2026/)
use the Indifi application data (316,719 complete applications, September 2015 to
November 2022, with payment records, applicant characteristics, credit-bureau
data, and outcomes) to study how cashless-payment data changes FinTech lending.

- **Cost:** not for sale. Confidential single-counterparty data.
- **Source:** Indifi Technologies (an Indian FinTech lender).
- **Coverage:** applicants to this one lender, not the Indian borrower population;
  loan-level applications and outcomes over the sample window.

## Access (restricted)

- **No public download and no resale.** The data is the lender's confidential
  application records; it cannot be purchased or redistributed.
- **Through a confidential agreement with the lender.** Access required a research
  agreement. There is no standing way for a third party to reach the same data.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **One lender is a selected applicant pool.** Applicants self-select into this
  FinTech and into its product (small-business lending with payment data); they are
  not the Indian credit market. Do not generalize approval or default patterns.
- **Outcomes are observed only for approved loans.** Repayment is seen for funded
  loans, so default analysis conditions on approval; correcting for the
  approval decision requires the application stage. Model selection into funding.
- **Payment history is platform-specific.** The cashless-payment signal reflects
  flows the lender observes, which depend on the merchant's payment setup; coverage
  of a borrower's total turnover is partial. Do not read it as full revenue.
- **Credit-bureau merge.** Bureau (Cibil) scores are merged in and follow Indian
  bureau definitions and history-length rules; the merge has match error. Verify
  the join.
- **Policy and platform changes over 2015 to 2022.** Demonetization, the growth of
  digital payments, and the lender's own model changes shift the data-generating
  process across the window. Control for regime and vintage.
- **Not redistributable.** Results can be reported but the microdata cannot be
  shared and cannot be re-pulled by others. Plan for non-reproducibility of the raw
  inputs.

## Citation

Cite the source and arrangement, e.g.: *Indifi loan-application records (Indifi
Technologies, India), confidential; used under agreement, YYYY-MM-DD.* State the
sample window, the approved-versus-applied scope, and the credit-bureau merge.


==============================================================================
# INSEE DADS: French matched employer-employee data (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/insee-dads/
# DADS is the French administrative matched employer-employee dataset: annual social declarations linking workers to establishments, with earnings, occupation, and hours. It is restricted administrative microdata reached through the CASD secure data centre. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: labor, matched-employer-employee, france, administrative, data:insee-dads
==============================================================================

:::caution[Restricted administrative microdata: not exercised here]
**DADS is confidential French administrative microdata** (INSEE). It is **not for
sale and not publicly downloadable**: it carries **no provenance badge** because
there is no access path we can run here. Researchers reach it only through an
approved project at the **CASD** (Centre d'acces securise aux donnees) secure
data centre, with output subject to statistical-confidentiality review. The page
documents what the collection is and the gotchas; treat it as unverified. This is
the honest grade under the institute's Verified discipline.
:::

**DADS** (Declarations Annuelles des Donnees Sociales) is the French
administrative matched employer-employee dataset built from the annual social
declarations that employers must file. It links **workers to establishments**
with earnings, occupation (the detailed PCS occupation codes), hours, and
contract information, covering essentially the universe of salaried employment.
It is the French analog of the matched employer-employee panels used in labor
and firm research. A paper we distill uses it:
[Beaumont, Hebert & Lyonnet](/wiki/papers/rfs/2025/beaumont-build-buy-human-capital-2025/)
use DADS for firm-level workforce composition by 4-digit occupation code (414
codes) to compute a human-capital distance between firms, alongside
[LIFI](/wiki/confidential/insee-lifi/) ownership links and
[firm tax files](/wiki/confidential/insee-tax-files/).

- **Cost:** not for sale. Restricted administrative microdata.
- **Producer:** INSEE (the French national statistical institute); access via
  CASD.
- **Coverage:** essentially the universe of salaried jobs in France over a long
  annual panel; establishment and worker identifiers allow matching across years.

## Access (restricted)

- **Through the CASD secure data centre.** Approved researchers connect to CASD
  via a dedicated secure device (the SD-Box); the data never leaves the secure
  environment, and outputs are checked before release.
- **Project approval required.** Access requires an approved research project and
  the relevant French confidentiality authorization; it is granted per project,
  not as an open download.
- INSEE publishes documentation of the DADS variables and concepts, so the schema
  can be read even though the data cannot be pulled.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Identifiers and their changes are the whole game.** Worker (NIR-derived) and
  establishment (SIRET) identifiers are pseudonymized and can change across
  vintages and reforms; a naive year-over-year join loses spells. Use the
  official panel identifiers and watch identifier breaks.
- **DADS has had major format changes.** The collection moved through the DADS
  "postes," DADS-U, and later DSN regimes; variable definitions and the unit of
  observation differ across these. Read the documentation for your exact years.
- **Occupation codes are detailed but revised.** The PCS occupation
  classification has versions; a 4-digit code is not stable across reclassifications.
  Pin the classification vintage when building occupation-based measures.
- **Establishment, not firm.** The base unit is the establishment (SIRET); rolling
  up to the firm (SIREN) requires the establishment-to-firm mapping and care with
  multi-establishment firms. Do not treat an establishment as a firm.
- **Censoring and truncation in earnings.** Earnings concepts and any top/bottom
  treatment follow the declaration rules; do not assume a clean uncensored wage.
- **Output is disclosure-reviewed and cannot be redistributed.** Results leave
  CASD only after confidentiality checks, and the microdata itself cannot be
  shared.

## Citation

Cite the producer and access route, e.g.: *DADS (INSEE), accessed via the CASD
secure data centre under approved project, YYYY-MM-DD.* State the years, the unit
(establishment or firm), and the occupation-classification vintage.


==============================================================================
# INSEE LIFI: French inter-firm ownership links (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/insee-lifi/
# LIFI is INSEE's administrative determination of business-group structure in France: which firms control which, used to assemble corporate groups from ownership links. It is restricted administrative microdata reached through the CASD secure data centre. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: ownership, business-groups, france, administrative, data:insee-lifi
==============================================================================

:::caution[Restricted administrative microdata: not exercised here]
**LIFI is confidential French administrative microdata** (INSEE). It is **not for
sale and not publicly downloadable**: it carries **no provenance badge** because
there is no access path we can run here. Researchers reach it only through an
approved project at the **CASD** secure data centre, with output subject to
statistical-confidentiality review. The page documents what the collection is and
the gotchas; treat it as unverified. This is the honest grade under the
institute's Verified discipline.
:::

**LIFI** (Enquete sur les Liaisons Financieres entre societes) is INSEE's
determination of the **financial links between firms** in France: who owns and
controls whom, used to reconstruct business-group (groupe) structure and to
assign each legal unit to a controlling group. It is the standard source for
defining French corporate groups and parent-subsidiary relationships. A paper we
distill uses it:
[Beaumont, Hebert & Lyonnet](/wiki/papers/rfs/2025/beaumont-build-buy-human-capital-2025/)
use LIFI to identify business-group structure and subsidiaries and to link firms
to their M&A targets, alongside [DADS](/wiki/confidential/insee-dads/) workforce data
and [firm tax files](/wiki/confidential/insee-tax-files/).

- **Cost:** not for sale. Restricted administrative microdata.
- **Producer:** INSEE; access via CASD.
- **Coverage:** ownership and control links among French firms, used to build
  group perimeters; annual, with the linked-unit population evolving over time.

## Access (restricted)

- **Through the CASD secure data centre.** Approved researchers connect to CASD
  via the secure SD-Box; the data never leaves the secure environment and outputs
  are checked before release.
- **Project approval required.** Access requires an approved research project and
  the relevant French confidentiality authorization, granted per project.
- INSEE publishes documentation of the LIFI concepts (control thresholds, group
  definitions), so the methodology can be read even though the data cannot be
  pulled.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **A "group" is a constructed perimeter, not a raw field.** LIFI assembles
  groups from ownership and control links using thresholds; the group boundary
  depends on those rules, and they have evolved. Use the LIFI-defined perimeter
  rather than inventing your own from raw stakes.
- **Control versus ownership.** The link concept is about control, which can
  differ from cash-flow ownership; do not read a control link as a precise equity
  share.
- **Coverage thresholds and survey design.** LIFI combines a survey with
  administrative sources and has size and reporting thresholds; small links can be
  missing. Do not read an absent link as proof of independence.
- **Identifier joins.** Firms are keyed by SIREN; linking LIFI to
  [DADS](/wiki/confidential/insee-dads/) (SIRET establishments) and to
  [tax files](/wiki/confidential/insee-tax-files/) requires the SIREN/SIRET mapping
  and care with restructurings. Verify the join.
- **Time consistency.** Group perimeters change year to year as links are added or
  dropped; a firm can enter and leave a group for reporting reasons. Build
  point-in-time perimeters.
- **Output is disclosure-reviewed and cannot be redistributed.** Results leave
  CASD only after confidentiality checks, and the microdata itself cannot be
  shared.

## Citation

Cite the producer and access route, e.g.: *LIFI (INSEE), accessed via the CASD
secure data centre under approved project, YYYY-MM-DD.* State the years and the
group-perimeter definition used.


==============================================================================
# INSEE firm tax and accounting files (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/insee-tax-files/
# The French firm tax and accounting files (the BIC/FICUS-FARE lineage) are administrative firm-level balance sheets and income statements compiled by INSEE from DGFiP tax filings. They are restricted administrative microdata reached through the CASD secure data centre. This page documents what they are and the gotchas, but they were not exercised here.
# Tags: accounting, firm-financials, france, administrative, data:insee-tax-files
==============================================================================

:::caution[Restricted administrative microdata: not exercised here]
**The French firm tax/accounting files are confidential administrative microdata**
(INSEE, from DGFiP tax filings). They are **not for sale and not publicly
downloadable**: they carry **no provenance badge** because there is no access
path we can run here. Researchers reach them only through an approved project at
the **CASD** secure data centre, with output subject to statistical-confidentiality
review. The page documents what the collection is and the gotchas; treat it as
unverified. This is the honest grade under the institute's Verified discipline.
:::

**The French firm tax and accounting files** are the administrative firm-level
financial statements that INSEE compiles from corporate tax filings collected by
the DGFiP (the French tax authority). In the **BIC** (Benefices Industriels et
Commerciaux) regime and its INSEE products (historically BRN/FICUS, later FARE),
they provide balance sheets, income statements, and standard accounting items for
essentially the universe of French firms. They are the French administrative
analog of a firm-fundamentals database, with near-complete coverage rather than a
listed-firm sample. A paper we distill uses them:
[Beaumont, Hebert & Lyonnet](/wiki/papers/rfs/2025/beaumont-build-buy-human-capital-2025/)
use the French tax files (BIC) for subsidiary-level balance sheets and income
statements as firm-level controls, alongside [DADS](/wiki/confidential/insee-dads/)
workforce data and [LIFI](/wiki/confidential/insee-lifi/) ownership links.

- **Cost:** not for sale. Restricted administrative microdata.
- **Producer:** INSEE, compiled from DGFiP tax filings; access via CASD.
- **Coverage:** essentially all French firms filing the relevant tax regimes,
  annual; the product name and exact scope change across the FICUS/FARE vintages.

## Access (restricted)

- **Through the CASD secure data centre.** Approved researchers connect via the
  secure SD-Box; the data never leaves the secure environment and outputs are
  checked before release.
- **Project approval required.** Access requires an approved research project and
  the relevant French confidentiality authorization, granted per project.
- INSEE publishes documentation of the accounting variables and the
  FICUS-to-FARE transition, so the schema can be read even though the data cannot
  be pulled.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **A regime break sits in the middle of the panel.** INSEE switched from the
  FICUS product to FARE around the early 2010s, with definitional and
  source changes; a series spanning the break needs explicit reconciliation. Do
  not splice naively.
- **Tax regime determines who appears.** Firms in different tax regimes (BIC
  versus others, micro-regimes, simplified filings) report different items or are
  absent; the sample depends on the regime, not on economic size alone. Check the
  regime for your population.
- **Legal unit, not group.** Statements are at the legal-unit (SIREN) level; to
  study a group you must consolidate using [LIFI](/wiki/confidential/insee-lifi/), or
  you will double-count intra-group flows. Decide consolidation explicitly.
- **Accounting concepts are French GAAP.** Items follow the French chart of
  accounts; mapping to Compustat-style fields is not one-to-one. Map carefully
  before any cross-country comparison.
- **Identifier joins.** Linking to [DADS](/wiki/confidential/insee-dads/) (SIRET) and
  LIFI (SIREN) requires the SIREN/SIRET mapping and care with restructurings.
  Verify the join.
- **Output is disclosure-reviewed and cannot be redistributed.** Results leave
  CASD only after confidentiality checks, and the microdata itself cannot be
  shared.

## Citation

Cite the producer and access route, e.g.: *French firm tax/accounting files
(INSEE, from DGFiP), accessed via the CASD secure data centre under approved
project, YYYY-MM-DD.* State the product (for example FICUS or FARE), the years,
and the consolidation level.


==============================================================================
# LISA: Swedish longitudinal population register (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/lisa-sweden/
# LISA is Statistics Sweden's individual-level longitudinal register covering the entire resident population: annual labor-market, income, transfer, education, and family records, with the Wealth Register accessed under the same terms. It is restricted administrative microdata. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: labor, income, wealth, population-register, sweden, administrative, data:lisa-sweden
==============================================================================

:::caution[Restricted administrative microdata: not exercised here]
**LISA is confidential Swedish administrative microdata** (Statistics Sweden /
SCB). It is **not for sale and not publicly downloadable**: it carries **no
provenance badge** because there is no access path we can run here. Researchers
reach it only through an approved project with an ethics approval, accessed
remotely in SCB's secure environment (MONA), with output subject to
confidentiality review. The page documents what the collection is and the
gotchas; treat it as unverified. This is the honest grade under the institute's
Verified discipline.
:::

**LISA** (the longitudinal integrated database for health insurance and
labor-market studies) is Statistics Sweden's register linking **every resident**
aged 16 and over across years (the lower bound was later extended to include
15-year-olds), with annual records on employment, wages, unemployment, transfers,
family structure, and education. Combined with the **Wealth Register** (accessed
under the same terms, covering 1999 to 2007, when the wealth tax was abolished),
it supports population-wide labor and household-finance research. A paper we
distill uses it:
[Olsson & Tag](/wiki/papers/jf/2025/olsson-what-cost-privatization-workers-2025/)
use LISA for individual-level wages, unemployment, transfers, family structure,
education, and wealth for all Swedish residents (the paper states aged 15 and
over), 1990 to 2017, alongside the firm-level [FEK](/wiki/confidential/fek-sweden/)
statistics.

- **Cost:** not for sale. Restricted administrative microdata.
- **Producer:** Statistics Sweden (SCB).
- **Coverage:** the full resident population on an annual panel, with linkable
  individual, family, firm, and establishment identifiers; the Wealth Register
  component ends in 2007.

## Access (restricted)

- **Through SCB's secure remote environment (MONA).** Approved researchers
  access pseudonymized extracts remotely; the microdata does not leave SCB, and
  outputs are checked before release.
- **Project and ethics approval required.** Access requires an approved research
  project and a Swedish ethics review; data is delivered as a project-specific
  extract, not an open download.
- SCB publishes documentation of the LISA and Wealth Register variables, so the
  schema can be read even though the data cannot be pulled.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **The Wealth Register ends in 2007.** Sweden abolished the wealth tax effective
  2007, so the register covers 1999 to 2007 and individual wealth, risky assets,
  and debt are not available afterward; any later wealth analysis must use a
  different source or stop there. Do not extrapolate wealth past the register's
  end, and mind the income-year versus file-year distinction when pinning the last
  usable year.
- **Pseudonymized linkage keys, project-specific.** Individual and firm keys are
  pseudonyms created per delivery; you cannot merge across separately delivered
  extracts, and keys are not comparable to other projects. Plan all links inside
  one delivery.
- **Annual snapshots miss within-year dynamics.** LISA records an annual
  position (often a reference week or year-end); job changes and spells within
  the year are coarsened. Match your timing to the annual structure.
- **Variable definitions and the population frame change.** Income, employment,
  and education concepts have been revised, and the lower age cutoff and frame
  have changed over time; a long panel is not perfectly consistent. Read the
  documentation for your years.
- **Firm/establishment links need the right registers.** Linking individuals to
  employers requires the establishment identifier and the matched registers;
  joining to [FEK](/wiki/confidential/fek-sweden/) firm data must respect the
  firm-versus-establishment distinction.
- **Output is disclosure-reviewed and cannot be redistributed.** Results leave the
  secure environment only after review, and the microdata itself cannot be shared.

## Citation

Cite the producer and access route, e.g.: *LISA and the Wealth Register
(Statistics Sweden), accessed via MONA under approved project, YYYY-MM-DD.* State
the years, the population frame, and whether the Wealth Register (through 2007)
was used.


==============================================================================
# NMLS Mortgage Call Report (company-level, restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/mcr-nmls/
# The NMLS Mortgage Call Report collects loan-origination and financial-condition data from state-licensed mortgage companies. Aggregate statistics are published; the company-level data used in research is restricted. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: mortgages, nonbank-lenders, lending, regulatory, data:mcr-nmls
==============================================================================

:::caution[Restricted regulatory data: not exercised here]
**The company-level Mortgage Call Report is restricted.** It is collected through
the Nationwide Multistate Licensing System (NMLS) by state regulators and the
Conference of State Bank Supervisors (CSBS). Aggregate statistics are published,
but the **company-level** records used in research carry **no provenance badge**:
there is no open access path we can run here. They are obtained under restricted
regulatory terms (a research agreement with the regulator/CSBS). The page
documents what the collection is and the gotchas; treat it as unverified. This is
the honest grade under the institute's Verified discipline.
:::

**The Mortgage Call Report (MCR)** is the standardized report that state-licensed
mortgage companies (most importantly **nonbank lenders**) file through the
**NMLS**. It contains residential mortgage loan-origination activity (application,
closing, and pipeline volumes by loan type and state) and a financial-condition
component (income, expenses, assets, and liabilities of the licensed company). It
is a leading source for studying nonbank mortgage lenders, which are largely
outside the bank Call Report system. A paper we distill uses it:
[Bhutta, Fuster & Hizmo](/wiki/papers/jf/2026/bhutta-mortgage-overpayment-borrower-sophistication-2026/)
use company-level MCR data (2015:Q1 to 2019:Q4, 162 unique lenders) for nonbank
lender income, expenses, and profitability, merged with Optimal Blue to study how
rate premiums translate into lender margins and how competition moderates
borrower overpayment.

- **Cost:** aggregate statistics are published free; the company-level data is
  obtained under a restricted regulatory arrangement.
- **Collector:** state regulators via the NMLS, administered by the Conference of
  State Bank Supervisors (CSBS).
- **Coverage:** state-licensed mortgage companies (nonbanks and some others);
  origination and financial-condition components, quarterly. Depository
  institutions report mortgage data elsewhere, so the MCR is mainly a nonbank
  view.

## Access (restricted)

- **Aggregates are public; company-level data is not.** The NMLS publishes
  aggregate Mortgage Call Report statistics; the company-level records are not
  openly downloadable.
- **Company-level data is obtained under a regulatory arrangement.** Research use
  of the company-level MCR is granted under restricted terms by the relevant
  regulator or CSBS; it is not a commercial product and not an open download.
- The reporting **form and instructions** are public, so the field definitions
  can be read even though the company-level data cannot be pulled openly.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **It is a licensee view, not all mortgage lending.** Depository institutions
  report mortgage activity through other channels; the MCR is mainly state-
  licensed nonbanks. Do not read it as the whole mortgage market.
- **Two components with different bases.** The residential-activity component
  (loan volumes) and the financial-condition component (company financials) are
  reported on different bases; do not mix a flow line with a balance-sheet line
  without checking definitions.
- **State-by-state reporting and licensing.** A company licensed in many states
  reports activity by state; aggregation and de-duplication across state lines is
  required to get a clean company total. Watch for double counting.
- **Company identity and entity changes.** Licensed entities merge, rebrand, and
  surrender licenses; the company identifier is not stable across reorganizations.
  Track entity changes when building a panel.
- **Self-reported financials.** The financial-condition data is self-reported by
  licensees and is not audited to bank-Call-Report standards; treat profitability
  measures with appropriate caution.
- **Restricted use.** Company-level results are governed by the regulatory
  agreement; plan for the disclosure and use restrictions of that arrangement.

## Citation

Cite the collection and administrator, e.g.: *NMLS Mortgage Call Report
(company-level), Conference of State Bank Supervisors / state regulators;
obtained under restricted research arrangement, YYYY-MM-DD.* For the public
series, cite the NMLS aggregate Mortgage Call Report statistics. State the
component (activity or financial condition), the sample window, and the lender
universe.


==============================================================================
# NIC supervisory data: CAMELS ratings and BHC structure (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/nic-fed/
# The Federal Reserve's National Information Center holds confidential CAMELS supervisory ratings alongside public bank holding company structure and ownership history. The ratings are confidential; the structure data is public. This page documents what it is and the gotchas, but the ratings were not exercised here.
# Tags: banks, supervision, bank-structure, banking, federal-reserve, data:nic-fed
==============================================================================

:::caution[Mixed: structure public, CAMELS ratings confidential]
**The NIC structure and ownership data is public**, but the **CAMELS supervisory
ratings** held alongside it are confidential. This page carries **no provenance
badge**: the ratings have no access path we can run here, and the public
structure data was not exercised in this session. CAMELS ratings are reached only
through a Federal Reserve, FDIC, or OCC affiliation or an approved restricted-data
arrangement, inside a secure environment, with output subject to disclosure
review. The page documents what the collection is and the gotchas; treat it as
unverified. This is the honest grade under the institute's Verified discipline.
:::

**The National Information Center (NIC)** is the Federal Reserve's repository of
data on banking organizations. Its **structure** side is public: the
organizational hierarchy, ownership, and history of bank holding companies
(formations, mergers, RSSD identifiers, listing history). Held alongside it,
**CAMELS** composite and component supervisory ratings (Capital, Asset quality,
Management, Earnings, Liquidity, Sensitivity, on a 1 to 5 scale) are confidential
supervisory assessments. A paper we distill uses both:
[Falato & Scharfstein](/wiki/papers/jf/2025/falato-stock-market-bank-risk-2025/)
use confidential CAMELS composite and component ratings as the bank-risk outcome,
with NIC ownership and listing history to identify which BHCs are publicly listed.

- **Cost:** structure data is free and public; CAMELS ratings are not for sale.
- **Source:** Federal Reserve (the National Information Center).
- **Coverage:** banking organizations supervised in the U.S.; structure history
  is broad, CAMELS ratings cover examined institutions over their exam cycle.

## Access (restricted)

- **Structure data is public.** The NIC website exposes institution search,
  organizational hierarchies, and RSSD identifiers; this is downloadable.
- **CAMELS ratings are confidential.** Access to supervisory ratings is limited
  to researchers at the Federal Reserve System, the FDIC, or the OCC, or others
  granted entry to the restricted data, worked inside a secure environment with
  output subject to disclosure review.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **CAMELS is assigned on an exam cycle, not continuously.** A rating is set at
  an examination and carried until the next exam; the timestamp that matters is
  the exam date, not the report date. Do not treat ratings as a smooth monthly or
  quarterly series.
- **Ratings are confidential and their disclosure is restricted.** A bank's CAMELS
  rating is non-public supervisory information; results using it are disclosure-
  reviewed and the rating itself cannot be published at the bank level.
- **Composite versus component.** The composite rating and the six components
  carry different information; using the composite alone discards component-level
  signal (and vice versa). Be explicit about which you use.
- **Lead-supervisor and charter-type differences.** Banks are supervised by the
  Fed, OCC, FDIC, or state regulators, and rating practices can differ; pooling
  across supervisors mixes assessment styles. Note the supervisor.
- **Linking structure to ratings and to market data.** RSSD identifiers must be
  linked across the BHC hierarchy and to market data (for listed parents);
  mergers and reorganizations break the link. Verify the crosswalk and the
  parent-subsidiary mapping.
- **Output is disclosure-reviewed and cannot be redistributed.** Rating-level
  results leave the secure environment only after review.

## Citation

Cite the source, e.g.: *National Information Center (Federal Reserve): structure
data, accessed YYYY-MM-DD; CAMELS supervisory ratings, confidential, accessed
under restricted-data arrangement.* State the rating (composite or component),
the exam-date basis, and the sample window.


==============================================================================
# German online broker retail investor data (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/online-broker/
# Individual-level holdings, trades, and returns for retail investors at one anonymous German online broker, used in household-finance research. It is confidential single-counterparty data. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: household-finance, retail-investors, portfolios, germany, data:online-broker
==============================================================================

:::caution[Confidential single-counterparty data: not exercised here]
**This is confidential data from a single online broker.** It is **not for sale,
not publicly downloadable, and not re-obtainable without the same agreement**: it
carries **no provenance badge** because there is no access path we can run here.
Researchers obtained it under a confidential agreement with the broker, which is
kept anonymous. The page documents what the data is and the gotchas; treat it as
unverified. This is the honest grade under the institute's Verified discipline.
:::

**The online-broker data** is the account-level record of retail investors at one
anonymous German online brokerage: security holdings, trades, portfolio
characteristics, and returns over a multi-year window. It is one of the
recurring sources for German retail-investor behavior, valued because it observes
actual holdings and trades rather than survey-reported intentions. A paper we
distill uses it:
[Laudenbach, Malmendier & Niessen-Ruenzi](/wiki/papers/jf/2026/laudenbach-communism-attitudes-2026/)
use the online-broker data (839,292 investor-year observations for 230,229 retail
investors, 2004 to 2012; holdings, portfolio characteristics, and returns, with
[Datastream](/wiki/commercial/datastream/) returns merged in) to study how growing
up under communism shapes financial attitudes.

- **Cost:** not for sale. Confidential single-counterparty data.
- **Source:** an anonymous German online broker.
- **Coverage:** the broker's own clients, not the German population; account-level
  holdings, trades, and returns over the sample window.

## Access (restricted)

- **No public download and no resale.** The data is the broker's confidential
  client records; it cannot be purchased or redistributed.
- **Through a confidential agreement with the broker.** Access required a research
  agreement, and the broker's identity is not disclosed. There is no standing way
  for a third party to reach the same data.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **One broker is a selected clientele.** Online-broker clients self-select on
  being active, online, and direct investors; they are not representative of all
  German households or investors. Do not generalize portfolio behavior to the
  population.
- **Holdings at this broker, not total wealth.** The account shows assets held at
  this broker, not the investor's whole portfolio or non-traded wealth; apparent
  concentration or risk-taking can be an artifact of partial coverage. Treat it as
  one account.
- **Returns need an external price source.** Position-level returns are computed by
  merging an external price/return feed (here [Datastream](/wiki/commercial/datastream/));
  corporate actions and mergers must be handled or returns are mismeasured.
- **Demographics are limited and sometimes imputed.** Investor characteristics are
  thinner than in administrative registers and may be merged from municipality-level
  sources; do not treat imputed demographics as individual ground truth.
- **Not redistributable and broker stays anonymous.** Results can be reported but
  the microdata cannot be shared and the broker is not named, limiting external
  replication. Plan for non-reproducibility of the raw inputs.

## Citation

Cite the source and arrangement, e.g.: *Retail online-broker account data
(anonymous German broker), confidential; used under agreement, YYYY-MM-DD.* State
the sample window, the unit (investor-year), and the external return source used.


==============================================================================
# OSFI federally regulated lender data (Canada, restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/osfi-canada/
# Contract-level mortgage records for federally regulated Canadian lenders, collected by OSFI and reached through the Bank of Canada: lender identity, loan size, rate, amortization, LTV, and debt-service ratio. It is restricted supervisory data. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: mortgages, banks, supervision, canada, data:osfi-canada
==============================================================================

:::caution[Restricted supervisory data: not exercised here]
**OSFI federally regulated lender data is confidential supervisory data.** It is
**not for sale and not publicly downloadable**: it carries **no provenance badge**
because there is no access path we can run here. Researchers reach it only through
an approved restricted-data arrangement (in practice through the Bank of Canada),
inside a secure environment, with output subject to disclosure review. The page
documents what the collection is and the gotchas; treat it as unverified. This is
the honest grade under the institute's Verified discipline.
:::

**OSFI federally regulated lender data** is the contract-level mortgage reporting
that federally regulated Canadian lenders submit to the **Office of the
Superintendent of Financial Institutions (OSFI)**, the prudential supervisor of
federal banks and insurers. Each record carries lender identity, loan size,
funding date, monthly payment, outstanding balance, mortgage rate, amortization,
loan-to-value (LTV), and total debt-service ratio. It is a near-complete view of
new federally regulated mortgage lending, used for studies of mortgage pricing and
competition. A paper we distill uses it:
[Allen & Li](/wiki/papers/jf/2025/allen-dynamic-competition-negotiated-price-2025/)
use the OSFI contract-level mortgage data (lender identity, loan size, rate,
amortization, LTV, debt-service ratio) together with Canadian credit-bureau
records to study dynamic competition in negotiated mortgage pricing.

- **Cost:** not for sale. Restricted-access confidential supervisory data.
- **Collector:** OSFI; research access in practice through the Bank of Canada.
- **Coverage:** mortgages originated by federally regulated lenders;
  provincially regulated lenders and credit unions are outside the federal
  perimeter.

## Access (restricted)

- **No public download.** The contract-level microdata is confidential and is not
  posted.
- **Through an approved arrangement, in practice via the Bank of Canada.** Access
  is limited to researchers granted entry to the restricted data, worked inside a
  secure environment with output subject to disclosure review.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Federally regulated only, so it is not all mortgages.** Provincially regulated
  lenders, credit unions, and some private lenders are outside the OSFI perimeter;
  the data is the federal slice, not the whole Canadian mortgage market. Do not
  generalize to all lenders.
- **New originations, not the full stock.** The reporting is oriented to newly
  funded mortgages and their terms; it is not a complete outstanding-stock panel.
  Match the new-origination scope to your question.
- **Debt-service and LTV are reported metrics.** TDS/GDS ratios and LTV follow
  the reporting and underwriting definitions, which interact with regulatory
  guidelines (for example the B-20 stress-test rules) that changed over the
  period; a ratio is not a fixed concept across years. Read the definition for
  your window.
- **Lender identity needs linking.** Joining to credit-bureau records and to
  house-price or demographic data requires borrower- and lender-level crosswalks
  under pseudonyms; plan links inside one secure delivery.
- **Regulatory regime breaks.** Mortgage-insurance rules and qualifying-rate
  guidelines shifted over the sample; loan terms respond to policy, not only to
  market conditions. Control for the regulatory regime.
- **Output is disclosure-reviewed and cannot be redistributed.** Results leave the
  secure environment only after review, and the microdata itself cannot be shared.

## Citation

Cite the source and access route, e.g.: *OSFI federally regulated lender mortgage
data (Canada), confidential; accessed via the Bank of Canada under restricted-data
arrangement, YYYY-MM-DD.* State the origination window, the LTV/debt-service
definitions, and the regulatory regime in force.


==============================================================================
# SLI private meeting notes and fund records (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/sli-private-meetings/
# Internal records of one asset manager (Standard Life Investments / abrdn): private-meeting notes, analyst ratings and recommendations, fund holdings, and daily trades. It is confidential single-counterparty data. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: asset-management, private-meetings, text, holdings, data:sli-private-meetings
==============================================================================

:::caution[Confidential single-counterparty data: not exercised here]
**This is confidential internal data from a single asset manager** (Standard
Life Investments / abrdn). It is **not for sale, not publicly downloadable, and
not re-obtainable without the same agreement**: it carries **no provenance
badge** because there is no access path we can run here. Researchers obtained it
under a confidential agreement with the firm. The page documents what the data
is and the gotchas; treat it as unverified. This is the honest grade under the
institute's Verified discipline.
:::

**The SLI data** is the internal record of one large asset manager: notes from
**private meetings** between the manager's analysts and the companies it invests
in, internal analyst ratings and Buy/Hold/Sell recommendations, fund holdings,
and daily trades. The meeting notes are the distinguishing asset, because they
capture private-access information flow that is otherwise unobservable. A paper we
distill uses it:
[Becht, Franks & Wagner](/wiki/papers/jf/2026/becht-private-meetings-portfolio-firms-2026/)
use 4,700 SLI meeting notes (with attendee info and analyst ratings), daily
fund holdings and trades across roughly 40 to 50 funds covering FTSE All-Share
stocks (2007 to 2015), and internal recommendations to study the value of private
access to portfolio firms.

- **Cost:** not for sale. Confidential single-counterparty data.
- **Source:** Standard Life Investments / abrdn (one asset manager).
- **Coverage:** this manager's own meetings, ratings, holdings, and trades over
  the sample window, focused on its UK (FTSE All-Share) investment universe.

## Access (restricted)

- **No public download and no resale.** The records are the firm's confidential
  internal documents; they cannot be purchased or redistributed.
- **Through a confidential agreement with the firm.** Access required a research
  agreement. There is no standing way for a third party to reach the same data.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **One manager is not the industry.** The notes, ratings, and trades reflect a
  single firm's process, universe, and house style; external validity to other
  managers is limited. Do not generalize the value of access from one firm.
- **Meeting notes are unstructured and analyst-written.** Notes vary in length,
  detail, and candor by author and over time; text measures (tone, content) inherit
  that heterogeneity. Validate any text classification against the raw notes.
- **Which meetings are recorded is selective.** A note exists when the analyst held
  and logged a meeting; absence is not evidence of no contact, and meeting selection
  is endogenous to interest in the stock. Treat note presence as a choice.
- **Holdings and trades need careful alignment to notes.** Linking a meeting to
  subsequent trades requires precise timestamps and a position panel; the
  10-million-plus fund-stock-day structure demands clustering. Align timing
  deliberately.
- **MNPI and disclosure constraints.** The notes can contain material non-public
  information; the firm's sharing is bounded and results are constrained
  accordingly. Respect the disclosure limits in what is reported.
- **Not redistributable.** Results can be reported but the microdata cannot be
  shared and cannot be re-pulled by others. Plan for non-reproducibility of the raw
  inputs.

## Citation

Cite the source and arrangement, e.g.: *SLI private-meeting notes and fund records
(Standard Life Investments / abrdn), confidential; used under agreement,
YYYY-MM-DD.* State the sample window, the note count, and the fund universe.


==============================================================================
# Spain CIR: Central de Informacion de Riesgos credit register (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/spain-cir/
# The CIR is the Banco de Espana's confidential loan-level credit register covering corporate loans by Spanish banks, with bank supervisory data matched to it. It is confidential supervisory data. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: credit-register, banks, credit, spain, central-bank, data:spain-cir
==============================================================================

:::caution[Confidential supervisory data: not exercised here]
**The CIR is confidential supervisory data** held by the Banco de Espana. It is
**not for sale and not publicly downloadable**: it carries **no provenance badge**
because there is no access path we can run here. Researchers reach it only through
a Banco de Espana affiliation or an approved restricted-data arrangement, inside
a secure environment, with output subject to disclosure review. The page documents
what the collection is and the gotchas; treat it as unverified. This is the honest
grade under the institute's Verified discipline.
:::

**The CIR** (Central de Informacion de Riesgos) is the Banco de Espana's credit
register: a near-complete loan-level record of corporate lending by Spanish banks,
with credit volumes, maturities, and default/non-performance status, matched to
**bank-level supervisory data** (balance sheet, capital, liquidity, ROA, NPL
ratios). It is one of the longest-running credit registers and a workhorse for
Spanish credit-supply and monetary-transmission research. A paper we distill uses
it:
[Jimenez, Kuvshinov, Peydro & Richter](/wiki/papers/jf/2026/jimenez-monetary-policy-inflation-crises-2026/)
use the CIR for loan-level monthly data on all corporate loans by Spanish banks,
1984 to 2008Q3 (a 10% random sample), with credit volumes, maturities, and
defaults, matched to Banco de Espana bank supervisory characteristics.

- **Cost:** not for sale. Restricted-access confidential supervisory data.
- **Collector:** Banco de Espana (the CIR; matched supervisory reports).
- **Coverage:** corporate loans by Spanish banks above the reporting threshold,
  monthly, with a long history; the threshold and scope have changed over time.

## Access (restricted)

- **No public download.** The microdata is confidential and is not posted.
- **Through a Banco de Espana affiliation or approved program.** Access is limited
  to researchers at the Banco de Espana or others granted entry to the restricted
  data, worked inside a secure environment with output subject to disclosure
  review.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **The reporting threshold censors small loans.** Only loans above the threshold
  are reported, and it has changed over the register's life; a borrower can enter
  or leave for reporting reasons. Do not read a missing loan as no lending.
- **A random sample is sometimes used, not the full register.** Some projects (and
  the distilled paper) draw a random sample (for example 10%) for tractability;
  sampling changes the standard errors and the multi-bank-firm structure. Be
  explicit about whether you use the full register or a sample.
- **Drawn versus committed and default timing.** Credit drawn, credit committed,
  and the date a loan is flagged non-performing are distinct; using the wrong one
  mismeasures supply or risk. Match the concept and the default definition to your
  question.
- **Identifier joins and bank mergers.** Firms and banks are linked across the CIR,
  the supervisory data, and firm-financials sources (the Mercantile Register);
  Spain's heavy bank consolidation breaks bank identifiers over time. Verify the
  crosswalk and track mergers.
- **Multi-bank firms drive the identification.** The register's value is the
  within-firm cross-bank variation; single-bank firms cannot be used for
  borrower-fixed-effect designs. Keep the multi-bank structure in mind.
- **Output is disclosure-reviewed and cannot be redistributed.** Results leave the
  secure environment only after review, and the microdata itself cannot be shared.

## Citation

Cite the source, e.g.: *Central de Informacion de Riesgos (Banco de Espana),
confidential supervisory data; accessed under restricted-data arrangement,
YYYY-MM-DD.* State the sample window, whether the full register or a random sample
was used, and the default/non-performance definition.


==============================================================================
# Statistics Norway administrative registers (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/statistics-norway/
# Statistics Norway (SSB) maintains linked individual- and firm-level administrative registers: demographics, income, wealth, and balance sheets from tax records. Aggregate tables are public; the linked microdata is restricted. This page documents what it is and the gotchas, but the microdata was not exercised here.
# Tags: income, wealth, population-register, norway, administrative, data:statistics-norway
==============================================================================

:::caution[Mixed: aggregates public, microdata restricted]
**Statistics Norway publishes aggregate tables publicly**, but the **linked
individual- and firm-level microdata is restricted**. This page carries **no
provenance badge**: the microdata has no access path we can run here, and the
public aggregates were not exercised in this session. Microdata is reached only
through an approved project in SSB's secure environment (the microdata access
service), with output subject to confidentiality review. The page documents what
the registers are and the gotchas; treat it as unverified. This is the honest
grade under the institute's Verified discipline.
:::

**Statistics Norway** (Statistisk sentralbyra, SSB) maintains the Norwegian
administrative registers: population demographics, employment, income, and the
tax-based household balance sheet (financial assets, real assets, and debt), all
linkable at the individual level through national identifiers and to firms. The
tax-record wealth data is especially detailed because Norway levies a wealth tax,
so household balance sheets are reported annually. A paper we distill uses it:
[Betermier, Calvet, Knupfer & Kvaerner](/wiki/papers/jf/2025/betermier-investor-factors-2025/)
use SSB tax records for investor demographics, balance sheets, income, and wealth
(annual, 1997 to 2017), linked to individual stockholdings from
[VPS](/wiki/confidential/vps-norway/).

- **Cost:** aggregate tables are free; the linked microdata is not for sale.
- **Producer:** Statistics Norway (SSB).
- **Coverage:** the full resident population and Norwegian firms on annual panels,
  with national identifiers enabling individual, household, and firm linkage.

## Access (restricted)

- **Aggregates are public; microdata is not.** SSB publishes statistics and tables
  openly; the linked individual/firm microdata is confidential.
- **Through SSB's secure microdata service.** Approved researchers access
  pseudonymized extracts in a secure environment; the microdata does not leave
  SSB, and outputs are checked before release.
- **Project approval required.** Access requires an approved research project and
  the relevant authorization; data is a project-specific extract, not an open
  download.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Tax-record wealth is a year-end snapshot at assessed values.** Balance-sheet
  items are reported for tax purposes at year-end and at assessed (not always
  market) values; this affects real estate and unlisted assets especially. Do not
  treat assessed values as market values without adjustment.
- **Linkage keys are pseudonymized per project.** Individual and firm keys are
  pseudonyms tied to a delivery; you cannot merge across separately delivered
  extracts. Plan all links inside one delivery.
- **Household versus individual.** Many outcomes are individual but tax and wealth
  concepts can be household-level; mixing the two double-counts or misattributes.
  Fix the unit before aggregating.
- **Definitional changes and the wealth-tax regime.** Income and wealth concepts
  follow tax rules that change over time; a long panel is not perfectly
  consistent. Read the documentation for your years.
- **Coverage of assets.** Listed-equity and bank-reported items are well measured;
  some assets (private business, foreign holdings) are less complete. Do not read
  a missing asset as zero.
- **Output is disclosure-reviewed and cannot be redistributed.** Microdata
  results leave the secure environment only after review.

## Citation

Cite the producer and access route, e.g.: *Statistics Norway (SSB) administrative
registers, accessed via the SSB secure microdata service under approved project,
YYYY-MM-DD.* For public figures, cite the relevant SSB statistics. State the
years, the unit (individual or household), and the registers linked.


==============================================================================
# STBL: Survey of Terms of Business Lending (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/stbl-fed/
# The Federal Reserve's Survey of Terms of Business Lending collected loan-level commercial-and-industrial loan terms and internal risk ratings from reporting banks. Aggregates were published (E.2); the loan-level data is confidential. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: banks, credit, business-lending, banking, federal-reserve, data:stbl-fed
==============================================================================

:::caution[Confidential supervisory data: not exercised here]
**STBL loan-level data is confidential** and was collected by the Federal
Reserve. Published aggregates (the E.2 release) are public, but the
**loan-level** records carry **no provenance badge**: there is no access path we
can run here. Researchers reach the loan-level data only through a Federal
Reserve affiliation or an approved restricted-data arrangement, inside a secure
environment, with output subject to disclosure review. The page documents what
the collection is and the gotchas; treat it as unverified. This is the honest
grade under the institute's Verified discipline.
:::

**STBL** (the Survey of Terms of Business Lending) was a quarterly Federal
Reserve survey in which a sample of banks reported the terms of a sample of new
**commercial-and-industrial (C&I) loans** made during a survey week: amount,
rate, maturity, collateral, commitment status, and the bank's **internal risk
rating** (a 1 to 5 scale). It is a standard source for loan pricing and loan-risk
research. The survey ran from roughly 1977 to 2017 (when the Fed discontinued the
E.2 release); the internal risk-rating component was added in 1997, so the
risk-rating studies start there. A paper we distill uses it:
[Falato & Scharfstein](/wiki/papers/jf/2025/falato-stock-market-bank-risk-2025/)
use the STBL loan risk rating as a measure of the riskiness of new lending when
studying how going public affects bank risk-taking (their sample runs through
2012, the end of their broader panel, not the end of the survey).

- **Cost:** not for sale at the loan level. The E.2 aggregates were published
  free.
- **Collector:** Federal Reserve Board (the STBL collection; E.2 release).
- **Coverage:** a sample of banks reporting a sample of new C&I loans, quarterly.
  The survey ran from roughly 1977 to 2017; the internal risk rating is available
  from 1997 on. It is a sample of loans, not a census.

## Access (restricted)

- **The E.2 aggregates are public; the loan-level data is not.** The Fed
  published the aggregate Survey of Terms of Business Lending (E.2) release; the
  loan-level micro-records are confidential.
- **Through a Federal Reserve affiliation or approved program.** Loan-level
  access is limited to researchers at the Federal Reserve System or others
  granted entry to the restricted data, worked inside a secure environment with
  output subject to disclosure review.
- The survey was **discontinued in 2017**, so it covers a closed historical
  window; there is no ongoing collection.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **It is a sample of loans in a survey week, not all lending.** A reporting bank
  submits a sample of loans made during the survey week, so the data is a
  stratified sample, not the bank's full loan book. Use the survey weights; do
  not treat counts as totals.
- **Internal risk ratings are bank-specific.** The 1 to 5 risk rating is mapped
  from each bank's own internal grades, so the scale is not strictly comparable
  across banks. Treat cross-bank rating comparisons with care.
- **The survey ended in 2017 and the risk rating starts in 1997.** Risk-rating
  analysis is bounded below by 1997 (when the rating was added) and above by 2017
  (when the E.2 release was discontinued); there is no continuation past then. Do
  not extrapolate the series outside that window, and do not confuse a paper's own
  sample end with the survey's end.
- **New loans only, C&I only.** The survey captures terms of newly made C&I
  loans, not outstanding balances or other loan types. Match your question to the
  new-loan, C&I scope.
- **Definitions and the sampling frame changed.** Reporting panels and
  definitions were revised across the survey's life; a field can change basis
  between vintages. Read the instructions for your period.
- **Output is disclosure-reviewed and cannot be redistributed.** Loan-level
  results leave the secure environment only after review.

## Citation

Cite the collection and collector, e.g.: *Survey of Terms of Business Lending
(STBL), Federal Reserve Board, confidential loan-level data; accessed under
restricted-data arrangement, YYYY-MM-DD.* For the public series, cite the E.2
release. State the sample window (the risk rating is available from 1997; the
survey ended in 2017) and the use of survey weights.


==============================================================================
# TransUnion credit bureau data (Canada, restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/transunion-canada/
# Monthly, population-wide individual credit-bureau records for Canada from TransUnion: borrower characteristics, mortgage identity, switching activity, and inquiries. It is restricted research microdata, not an off-the-shelf purchase. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: consumer-credit, mortgages, credit-bureau, canada, data:transunion-canada
==============================================================================

:::caution[Restricted individual-level data: not exercised here]
**This is restricted TransUnion Canada credit-bureau microdata** (personally
identifiable in raw form). It is **not an off-the-shelf research purchase and not
publicly downloadable**: it carries **no provenance badge** because there is no
access path we can run here. Researchers reach it only under a restricted research
agreement, anonymized. The page documents what the data is and the gotchas; treat
it as unverified. This is the honest grade under the institute's Verified
discipline.
:::

**TransUnion Canada credit-bureau data** is the monthly, near population-wide
individual credit record for Canada: borrower characteristics (age, credit score,
location at the forward-sortation-area level, non-mortgage debt), mortgage
identity, switching activity between lenders, and credit inquiries. Its value for
research is the combination of full population coverage and the ability to follow
borrowers across lenders over time. A paper we distill uses it:
[Allen & Li](/wiki/papers/jf/2025/allen-dynamic-competition-negotiated-price-2025/)
use TransUnion monthly credit-bureau records for borrower characteristics,
mortgage identity, switching, and inquiries, linked to
[OSFI](/wiki/confidential/osfi-canada/) contract-level mortgage data, to study dynamic
competition in negotiated mortgage pricing.

- **Cost:** not an off-the-shelf research purchase. Restricted-access individual
  microdata.
- **Source:** TransUnion Canada.
- **Coverage:** near population-wide consumers with a credit file in Canada,
  monthly; location is typically at the forward-sortation-area (FSA) level, not the
  exact address.

## Access (restricted)

- **No public download; PII in raw form.** Raw records are personally identifiable
  and cannot be posted; research uses anonymized extracts.
- **Through a restricted research agreement.** Access is granted per project under
  strict privacy terms; there is no standing way for a third party to reach the
  same data.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Only the credit-visible population.** Consumers without a credit file are
  absent; the panel is not literally everyone. Do not generalize to the
  credit-invisible.
- **Geography is coarse (FSA).** Location is at the forward-sortation-area level,
  not the address; merging to fine-grained local data (house prices, demographics)
  inherits that coarseness. Match at the FSA level deliberately.
- **Scores and tradelines follow bureau definitions.** Credit scores and
  account/tradeline fields are TransUnion constructs that change over time and can
  differ from Equifax Canada; a score is not a fixed cross-vintage or cross-bureau
  concept. Pin the model and codebook.
- **Mortgage identity and switching need careful linking.** Identifying the same
  mortgage across months and a switch between lenders relies on the bureau's
  tradeline matching, which has error; verify the switching definition against the
  lender-side ([OSFI](/wiki/confidential/osfi-canada/)) data.
- **Reporting lags and updates.** Tradelines update on lender reporting cycles, so
  a month's snapshot can lag actual events; be explicit about the reporting-date
  basis.
- **Not redistributable.** Results can be reported but the microdata cannot be
  shared or re-pulled by others. Plan for non-reproducibility of the raw inputs.

## Citation

Cite the source and arrangement, e.g.: *TransUnion Canada credit-bureau records,
restricted; accessed under research agreement, YYYY-MM-DD.* State the sample
window, the geographic level (FSA), and the score/codebook vintage.


==============================================================================
# Technology service provider user list (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/tsp-user-list/
# A confidential list identifying which banks used a third-party technology service provider that was the target of a cyberattack, used as a treatment indicator. It is confidential single-source data. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: banks, cyber-risk, third-party-risk, payments, data:tsp-user-list
==============================================================================

:::caution[Confidential single-source data: not exercised here]
**This is a confidential list of bank-vendor relationships.** It is **not for sale
and not publicly downloadable**: it carries **no provenance badge** because there
is no access path we can run here. Identifying which banks used a particular
third-party technology service provider (TSP) is sensitive supervisory/operational
information, obtained under a confidential arrangement. The page documents what
the data is and the gotchas; treat it as unverified. This is the honest grade
under the institute's Verified discipline.
:::

**The TSP user list** identifies which banks were customers of a specific
third-party **technology service provider** that suffered a cyberattack. By
itself it is a small but pivotal dataset: it is the **treatment indicator** that
separates banks that lost access to the provider's services from those that did
not, which is what makes a clean event study of cyber propagation possible. A
paper we distill uses it:
[Kotidis & Schreft](/wiki/papers/jf/2025/kotidis-propagation-cyberattacks-financial-system-2025/)
use the confidential user list as the treatment-group indicator (user versus
nonuser banks) when tracing how a cyberattack propagates through the
[Fedwire](/wiki/confidential/fedwire/) payment system.

- **Cost:** not for sale. Confidential single-source data.
- **Source:** confidential records of which banks used the attacked TSP (a
  vendor-relationship mapping, not a market product).
- **Coverage:** the set of banks linked to one technology service provider around
  the attack event; it is an identification key, not a broad panel.

## Access (restricted)

- **No public download and no resale.** Vendor-relationship identities are
  sensitive and are not published; the list was obtained under a confidential
  arrangement.
- **Event-specific and supervisory in nature.** The list pertains to one provider
  and one episode; it is paired with confidential payment data inside a secure
  environment. There is no standing way for a third party to reach it.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **It is a treatment indicator, not an outcome.** The list only says who used the
  TSP; all the economics comes from pairing it with payment and borrowing data.
  Without the linked outcomes it carries no information on its own.
- **TSP use is not random.** Banks chose this provider for reasons (size, region,
  business model) that may correlate with outcomes; the user/nonuser split is not a
  clean experiment by itself. The design must address selection.
- **Membership can be fuzzy at the edges.** Some banks use a provider partially, or
  through an intermediary; binary user/nonuser coding hides partial exposure.
  Check how borderline relationships are classified.
- **Identifier joins.** Banks on the list must be linked to
  [Fedwire](/wiki/confidential/fedwire/) participants and to
  [Call Reports](/wiki/datasets/call-reports/) (RSSD) for outcomes and controls;
  mergers break the link. Verify the crosswalk.
- **Not redistributable, and the provider is not named.** The vendor and the bank
  identities are confidential; results are reported in aggregate. Plan for
  non-reproducibility of the raw inputs.

## Citation

Cite the source and arrangement, e.g.: *Confidential user list for the attacked
technology service provider; used under arrangement, YYYY-MM-DD.* State the event,
the user/nonuser coding, and the data it was linked to.


==============================================================================
# VPS: Norwegian securities depository holdings (restricted access)
# https://instituteforautomatedresearch.org/wiki/confidential/vps-norway/
# VPS is the Norwegian central securities depository; its records give complete individual-level securities holdings for Norwegian investors. It is restricted research microdata. This page documents what it is and the gotchas, but it was not exercised here.
# Tags: holdings, household-finance, equities, norway, administrative, data:vps-norway
==============================================================================

:::caution[Restricted holdings microdata: not exercised here]
**VPS holdings data is confidential individual-level microdata** (Euronext
Securities Oslo, the former Verdipapirsentralen). It is **not for sale and not
publicly downloadable**: it carries **no provenance badge** because there is no
access path we can run here. Researchers reach it only through an approved
project under restricted research terms, typically in a secure environment, with
output subject to confidentiality review. The page documents what the collection
is and the gotchas; treat it as unverified. This is the honest grade under the
institute's Verified discipline.
:::

**VPS** (Verdipapirsentralen, now Euronext Securities Oslo) is Norway's central
securities depository. Because Norwegian securities are registered there at the
beneficial-owner level, VPS records give **complete individual-level holdings** of
listed stocks and other registered securities, at monthly (or finer) frequency.
This makes it one of the few datasets with the full cross-section of a market's
retail and institutional holdings, not a sample. A paper we distill uses it:
[Betermier, Calvet, Knupfer & Kvaerner](/wiki/papers/jf/2025/betermier-investor-factors-2025/)
use VPS for complete individual investor stockholdings at monthly frequency
(300,000+ investors), linked to investor characteristics from
[Statistics Norway](/wiki/confidential/statistics-norway/) tax records and to OSE
prices from the Titlon database.

- **Cost:** not for sale. Restricted research microdata.
- **Source:** the Norwegian CSD (VPS / Euronext Securities Oslo).
- **Coverage:** registered holdings of Norwegian securities at the beneficial-owner
  level; complete for registered instruments, monthly or finer.

## Access (restricted)

- **Through an approved research arrangement.** Access to the holdings microdata
  is granted per project under restricted terms; it is not an open or commercial
  download.
- **Linked in a secure environment.** In practice VPS holdings are merged with
  [Statistics Norway](/wiki/confidential/statistics-norway/) registers under
  pseudonymized identifiers inside a secure setting, with outputs checked before
  release.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; they are documented, not verified here.

- **Registered holdings, not all wealth.** VPS covers securities registered in the
  depository; assets held abroad, through foreign brokers, or in funds not
  registered there are outside it. Do not read VPS as the investor's whole
  portfolio.
- **Nominee and omnibus accounts hide beneficial owners.** Holdings via nominee
  or omnibus accounts (often foreign) are not resolved to the ultimate owner; the
  beneficial-owner completeness is strongest for domestic direct holders. Treat
  nominee-held shares carefully.
- **Account-to-investor mapping.** An investor can have multiple VPS accounts;
  aggregating to the person requires the account-to-owner mapping. Do not treat an
  account as an investor.
- **Linkage keys are pseudonymized per project.** The keys joining VPS to SSB
  registers are project-specific pseudonyms; you cannot merge across separate
  deliveries. Plan all links inside one delivery.
- **Corporate actions and identifier changes.** Splits, mergers, and ISIN changes
  alter holdings mechanically; without adjusting for corporate actions, holdings
  changes are mismeasured. Apply a corporate-action map.
- **Output is disclosure-reviewed and cannot be redistributed.** Results leave the
  secure environment only after review, and the microdata itself cannot be shared.

## Citation

Cite the source and access route, e.g.: *VPS holdings (Euronext Securities Oslo /
Verdipapirsentralen), accessed under approved research arrangement, YYYY-MM-DD.*
State the years, the holdings frequency, and how accounts were aggregated to
investors.


==============================================================================
# American Community Survey (ACS)
# https://instituteforautomatedresearch.org/wiki/datasets/acs/
# How to pull American Community Survey estimates from the Census Bureau API (free key) or the no-key bulk files, and the gotchas that bite pipelines (1-year vs 5-year, every estimate has a margin of error, table-code churn, geographies are vintaged, it is a sample not a count).
# Verified 2026-06-22 · tested with live Census API pull of ACS 5-year (2022 acs5, B01001_001E total population for California, returned 39,356,104)
# Tags: census, demographics, labor, survey, free, data:acs
==============================================================================

**The American Community Survey (ACS)** is the U.S. Census Bureau's rolling
household survey of demographic, social, economic, and housing characteristics:
income, occupation, education, commuting, rent, ancestry, and more. It replaced
the decennial census "long form" and is the standard source for small-area
socioeconomic data between censuses. The data is free and public. Used in, for
example,
[Allcott, Montanari, Ozaltun & Tan](/wiki/papers/jf/2026/allcott-corporate-social-impact-2026/),
where ACS occupation-by-county employment distributions (2010-2019) drive the
worker-surplus estimation.

- **Cost:** free, public.
- **API key:** the Census Data API needs a **free** API key
  (<https://api.census.gov/data/key_signup.html>); the bulk download files do
  not.
- **Coverage:** annual. **1-year** estimates for areas with population >= 65,000
  (from 2005); **5-year** estimates for all geographies down to block group
  (from the 2005-2009 release). The Census Bureau page covers other Census
  products; this page is the ACS specifically.
- **Home:** <https://www.census.gov/programs-surveys/acs>

## Access

The Census Data API serves ACS as JSON. The endpoint encodes the year and
dataset (`acs/acs5` = 5-year, `acs/acs1` = 1-year), and you select variables and
a geography:

```bash
# ACS 5-year 2022: total population by state (free key required)
curl -sL "https://api.census.gov/data/2022/acs/acs5?get=NAME,B01001_001E&for=state:*&key=$CENSUS_API_KEY"

# Occupation/employment table at county level (detail table B-series), one state
curl -sL "https://api.census.gov/data/2019/acs/acs5?get=NAME,group(B24010)&for=county:*&in=state:06&key=$CENSUS_API_KEY"
```

Variable codes follow the table convention: `B01001_001E` is the *estimate* (`E`)
for line 001 of detail table `B01001`; the matching margin of error is
`B01001_001M`. The bulk **summary files** (no key) live under
`www2.census.gov/programs-surveys/acs/summary_file/<year>/` for whole-table or
whole-state pulls.

### Load in Python

```python
import pandas as pd, requests, os

url = ("https://api.census.gov/data/2022/acs/acs5"
       "?get=NAME,B01001_001E,B01001_001M&for=county:*&in=state:06"
       f"&key={os.environ['CENSUS_API_KEY']}")
rows = requests.get(url, timeout=60).json()
df = pd.DataFrame(rows[1:], columns=rows[0])          # row 0 is the header
df["pop"] = df["B01001_001E"].astype(float)            # estimate
df["pop_moe"] = df["B01001_001M"].astype(float)        # 90% margin of error
```

## Gotchas (the ones that bite pipelines)

- **1-year and 5-year are different products; never mix or overlap them.** The
  1-year estimates are current but only for areas >= 65,000 population; the
  5-year estimates cover everywhere but are a **pooled five-year average**
  centered on the period, so "2019 acs5" means 2015-2019 data. Overlapping
  5-year releases (2015-2019 vs 2016-2020) share four years of sample and must
  **not** be treated as independent annual observations. Pick one product and
  state the period.
- **Every estimate ships with a margin of error; small areas are noisy.** ACS is
  a sample, so each `..._E` has a 90% `..._M` margin of error that can be large
  for small geographies or small subgroups. Ignoring the MOE and treating
  estimates as exact counts produces false precision and spurious cross-area
  differences. Carry the `M` columns and propagate them, especially for derived
  ratios.
- **It is a survey estimate, not a population count.** ACS does not enumerate
  everyone; the decennial census does. Do not use ACS for legal population
  counts, and expect ACS totals to differ from the decennial and from Census
  population estimates.
- **Table codes and universes change between years.** Detail-table IDs are added,
  dropped, and re-specified, and a measure's universe can shift. A code that
  exists in 2019 may be absent or redefined in 2012. Verify each
  `<year>/acs5/variables.json` rather than assuming a code is stable across the
  whole 2010-2019 panel.
- **The 2020 ACS 1-year was not released as standard estimates.** COVID-19
  disrupted 2020 data collection, so the Census Bureau published only
  experimental 2020 1-year estimates, not the standard `acs/acs1` tables. A loop
  iterating years over `acs/acs1` will hit a gap at 2020 (the 5-year series is
  unaffected). Skip 2020 for the 1-year product or handle the missing year
  explicitly.
- **Geographies are vintaged.** County, tract, and place boundaries (and FIPS
  codes) follow the vintage of each release; tract definitions change at the
  decennial boundary. A county-by-year panel needs a geographic crosswalk across
  vintages, not a naive FIPS merge.
- **Geography is hierarchical in the API.** Sub-state geographies require an `in=`
  clause (county needs `in=state:`, tract needs `in=state:...+county:...`).
  Pulling all tracts nationally in one call is not allowed; loop over states.
  `for=...:*` is the wildcard within the parent.
- **Dollar figures are nominal to the survey period.** Income and rent are in the
  dollars of the (multi-year) collection period and are not inflation-adjusted
  across releases by default. Deflate explicitly when building a real series.

## Products and grains

| Dataset | API path | Geographies | Note |
|---|---|---|---|
| ACS 1-year | `acs/acs1` | areas >= 65,000 pop | Most current; coarse geography |
| ACS 5-year | `acs/acs5` | down to block group | Pooled 5-year average; all areas |
| ACS 1-year supplemental | `acs/acsse` | areas >= 20,000 pop | Simplified tables |
| PUMS microdata | `acs/acs5/pums` | PUMA (>= 100,000 pop) | Person/household records for custom tabs |

## Citation

Cite the U.S. Census Bureau, the ACS, the product and period, the table, and the
access date, for example: *U.S. Census Bureau, American Community Survey,
5-year estimates 2018-2022, table B01001, retrieved from the Census Data API
(https://api.census.gov/data), accessed YYYY-MM-DD.* Record the product
(1-year/5-year), the period, and the table/variable codes so the estimate is
reproducible.


==============================================================================
# Amsterdam historical housing prices and rents
# https://instituteforautomatedresearch.org/wiki/datasets/amsterdam-housing-transactions/
# How to pull the long-run Amsterdam house-price and rent series compiled by Eichholtz, Korevaar, Francke and co-authors as no-login Excel files, plus the gotchas (the compiled panels are separate from the raw City Archives, the hosting is personal Google Drive with link rot, and several distinct series must not be spliced).
# Verified 2026-06-22 · tested with live no-login Google Drive download of the Amsterdam 1620-2019 house-price index .xlsx (42,591 bytes, sheets Sources/Data) and the RFS 2021 headline series .xlsx (29,096 bytes, Amsterdam 1900-1979 plus Paris 1809-1943), from Matthijs Korevaar's data page
# Tags: housing, prices, rents, historical, real-estate, time-series, free, no-api-key, academic, data:amsterdam-housing-transactions
==============================================================================

**The Amsterdam historical housing data** is a set of long-run house-price and
rent series for Amsterdam compiled from the city's archival records by Piet
Eichholtz, Matthijs Korevaar, Marc Francke, Thies Lindenthal and co-authors.
The headline product is a repeat-sales house-price index reaching back to the
seventeenth century, built from mandatory transaction registrations and, for the
nineteenth century, repeat sales on the Herengracht canal. It is the standard
free source for multi-century Dutch housing-return work. Used in, for example,
[Francke (2025)](/wiki/papers/jf/2025/francke-baby-booms-asset-booms-2025/),
where the Amsterdam house-price index (Bayesian repeat-sales) anchors the
long-run price series tested against demographic cohorts.

- **Cost:** free, no account, no key (Excel files on the authors' data page).
- **Source records:** Amsterdam City Archives (Stadsarchief Amsterdam): aldermen
  transaction registrations (1620 to 1811), nineteenth-century Herengracht repeat
  sales, the 1805 rental census, and civil registers.
- **Coverage:** the compiled house-price index runs 1620 to 2019; a separate
  total-return series covers 1900 to 1979; eighteenth-century rents and yields
  are a further file.
- **Data page:** Matthijs Korevaar,
  <https://sites.google.com/view/matthijskorevaar/data>

## Access

The compiled panels are posted as Excel workbooks on the authors' data page and
download with no login.

```bash
# Amsterdam house-price index 1620-2019 (Korevaar, Eichholtz & Francke 2021, ESB);
# 42,591-byte .xlsx, sheets "Sources / Citation" and "Data":
curl -sL "https://drive.usercontent.google.com/download?id=1d5nOAuyFcD7KhGrgRj3jhlvAgvEfCXt0&export=download" -o amsterdam_index_1620_2019.xlsx

# RFS 2021 headline total-return series (Amsterdam 1900-1979 + Paris 1809-1943);
# 29,096-byte .xlsx, sheets "Sources" and "Headline Series":
curl -sL "https://drive.usercontent.google.com/download?id=1u-cYs7TiB6-rEhcDqCiGkfXsHdP9mWx1&export=download" -o rfs2021_headline.xlsx

# Full RFS 2021 microdata replication package (large): the plain link returns a
# Google Drive virus-scan interstitial, so add the confirm token:
curl -sL "https://drive.usercontent.google.com/download?id=1G6X2034aQZCSVprT-7srBEWhMPPtgIxW&export=download&confirm=t" -o rfs2021_replication.zip
```

The eighteenth-century Amsterdam rents and yields (Korevaar 2023, JFE) are on
Mendeley Data: <https://data.mendeley.com/datasets/db9nz3yhdr/1>.

## Gotchas (the ones that bite pipelines)

- **The compiled panel is not the raw archive.** The fetchable value is the
  authors' Excel series. The underlying Stadsarchief Amsterdam records (notarial
  and transaction registers, in Dutch, largely un-digitized) are a separate,
  much harder source and are not a clean panel. Do not expect transaction-level
  microdata from the index file.
- **Personal Google Drive hosting, so link rot is real.** The files are served
  from id-based Drive links with no DOI and no versioned permanent URL. One link
  on the data page (a "Baby Booms" surfdrive link) is already dead (404).
  Snapshot the file and record the Drive id plus retrieval date.
- **Large packages need the Drive confirm token.** The microdata zip returns an
  HTML virus-scan interstitial on the plain link; add `&confirm=t` and verify you
  received a zip or xlsx, not an HTML page.
- **Several distinct series, do not splice blindly.** The 1620-2019 index, the
  1900-1979 total-return series, and the eighteenth-century rents/yields are
  built with different methods and scopes. Pick the file matching your period;
  joining them into one spliced series needs care about method breaks.
- **Workbooks are multi-sheet and partly Dutch.** Each file carries a Sources or
  Citation sheet plus a Data sheet; some labels are in Dutch. Read the Sources
  sheet for the exact definitions and the citation the authors ask for.
- **The index is a level, not transactions.** The headline product is a
  repeat-sales index (Bayesian, Francke method); it is a price level over time,
  not individual sales. For transaction-level analysis you must go back to the
  archive.

## Reference

| Field | Value |
|-------|-------|
| Data page | `sites.google.com/view/matthijskorevaar/data` |
| House-price index | Amsterdam 1620-2019, .xlsx (Drive id `1d5nO...CXt0`) |
| Total-return series | Amsterdam 1900-1979 + Paris 1809-1943, .xlsx (Drive id `1u-cYs...mWx1`) |
| 18th-century rents/yields | Mendeley Data `db9nz3yhdr` (Korevaar 2023) |
| Raw source | Amsterdam City Archives (Stadsarchief), separate, un-digitized |
| Format | Multi-sheet Excel (Sources + Data) |
| Key required | No |

## Citation

Cite the specific compiled series and its paper, plus the Drive file (id) and
retrieval date: Korevaar, Eichholtz & Francke (2021, ESB) for the 1620-2019
house-price index; Eichholtz, Korevaar, Lindenthal & Tallec (2021),
"The Total Return and Risk to Residential Real Estate," *Review of Financial
Studies* 34(8), 3608-3646, for the 1900-1979 total-return series; Korevaar
(2023, *Journal of Financial Economics*) for the eighteenth-century rents. If you
use the underlying records, cite Stadsarchief Amsterdam separately. Record the
Drive id and access date, since the files are not versioned.


==============================================================================
# Barro-Ursua macroeconomic database
# https://instituteforautomatedresearch.org/wiki/datasets/barro-ursua/
# The Barro-Ursua database is a long-run cross-country panel of annual real per-capita GDP and consumption, assembled to study macroeconomic disasters. It is a free academic dataset; the canonical host blocked automated fetches from this session, so the download was not exercised here.
# Tags: macro, historical, free, academic, time-series, data:barro-ursua
==============================================================================

:::note[Free academic data, not exercised here]
The Barro-Ursua database is a **free** academic download (no account, no key),
but the canonical host (Robert Barro's university page) returned automated
requests with a block from this session, so the end-to-end download was **not**
run here. The page therefore carries **no provenance badge**: it documents the
dataset and the access path, but the pull is unverified under the institute's
Verified discipline. Anyone with a browser can download the file directly.
:::

**The Barro-Ursua macroeconomic database** (Robert J. Barro and Jose F. Ursua)
is a long-run, cross-country panel of **annual real per-capita GDP and personal
consumer expenditure**, built to study rare macroeconomic disasters: large
cumulative contractions in output or consumption. For some economies the series
reach back into the nineteenth century. It is used in, for example,
[Krishnamurthy & Muir](/wiki/papers/jf/2025/krishnamurthy-credit-cycles-financial-crisis-2025/)
for long historical real per-capita GDP series for advanced economies.

- **Cost:** free, academic release (no key, no account).
- **Authors / source:** Robert J. Barro and Jose F. Ursua; hosted on the
  authors' academic pages and referenced by the NBER.
- **Coverage:** roughly 40 economies, annual; GDP for more countries than
  consumption; history extends to the 1800s for several countries, ending at
  the published update vintage.

## Access

- **Author download.** The dataset is published as a spreadsheet workbook on
  Robert Barro's academic page (consumption and GDP series, plus the
  disaster-dating worksheets). It is a direct file download with no key or
  account.
- **Not exercised here.** Automated requests from this session were blocked by
  the host, so confirm the current download URL in a browser. The dataset is a
  static academic release, so a cached copy is a reasonable fallback.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; the download itself was not run here.

- **Long series are spliced from heterogeneous national sources.** Pre-WWII
  observations stitch together different historical sources by country, so data
  quality and definitions vary across eras; treat the deep history as more
  uncertain than the post-war sample.
- **Consumption and GDP coverage differ.** More countries have a GDP series
  than a consumption series, and start dates vary by country, so a balanced
  panel is much shorter than the full file. Decide on the series and the
  balanced window deliberately.
- **The disaster definition is a construct, not a field.** "Disasters" are
  derived by applying a cumulative-contraction threshold (the authors use a
  fractional peak-to-trough decline) to the series; the raw data does not label
  them. Reproduce the threshold to match the papers.
- **It is a periodic academic release, not a maintained feed.** The data ends
  at the publication update vintage and is not refreshed in real time; do not
  expect recent years. Pin the vintage you used.
- **Currency basis and population denominators vary.** Real-terms basis and the
  per-capita denominator are country-specific; check the definitions before
  pooling across countries.

## Citation

Cite the Barro-Ursua macroeconomic database (Barro and Ursua), stating the
series used (real per-capita GDP or consumption), the country sample and
window, the release vintage, and the disaster threshold applied. The companion
references are Barro and Ursua, "Macroeconomic Crises since 1870" (2008) and
related work.


==============================================================================
# BEA Input-Output Accounts
# https://instituteforautomatedresearch.org/wiki/datasets/bea-io/
# How to pull the BEA Input-Output Accounts (Use, Make/Supply, and Requirements tables) for free with a registered API key, the table IDs you actually need to build upstreamness and production-network measures, and the gotchas that bite pipelines.
# Verified 2026-06-22 · tested with live keyed BEA API pull, InputOutput dataset, Use table (TableID 259, year 2022, 4642 rows) and total-requirements table (TableID 56)
# Tags: production-networks, supply-chains, industry, free, bea, data:bea-io
==============================================================================

The **BEA Input-Output (I-O) Accounts** record, in dollars, how each US industry
buys from and sells to every other industry: which commodities each industry
uses as inputs (the **Use** table), which commodities each industry produces
(the **Make / Supply** table), and the **Requirements** tables that invert those
into the total output each industry needs from every other to deliver a dollar
of final demand. They are the standard source for production-network and
supply-chain measures, including the upstreamness/downstreamness index of Antras
and Chor (2018). This page is the distilled access recipe.

- **Cost:** free, no paywall.
- **API key:** free but **required** (register for a `UserID`; bare requests are
  rejected).
- **Coverage:** through this API, Sector (~15 industries) and Summary (~71)
  tables, annual from **1997 to 2024** (confirmed live). The Detail (~400
  industry) tables are published only for benchmark years (most recently 2017)
  and come as separate downloadable files, not through this endpoint.
- **Home:** <https://www.bea.gov/industry/input-output-accounts-data> ·
  **API docs:** <https://apps.bea.gov/api/>

## Access

The I-O tables are not in the no-key NIPA static files. They come through the
BEA Data API, which needs a free key.

### Register a key

Get a free `UserID` at <https://apps.bea.gov/API/signup/>. BEA emails the key
**plus an activation step**: the key stays inert until the account is activated,
so a freshly issued key can return `"This UserId is not active"` until you
complete it. Store it in the environment (e.g. `.env` as `BEA_API_KEY=...`);
never hard-code it.

### Pull a table

The `InputOutput` dataset endpoint takes a `TableID` and a `Year` (or
`Year=ALL`). The same `UserID` also serves NIPA, Regional, GDPbyIndustry, and
the other BEA datasets.

```python
import os, requests
key = os.environ["BEA_API_KEY"]
r = requests.get("https://apps.bea.gov/api/data", params={
    "UserID": key, "method": "GetData", "datasetname": "InputOutput",
    "TableID": "259", "Year": "2022", "ResultFormat": "JSON",
})
rows = r.json()["BEAAPI"]["Results"]["Data"]   # long format: RowCode/ColCode/DataValue
```

Discover valid table IDs and years before pulling, since they are not a tidy
1..N range:

```python
requests.get("https://apps.bea.gov/api/data", params={
    "UserID": key, "method": "GetParameterValues",
    "datasetname": "InputOutput", "ParameterName": "TableID",
    "ResultFormat": "JSON",
})
```

## Gotchas (the ones that bite pipelines)

The reason to read this page rather than the BEA docs. Verified against live
keyed pulls on the date above.

- **`"This UserId is not active"` is BEA's generic error, not always an
  activation problem.** Error code `4` carries this message for *any* malformed
  `GetData` call, including a wrong `TableID`. If a metadata call
  (`GetParameterValues`) on the same dataset succeeds with the same key, the key
  is active and the real fault is the request. Check the `TableID` against the
  valid list before assuming the key is dead.
- **Table IDs are an unordered set, not 1..N.** The I-O tables are numbered with
  values like `56`-`61` (Requirements) and `258`-`262` (Use, Make/Supply). A
  guessed small integer such as `TableID=2` returns the "not active" error.
  Always pull the `TableID` parameter list first.
- **Two API levels, plus a Detail tier off-API.** This API serves Sector (~15)
  and Summary (~71) tables; a measure built at one level is not comparable to the
  same measure at the other. The Detail (~400 industry) tables exist only for
  benchmark years and are downloaded as separate files, not through this dataset
  endpoint.
- **Use and Make are different objects.** Upstreamness/downstreamness and
  total-requirements measures are derived from the **Use** and **Make/Supply**
  tables (and their Leontief inverse), not from a single matrix. Confirm which
  table a method needs before pulling: Antras-Chor (2018) upstreamness is built
  from the Use/requirements structure.
- **Long format, paired codes.** Rows return as `RowCode`/`RowDescr` (the
  using/producing industry or commodity) crossed with `ColCode`/`ColDescr`; you
  pivot to a matrix yourself. `RowType`/`ColType` flag whether each axis is an
  Industry or a Commodity, which differs by table.
- **Benchmark vs annual revisions.** Detail benchmark tables (every five years)
  are revised when a new benchmark lands and can reweight history; annual
  Summary tables are themselves revised. State the vintage year for any
  reproducible network measure.

## Tables you actually need

Pulled live from the `TableID` parameter list. Verify against the live list, as
IDs can change between releases.

| TableID | Table | Use |
|---|---|---|
| `259` | Use of Commodities by Industries, Summary | input flows; upstreamness |
| `258` | Use of Commodities by Industries, Sector | coarse input flows |
| `262` | Domestic Supply of Commodities by Industries, Summary | Make/Supply side |
| `261` | Domestic Supply of Commodities by Industries, Sector | coarse supply |
| `56` | Total Requirements, Industry-by-Commodity, Sector | Leontief inverse |
| `57` | Total Requirements, Industry-by-Commodity, Summary | Leontief inverse |
| `60` | Total Requirements, Industry-by-Industry, Sector | industry linkages |
| `61` | Total Requirements, Industry-by-Industry, Summary | industry linkages |

## Standard operations

- **Upstreamness / position in the chain:** build from the Use and total
  requirements tables following Antras-Chor (2018); pick one industry-detail
  level and hold it fixed across years.
- **Network centrality / propagation:** use the requirements (Leontief inverse)
  tables, not raw Use dollars, when you need how much output each industry
  ultimately draws from every other.
- **Pivot before analysis:** the API returns long `Row x Col` records; pivot to
  a square industry-by-commodity (or industry-by-industry) matrix, watching
  `RowType`/`ColType`.
- **Pair with NIPA / GDP-by-Industry:** for value-added and output deflators per
  industry, combine with [NIPA](/wiki/datasets/nipa/) and the BEA
  GDP-by-Industry dataset (same key).
- **State the vintage year and detail level** with any network measure; the I-O
  tables are revised and exist at three granularities.

## Citation

U.S. Bureau of Economic Analysis. *Input-Output Accounts.* Suitland, MD: BEA.
Cite the specific table and year and your access date, e.g. *Bureau of Economic
Analysis, Use of Commodities by Industries (Summary), 2022, accessed
YYYY-MM-DD.*


==============================================================================
# BIS Effective Exchange Rate Indices (EER)
# https://instituteforautomatedresearch.org/wiki/datasets/bis-rer/
# How to pull the BIS nominal and real effective exchange rate indices from the no-key BIS statistics API, and the gotchas that bite pipelines (real vs nominal, narrow vs broad basket, an up-move means appreciation, the index is rebased not a level).
# Verified 2026-06-22 · tested with live no-key BIS stats API pull of WS_EER (real & nominal, broad-64 and narrow-27 baskets, US, monthly)
# Tags: macro, exchange-rates, international, time-series, free, no-api-key, data:bis-rer
==============================================================================

**The BIS Effective Exchange Rate (EER) indices** are trade-weighted exchange
rate indices published by the Bank for International Settlements for around 60
economies. Each index averages a currency against a basket of trading partners,
weighted by bilateral trade. They come in **nominal** (NEER) and **real**
(REER, CPI-deflated) flavors, and in a **broad** basket (64 economies) and a
**narrow** basket (27 economies). The data is free and public. Used in, for
example,
[Kremens, Martin & Varela](/wiki/papers/jf/2025/kremens-long-horizon-exchange-rate-2025/),
where the BIS real effective exchange rate is the RER predictor variable in the
long-horizon exchange-rate regressions.

- **Cost:** free, public.
- **API key:** none required.
- **Coverage:** ~60 economies. Narrow basket monthly from 1964; broad basket
  monthly from 1994. Real (CPI-based) and nominal.
- **Home:** <https://www.bis.org/statistics/eer.htm>

## Access

The BIS statistics API serves the EER dataflow (`WS_EER`) as CSV with no
authentication. The series key is `FREQ.EER_TYPE.EER_BASKET.REF_AREA`:

```bash
# Real, broad (64-economy) basket, United States, monthly, last 12 obs
curl -sL \
  "https://stats.bis.org/api/v2/data/dataflow/BIS/WS_EER/1.0/M.R.B.US?lastNObservations=12&format=csv"

# Nominal, narrow (27-economy) basket, full history
curl -sL \
  "https://stats.bis.org/api/v2/data/dataflow/BIS/WS_EER/1.0/M.N.N.US?format=csv"
```

The dimension codes are: `FREQ` = `M` (monthly); `EER_TYPE` = `R` (real,
CPI-deflated) or `N` (nominal); `EER_BASKET` = `B` (broad, 64 economies) or `N`
(narrow, 27 economies); `REF_AREA` = the ISO-style economy code (`US`, `XM` for
the euro area, and so on). Returned rows carry `TIME_PERIOD` (`YYYY-MM`) and
`OBS_VALUE` (the index level). Bulk CSV for the whole dataflow is also linked
from the EER landing page.

### Load in Python

```python
import pandas as pd, io, requests

url = ("https://stats.bis.org/api/v2/data/dataflow/BIS/WS_EER/1.0/"
       "M.R.B.US?format=csv")
df = pd.read_csv(io.StringIO(requests.get(url, timeout=60).text))
df["TIME_PERIOD"] = pd.PeriodIndex(df["TIME_PERIOD"], freq="M")
reer = df.set_index("TIME_PERIOD")["OBS_VALUE"]  # real broad US, index level
```

## Gotchas (the ones that bite pipelines)

- **Real and nominal are different series; pick deliberately.** The real EER
  deflates the nominal index by relative CPI, so it moves with inflation
  differentials, not just nominal FX. A paper using "the BIS RER" wants
  `EER_TYPE=R`; do not substitute the nominal index. They diverge materially
  over long horizons.
- **Broad versus narrow basket changes both the level and the history.** The
  broad (64-economy) basket starts in 1994; the narrow (27-economy) basket runs
  back to 1964 but covers fewer partners. They are separate indices with
  different weights and base periods, so do not splice the narrow history onto
  the broad series to get a longer sample. Choose one basket and state which.
- **An increase means appreciation.** The BIS convention is that a *rise* in the
  index is an appreciation of the reference economy's currency (it buys more
  foreign goods/currency, trade-weighted). This is the opposite sign of some
  bilateral quote conventions; confirm the direction before signing your
  regressor.
- **It is a rebased index, not a price or a rate.** Values are index numbers
  normalized to a base year (currently averaging 100 over a recent reference
  period), and BIS periodically rebases and re-weights. Work in log changes or
  ratios, and re-pull rather than caching a level that may be rebased on the
  next release.
- **Weights are updated in vintages.** Trade weights are revised roughly every
  three years and applied with a chaining method, so the historical index is not
  frozen. For exact reproduction, record the access date and the basket; a pull
  a year later can differ in back-history.
- **Coverage and start dates differ by economy.** Not every economy has both
  baskets or the full history; emerging-market series often start later. Check
  the first non-missing `TIME_PERIOD` for your `REF_AREA` rather than assuming
  the basket's nominal start date.

## Dimensions (WS_EER)

| Dimension | Code | Meaning |
|---|---|---|
| `FREQ` | `M` | Monthly (the only frequency) |
| `EER_TYPE` | `R` / `N` | Real (CPI-deflated) / Nominal |
| `EER_BASKET` | `B` / `N` | Broad (64 economies) / Narrow (27 economies) |
| `REF_AREA` | e.g. `US`, `XM`, `JP` | Reference economy |

## Citation

Cite the BIS, the EER series (type, basket, economy), the URL, and the access
date, for example: *Bank for International Settlements, Effective Exchange Rate
Indices (real, broad basket, United States), retrieved from
https://www.bis.org/statistics/eer.htm, accessed YYYY-MM-DD.* Record the
type (real/nominal), basket (broad/narrow), and the vintage so the index is
reproducible.


==============================================================================
# Bureau of Labor Statistics (BLS)
# https://instituteforautomatedresearch.org/wiki/datasets/bls/
# How to pull BLS labor-force, employment, wage, and price series from the public data API with no key, plus the QCEW county wage files, and the series-ID and revision gotchas that bite pipelines.
# Verified 2026-06-09 · tested with live no-key BLS API v2 pull (unemployment rate LNS14000000) and no-key QCEW county-wage CSV (data.bls.gov/cew, area 06037, 2022 Q1)
# Tags: labor, employment, wages, prices, inflation, time-series, free, no-api-key, data:bls
==============================================================================

**The Bureau of Labor Statistics (BLS)** publishes U.S. labor-force,
employment, wage, and price statistics: the labor-force participation rate,
the unemployment rate, payroll employment, the Consumer Price Index (CPI) and
Producer Price Index (PPI), job openings (JOLTS), and county-by-industry
employment and wages (the Quarterly Census of Employment and Wages, QCEW). The
public data API and the QCEW files are free and need no key. Used in, for
example,
[Beyhaghi, Fracassi & Weitzner](/wiki/papers/jf/2026/beyhaghi-adverse-selection-corporate-loans-2026/)
for county-level and financial-industry wages, and the CPI-U series underlies
[Dittmar, Hsu, Roussellet & Simasek](/wiki/papers/jf/2026/dittmar-default-risk-sovereign-bonds-2026/)
as the inflation measure (obtained there through [FRED](/wiki/datasets/fred/),
which re-publishes BLS series).

- **Cost:** free, public.
- **API key:** none required. The public API v2 works with no key at lower
  limits (25 queries/day, up to 10 years, up to 25 series per call). A free
  `BLS_API_KEY` raises the limits (500 queries/day, 20 years, 50 series).
- **Coverage:** national labor-force and price series (monthly, from the
  mid-twentieth century for the headline series); QCEW county-by-industry
  employment and wages (quarterly, from 1990).
- **Home:** <https://www.bls.gov/data/> ·
  **API:** <https://api.bls.gov/publicAPI/v2/timeseries/data/> ·
  **QCEW:** <https://www.bls.gov/cew/>

## Access

### Time series via the public API (no key)

The API takes a JSON POST with a list of series IDs and a year range. No key is
needed for low-volume use:

```bash
# Unemployment rate (LNS14000000), 2023-2024, no key
curl -s -X POST "https://api.bls.gov/publicAPI/v2/timeseries/data/" \
  -H "Content-Type: application/json" \
  -d '{"seriesid":["LNS14000000"],"startyear":"2023","endyear":"2024"}'
```

Add `"registrationkey":"<BLS_API_KEY>"` to the JSON body to use a free key and
the higher limits.

```python
import requests

def bls_series(series_ids, start, end, key=None):
    body = {"seriesid": list(series_ids), "startyear": str(start),
            "endyear": str(end)}
    if key:
        body["registrationkey"] = key
    r = requests.post(
        "https://api.bls.gov/publicAPI/v2/timeseries/data/", json=body)
    r.raise_for_status()
    out = r.json()
    assert out["status"] == "REQUEST_SUCCEEDED", out.get("message")
    return out["Results"]["series"]   # each: seriesID + data[] (year, period, value)
```

### QCEW county wages (no key)

County-by-industry employment and wages come as direct CSV downloads keyed by
the 5-digit area FIPS code. No key, no API body:

```bash
# Los Angeles County (06037), 2022 Q1, all industries: establishments,
# employment, and wages by industry/ownership
curl -s -o qcew_06037_2022q1.csv \
  "https://data.bls.gov/cew/data/api/2022/1/area/06037.csv"
```

The header row names the fields (`area_fips`, `own_code`, `industry_code`,
`qtrly_estabs`, employment columns, `total_qtrly_wages`, `avg_wkly_wage`, and a
`disclosure_code` flag).

### Flat files

The full published tables, including discontinued and detailed series, live as
flat files under <https://download.bls.gov/pub/time.series/>, organized by
survey prefix (for example `ln/` for the CPS labor-force series, `cu/` for
CPI-U). These need no key.

## Gotchas (the ones that bite pipelines)

- **The seasonal-adjustment flag is hidden in the series ID.** For the CPS
  labor series the third letter sets it: `LNS...` is seasonally adjusted,
  `LNU...` is not. They are the same concept with different numbers. The same
  split holds across surveys (`CES`/`CEU`, `CUUR`/`CUSR`, `WPU`/`WPS`). Pick one
  deliberately and state which you used.
- **The no-key tier is rate-limited.** Without a key you get 25 queries per day,
  10 years, and 25 series per call. A batch backfill will hit that wall; either
  register a free key or cache aggressively. Closed-period data is immutable, so
  cache hits are safe.
- **Series are revised.** Monthly CES employment and CPI figures are revised
  (CES has preliminary, second, and final estimates, plus annual benchmark
  revisions). Pin the vintage you pulled; a re-pull of a recent month can change
  the number.
- **QCEW cells are suppressed for confidentiality.** A non-empty
  `disclosure_code` marks a cell withheld to protect an identifiable employer,
  so employment and wage figures are blanked there. A blank is not a zero;
  handle the disclosure flag.
- **QCEW counts jobs and wages, not firm financials.** It is administrative
  unemployment-insurance data covering covered employment by establishment
  location and industry, not a firm-level financial panel.
- **The API is the same series FRED re-publishes, not a different number.** FRED
  mirrors many headline BLS series (for example seasonally adjusted CPI-U
  `CUSR0000SA0` as `CPIAUCSL`; the not-seasonally-adjusted `CUUR0000SA0` maps to
  `CPIAUCNS` instead). Cite the source you actually pulled from; do not imply two
  independent measurements.

## Series-ID prefixes

| Prefix | Survey | Common use |
|--------|--------|------------|
| `LNS` / `LNU` | CPS (household), SA / NSA | LFPR, employment-population ratio, unemployment |
| `CES` / `CEU` | CES (establishment), SA / NSA | Payroll employment, hours, earnings |
| `CUUR` / `CUSR` | CPI-U, NSA / SA | Consumer prices |
| `WPU` / `WPS` | PPI, NSA / SA | Producer prices |
| `JTU` | JOLTS | Job openings, hires, quits, layoffs |
| `ENU` | QCEW | County by industry employment and wages |
| `SMU` / `SMS` | State and area CES, NSA / SA | Sub-national payrolls |

Discover exact series IDs with the BLS data finder
(<https://www.bls.gov/data/>) or the per-survey "Series ID formats" pages.

## Citation

Cite the U.S. Bureau of Labor Statistics, the survey and series IDs used (with
the seasonal-adjustment choice), the period range, the retrieval URL, and the
access date. For QCEW, record the area FIPS, year, quarter, and ownership and
industry codes so the pull is reproducible.


==============================================================================
# FFIEC Call Reports: US bank condition and income filings
# https://instituteforautomatedresearch.org/wiki/datasets/call-reports/
# How to pull US bank Call Report data for free: the FDIC financials JSON API (with the host move and amounts-in-thousands trap), the FFIEC bulk schedules, the CERT/RSSD identifiers, the RCFD-vs-RCON codes, and the YTD-income trap.
# Verified 2026-06-09 · tested with live api.fdic.gov/banks fetch (financials REPDTE:20231231, CERT 3511 Q1/Q2 2023, institutions CERT->FED_RSSD) + host-move 301 confirmed
# Tags: banks, filings, panel-data, free, no-api-key, federal-reserve, fdic, data:call-reports
==============================================================================

**Call Reports** (FFIEC Reports of Condition and Income, forms FFIEC 031 / 041 /
051) are the quarterly regulatory filings every US insured depository
institution submits. They are the canonical free, bank-level panel for
balance sheet, income statement, loan composition, deposits, capital, and past
due items, going back decades. This page is the access recipe for the two public
paths: the FDIC financials API (clean JSON) and the FFIEC bulk schedules (raw
MDRM items).

- **Cost:** free, no auth, no API key.
- **Coverage:** every FDIC-insured bank and thrift, quarterly. FDIC financials
  API reaches back to 1992; FFIEC bulk Call Reports go back further per form.
- **Granularity:** one filing per (institution, quarter-end).
- **Home:** FDIC API <https://api.fdic.gov/banks/> ·
  FFIEC bulk <https://cdr.ffiec.gov/public/>

## Access

### Option 1: FDIC financials API (JSON, no auth, cleaned series)

A no-key JSON API serving FDIC-computed financials derived from the Call
Reports. Filter on `REPDTE` (the quarter-end report date, `YYYYMMDD`) and pick
fields:

```python
import requests

def fdic_financials(repdte="20231231", limit=10000, fields="CERT,NAME,ASSET,DEP,NETINC,REPDTE"):
    r = requests.get(
        "https://api.fdic.gov/banks/financials",
        params={"filters": f"REPDTE:{repdte}", "fields": fields,
                "limit": limit, "format": "json"},
        timeout=60,
    )
    r.raise_for_status()
    return [row["data"] for row in r.json()["data"]]

q4 = fdic_financials("20231231")   # ~4,660 banks that quarter
```

The companion `/banks/institutions` endpoint is the bank master (name, state,
charter, and the identifier bridge): pull it to map `CERT` to the Fed `RSSD`.

```python
import requests
inst = requests.get(
    "https://api.fdic.gov/banks/institutions",
    params={"filters": "CERT:3511", "fields": "CERT,NAME,FED_RSSD,STALP"},
    timeout=30,
).json()["data"][0]["data"]
# {'CERT': 3511, 'FED_RSSD': '451965', 'NAME': 'Wells Fargo Bank, ...'}
```

### Option 2: FFIEC bulk schedules (raw MDRM items)

For items the FDIC series does not expose, download the raw Call Report
schedules at
<https://cdr.ffiec.gov/public/PWS/DownloadBulkData.aspx>: pick "Call Reports --
Single Period" (or balance-sheet / income / past-due subsets), choose a
quarter, and get a ZIP of tab-delimited schedule files keyed by `IDRSSD`. Each
column is an MDRM code (see gotchas). Programmatic pulls of this path go through
the FFIEC Public Web Service (PWS), which needs a free registered account token;
the bulk web download itself is an ASP.NET postback, not a clean GET.

## Gotchas (the ones that bite pipelines)

The reason to read this page rather than the FDIC/FFIEC sites. Verified live on
the date above (financials and institutions endpoints pulled; the host 301 and
the YTD-income behavior were reproduced).

- **The FDIC API host moved.** `banks.data.fdic.gov` now issues a `301` to
  `api.fdic.gov/banks/...` (confirmed live). A client that does not resend the
  query on redirect, or a hard-coded old host, silently breaks. Request
  `api.fdic.gov/banks` directly.
- **Amounts are in thousands of dollars.** `ASSET` of `1687507000` is
  $1.687 trillion, not $1.7 billion (confirmed: Wells Fargo, CERT 3511,
  2023-Q1). Scale by 1,000 before reporting dollars.
- **Income items are year-to-date, not quarterly.** `NETINC` is cumulative
  within the calendar year: Wells Fargo CERT 3511 reads 6,017,000 at
  2023-03-31 and 11,455,000 at 2023-06-30 (confirmed live). The Q2 *quarterly*
  flow is the difference (11,455,000 - 6,017,000), and Q1 YTD equals Q1
  quarterly. Differencing the wrong way double-counts Q1.
- **Three identifiers, do not confuse them.** `CERT` (FDIC certificate) keys the
  FDIC API; `IDRSSD` / `RSSD` (Federal Reserve) keys the raw FFIEC bulk
  schedules and NIC; the OCC/OTS charter is a third. Join FDIC-API data to raw
  FFIEC items via `FED_RSSD` from the institutions endpoint (CERT 3511 ->
  RSSD 451965).
- **The form differs by bank, so a field may not exist for every filer.**
  FFIEC 031 (banks with foreign offices), 041 (domestic only), and 051 (small
  banks under $5B, streamlined with fewer schedules) carry different item sets.
  Do not assume a schedule or MDRM code is present for every institution.
- **RCFD vs RCON prefix.** In the raw schedules, the same concept appears as a
  consolidated item (`RCFD...`, includes foreign offices) and a domestic item
  (`RCON...`). Total assets is `RCFD2170` consolidated vs `RCON2170` domestic;
  mixing them across banks under-counts or double-counts.
- **Mergers break naive panels.** Banks enter and exit via M&A; a `CERT`/`RSSD`
  can be retired. Build the panel with the NIC structure/transformations table,
  not by assuming identifier continuity.
- **Quarter-end only, and amended filings exist.** `REPDTE` is always a
  quarter-end (`0331`, `0630`, `0930`, `1231`). Banks file late amendments, so
  "as filed" and "amended" can differ; the FDIC API serves a cleaned series.
- **The confidential supervisory overlay is a distinct source.** Holding-company
  filings (FR Y-9C) and the stress-test panel (FR Y-14) are restricted and
  tracked under separate slugs. Call Reports are the public bank-level layer.

## Identifiers and key fields

| Field | Meaning |
|---|---|
| `CERT` | FDIC certificate number; the FDIC-API key |
| `FED_RSSD` / `IDRSSD` | Federal Reserve RSSD ID; keys the FFIEC bulk schedules and NIC |
| `REPDTE` | Report date, always a quarter-end (`YYYYMMDD`) |
| `ASSET` | Total assets (thousands of dollars) |
| `DEP` | Total deposits (thousands) |
| `NETINC` | Net income, year-to-date (thousands) |
| `RCFD2170` / `RCON2170` | Total assets, consolidated / domestic (raw MDRM) |

## Standard operations

- **Bank-quarter panel:** stack `/banks/financials` across `REPDTE`, key on
  `(CERT, REPDTE)`; map to `RSSD` once via the institutions endpoint.
- **Quarterly flows:** difference YTD income/flow items within the calendar year
  (Q_n quarterly = Q_n YTD - Q_{n-1} YTD; Q1 quarterly = Q1 YTD).
- **Ratios:** compute ROA, capital, and loan-mix ratios from levels you pulled,
  rather than trusting a precomputed field across form changes.
- **Raw items:** when the FDIC series lacks an item, pull the FFIEC schedule by
  MDRM code, choosing `RCFD` vs `RCON` deliberately.
- **Always state** the report quarter, the form (031/041/051) if it matters, and
  whether a figure is a level or a YTD/quarterly flow.

## Citation

*Federal Financial Institutions Examination Council (FFIEC), Consolidated
Reports of Condition and Income (Call Reports); retrieved via the FDIC
financials API (https://api.fdic.gov/banks/) / FFIEC CDR
(https://cdr.ffiec.gov/public/), accessed YYYY-MM-DD.*


==============================================================================
# U.S. Census Bureau public data
# https://instituteforautomatedresearch.org/wiki/datasets/census/
# How to pull U.S. Census Bureau public data products (BDS, QWI, ACS, CBP, population estimates) for free, covering the bulk no-key flat-file path and the api.census.gov API, with the gotchas that bite automated pipelines.
# Verified 2026-06-09 · tested with live no-key Census BDS bulk CSV (www2.census.gov bds time-series)
# Tags: firms, demographics, free, data:census
==============================================================================

**U.S. Census Bureau public data** covers a family of distinct statistical
products used in finance and economics research: Business Dynamics Statistics
(BDS: firm and establishment counts, entry and exit, job creation and
destruction, by age and size), Quarterly Workforce Indicators (QWI: local
employment and earnings flows), the American Community Survey (ACS: demographic
and economic characteristics down to ZIP code and census tract), County Business
Patterns (CBP: establishment counts and payroll by industry and county), and
official population estimates. These products are cited in, for example,
[Barkai & Panageas](/wiki/papers/jf/2025/barkai-value-employment-2025/) for
Business Dynamics Statistics (young-firm employment share, firm-size
distribution, job creation and destruction), and
[Beyhaghi, Fracassi & Weitzner](/wiki/papers/jf/2026/beyhaghi-adverse-selection-corporate-loans-2026/)
for county-level population estimates.

- **Cost:** free, no paywall.
- **API key:** free key required for the api.census.gov API (returns "Missing
  Key" otherwise); the bulk flat files on www2.census.gov need no key.
- **Coverage:** U.S. firms, establishments, employment flows, demographics, and
  population, from the 1970s onward depending on the product.
- **Home:** <https://www.census.gov/data.html> ·
  **API docs:** <https://www.census.gov/data/developers.html>

## Access

### Option 1: No-key bulk flat files (verified path)

The BDS time-series CSVs are published as flat files on www2.census.gov and
need no authentication. The 2022 release is at:

```
https://www2.census.gov/programs-surveys/bds/tables/time-series/2022/bds2022.csv
```

Key columns: `year`, `firms`, `estabs`, `emp`, `estabs_entry`,
`estabs_exit`, `job_creation`, `job_destruction`, `job_creation_rate`,
`job_destruction_rate`.

```python
import pandas as pd

bds = pd.read_csv(
    "https://www2.census.gov/programs-surveys/bds/tables/time-series/2022/bds2022.csv",
    dtype=str,   # suppress int-parsing of suppressed cells marked "S"
)
# Convert numeric columns explicitly after handling suppression flags
numeric_cols = ["year", "firms", "estabs", "emp",
                "estabs_entry", "estabs_exit",
                "job_creation", "job_destruction"]
bds[numeric_cols] = bds[numeric_cols].apply(pd.to_numeric, errors="coerce")
```

This path is the genuinely no-key route and is what the verified stamp above
covers. Analogous bulk files exist for CBP and population estimates under
`www2.census.gov/programs-surveys/`.

### Option 2: api.census.gov (key required)

Get a free key at <https://www.census.gov/data/developers.html> and store it
in your environment (e.g. `CENSUS_API_KEY=...`); never hard-code it.

Example: ACS 5-year estimates for median household income (`B19013_001E`) by
county in California:

```python
import os, requests, pandas as pd

key = os.environ["CENSUS_API_KEY"]
url = (
    "https://api.census.gov/data/2022/acs/acs5"
    "?get=NAME,B19013_001E"
    "&for=county:*"
    "&in=state:06"
    f"&key={key}"
)
data = requests.get(url).json()
df = pd.DataFrame(data[1:], columns=data[0])
df["B19013_001E"] = pd.to_numeric(df["B19013_001E"], errors="coerce")
```

The API supports BDS, ACS, CBP, QWI, and population estimates with consistent
`?get=<variables>&for=<geography>&key=<key>` syntax. Variable and geography
codes are listed in the API discovery endpoint at
`https://api.census.gov/data/2022/acs/acs5/variables.json`.

## Gotchas (the ones that bite pipelines)

- **The API requires a free key; the bulk files do not.** The api.census.gov
  endpoint returns a plain-text "Missing Key" error (not a JSON error object)
  when called without a key. The genuinely no-key path is the bulk flat files
  on www2.census.gov as shown in Option 1 above.
- **"Census data" is not one dataset.** BDS, QWI, ACS, CBP, and population
  estimates have different geographies, frequencies, disclosure rules, and
  update cadences. Code that treats them interchangeably will silently produce
  nonsense. Pick the specific product and vintage for each research use, and
  document it.
- **Disclosure avoidance and noise injection.** The post-2020 differential
  privacy regime changed how small-cell counts are released. Suppressed cells
  appear as `"S"` or `"D"` in bulk files, and noise-injected counts in
  small geographies can be negative or implausibly volatile. This regime is
  a break from earlier products and affects panel comparability.
- **ACS margins of error must be propagated.** Every ACS estimate ships with a
  margin-of-error column (suffix `MA` or `M`). Ignoring it and treating point
  estimates as exact is incorrect, especially for small geographies or small
  subgroups. Also: 1-year ACS (larger geographies, current) and 5-year ACS
  (smaller geographies, averaged) are different products and must not be mixed
  in a panel without adjustment.
- **NAICS and geography vintages change.** Industry codes are revised across
  NAICS releases (1997, 2002, 2007, 2012, 2017, 2022), and county FIPS codes
  change when counties are created, dissolved, or renamed. Concatenating CBP or
  BDS panels across NAICS transitions or county changes requires crosswalks,
  which the Census provides but which must be applied explicitly.
- **BDS and QWI measure jobs and establishments, not firm financials.** These
  products describe entry, exit, employment flows, and earnings, not revenue,
  assets, or profitability. They complement Compustat or CRSP for
  macro-finance calibration but do not substitute for firm-level financial data.

## Products at a glance

| Product | Unit | Geography | Frequency | Key use in finance research |
|---------|------|-----------|-----------|------------------------------|
| BDS | Firm / establishment | Nation, state, metro, county | Annual | Entry/exit rates, firm-age distribution, job flows |
| QWI | Worker / establishment | State, metro, county, WIA | Quarterly | Local labor market dynamics, earnings by firm age |
| ACS | Household / person | Nation down to tract/ZIP (5-yr) | Annual | Demographics, income, education for local-area work |
| CBP | Establishment | Nation, state, county, ZIP | Annual | Industry structure, payroll, establishment counts |
| Population estimates | Person | Nation, state, county | Annual | County-level controls, normalization denominators |

## Citation

Cite the specific Census product and vintage, not just "Census." Examples:

- *U.S. Census Bureau, Business Dynamics Statistics, 2022 release,
  https://www.census.gov/programs-surveys/bds.html, accessed YYYY-MM-DD.*
- *U.S. Census Bureau, American Community Survey 5-Year Estimates, 2022,
  Table B19013, accessed via api.census.gov, accessed YYYY-MM-DD.*
- *U.S. Census Bureau, County Business Patterns, 2021,
  https://www.census.gov/programs-surveys/cbp.html, accessed YYYY-MM-DD.*

Each product's landing page lists the suggested citation format and the
appropriate technical documentation to reference.


==============================================================================
# CFTC Commitments of Traders (COT)
# https://instituteforautomatedresearch.org/wiki/datasets/cftc-cot/
# How to pull weekly aggregate futures positions by trader category from the CFTC, including the Traders in Financial Futures report, the no-key history-file download, and the gotchas that bite pipelines.
# Verified 2026-06-09 · tested with live no-key download of the CFTC Traders in Financial Futures history file (cftc.gov, fut_fin_txt_2023.zip -> FinFutYY.txt)
# Tags: futures, positioning, derivatives, free, no-api-key, time-series, data:cftc-cot
==============================================================================

**CFTC Commitments of Traders (COT)** is a weekly report published by the U.S.
Commodity Futures Trading Commission (CFTC). Each release gives aggregate open
futures positions (a separate version covers futures and options combined),
broken down by trader category for each reported market. Positions are measured
as of Tuesday and released the following Friday. The data is free, public, and
needs no API key. Used in, for example,
[Siriwardane, Sunderam & Wallen](/wiki/papers/jf/2025/siriwardane-segmented-arbitrage-2025/),
where it supplies weekly position quantities by dealer, hedge fund, and
asset-manager type for futures-based trades.

- **Cost:** free, public.
- **API key:** none required.
- **Coverage:** weekly aggregate positions by trader category. The Traders in
  Financial Futures (TFF) report runs from June 2010 onward; the Legacy report
  goes back to 1986, with some series earlier.
- **Home:** <https://www.cftc.gov/MarketReports/CommitmentsofTraders/index.htm>

## Access

Historical annual files are compressed and live under
<https://www.cftc.gov/files/dea/history/>. The paper above uses the Traders in
Financial Futures (TFF) report, futures-only. Download one year directly, no
authentication:

```bash
# Financial-futures (TFF), futures-only, 2023 history file, no key
curl -sL -o fut_fin_txt_2023.zip \
  "https://www.cftc.gov/files/dea/history/fut_fin_txt_2023.zip"
unzip fut_fin_txt_2023.zip   # -> FinFutYY.txt  (~2.1 MB)
```

The archive unzips to a single text file, `FinFutYY.txt`, holding the
financial-futures, futures-only data for that year.

The filename encodes the report family and version: `fut_fin_txt_<year>.zip`
is financial-futures (TFF), futures-only. The combined (futures and options)
version and the other report families (Legacy, Disaggregated, Supplemental)
use different prefixes under the same `/files/dea/history/` directory. List
that directory to find the exact file you need before downloading.

### Load in Python

```python
import pandas as pd

df = pd.read_csv(
    "FinFutYY.txt",
    dtype=str,           # read as str first; cast after inspecting the schema
    low_memory=False,
)

# The TFF position columns hold counts of contracts by trader category, e.g.
# Dealer/Intermediary, Asset Manager/Institutional, Leveraged Funds,
# Other Reportables, plus a Nonreportable residual.
```

Inspect the header row to confirm the exact column names for the vintage you
pulled; the layout differs across report families.

## Gotchas (the ones that bite pipelines)

- **Futures-only versus futures-and-options-combined are different files.**
  They live under different filenames in the same directory. Pick one
  deliberately and do not mix the two in a single series.
- **Dated Tuesday, released Friday.** The snapshot is measured as of Tuesday
  but published the following Friday. Align event windows to the Tuesday
  "as of" date, not the release date, or you bake in a reporting lag.
- **Nonreportable is a residual, not a measured group.** The report counts
  only reportable positions. The Nonreportable figure is what is left over
  after the reportable categories, not a surveyed trader type.
- **Classification is self-reported and can change.** Traders report their own
  category to the CFTC and can be reclassified over time. A category's series
  is not a fixed set of firms across the sample.
- **Each row aggregates across all contract months.** A row covers a whole
  market, not a single contract month, so the data is not contract-month-level.
- **TFF covers financial futures only.** Interest rates, FX, and equity indices
  are in TFF. Physical commodity markets (energy, metals, agriculture) are in
  the Disaggregated report, not TFF. Use the report family that matches the
  market.
- **Market identifiers differ across report families.** Contract-market names
  and codes are not consistent between Legacy, Disaggregated, and TFF; build a
  mapping before joining across families or to other datasets.

## Report families

| Family | Trader categories | History |
|--------|-------------------|---------|
| Legacy | Commercial, Non-Commercial, Nonreportable | from 1986, some series earlier |
| Disaggregated | Producer/Merchant, Swap Dealers, Managed Money, Other Reportables (physical commodity markets) | from 2006 |
| Traders in Financial Futures (TFF) | Dealer/Intermediary, Asset Manager/Institutional, Leveraged Funds, Other Reportables, plus Nonreportable residual (financial futures) | from June 2010 |
| Supplemental | Index trader breakout for selected agricultural markets | from 2006 |

## Citation

Cite the CFTC, the COT report family used, the year or history file, the URL,
and the access date, for example: *U.S. Commodity Futures Trading Commission,
Commitments of Traders, Traders in Financial Futures (futures-only), 2023
history file, retrieved from https://www.cftc.gov/files/dea/history/, accessed
YYYY-MM-DD.* Record the report family and futures-only-versus-combined choice
so the pull is reproducible.


==============================================================================
# CMS Hospital Quality & Patient-Outcome Metrics (Care Compare)
# https://instituteforautomatedresearch.org/wiki/datasets/cms-quality/
# How to pull CMS hospital quality measures (mortality, readmissions, complications, HCAHPS patient satisfaction) from the no-key Provider Data Catalog API, and the gotchas that bite pipelines (risk-adjusted not raw, suppressed small cells, measures and vintages change, footnote codes).
# Verified 2026-06-22 · tested with live no-key pull of the CMS Provider Data Catalog API (metastore) + the "Complications and Deaths - Hospital" CSV (95,840 rows)
# Tags: healthcare, hospitals, quality, free, no-api-key, data:cms-quality
==============================================================================

**CMS hospital quality and patient-outcome metrics** are the publicly reported
performance measures behind Medicare's **Care Compare** (formerly Hospital
Compare): risk-adjusted mortality and readmission rates, complication rates,
healthcare-associated infections, timely/effective-care process measures, and
**HCAHPS** patient-satisfaction survey results, reported per hospital. The data
is free and public. Used in, for example,
[Lewellen](/wiki/papers/jf/2025/lewellen-women-charge-evidence-hospitals-2025/),
where CMS patient-outcome metrics (mortality, readmissions, patient
satisfaction) are nonfinancial hospital-performance proxies in supplementary
tests.

- **Cost:** free, public.
- **API key:** none required.
- **Coverage:** all Medicare-certified hospitals (~4,000-5,000 facilities);
  many measures published from 2008-2009 onward, refreshed roughly quarterly.
- **Home:** <https://data.cms.gov/provider-data/>

## Access

The Provider Data Catalog exposes a DKAN/metastore API. List datasets, then pull
a dataset's CSV distribution. No authentication:

```bash
# 1. List all provider-data datasets (titles + identifiers)
curl -sL "https://data.cms.gov/provider-data/api/1/metastore/schemas/dataset/items?show-reference-ids=false"

# 2. Resolve one dataset's CSV download URL (e.g. Complications and Deaths - Hospital)
curl -sL "https://data.cms.gov/provider-data/api/1/metastore/schemas/dataset/items/ynj2-r877?show-reference-ids=true"

# 3. Download the CSV using the downloadURL from the step-2 response.
#    The resources/<hash>_<id>/ subdirectory is in that response and changes
#    each refresh, so read it from the API rather than hard-coding it:
curl -sL -o complications_deaths_hospital.csv \
  "https://data.cms.gov/provider-data/sites/default/files/resources/6af7c44d77436e5a1caac3ce39a83fe9_1777413950/Complications_and_Deaths-Hospital.csv"
```

Each measure family is published at three grains as separate datasets, suffixed
`- Hospital`, `- State`, and `- National` (for example `ynj2-r877` Complications
and Deaths - Hospital, `dgck-syfz` Patient survey (HCAHPS) - Hospital,
`77hc-ibv8` Healthcare Associated Infections - Hospital). A hospital row carries
`Facility ID` (the CMS Certification Number, CCN), `Measure ID`, `Score`,
`Denominator`, `Lower Estimate`, `Higher Estimate`, and a `Compared to National`
flag.

### Load in Python

```python
import pandas as pd

df = pd.read_csv("complications_deaths_hospital.csv", dtype={"Facility ID": str})
# Long format: one row per (Facility ID, Measure ID). Pivot to a hospital panel:
wide = df.pivot_table(index="Facility ID", columns="Measure ID",
                      values="Score", aggfunc="first")
```

## Gotchas (the ones that bite pipelines)

- **Scores are risk-adjusted, not raw rates.** Mortality, readmission, and
  complication measures are CMS risk-standardized rates (RSMR/RSRR), modeled to
  adjust for patient case mix. They are **not** raw counts divided by
  admissions, and the adjustment model changes over time. Do not compare a CMS
  risk-adjusted rate to a hand-computed raw rate as if they measure the same
  thing.
- **`Score` is text with footnote codes, not always a number.** Cells for
  suppressed or not-applicable measures hold footnote codes (for example "Not
  Available", "Number of Cases Too Small") rather than a value. Parse the
  companion footnote columns and coerce `Score` to numeric explicitly; a naive
  `float()` will choke or silently drop rows.
- **Small-volume hospitals are suppressed.** Measures are withheld when the
  denominator is below a threshold (commonly 25 cases). This censors small and
  rural hospitals non-randomly, so a sample of hospitals with non-missing scores
  is selected toward larger facilities. Account for this before treating missing
  as random.
- **Measures and reporting periods change across refreshes.** CMS adds, retires,
  and re-specifies measures, and each measure covers a rolling multi-year
  window (often three years of discharges) that is **not** the calendar year of
  the file. Two vintages can have different `Measure ID` sets and different
  underlying periods; pin the refresh date and read the measure's data-collection
  window, do not assume the file year is the data year.
- **Join key is the CCN, and it is a string.** `Facility ID` is the 6-character
  CMS Certification Number. For the acute-care, critical-access, and psychiatric
  hospitals in Care Compare it is all digits with leading zeros (for example
  `140010`), so read it as a string or the leading zeros are lost. It is not the
  NPI, not the AHA ID, and not an EIN; crosswalking to AHA or HCRIS requires the
  CCN explicitly.
- **HCAHPS is survey-based and differently scaled.** Patient-satisfaction
  (HCAHPS) measures are "top-box" percentages from a sampled survey with their
  own response-rate and mode adjustments, on a different scale from the clinical
  outcome measures. Do not pool HCAHPS percentages with risk-adjusted rates
  without rescaling.

## Measure families (per-hospital datasets)

| Family | Example dataset id | What it reports |
|---|---|---|
| Complications and Deaths | `ynj2-r877` | Risk-adjusted mortality and complication rates |
| Unplanned Hospital Visits | (readmissions) | Risk-adjusted readmission / return rates |
| Healthcare Associated Infections | `77hc-ibv8` | CLABSI, CAUTI, SSI, MRSA, C. diff |
| Patient survey (HCAHPS) | `dgck-syfz` | Patient-experience top-box percentages |
| Timely and Effective Care | (process) | Process-of-care / timeliness measures |

(Each also has `- State` and `- National` siblings for the benchmark levels.)

## Citation

Cite CMS, Care Compare / Provider Data Catalog, the dataset name and identifier,
the refresh date, and the access date, for example: *Centers for Medicare and
Medicaid Services, Care Compare: Complications and Deaths - Hospital
(data.cms.gov/provider-data, dataset ynj2-r877), accessed YYYY-MM-DD.* Record
the dataset identifier and refresh date, since measures and periods change
between vintages.


==============================================================================
# CRA disclosure data (FFIEC)
# https://instituteforautomatedresearch.org/wiki/datasets/cra-ffiec/
# How the FFIEC Community Reinvestment Act small-business, small-farm, and community-development lending files are structured, plus why a pipeline cannot fetch them no-key (the FFIEC host returns a Cloudflare challenge to automated requests), the fixed-width record-type layouts, and the disclosure vs aggregate vs transmittal split.
# Tags: banking, lending, small-business, filings, panel-data, free, no-api-key, academic, data:cra-ffiec
==============================================================================

:::note[Open data, but not fetchable no-key from this session]
CRA disclosure data is **free and public** (no account, no key, no fee), but
every FFIEC host path (`www.ffiec.gov`) returned a **Cloudflare bot challenge**
(HTTP 403 with a ~244 KB HTML body) to automated requests from this session, and
no non-blocked mirror of the actual data files could be confirmed. The page
therefore carries **no provenance badge**: it documents the dataset and access
path, but the end-to-end pull was **not** exercised here. A human browser can
download the files directly; an unattended pipeline needs a headless browser that
solves the JS challenge, or a manually cached copy.
:::

**CRA disclosure data** is the FFIEC's annual record of bank lending to small
businesses, small farms, and communities under the Community Reinvestment Act.
For each reporting institution it gives the count and dollar volume of
small-business and small-farm loans by census tract, plus community-development
lending. It is the standard free source for studying where banks lend to small
firms, distinct from the mortgage-focused [HMDA](/wiki/datasets/hmda/). Used in,
for example,
[Rehbein & Rother (2025)](/wiki/papers/rfs/2025/rehbein-social-connectedness-bank-lending-2025/),
where county-to-county CRA small-business loan volumes are the dependent
variable linked to social connectedness.

- **Cost:** free, public (no account, no key, no fee).
- **API key:** none (the data are flat files, not an API).
- **Coverage:** annual, roughly 1996 to the latest release; institution-level
  disclosure files plus MSA/county aggregate files.
- **Content:** small-business, small-farm, and community-development lending
  (NOT home mortgages; that is HMDA).
- **Home (browser only):** <https://www.ffiec.gov/data/cra>

## Access

The files live under `https://www.ffiec.gov/data/cra` (Data Products), with the
bulk flat files at `.../cra/flat-files` and interactive query tools at
`https://www.ffiec.gov/craadweb/`. All sit on the same Cloudflare-gated host, so
the path below works in a browser but is blocked for unattended automation.

```bash
# Reachable in a browser; an automated GET returns a Cloudflare challenge
# page (HTTP 403, HTML body) rather than the zip:
#   https://www.ffiec.gov/data/cra/flat-files
# Disclosure (institution-by-institution) and Aggregate (MSA/county totals)
# flat files are released per year, alongside the fixed-width layout specs:
#   .../flat-files/<yy>FlatDiscSpecs.pdf   (Disclosure layout)
#   .../flat-files/<yy>FlatAggSpecs.pdf    (Aggregate layout)
```

Validate any download by magic bytes before trusting it: a real flat-file zip
starts with `PK\x03\x04`; the Cloudflare challenge is HTML even when the request
named a `.zip`.

## Gotchas (the ones that bite pipelines)

- **The host bot-blocks, and the block masquerades as data.** A naive fetch
  gets HTTP 403 with a ~244 KB HTML challenge body. A pipeline that only checks
  the status code, or that saves the body to a `.zip` without checking the
  magic bytes, silently captures the challenge page. Check `PK\x03\x04`.
- **Fixed-width flat files, not CSV.** Each line is parsed by the published
  column-position layout in the spec PDF; there are no delimiters and no header.
- **Three distinct file families, do not conflate.** Transmittal Sheet (reporter
  roster and identity), Disclosure (institution-by-institution lending), and
  Aggregate (MSA/county totals across all reporters). Pick by analysis level.
- **Multiple record types interleaved in one file.** A single file mixes several
  record formats (e.g. table D1-1, D1-2, D2, ...), keyed by a leading
  record-identifier code; branch the parser on that code per line.
- **Reporting year is not activity year.** Small-business and small-farm lending
  is reported the year after it occurs, so the latest release lags the activity
  it covers.
- **CRA is not HMDA.** CRA here is small-business, small-farm, and
  community-development lending; home mortgages are the separate (and
  no-key-fetchable) [HMDA](/wiki/datasets/hmda/) dataset, now CFPB-hosted.
- **Geography keys need a join.** Records carry FIPS state/county plus MSA/MD
  codes; tract characteristics come from the separate FFIEC Census flat files on
  the same blocked host.

## Reference

| Field | Value |
|-------|-------|
| Host | FFIEC Data Products: `www.ffiec.gov/data/cra` (browser only) |
| File families | Transmittal, Disclosure (institution), Aggregate (MSA/county) |
| Format | Fixed-width flat files, multiple record types per file |
| Layout specs | `<yy>FlatDiscSpecs.pdf`, `<yy>FlatAggSpecs.pdf` |
| Coverage | Annual, roughly 1996 to latest |
| Content | Small-business, small-farm, community-development lending |
| Key required | No (but host bot-blocks automation) |

## Citation

Cite the FFIEC as the source: Federal Financial Institutions Examination Council,
"Community Reinvestment Act (CRA) data," with the specific year, file family
(Disclosure or Aggregate), and retrieval date. Note in a methods appendix that
the files were retrieved through a browser session, since the host blocks
automated access.


==============================================================================
# Dartmouth Atlas of Health Care
# https://instituteforautomatedresearch.org/wiki/datasets/dartmouth-atlas/
# How to pull the Dartmouth Atlas ZIP-to-HSA-to-HRR geographic crosswalk and regional health-care utilization measures from the no-key data downloads, and the gotchas that bite pipelines (HSA/HRR are care markets not Census geographies, the crosswalk vintage matters, utilization is age-sex-adjusted).
# Verified 2026-06-22 · tested with live no-key download of the Dartmouth Atlas ZIP-to-HSA-to-HRR crosswalk (ZipHsaHrr19.csv.zip, ZipHsaHrr19.csv ~1.7 MB)
# Tags: healthcare, geography, hospitals, free, no-api-key, data:dartmouth-atlas
==============================================================================

**The Dartmouth Atlas of Health Care** maps how medical resources are
distributed and used across the United States. Its lasting contribution is two
geographic units built from Medicare patient flows: the **Hospital Service Area
(HSA)**, a local market of ZIP codes whose residents mostly use a given
hospital, and the **Hospital Referral Region (HRR)**, a larger region organized
around tertiary referral care. These are the standard way to define a hospital's
local market. The Atlas also publishes regional utilization, spending, and
workforce measures. The data is free and public. Used in, for example,
[Lewellen](/wiki/papers/jf/2025/lewellen-women-charge-evidence-hospitals-2025/),
where Dartmouth Atlas HSAs supply the geographic market definitions for
hospitals (with demographics drawn from the 2010 Census).

- **Cost:** free, public.
- **API key:** none required.
- **Coverage:** national; ZIP-to-HSA-to-HRR crosswalk plus regional utilization
  and spending measures, derived from Medicare claims.
- **Home:** <https://www.dartmouthatlas.org/> (data at
  <https://data.dartmouthatlas.org/>)

## Access

The geographic crosswalk and the measure files are plain downloads, no
authentication:

```bash
# ZIP -> HSA -> HRR crosswalk (the workhorse file; vintage in the name)
curl -sL -o ZipHsaHrr19.csv.zip \
  "https://data.dartmouthatlas.org/downloads/geography/ZipHsaHrr19.csv.zip"
unzip ZipHsaHrr19.csv.zip   # -> ZipHsaHrr19.csv
```

The crosswalk columns are `zipcode19, hsanum, hsacity, hsastate, hrrnum,
hrrcity, hrrstate`: every ZIP maps to exactly one HSA and one HRR, labelled by
the dominant city and state. Utilization, spending, and post-discharge measure
files (by HSA, HRR, hospital, or state) are downloadable from the same data
portal under their topic sections.

### Load in Python

```python
import pandas as pd

xw = pd.read_csv("ZipHsaHrr19.csv", dtype={"zipcode19": str})
# Map a hospital's ZIP to its market:
zip_to_hrr = xw.set_index("zipcode19")["hrrnum"].to_dict()
# ~3,400 HSAs nest into 306 HRRs nationally.
```

## Gotchas (the ones that bite pipelines)

- **HSAs and HRRs are care markets, not Census or postal geographies.** They are
  built from where Medicare patients actually go for care, so they cross county
  and state lines and do not nest inside CBSAs, counties, or ZIP-code tabulation
  areas. Do not treat an HRR as a metro area or merge it to county-level Census
  data without an explicit, area-weighted crosswalk.
- **The crosswalk is versioned; match the vintage to your sample.** The file is
  named for its ZIP vintage (`ZipHsaHrr19` = 2019 ZIPs). ZIP codes are added,
  retired, and reassigned over time, and the HSA/HRR boundaries were defined from
  a specific Medicare era (the canonical 1990s definitions, periodically
  refreshed). Using a 2019 crosswalk on 2005 ZIPs (or vice versa) silently
  drops or misassigns codes. Pick the crosswalk vintage closest to your data.
- **It is built from Medicare (elderly) flows.** The market definitions and most
  utilization measures derive from fee-for-service Medicare claims, so they
  describe the over-65 population's care-seeking. Applying them to commercially
  insured or pediatric populations imports an age-skewed market definition; note
  the assumption.
- **Utilization and spending measures are age-sex-(race-)adjusted.** The Atlas
  reports rates standardized for demographic mix (and price-adjusted for
  spending), not raw counts. Two regions' "spending" are comparable by design
  but are not the dollars actually spent; read each measure's adjustment notes
  before using it as a level.
- **One ZIP, one HSA, one HRR; population is not split.** The crosswalk assigns
  each ZIP wholly to a single HSA/HRR by dominant flow, with no fractional
  allocation. A ZIP that straddles two markets is assigned to one, so
  aggregating ZIP-level counts to HRRs is exact but the boundary is approximate.
- **`hrrnum`/`hsanum` are the join keys, not the city labels.** Match on the
  numeric `hrrnum` (306 HRRs) and `hsanum` codes; the `hrrcity`/`hsacity` labels
  are not unique and change formatting across files.

## Geographic units

| Unit | Count (national) | Definition |
|---|---|---|
| ZIP code | ~40,000 | Postal ZIP (vintage-dated in the file name) |
| HSA (Hospital Service Area) | ~3,400 | Local hospital market: ZIPs whose residents use a given hospital |
| HRR (Hospital Referral Region) | 306 | Tertiary-care referral region (nests HSAs) |

## Citation

Cite the Dartmouth Atlas, the file and its vintage, the URL, and the access
date, for example: *The Dartmouth Atlas of Health Care, ZIP-to-HSA-to-HRR
crosswalk (ZipHsaHrr19), retrieved from https://data.dartmouthatlas.org/,
accessed YYYY-MM-DD.* Record the crosswalk vintage and, for measures, the
reporting year and adjustment, so the geography and rates are reproducible.


==============================================================================
# DFA: Distributional Financial Accounts (Federal Reserve)
# https://instituteforautomatedresearch.org/wiki/datasets/dfa/
# How to download the Federal Reserve Board's Distributional Financial Accounts, which give quarterly estimates of US household wealth distribution by wealth percentile, generation, education, and race, reconciled to Z.1 aggregates, with no key required.
# Verified 2026-06-09 · tested with live no-key download of dfa.zip CSV tables from federalreserve.gov
# Tags: macro, wealth-distribution, free, no-api-key, federal-reserve, time-series, data:dfa
==============================================================================

**Note:** DFA here refers to the Federal Reserve Board's Distributional Financial Accounts, not Dimensional Fund Advisors.

**The Distributional Financial Accounts (DFA)** are the Federal Reserve Board's quarterly estimates of the distribution of US household wealth (assets, liabilities, and net worth) across groups defined by wealth percentile, income, generation (age cohort), education, and race. The series are reconciled to the Financial Accounts of the United States (Z.1) totals. Used in, for example, [Catherine, Miller & Sarin](/wiki/papers/jf/2025/catherine-social-security-trends-wealth-2025/) for the aggregate value of defined-benefit pension obligations by wealth group.

- **Cost:** free, no paywall, no key.
- **Coverage:** US households, quarterly, from 1989 Q3.
- **Breakdowns:** wealth percentile (top 0.1%, 1%, 10%, next 40%, bottom 50%), generation (Silent and earlier, Baby Boomer, Gen X, Millennial, Gen Z and younger), education (college degree, some college, no college), race (White non-Hispanic, Black and African American, Hispanic, Other).
- **Home:** <https://www.federalreserve.gov/releases/z1/dataviz/dfa/>

## Access

The entire release is a single zip of CSV tables, downloadable with no authentication:

```
https://www.federalreserve.gov/releases/z1/dataviz/download/zips/dfa.zip
```

The zip contains files such as:

- `dfa-networth-levels-detail.csv` (net worth levels by wealth percentile)
- `dfa-generation-levels.csv` (levels by generation)
- `dfa-education-shares.csv` (shares by education)
- `dfa-race-shares.csv` (shares by race)
- `dfa-data-definitions.txt` (variable definitions and metadata)

### Shell

```bash
curl -O https://www.federalreserve.gov/releases/z1/dataviz/download/zips/dfa.zip
unzip dfa.zip -d dfa/
```

### Python

```python
import io, zipfile, urllib.request
import pandas as pd

url = "https://www.federalreserve.gov/releases/z1/dataviz/download/zips/dfa.zip"
with urllib.request.urlopen(url) as resp:
    zf = zipfile.ZipFile(io.BytesIO(resp.read()))

# list available tables
print(zf.namelist())

# load net worth levels by wealth percentile
nw = pd.read_csv(zf.open("dfa-networth-levels-detail.csv"))
print(nw.head())
```

An interactive data visualization tool is also available on the home page above, allowing point-and-click queries by category and date range without any download.

## Gotchas (the ones that bite pipelines)

- **Model-based distribution, not direct measurement.** The DFA interpolate the triennial Survey of Consumer Finances (SCF) onto quarterly Z.1 aggregates. Quarterly distributional movements between SCF waves are smoothed and imputed, not directly observed. Treat quarter-over-quarter changes in distributional shares with appropriate skepticism, especially in non-SCF years.
- **Definitions follow Z.1, not household intuition.** Because the DFA reconcile to Financial Accounts totals, category definitions (e.g. defined-benefit pension entitlements as a specific Z.1 line item) may differ from what a survey respondent or a household-level study would count. Read `dfa-data-definitions.txt` before constructing any derived series.
- **Wealth groups are contemporaneous, not cohorts.** The "top 1%" series tracks whichever households are in the top 1% in each period. It does not follow the same households over time. Do not interpret changes in group-level holdings as individual-level dynamics.
- **Breakdowns are not cross-tabulated.** The percentile breakdown, the generation breakdown, the education breakdown, and the race breakdown are separate files. There is no joint distribution (e.g. "top 10% by wealth among Black households") in the public release.
- **Vintage dependence.** The DFA are revised whenever the underlying SCF or Z.1 is revised. A replication that does not pin the download vintage may not reproduce exactly. Record the access date and, where possible, the release label printed in the data definitions file.

## Standard operations

- **Shares vs. levels:** the zip provides both. Use shares when comparing relative concentration across time; use levels when calibrating a model that requires dollar magnitudes. Never mix the two without rescaling.
- **Reconciliation check:** sum the wealth-percentile group levels; they should equal the Z.1 household net worth aggregate for that quarter (small rounding differences are expected).
- **Merging with Z.1:** the DFA are designed to be used alongside Z.1 tables. Match by date column and verify the aggregate before constructing group-level series derived from other Z.1 lines.
- **Always state the vintage and breakdown dimension** (percentile, generation, education, or race) in any result that relies on these data.

## Citation

Cite the Federal Reserve Board Distributional Financial Accounts with the release vintage date and the specific table (levels or shares, and the breakdown dimension) used. Example:

*Board of Governors of the Federal Reserve System, Distributional Financial Accounts: Net Worth Levels by Wealth Percentile [dfa-networth-levels-detail.csv], release dated YYYY-QQ, downloaded from https://www.federalreserve.gov/releases/z1/dataviz/dfa/, accessed YYYY-MM-DD.*

The home page lists the current release date; record it at time of download.


==============================================================================
# ECB Data Portal (Statistical Data Warehouse)
# https://instituteforautomatedresearch.org/wiki/datasets/ecb-data-warehouse/
# How to pull euro-area macro, monetary, and Eurosystem balance-sheet series from the ECB, including the no-key data-api CSV endpoint, series-key structure, and the gotchas that bite pipelines.
# Verified 2026-06-09 · tested with live no-key CSV pull of the ECB data-api (EXR/D.USD.EUR.SP00.A daily euro reference rate, csvdata format)
# Tags: macro, monetary-policy, euro-area, central-bank, free, no-api-key, time-series, data:ecb-data-warehouse
==============================================================================

**The ECB Data Portal** (formerly the Statistical Data Warehouse, SDW) is the
European Central Bank's public statistics service: euro-area monetary
aggregates, interest rates, exchange rates, bank balance sheets, securities
holdings, and the Eurosystem's own asset-purchase data (APP and PEPP). The data
is free and public, served over an SDMX REST API that needs no key. Used in, for
example,
[Costain, Nuno & Thomas](/wiki/papers/jf/2025/costain-term-structure-interest-rates-2025/),
where it supplies Eurosystem bond holdings and purchase paths to calibrate net
bond supply and the maturity distribution of holdings under APP and PEPP.

- **Cost:** free, public.
- **API key:** none required.
- **Coverage:** euro-area and member-state series, many from 1999 onward;
  exchange rates and some series earlier; daily, monthly, quarterly frequencies.
- **Home:** <https://data.ecb.europa.eu>

## Access

The data API serves any series by its dataset and series key. Append
`?format=csvdata` for a flat CSV, no authentication:

```bash
# Daily euro reference exchange rate, USD per EUR, last 3 observations, no key
curl -sL -H "Accept: text/csv" \
  "https://data-api.ecb.europa.eu/service/data/EXR/D.USD.EUR.SP00.A?lastNObservations=3&format=csvdata" \
  -o eur_usd.csv
```

The path is `service/data/<DATAFLOW>/<SERIES_KEY>`. The dataflow (here `EXR`,
exchange rates) sets the dimension order; the series key is the
dot-separated dimension values (`D.USD.EUR.SP00.A` = daily, USD, against EUR,
spot, average). Leave a dimension blank to fetch all of its values, or use `+`
to list several. Find the dataflow and its dimensions by browsing the dataset
on <https://data.ecb.europa.eu> and copying the series key it shows.

### Load in Python

```python
import requests, io, pandas as pd

r = requests.get(
    "https://data-api.ecb.europa.eu/service/data/EXR/D.USD.EUR.SP00.A",
    params={"format": "csvdata", "startPeriod": "2020-01-01"},
    headers={"Accept": "text/csv"},
    timeout=60,
)
df = pd.read_csv(io.StringIO(r.text))
# Key columns: KEY (series id), TIME_PERIOD, OBS_VALUE.
```

## Gotchas (the ones that bite pipelines)

- **The SDW was rebranded the ECB Data Portal; old SDW URLs may not resolve.**
  `sdw.ecb.europa.eu` links and the legacy SDW API have been superseded by
  `data.ecb.europa.eu` and `data-api.ecb.europa.eu`. Point pipelines at the
  current host.
- **Series keys are positional and dataflow-specific.** A key like
  `D.USD.EUR.SP00.A` only makes sense against the `EXR` dimension order; the
  same dotted string means nothing in another dataflow. Always pair a key with
  its dataflow, and read the dimension order from the dataset page.
- **Asset-purchase series are split across PSPP, CSPP, CBPP, ABSPP, and PEPP.**
  The APP is the sum of its component programmes; PEPP is reported separately.
  Adding holdings across programmes requires picking the right component series
  and not double-counting the APP total against its parts.
- **Book value, not market value.** Eurosystem securities-holdings series are
  generally reported at amortised book value, and purchase data is at
  settlement. Do not treat them as mark-to-market positions.
- **Blank dimension means "all", which can return a large multi-series file.**
  Omitting a dimension value fetches every series along that dimension. Filter
  deliberately, or a single call can pull thousands of series.
- **Revisions and breaks.** Monetary and balance-sheet statistics are revised,
  and methodology breaks exist across the sample. Record the access date and
  watch for series discontinuities before splicing.

## Common dataflows

| Dataflow | Content |
|----------|---------|
| `EXR` | Exchange rates (reference and effective) |
| `BSI` | Bank (MFI) balance-sheet items, monetary aggregates |
| `FM` | Financial-market and money-market interest rates |
| `ILM` | Eurosystem balance-sheet and liquidity items |
| `SEC` | Securities issues and holdings |
| `YC` | Euro-area yield curve (AAA and all-bond) |

Confirm the dataflow code and its dimensions on the portal before building a
series key; the list above is a starting point, not the full set.

## Citation

Cite the European Central Bank, the ECB Data Portal, the series key and
dataflow, the URL, and the access date, for example: *European Central Bank,
ECB Data Portal, series EXR/D.USD.EUR.SP00.A, retrieved from
https://data.ecb.europa.eu, accessed YYYY-MM-DD.* Record the series key and
frequency so the pull is reproducible.


==============================================================================
# SEC EDGAR: filings, financials, full-text search
# https://instituteforautomatedresearch.org/wiki/datasets/edgar/
# How to pull SEC filings, XBRL financial facts, insider trades, and institutional holdings from EDGAR for free: the User-Agent trap, the 10 req/s limit, and XBRL-vs-text, for automated pipelines.
# Verified 2026-05-29 · tested with live data.sec.gov + efts.sec.gov fetch (submissions, XBRL, full-text search, N-1A form rows)
# Tags: fundamentals, filings, event-data, free, no-api-key, sec, data:edgar
==============================================================================

**SEC EDGAR** is the free, authoritative source for US corporate disclosure:
10-K/10-Q/8-K filings, standardized **XBRL financial facts** (cross-company
comparable), insider trades (Form 4), institutional holdings (13F), proxy
statements, and full-text search across everything. No API key, just a
`User-Agent` header. It is the corporate-finance backbone the
[ZeroPaper](https://github.com/alejandroll10/zeropaper) pipeline uses for
governance, disclosure, and fundamentals work. This page is the distilled
access recipe.

- **Cost:** free, no paywall, no key.
- **Auth:** none, but a descriptive `User-Agent` header is **mandatory**.
- **Coverage:** all US public-company filings; XBRL facts from ~2009.
- **Home:** <https://www.sec.gov/edgar> ·
  **API:** <https://data.sec.gov> ·
  **Full-text:** <https://efts.sec.gov>

## Access

### Option 1: `edgartools` (preferred, structured data)

```python
# pip install edgartools
from edgar import Company, set_identity

# Identity = your User-Agent. Required by the SEC. Keep it in .env, not here.
set_identity("Your Name your@email.edu")

company = Company("AAPL")                 # ticker → CIK resolved for you
tenks = company.get_filings(form="10-K")
filing = tenks[0]
print(filing.filing_date, filing.accession_no)

facts = company.get_facts()               # XBRL, cross-company comparable
revenue = facts.to_pandas("us-gaap:Revenues")

form4 = company.get_filings(form="4")[0].obj()   # insider trades
```

### Option 2: Direct REST API (no package)

```python
import requests
headers = {"User-Agent": "Your Name your@email.edu"}  # mandatory

# All filings + metadata for a company (note 10-digit zero-padded CIK):
requests.get("https://data.sec.gov/submissions/CIK0000320193.json",
             headers=headers).json()

# One XBRL concept across time:
requests.get("https://data.sec.gov/api/xbrl/companyconcept/"
             "CIK0000320193/us-gaap/Revenues.json", headers=headers).json()

# Full-text search:
requests.get("https://efts.sec.gov/LATEST/search-index?"
             "q=%22climate+risk%22&forms=10-K", headers=headers).json()
```

Store the name/email in `.env` (e.g. `SEC_EDGAR_NAME`, `SEC_EDGAR_EMAIL`);
never hard-code them.

## Gotchas (the ones that bite pipelines)

The reason to read this page rather than the SEC docs. Verified against live
endpoints on the date above.

- **No `User-Agent` → HTTP 403.** This is the #1 EDGAR failure. A request
  with no (or a default `python-requests`) User-Agent is rejected outright;
  confirmed 403 live. Always send a descriptive `Name email` string.
- **Rate limit: 10 requests/second, hard.** `edgartools` throttles for you;
  for the direct API add `time.sleep(0.1)` between calls and never parallelize
  blindly; sustained bursts get the host IP blocked, not just throttled.
- **CIK must be 10-digit zero-padded** in `data.sec.gov` URLs
  (`CIK0000320193`, not `CIK320193` or `320193`). `edgartools`'
  `Company("TICKER")` hides this; the raw API does not.
- **Use XBRL for cross-company work, not filing text.** Narrative text and
  table formatting vary by filer and year; `us-gaap:*` XBRL facts are
  standardized and comparable. Only parse text when the datum isn't tagged.
- **XBRL coverage starts ~2009** and tag usage drifts: revenue may be
  `us-gaap:Revenues` *or*
  `us-gaap:RevenueFromContractWithCustomerExcludingAssessedTax`. Check both.
- **Cache aggressively.** Save XBRL facts to `data/*.parquet` and check before
  re-downloading; re-pulling companyfacts for a panel will blow the rate
  limit fast.
- **Amendments & restatements.** `10-K/A` supersedes `10-K`; a company can
  restate prior XBRL facts. Pin the accession number when reproducibility
  matters.

## Key filing types

| Form | Contents | Use for |
|------|----------|---------|
| `10-K` | Annual report | Financials, risk factors, business description |
| `10-Q` | Quarterly report | Interim financials |
| `8-K` | Current report | Material events (M&A, earnings, mgmt changes) |
| `DEF 14A` | Proxy statement | Executive comp, board, governance |
| `4` | Insider trades | Director/officer buy/sell transactions |
| `13F-HR` | Institutional holdings | Quarterly positions of large investors |
| `S-1` | IPO registration | Pre-IPO financials, risk factors |
| `SC 13D/G` | Beneficial ownership | >5% shareholder positions |
| `N-1A` | Open-end fund registration | Mutual fund / ETF prospectus, strategy, fees, classification |

## Common XBRL facts

| Concept | Tag |
|---------|-----|
| Revenue | `us-gaap:Revenues` / `us-gaap:RevenueFromContractWithCustomerExcludingAssessedTax` |
| Net income | `us-gaap:NetIncomeLoss` |
| Total assets | `us-gaap:Assets` |
| Total equity | `us-gaap:StockholdersEquity` |
| EPS (basic) | `us-gaap:EarningsPerShareBasic` |
| Shares outstanding | `us-gaap:CommonStockSharesOutstanding` |
| Cash | `us-gaap:CashAndCashEquivalentsAtCarryingValue` |
| Long-term debt | `us-gaap:LongTermDebt` |
| R&D expense | `us-gaap:ResearchAndDevelopmentExpense` |

## Standard recipes

**Panel of fundamentals across firms**: loop tickers, pull `get_facts()`,
take the XBRL concept you need, cache each to parquet, then assemble:

```python
from edgar import Company
import pandas as pd

rows = []
for t in ["AAPL", "MSFT", "GOOGL", "AMZN", "META"]:
    rev = Company(t).get_facts().to_pandas("us-gaap:Revenues")
    rows.append({"ticker": t, "rev_latest": rev.iloc[-1] if len(rev) else None})
df = pd.DataFrame(rows)
```

**Full-text search for a research topic**: `efts.sec.gov/LATEST/search-index`
returns hit counts and snippets; use it to scope a sample before downloading
filings (`data["hits"]["total"]["value"]`).

**Insider-trading study**: iterate a company's Form 4 filings and read
`.obj().transactions` (a DataFrame of trades) per filing.

## Form N-1A: open-end fund registration

`N-1A` is the registration statement and prospectus for **open-end
investment companies**: mutual funds and most ETFs. It is the EDGAR source
for what a fund *says it is*: investment objective, strategy, fee table,
share classes, adviser. Papers that classify funds (e.g. growth vs value,
active vs index) read N-1A prospectus text; the
[Kwan, Liu & Matthies](/wiki/papers/jf/2026/kwan-liu-matthies-2026/) attention paper
uses it for fund classification. Confirmed live on the verified date:
`efts.sec.gov` full-text search returns 3,473 `N-1A` hits, and
`data.sec.gov/submissions` returns `N-1A` and `N-1A/A` rows for a registrant.

Pull a fund's N-1A filings the same way as any other form:

```python
from edgar import Company
filings = Company("0002100194").get_filings(form="N-1A")   # also matches N-1A/A
```

Or by form across all registrants via full-text search
(`forms=N-1A` on `efts.sec.gov`).

### N-1A gotchas (fund filings are not company filings)

- **No `us-gaap` XBRL facts.** N-1A is a registration document, not a
  financial report; the XBRL on it is the **risk/return summary** taxonomy
  (`rr:*`), not `us-gaap:*`. Do not expect `get_facts()` financials here.
- **A fund family files under one registrant, many series and classes.** A
  single N-1A filer (the trust) can cover dozens of funds (**series**) each
  with multiple share **classes**, keyed by EDGAR `S######` / `C######`
  identifiers, not a ticker. Resolve series/class before attributing a
  prospectus to a fund.
- **`485BPOS` / `485APOS` carry the updates.** The initial `N-1A` is filed
  once; ongoing annual prospectus updates arrive as `485BPOS` (immediately
  effective) and `485APOS` (post-effective amendment). For a current
  prospectus, follow the 485 stream, not the original N-1A.
- **ETFs file N-1A too.** Most ETFs register as open-end funds, so they are
  N-1A filers; only a few structures (e.g. some commodity pools) are not.
- **Classification is text, not a tagged field.** The investment objective
  and strategy are prose in the prospectus; deriving a clean style label
  means parsing text or mapping the SEC series/class metadata, not reading a
  single field.

## Citation

Cite the filing and source, e.g.: *Apple Inc., Form 10-K, fiscal year 2024,
filed [date], accession [no.], U.S. Securities and Exchange Commission EDGAR;
https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=0000320193,
accessed YYYY-MM-DD.* For XBRL facts, state the concept tag and the filing the
value came from.


==============================================================================
# EIA Electricity Data
# https://instituteforautomatedresearch.org/wiki/datasets/eia-electricity/
# How to pull US Energy Information Administration electricity data (retail sales and prices, generation by fuel, plant-level operations, CO2 emission factors) from the EIA API v2 with a free key, plus the bracket-encoding, row-limit, and facet gotchas that bite pipelines.
# Verified 2026-06-22 · tested with live keyed EIA API v2 pulls, electricity/retail-sales (annual price) and electricity/electric-power-operational-data (annual coal generation, fueltype COW, 2022)
# Tags: energy, electricity, emissions, industry, time-series, free, eia, data:eia-electricity
==============================================================================

The **US Energy Information Administration (EIA)** publishes the official US
energy statistics. The **electricity** branch covers retail sales, prices, and
revenue by state and customer sector; net generation and fuel consumption by
fuel type; plant-level operating data; operable-generator inventory; and, via
the SEDS and CO2 branches, the emission factors used to convert generation into
carbon. It is the standard source for electricity-sector calibration, including
the brown-electricity emission factor in Pedersen (2026). This page is the
distilled access recipe.

- **Cost:** free, no paywall.
- **API key:** free but **required** (bare requests are rejected).
- **Coverage:** monthly/quarterly/annual electricity series by state and sector,
  generally from 2001; plant-level and hourly grid data on their own histories.
- **Home:** <https://www.eia.gov/electricity/> ·
  **API docs:** <https://www.eia.gov/opendata/documentation.php>

## Access

### Register a key

Get a free key at <https://www.eia.gov/opendata/register.php> (instant email).
Store it in the environment (e.g. `.env` as `EIA_API_KEY=...`); never hard-code
it. The same key works across all EIA branches (coal, natural gas, petroleum,
SEDS, ...), not just electricity.

### Pull a series (API v2)

API v2 is a browsable route tree. Each leaf `/data/` endpoint takes one or more
`data[]` metrics, optional `facets[...][]` filters, and a `frequency`. Query the
parent route (no `/data/`) to discover the available `data` columns, `facets`,
and `frequency` values first.

```python
import os, requests
key = os.environ["EIA_API_KEY"]
r = requests.get(
    "https://api.eia.gov/v2/electricity/retail-sales/data/",
    params={
        "api_key": key, "frequency": "annual",
        "data[0]": "price", "facets[stateid][]": "CA",
        "start": "2022", "end": "2023", "length": 5000,
    },
)
rows = r.json()["response"]["data"]   # period/stateid/sectorid/price/price-units
```

The electricity sub-routes: `retail-sales` (sales, revenue, price, customers by
state and sector), `electric-power-operational-data` (generation and fuel
consumption by fuel type), `rto` (daily/hourly grid operations), `facility-fuel`
(individual plants), `operating-generator-capacity`, and
`state-electricity-profiles`. For emission factors and CO2, use the `seds` and
`co2-emissions` branches.

## Gotchas (the ones that bite pipelines)

The reason to read this page rather than the EIA docs. Verified against live
keyed pulls on the date above.

- **The `data[0]` / `facets[...][]` brackets are real and must survive the
  shell.** With `curl`, the `[` and `]` are glob metacharacters: the request
  silently returns an empty body until you pass `-g` (or `--globoff`). In
  Python's `requests`, pass them as literal param keys (`"data[0]": "price"`) and
  it works. A blank response is almost always this, not an auth failure.
- **You must name at least one `data[]` metric.** A leaf `/data/` call with no
  `data[0]=` returns metadata-style output or an error, not the series. Pick the
  column (`price`, `sales`, `revenue`, `customers`, `generation`, ...) explicitly.
- **5000-row JSON cap, with a quiet warning.** Each call returns at most 5000
  rows; beyond that you get a `"warnings": [{"warning": "incomplete return"}]`
  block and silently truncated data. Constrain with `facets`, `start`/`end`, or
  paginate with `offset`; do not trust an unpaginated bulk pull.
- **Codes, not labels, in facets.** Fuel types are codes (`COW` = all coal,
  `NG` = natural gas, ...), sectors are numeric IDs, states are postal codes plus
  Census-region aggregates (e.g. `location: 90` Pacific). Pull the facet's value
  list from the parent route before filtering.
- **Sales price is cents per kWh, sector-split.** `retail-sales` `price` is in
  cents/kWh and is reported per `sectorid` (residential, commercial, industrial,
  transportation, and `ALL`); do not average sectors naively. Generation is in
  thousand megawatthours.
- **Some endpoints are deprecated mid-flight.** The CO2 aggregates endpoint, for
  example, is marked *deprecated, see SEDS*. Check the route's `name`/`description`
  for a redirect before standardizing a pipeline on it.

## Endpoints you actually need

| Route (under `/v2/electricity/`) | What it is | Key data columns |
|---|---|---|
| `retail-sales` | Sales, price, revenue, customers by state & sector | `price`, `sales`, `revenue`, `customers` |
| `electric-power-operational-data` | Net generation & fuel use by fuel type | `generation`, `total-consumption` |
| `facility-fuel` | Plant-level operations | `generation`, `consumption` |
| `operating-generator-capacity` | Operable generator inventory | `nameplate-capacity-mw` |
| `rto` | Daily/hourly grid demand & interchange | `value` |

Outside `/electricity/`: `/seds/` (State Energy Data System, including emission
coefficients) and `/co2-emissions/` for carbon.

## Standard operations

- **Electricity prices:** `retail-sales` `price` (cents/kWh); filter to one
  `sectorid` and state, and state the sample window.
- **Generation mix / emission factors:** `electric-power-operational-data`
  `generation` by `fueltypeid`; convert to CO2 with the SEDS coefficient for the
  fuel and vintage you need. Pedersen (2026) calibrates its brown-electricity
  factor at ~0.82 kg CO2/kWh; EIA's published coal factors run somewhat higher,
  so use the paper's number only as its stated calibration, not as the EIA coal
  factor.
- **Always paginate or constrain** to stay under the 5000-row cap, and keep the
  facet codes (fuel, sector, state) documented next to any aggregate.

## Citation

U.S. Energy Information Administration. *Electricity data.* Washington, DC: EIA.
Cite the specific series or table and your access date, e.g. *U.S. Energy
Information Administration, Electricity retail sales (annual, by state and
sector), accessed YYYY-MM-DD, https://www.eia.gov/opendata/.*


==============================================================================
# EPA Toxics Release Inventory (TRI)
# https://instituteforautomatedresearch.org/wiki/datasets/epa-tri/
# Facility-by-chemical annual reports of toxic chemical releases and waste management from the U.S. EPA, with the no-key bulk CSV and Envirofacts REST recipes and the gotchas that bite pipelines.
# Verified 2026-06-09 · tested with live no-key pulls of the EPA TRI bulk CSV (mv_tri_basic_download/2022_US) and Envirofacts tri_facility JSON (data.epa.gov)
# Tags: pollution, environmental, esg, free, no-api-key, panel-data, data:epa-tri
==============================================================================

The **Toxics Release Inventory (TRI)** is the U.S. EPA's annual program tracking
releases and other waste management of listed toxic chemicals. Facilities above
size and chemical-throughput thresholds in covered industry sectors must report
under Section 313 of the Emergency Planning and Community Right-to-Know Act
(EPCRA). Each record is at the facility x chemical x year level: on-site and
off-site releases (to air, water, land), plus recycling, energy recovery, and
treatment quantities. Reporting began for 1987. Used in, for example,
[Duchin, Gao & Xu](/wiki/papers/jf/2025/duchin-sustainability-greenwashing-evidence-asset-2025/)
for plant-chemical-level toxic emissions (total pollution, pollution intensity,
abatement), 1,056,361 plant-chemical-year observations, 2000-2020.

- **Cost:** free, public.
- **API key:** none required.
- **Coverage:** facility x chemical x year, reporting years from 1987 to the most
  recent processed year, for covered industry sectors and listed chemicals.
- **Home:** <https://www.epa.gov/toxics-release-inventory-tri-program>

## Access

All paths below need no key. The bulk path gives you a full year of national
data in one file; the Envirofacts REST service lets you slice individual TRI
tables. The TRI Basic Data Files (annual CSV/ZIP) are also published on the EPA
website if you prefer to download by hand.

### Option 1: bulk CSV (one year, national)

A materialized download view returns a full year of US data as a single CSV:

```bash
# (a) Bulk: one year, national, as CSV (no key)
curl -sL -o tri_2022_US.csv \
  "https://data.epa.gov/efservice/downloads/tri/mv_tri_basic_download/2022_US/csv"
```

The 2022 US file is roughly 25 MB. Load it as strings first, then cast the
columns you need:

```python
import pandas as pd

df = pd.read_csv(
    "tri_2022_US.csv",
    dtype=str,           # read everything as str first; cast after schema alignment
    low_memory=False,
)

# Cast key fields after inspecting column names and the unit column
df["total_releases"] = pd.to_numeric(df["TOTAL RELEASES"], errors="coerce")
```

Inspect the header before casting; column names and the unit column matter (see
Gotchas).

### Option 2: Envirofacts REST service

The Envirofacts service exposes TRI tables directly. The URL grammar is:

```
https://data.epa.gov/efservice/<table>/<column>/<value>/rows/<start>:<end>/<format>
```

`<format>` can be `JSON` or `CSV`. The `rows/<start>:<end>` window is required
for large tables.

```bash
# (b) Envirofacts REST: a windowed slice of a TRI table as JSON (no key)
curl -sL "https://data.epa.gov/efservice/tri_facility/state_abbr/CA/rows/0:1/JSON"
```

This returns facility records with fields such as `tri_facility_id`,
`facility_name`, `street_address`, `city_name`, `county_name`,
`state_county_fips_code`, `state_abbr`, `zip_code`, `region`, `fac_closed_ind`,
and `frs_id`. Page through a table with successive windows
(`rows/0:9999`, `rows/10000:19999`, and so on).

## Gotchas (the ones that bite pipelines)

- **Self-reported estimates, not measured emissions.** TRI quantities are
  estimates; estimation methods vary by facility, and reported values can be
  revised and back-filed in later releases. Pin the release year you pulled.
- **Chemical list and thresholds change across years.** A chemical added in a
  given year appears to jump from zero. That is a reporting-scope change, not a
  real increase. Check whether a chemical was reportable in the years you
  compare.
- **Covered sectors have expanded over time.** Do not read a sector first
  appearing in the data as a real onset of pollution; it can be a coverage
  change.
- **Units: pounds, except PBT chemicals in grams.** Quantities are in pounds,
  except certain highly toxic persistent bioaccumulative and toxic (PBT)
  chemicals (such as dioxin and dioxin-like compounds), which are reported in
  grams. Never sum the raw quantity column across chemicals without
  unit-aligning first; do not sum pounds and grams.
- **Facility identity drifts.** TRI Facility ID, FRS ID, and parent-company name
  do not map one-to-one, and ownership changes across years break a naive
  facility panel. This matters for any buyer/seller or plant-tracking analysis
  (as in the citing paper); link facilities deliberately across years.
- **On-site vs. off-site, released vs. managed.** Do not conflate on-site versus
  off-site, or "released" versus "managed as waste". These are distinct columns
  with distinct meaning.
- **Envirofacts needs the rows window.** The REST service requires the
  `rows/<start>:<end>` window for large tables, or the request times out. Page
  through with successive windows.

## Citation

Cite the U.S. EPA, the Toxics Release Inventory, the reporting year(s), and the
specific file or Envirofacts table used, with the EPA URL and the access date,
for example: *U.S. Environmental Protection Agency, Toxics Release Inventory
(TRI), reporting year 2022, TRI Basic Data Download
(mv_tri_basic_download/2022_US), retrieved from
https://data.epa.gov/efservice/downloads/tri/mv_tri_basic_download/2022_US/csv,
accessed YYYY-MM-DD.* Record the reporting year and the table or file so the
pull is reproducible.


==============================================================================
# Facebook Social Connectedness Index (SCI)
# https://instituteforautomatedresearch.org/wiki/datasets/facebook-sci/
# How to pull Meta's Social Connectedness Index (SCI) as no-key bulk CSVs from the Humanitarian Data Exchange, plus the gotchas that bite pipelines (it is a rescaled relative measure not a count, symmetric with both directions stored, the diagonal dominates, and region codes differ by file).
# Verified 2026-06-22 · tested with live no-key download from the Humanitarian Data Exchange of country.csv (31,684 country-pair rows) and us_counties.csv (245 MB, 10,265,616 county-pair rows; columns user_country/friend_country/user_region/friend_region/scaled_sci)
# Tags: social-network, geography, cross-section, panel, free, no-api-key, data:facebook-sci
==============================================================================

**The Social Connectedness Index (SCI)** is Meta's measure of how strongly two
places are linked by Facebook friendships. For a pair of locations, `scaled_sci`
is proportional to the relative probability that a Facebook user in one is
friends with a user in the other (friendship links between the two, divided by
the product of their user counts), built from a snapshot of anonymized,
aggregated friendships. It is distributed openly through the Humanitarian Data
Exchange (HDX). Used in, for example,
[Rehbein & Rother (2025)](/wiki/papers/rfs/2025/rehbein-social-connectedness-bank-lending-2025/),
where county-pair SCI proxies the social ties along which bank lending travels.

- **Cost:** free, public.
- **API key:** none required (bulk CSV download from HDX).
- **Coverage:** several geography levels, each its own file: country-country,
  sub-national admin level 1 and 2 (GADM), US counties, US ZCTAs, and European
  NUTS regions. A cross-section per release (the index has been re-snapshotted
  across several vintages), not a time series.
- **Home:** <https://data.humdata.org/dataset/social-connectedness-index>

## Access

Each geography level is a separate file on the HDX dataset page. The
country-country and US-county files are direct CSV downloads; some larger files
(GADM2, US ZCTA) are served from Google Drive links listed on the same page.

```bash
# Country-to-country SCI (small, ~0.5 MB):
curl -sL "https://data.humdata.org/dataset/e9988552-74e4-4ff4-943f-c782ac8bca87/resource/652cf9c9-541f-47de-8d53-ff818062bd0c/download/country.csv" -o sci_country.csv

# US county-to-county SCI (245 MB, ~10.3M rows):
curl -sL "https://data.humdata.org/dataset/e9988552-74e4-4ff4-943f-c782ac8bca87/resource/97dc352f-c9c5-47d6-a6ef-88709e14006c/download/us_counties.csv" -o sci_us_counties.csv
```

Every file has the same five columns: `user_country, friend_country,
user_region, friend_region, scaled_sci`. In the country file the region columns
repeat the country code; in the US-county file `user_region`/`friend_region`
are five-digit FIPS county codes.

### Load in Python

```python
import pandas as pd

sci = pd.read_csv("sci_us_counties.csv")            # FIPS-pair rows
sci = sci[sci.user_region != sci.friend_region]     # drop the diagonal
# undirected: keep each unordered pair once
sci = sci[sci.user_region < sci.friend_region]
```

## Gotchas (the ones that bite pipelines)

- **`scaled_sci` is a rescaled relative measure, not a count or a probability.**
  Within a release the values are rescaled to integers with the single
  most-connected pair set to 1,000,000,000. Only the ordering and ratios within
  one file are meaningful; the absolute number is arbitrary. Do not read it as a
  number of friendships.
- **Do not compare `scaled_sci` across files or vintages.** Each release and each
  geography file is rescaled independently to its own max of 1e9, so a value in
  one vintage's county file is not comparable to one in another vintage or in the
  country file. Normalize within the single file you use.
- **It is symmetric and both directions are stored.** `(i, j)` and `(j, i)` carry
  the same `scaled_sci`, so the file double-counts unordered pairs (the US-county
  file is roughly 3,200 squared = 10.3M rows). Deduplicate to one direction for
  pairwise analysis.
- **The diagonal is included and dominates.** Self-pairs (`user_region ==
  friend_region`) measure within-area connectedness and are far larger than any
  cross-area value (within-county friendships dominate). Drop the diagonal for
  between-area work.
- **Region codes differ by file.** US counties are five-digit FIPS; sub-national
  files use GADM codes; the European file uses NUTS; ZCTAs are US ZIP-code
  tabulation areas. The `user_country`/`friend_country` columns are still present
  in sub-national files (both `US` in the county file), so do not assume those
  columns vary.
- **It is a snapshot, not a panel.** Each release is built from Facebook
  friendships at one point in time; releases are occasional re-snapshots, not a
  continuous series. Treat a file as a single cross-section and record which
  vintage you pulled.
- **Built from anonymized, aggregated data with privacy protection.** The index
  comes from de-identified, aggregated friendship counts; very thin pairs can be
  floored or noised. It reflects Facebook's user base, which is not a uniform
  sample of the population, so connectedness is conditioned on who uses the
  platform.
- **The US-county file is large.** 245 MB and about 10.3M rows; read it filtered
  or in chunks rather than loading the raw symmetric table whole.

## Reference

| Field | Value |
|-------|-------|
| Host | Humanitarian Data Exchange: `data.humdata.org/dataset/social-connectedness-index` |
| Files | `country.csv`, `gadm1.csv`, `gadm2.zip`, US `us_counties.csv`, `us_zcta.zip`, `nuts_2024.zip`, geoboundaries variants |
| Columns | `user_country`, `friend_country`, `user_region`, `friend_region`, `scaled_sci` |
| US region code | Five-digit FIPS (county file) |
| Scaling | Integer, rescaled so the max pair = 1,000,000,000 within each file |
| Structure | Symmetric, both directions stored; diagonal included |
| Key required | No |

## Citation

Cite Bailey, Cao, Kuchler, Stroebel & Wong, "Social Connectedness: Measurement,
Determinants, and Effects," *Journal of Economic Perspectives* 32(3), 2018,
259-280, together with the specific SCI release and geography level, the
retrieval URL, and the access date. Record the vintage, since the index is
rescaled per release and the values are not comparable across versions.


==============================================================================
# Fannie Mae & Freddie Mac single-family loan-level data
# https://instituteforautomatedresearch.org/wiki/datasets/fannie-freddie/
# How the GSE single-family loan-level acquisition and performance datasets are structured and accessed, why a pipeline cannot pull them no-key (both are free but registration-gated behind a click-through), the acquisition vs performance split, the pipe-delimited vs CSV format difference, and the quarterly vintages.
# Tags: mortgages, housing, credit, loan-level, panel-data, free, academic, data:fannie-freddie
==============================================================================

:::note[Free, but registration-gated and not exercised here]
The GSE single-family loan-level datasets are **free** (no fee, including for
academic use), but bulk download from both Fannie Mae and Freddie Mac requires a
**no-cost account and a click-through terms acceptance**. The public landing
portals are reachable (HTTP 200), but the data files sit behind the login wall,
so the end-to-end pull was **not** exercised from this session. The page
therefore carries **no provenance badge**: it documents the datasets and the
access path. A researcher registers once (free) and downloads directly.
:::

**The Fannie Mae and Freddie Mac single-family loan-level datasets** are the two
government-sponsored enterprises' public releases of acquisition and monthly
performance records for the single-family mortgages they bought. Each loan has
origination characteristics (FICO, LTV, DTI, rate, term, purpose, geography) and
a monthly performance history (delinquency status, prepayment, default, losses).
They are the standard free microdata for US mortgage credit-risk and
performance research. Used in, for example,
[Rehbein & Rother (2025)](/wiki/papers/rfs/2025/rehbein-social-connectedness-bank-lending-2025/),
where 1.27 million GSE 30-year fixed-rate mortgages (2000 to 2008 originations,
observed to 2018) supply loan-level LTV, rate, delinquency, and default.

- **Cost:** free (no fee; free account and terms acceptance required to download).
- **API key:** none (bulk file download behind a registration portal).
- **Coverage:** Freddie covers roughly 55 million mortgages originated from 1999
  through the latest quarterly cutoff; Fannie's window is comparable. Both append
  new origination quarters over time.
- **Granularity:** one acquisition/origination row per loan, plus monthly
  performance rows per loan.
- **Portals (browser):** Fannie Data Dynamics
  <https://capitalmarkets.fanniemae.com/tools-applications/data-dynamics> ·
  Freddie via Clarity Data Intelligence
  <https://www.freddiemac.com/research/datasets/sf-loanlevel-dataset>

## Access

Both GSEs publish the data through a free registration portal. There is no
anonymous direct link to the bulk files.

- **Fannie Mae** distributes the Single-Family Loan Performance Data through
  Data Dynamics. After a free account and terms acceptance, the portal offers a
  one-click download of the full Acquisition file and the full Performance file,
  in CSV.
- **Freddie Mac** distributes the Single Family Loan-Level Dataset through
  Clarity Data Intelligence. After a free sign-in (the data is free for
  non-commercial and academic use), the download page offers the full dataset,
  the Standard dataset by year, the Non-Standard dataset, sample files, and an
  RPL mapping file. Files are pipe-delimited.

The public documentation (user guides, file-layout and data-dictionary PDFs,
FAQs) is generally readable without an account; the loan-level files and even
the sample files are behind the login.

## Gotchas (the ones that bite pipelines)

- **Two file types per GSE, joined on the loan id.** Acquisition (one static row
  per loan, captured at origination) and Performance (one row per loan per
  month). Do not expect a single combined table; the performance file is the
  large one.
- **Format differs between the two GSEs.** Freddie's files are pipe-delimited
  (`|`), no header, with a fixed column order set by the layout document. Fannie's
  Data Dynamics export is CSV. Column order, field set, and missing-value
  conventions differ, so do not reuse one schema for the other.
- **Click-through terms, so no anonymous fetch.** An automated pull must carry a
  session cookie obtained after accepting the terms; there is no static URL to the
  zips. Plan for a one-time human registration.
- **Quarterly vintages that grow.** Acquisition files are partitioned by
  origination quarter and new quarters are appended over time, so any hardcoded
  "latest quarter" logic rots. Record the vintage you pulled.
- **Freddie Standard vs Non-Standard.** The Non-Standard dataset adds loans
  excluded from Standard (for example certain modified, ARM, or balloon loans);
  the RPL mapping file tracks reperforming loans. Choose deliberately rather than
  mixing them.
- **Scale.** Tens of millions of loans and a monthly panel on top; the
  performance files are tens of gigabytes uncompressed. Stream or chunk; do not
  load whole.
- **Host bot-blocking on top of the login.** The Fannie capitalmarkets host
  returns HTTP 403 to automated requests even before the login, so scripted
  navigation needs a real browser session, not just an authenticated cookie.

## Reference

| Field | Value |
|-------|-------|
| Fannie portal | Data Dynamics (`capitalmarkets.fanniemae.com/tools-applications/data-dynamics`) |
| Freddie portal | Clarity Data Intelligence (`freddiemac.com/research/datasets/sf-loanlevel-dataset`) |
| File types | Acquisition (per loan) + Performance (per loan per month) |
| Fannie format | CSV |
| Freddie format | Pipe-delimited, no header |
| Freddie coverage | ~55M loans, 1999 to latest quarterly cutoff |
| Key required | No key, but free registration + terms acceptance |

## Citation

Cite each GSE dataset separately: Fannie Mae, "Single-Family Loan Performance
Data," and Freddie Mac, "Single Family Loan-Level Dataset," each with the vintage
(origination-quarter range and performance cutoff) and retrieval date. Note in a
methods appendix that access required free registration, since the files are not
anonymously downloadable.


==============================================================================
# FDIC Summary of Deposits (SOD)
# https://instituteforautomatedresearch.org/wiki/datasets/fdic-summary-of-deposits/
# How to pull branch-level deposit data from the FDIC Summary of Deposits for free, with no API key, including the headquarters-booking distortion and the other gotchas that bite branch-geography pipelines.
# Verified 2026-06-09 · tested with live no-key SOD API query (api.fdic.gov/banks/sod, YEAR:2022 branch deposits)
# Tags: banking, deposits, free, no-api-key, fdic, panel-data, data:fdic-summary-of-deposits
==============================================================================

The **FDIC Summary of Deposits (SOD)** is the annual survey of branch-office
deposits for all FDIC-insured institutions, recorded as of June 30 each year.
For each branch it reports the institution, the branch location (address, city,
county, state, and geocoordinates), and the deposits booked to that branch.
It is used in, for example,
[Di Maggio, Ma & Williams](/wiki/papers/jf/2025/maggio-red-overdrafts-payday-lending-2025/)
for bank branch locations and deposits of the 50 largest banks,
[Beyhaghi, Fracassi & Weitzner](/wiki/papers/jf/2026/beyhaghi-adverse-selection-corporate-loans-2026/)
for county-level branch counts, and
[Johnston-Ross, Ma & Puri](/wiki/papers/jf/2025/johnston-ross-private-equity-financial-stability-2025/)
for branch-level deposit balances and locations.

- **Cost:** free, no paywall.
- **API key:** none required.
- **Coverage:** every domestic branch of every FDIC-insured bank and thrift,
  annually since 1994.
- **Home:** <https://www.fdic.gov/resources/data-tools/> (Summary of Deposits)
  · **API:** <https://api.fdic.gov/banks/sod>

## Access

### Option 1: FDIC API (no key, JSON)

The FDIC public API serves the SOD at `https://api.fdic.gov/banks/sod`.
No key or authentication is needed.

```bash
curl -G "https://api.fdic.gov/banks/sod" \
  --data-urlencode "filters=YEAR:2022" \
  --data-urlencode "fields=NAMEFULL,STALPBR,CITYBR,DEPSUMBR,YEAR" \
  --data-urlencode "limit=10"
```

The response is JSON with a `data[]` array; each element is one branch-year
record. Pagination is controlled by `limit` and `offset` parameters.

```python
import requests, pandas as pd

params = {
    "filters": "YEAR:2022",
    "fields": "NAMEFULL,STALPBR,CITYBR,COUNTYNAME,DEPSUMBR,UNINUMBR,CERT,YEAR",
    "limit": 10000,
    "offset": 0,
}
r = requests.get("https://api.fdic.gov/banks/sod", params=params)
rows = [rec["data"] for rec in r.json()["data"]]
df = pd.DataFrame(rows)
```

Loop over `offset` in steps of 10,000 (the API maximum per call) to pull a
full year; a single year holds roughly 70,000-80,000 branch records.

### Option 2: Bulk download

The FDIC also publishes annual flat files at the Summary of Deposits page
(<https://www.fdic.gov/resources/data-tools/>). Each year's ZIP contains a
pipe-delimited `.csv` with all branches. This is faster than paging the API
for a full-year pull.

## Gotchas (the ones that bite pipelines)

- **Headquarters-booking distortion.** Deposits are booked to the branch of
  administrative assignment, not where the customer lives. Large banks
  concentrate a disproportionate share of deposits at one headquarters or main
  branch. Any branch-deposit geography measure is affected; papers that use
  SOD for deposit geography typically acknowledge or adjust for this.
- **Annual June-30 snapshot only.** There is no intra-year or quarterly
  frequency. The survey date is always June 30; do not treat it as a
  year-end figure.
- **API host change.** The API host moved from `banks.data.fdic.gov` to
  `api.fdic.gov` (the old host issues a 301 redirect). Pin the new host
  in your pipeline to avoid the redirect hop.
- **Identifier reassignment at M&A.** Branch identifiers (`UNINUMBR`) and
  institution certificate numbers (`CERT`) are reassigned at mergers and
  acquisitions. A naive year-over-year branch panel breaks at M&A events:
  track the acquirer's `CERT` to link branches across a merger.
- **Credit unions are absent.** SOD covers only FDIC-insured banks and
  thrifts. Credit unions are not included; they are reported separately by
  the NCUA.

## Citation

Cite the FDIC Summary of Deposits, the survey year (June 30 as-of date), and
note whether the deposit geography was adjusted for the headquarters-booking
issue, e.g.: *Federal Deposit Insurance Corporation, Summary of Deposits,
[year] (as of June 30); <https://www.fdic.gov/resources/data-tools/>, accessed
YYYY-MM-DD. Branch deposits not adjusted / adjusted for headquarters-booking
concentration.*


==============================================================================
# FDIC Quarterly Banking Profile & BankFind financials
# https://instituteforautomatedresearch.org/wiki/datasets/fdic/
# How to read the FDIC Quarterly Banking Profile (aggregate quarterly bank condition and income) and pull the institution-level BankFind financials API behind it, with the no-key access recipe and the gotchas that bite pipelines.
# Verified 2026-06-09 · tested with live no-key FDIC BankFind financials API call (banks.data.fdic.gov /api/financials)
# Tags: banking, free, no-api-key, fdic, time-series, data:fdic
==============================================================================

This page covers two linked FDIC products. The **Quarterly Banking Profile
(QBP)** is the FDIC's quarterly report on the aggregate condition and income of
all FDIC-insured commercial banks and savings institutions: aggregate assets,
loans, deposits, net income, net interest margin, asset quality, and capital,
by quarter and by group. The **institution-level data behind it** is exposed
through the FDIC BankFind Suite REST API at
<https://banks.data.fdic.gov/api>, whose endpoints include `/financials`
(quarterly financial items and ratios per institution `CERT`), `/institutions`,
`/history`, and `/failures`. This page is the access recipe for both. Used in,
for example, [Wang](/wiki/papers/jf/2025/wang-banks-low-interest-rates-2025/),
where the FDIC Quarterly Banking Profile cross-checks the Call Report series
for aggregate U.S. commercial bank data (sample 1997Q2-2018Q2).

This page is distinct from two siblings. The
[FDIC Summary of Deposits](/wiki/datasets/fdic-summary-of-deposits/) (annual
branch-level deposits) has its own page. The
[FFIEC Call Reports](/wiki/datasets/call-reports/) page covers the underlying
bank regulatory filings. The financials API items here are derived from Call
Reports and Thrift Financial Reports; this page is the FDIC aggregate and
financials view, not the raw Call Report schedules.

- **Cost:** free, no paywall.
- **API key:** none required.
- **Coverage:** institution-level financials from 1992 onward, quarterly; the
  QBP narrative and aggregates have been published for decades.
- **Home:** QBP at <https://www.fdic.gov/analysis/quarterly-banking-profile/> ·
  **API docs:** <https://banks.data.fdic.gov/docs/>

## Access

### Institution-level financials (no key, JSON)

The `/financials` endpoint returns quarterly financial items and ratios per
institution. Filters use `field:value` colon syntax, `fields` selects columns,
and `limit`/`offset` paginate.

```bash
# Institution-level financials for one quarter, selected fields, no key
curl -sL "https://banks.data.fdic.gov/api/financials?filters=REPDTE:20231231&fields=CERT,NAME,ASSET&limit=1"
```

`REPDTE` is the quarter-end report date in `YYYYMMDD` form. The response is JSON
with a `meta` block (including an index `createTimestamp`), a `data` array, and
a `totals` block. For `REPDTE:20231231` the query above returned
`totals.count` of 4658 institutions, and the first record was `CERT` 10004,
`NAME` "ERGO BANK", `ASSET` 240478 (in thousands of dollars).

### Pull a full quarter and build a panel

A full quarter is thousands of institutions, so loop `offset` until you have
`totals.count` rows, then iterate `REPDTE` across quarter-ends to build a panel.

```python
import requests, pandas as pd

BASE = "https://banks.data.fdic.gov/api/financials"

def pull_quarter(repdte, fields="CERT,NAME,ASSET,DEP,NETINC", page=2000):
    rows, offset = [], 0
    while True:
        params = {
            "filters": f"REPDTE:{repdte}",
            "fields": fields,
            "limit": page,
            "offset": offset,
        }
        r = requests.get(BASE, params=params)
        r.raise_for_status()
        payload = r.json()
        batch = [rec["data"] for rec in payload["data"]]
        rows.extend(batch)
        total = payload["totals"]["count"]
        offset += page
        if offset >= total:
            break
    return pd.DataFrame(rows)

# One quarter
q = pull_quarter("20231231")

# A panel across several quarter-ends
panel = pd.concat(
    pull_quarter(d).assign(REPDTE=d)
    for d in ["20221231", "20230331", "20230630", "20230930", "20231231"]
)
```

`ASSET` and other dollar items are reported in thousands of dollars; scale
them (multiply by 1,000) before reporting absolute amounts.

## Gotchas (the ones that bite pipelines)

- **Filters use colon syntax, not equals.** Write `filters=REPDTE:20231231`,
  not `REPDTE=20231231`. `REPDTE` is the quarter-end report date as `YYYYMMDD`.
- **Pagination is mandatory.** Results page with `limit`/`offset`. A single
  quarter is thousands of institutions (4658 at `REPDTE:20231231`), so loop
  `offset` until you have `totals.count` rows. Do not assume the default page
  is the whole population.
- **Dollar items are in thousands.** Institution-level dollar fields such as
  `ASSET` are in thousands of dollars; scale before reporting.
- **Item definitions change with regulatory form revisions.** The financials
  items are derived from Call Reports (FFIEC 031/041) and Thrift Financial
  Reports. What is rolled into a given series changes when the underlying form
  is revised, so a long quarterly panel is not on a fixed schema. Check item
  definitions for the vintage you are using.
- **CERT-level series start and stop.** Bank mergers and failures mean an
  institution's `CERT` series begins and ends over time. Track entity changes
  through the `/history` endpoint rather than assuming a balanced panel.
- **QBP aggregates need not equal a sum over /financials.** The published QBP
  aggregates use the FDIC's own groupings and adjustments (de novo
  institutions, prior-period amendments), so they may not equal a naive sum
  over `/financials`. If you need the official aggregate, use the QBP release;
  if you need institution detail, use `/financials`; do not expect the two to
  reconcile to the dollar.
- **Record the index timestamp for reproducibility.** The API `meta` block
  carries an index `createTimestamp`. The same query can return slightly
  different totals as the index is rebuilt, so capture the timestamp alongside
  any pull you need to reproduce.

## Citation

Cite the FDIC, the Quarterly Banking Profile and/or the BankFind financials
API, the report date(s) (`REPDTE`), the FDIC URL, and the access date, for
example: *Federal Deposit Insurance Corporation, Quarterly Banking Profile and
BankFind financials, report date 2023-12-31 (REPDTE 20231231); retrieved from
<https://banks.data.fdic.gov/api> and
<https://www.fdic.gov/analysis/quarterly-banking-profile/>, accessed
YYYY-MM-DD.*


==============================================================================
# Federal Register
# https://instituteforautomatedresearch.org/wiki/datasets/federal-register/
# The daily journal of U.S. federal agency Rules, Proposed Rules, and Notices, with full text from 1994 via the federalregister.gov API. Covers the no-key access recipe and the gotchas that bite pipelines.
# Verified 2026-06-09 · tested with live no-key Federal Register API call (federalregister.gov /api/v1/documents.json)
# Tags: regulation, government, text-as-data, free, no-api-key, data:federal-register
==============================================================================

The **Federal Register** is the daily journal of the U.S. federal government,
publishing agency Rules, Proposed Rules, and Notices, plus Presidential
Documents. It is produced by the Office of the Federal Register (part of the
National Archives, NARA) and printed through the GPO. Full text from 1994 to
the present is available through the federalregister.gov REST API and the
GovInfo bulk repository. Used in, for example,
[Kalmenovitz, Lowry & Volkova](/wiki/papers/jf/2025/kalmenovitz-regulatory-fragmentation-2025/),
where the full text of 783,950 documents (1994-2019) trains an LDA topic model
measuring regulatory fragmentation.

- **Cost:** free, public.
- **API key:** none required.
- **Coverage:** full text 1994 to present on federalregister.gov; older
  material back to 1936 is on GovInfo as scanned or text.
- **Home:** <https://www.federalregister.gov/developers/documentation/api/v1>

## Access

The API base is `https://www.federalregister.gov/api/v1`. Endpoints include
`/documents.json` (search with field selection), `/documents/{document_number}.json`
(one document), and `/agencies`. No authentication is needed.

### Step 1: search, selecting fields

```bash
# Search the Federal Register, selecting fields, no key
curl -sL "https://www.federalregister.gov/api/v1/documents.json?per_page=1&fields[]=document_number&fields[]=type&fields[]=agencies&fields[]=publication_date&fields[]=title"
```

The response carries `description`, `count`, `total_pages`, `next_page_url`,
and a `results` array. Each result includes `document_number`, `type` (one of
`Rule`, `Proposed Rule`, `Notice`, `Presidential Document`), `agencies` (a
list, each with `id`, `name`, `slug`), `publication_date`, `title`, `abstract`,
CFR references, and links to the body via `raw_text_url` and
`full_text_xml_url`.

### Step 2: page through a date window in Python

The search endpoint returns metadata only. Partition the query by a date
window so each slice stays under the page cap (see Gotchas).

```python
import requests

BASE = "https://www.federalregister.gov/api/v1/documents.json"

def fetch_window(start, end, per_page=1000):
    """All documents published in [start, end] (YYYY-MM-DD)."""
    params = {
        "per_page": per_page,
        "conditions[publication_date][gte]": start,
        "conditions[publication_date][lte]": end,
        "fields[]": [
            "document_number", "type", "agencies",
            "publication_date", "title", "raw_text_url",
        ],
    }
    url, docs = BASE, []
    while url:
        r = requests.get(url, params=params if url == BASE else None, timeout=60)
        r.raise_for_status()
        page = r.json()
        docs.extend(page["results"])
        url = page.get("next_page_url")
    return docs

docs = fetch_window("2019-01-01", "2019-01-31")
```

### Step 3: fetch a document's body text

The body comes from `raw_text_url` (plain text) or `full_text_xml_url` (XML),
one extra request per document:

```python
body = requests.get(docs[0]["raw_text_url"], timeout=60).text
```

## Gotchas (the ones that bite pipelines)

- **The search API caps a single query at 50 pages.** Confirmed live:
  `total_pages` was 50 even when `count` reported 10000. With the per-page
  maximum you cannot page through an unbounded match in one query. To retrieve
  a full corpus, partition by date window (or by agency) so each slice stays
  under the cap; do not trust that one query returns every matching document.
- **`count` overstates what you can retrieve.** The `count` field can report
  the full match count while only 50 pages are actually retrievable. Reconcile
  your downloaded count against the slice boundaries, not against `count`.
- **Search returns metadata only.** The document body must be fetched
  separately via `raw_text_url` or `full_text_xml_url`, one extra request per
  document. Budget request volume and caching accordingly.
- **`agencies` is a list (many-to-many).** A single document can be issued
  jointly by several agencies (confirmed: one result listed both the
  Transportation Department and the Federal Aviation Administration). Do not
  assume one agency per document.
- **Use GovInfo bulk data for a full historical pull.** For an entire-corpus
  download, the GovInfo bulk data repository is the route, not thousands of
  per-document API calls. Use the API for targeted or incremental queries.
- **Publication is uneven across the calendar.** There is no issue on federal
  holidays, so daily document counts are not a smooth series; do not treat the
  per-day count as a regular time series without accounting for non-publication
  days.

## Key document fields

| Field | Meaning |
|-------|---------|
| `document_number` | Unique ID, e.g. `2026-11559` |
| `type` | `Rule`, `Proposed Rule`, `Notice`, or `Presidential Document` |
| `agencies` | List of issuing agencies (`id`, `name`, `slug`) |
| `publication_date` | Date printed in the Register |
| `title` | Document title |
| `abstract` | Summary text, when present |
| `raw_text_url` | Plain-text body of the document |
| `full_text_xml_url` | XML body of the document |

## Citation

Cite the Office of the Federal Register / National Archives, the Federal
Register, the document number(s) or query used, the federalregister.gov URL,
and the access date, for example: *Office of the Federal Register, National
Archives and Records Administration, Federal Register, document no. 2026-11559,
retrieved from https://www.federalregister.gov/, accessed YYYY-MM-DD.* Record
the document number(s) or the exact query and date window so the pull is
reproducible.


==============================================================================
# FHFA House Price Index (HPI)
# https://instituteforautomatedresearch.org/wiki/datasets/fhfa-hpi/
# How to pull the FHFA repeat-sales house price index from the no-key master CSV, the ZIP-code developmental indexes, and the gotchas that bite pipelines (purchase-only vs all-transactions, NSA vs SA, the ZIP files are annual).
# Verified 2026-06-09 · tested with live no-key download of the FHFA HPI master file (fhfa.gov/hpi/download/monthly/hpi_master.csv, ~17 MB, 184,807 rows)
# Tags: housing, prices, mortgage, free, no-api-key, time-series, data:fhfa-hpi
==============================================================================

**The FHFA House Price Index (HPI)** is a repeat-sales measure of single-family
house price changes published by the Federal Housing Finance Agency, built from
repeat mortgage transactions on the same properties. It is reported at the
national, census-division, state, metro-area, county, and ZIP-code levels, in
purchase-only and all-transactions flavors. The data is free and public. Used
in, for example,
[Heitz, Martin & Ufier](/wiki/papers/jf/2026/heitz-bank-monitoring-onsite-inspections-2026/),
where a ZIP-level HPI serves as a time-varying collateral-value proxy in the
moral-hazard analysis (Tables IV and V).

- **Cost:** free, public.
- **API key:** none required.
- **Coverage:** USA and census divisions from 1991 (monthly) and 1975
  (quarterly); states, metros, counties; ZIP-code developmental indexes
  (annual). Purchase-only and all-transactions; seasonally adjusted and not.
- **Home:** <https://www.fhfa.gov/data/hpi>

## Access

The master file appends every quarterly and monthly series into one CSV, no
authentication:

```bash
# FHFA HPI master file: all geographies and flavors, no key
curl -sL -o hpi_master.csv \
  "https://www.fhfa.gov/hpi/download/monthly/hpi_master.csv"
```

The columns are `hpi_type, hpi_flavor, frequency, level, place_name, place_id,
yr, period, index_nsa, index_sa, rstderr, note`. Filter on `hpi_flavor`
(`purchase-only`, `all-transactions`, `expanded-data`), `frequency` (`monthly`,
`quarterly`), and `level` (`USA or Census Division`, `State`, `MSA`, and so on)
to pull the series you want.

The **ZIP-code indexes** are separate developmental files, and they are
**annual**, not monthly or quarterly:

```bash
# Five-digit ZIP, all-transactions developmental index (annual, NSA)
curl -sL -o hpi_at_zip5.xlsx \
  "https://www.fhfa.gov/hpi/download/annual/hpi_at_zip5.xlsx"
# Three-digit ZIP variant:
curl -sL -o hpi_at_zip3.xlsx \
  "https://www.fhfa.gov/hpi/download/annual/hpi_at_zip3.xlsx"
```

### Load in Python

```python
import pandas as pd

df = pd.read_csv("hpi_master.csv")
# Purchase-only, monthly, national series:
us = df[(df.hpi_flavor == "purchase-only")
        & (df.frequency == "monthly")
        & (df.place_name == "United States")]
# index_nsa is the level; build returns within a (place_id, flavor, frequency).
```

## Gotchas (the ones that bite pipelines)

- **Purchase-only and all-transactions are different indexes.** Purchase-only
  uses sales pairs from Fannie/Freddie purchase mortgages; all-transactions
  adds appraisal values from refinances and runs at lower frequency for finer
  geographies. They diverge, especially in stressed periods. Pick one flavor
  deliberately and do not splice them.
- **The ZIP-code indexes are annual and developmental.** FHFA publishes
  five-digit and three-digit ZIP indexes only at **annual** frequency, labelled
  developmental and not seasonally adjusted. There is no official monthly
  five-digit-ZIP FHFA series; a paper using "monthly ZIP-level HPI" has either
  interpolated the annual index or mapped a higher-frequency geography to ZIPs.
  Do not expect to download a monthly ZIP file from FHFA.
- **NSA versus SA columns.** The master file carries both `index_nsa` and
  `index_sa`; the seasonally adjusted column is populated only for some series.
  Choose the column to match your use and check it is non-empty for that series.
- **It is an index, not a price level.** Values are index numbers normalized to
  a base period (the purchase-only series is set to 100 at 1991Q1; other
  flavors and vintages use a different base), not dollar prices. Work in returns
  or ratios, and confirm the base before comparing across series.
- **Repeat-sales coverage skews to conforming mortgages.** The index is built
  from Fannie Mae and Freddie Mac loan data (and, for all-transactions,
  appraisals), so it underweights cash, jumbo, and non-conforming activity.
  This biases coverage in high-price and investor-heavy areas.
- **Small geographies are noisy and revised.** County and ZIP indexes rest on
  thin transaction counts; `rstderr` in the master file flags the standard
  error, and back-history is revised as new pairs arrive. Carry the standard
  error and re-pull rather than caching a small-area series indefinitely.

## Geography levels (master file)

| `level` value | Geography |
|---------------|-----------|
| `USA or Census Division` | National and the nine census divisions |
| `State` | 50 states + DC |
| `MSA` | Metropolitan statistical areas |
| `County` | Counties (all-transactions, annual) |
| ZIP files (separate) | Three- and five-digit ZIP (annual developmental) |

## Citation

Cite the FHFA, the House Price Index, the flavor and geography, the URL, and the
access date, for example: *Federal Housing Finance Agency, House Price Index,
purchase-only monthly series, retrieved from
https://www.fhfa.gov/data/hpi, accessed YYYY-MM-DD.* Record the flavor,
frequency, geography level, and NSA/SA choice so the series is reproducible.


==============================================================================
# Flexible data-mining strategies (Chen-Lopez-Lira-Zimmermann)
# https://instituteforautomatedresearch.org/wiki/datasets/flex-mining/
# How to get ~30,000 data-mined long-short strategies and the signal-theory classification for free: the gdown-for-Drive trap, the size trap, and start-from-the-small-file tip, for automated pipelines.
# Verified 2026-05-16 · tested with live GitHub raw fetch (SignalsTheoryChecked.csv, 200, real columns)
# Tags: asset-pricing, factors, anomalies, free, academic, data:flex-mining
==============================================================================

The **flexible data-mining** dataset (Chen, Lopez-Lira & Zimmermann, *"Peer-
reviewed theory does not help predict the cross-section of stock returns"*)
is the free benchmark of **~30,000 long-short strategies** built from every
constructible CRSP/Compustat accounting ratio, plus a classification of which
*published* signals are theory-motivated vs. purely empirical. It is what the
[ZeroPaper](https://github.com/alejandroll10/zeropaper) pipeline uses to ask
whether a "novel" predictor is actually distinguishable from data mining.

- **Cost:** free, no auth (public GitHub + public Google Drive).
- **Code:** <https://github.com/chenandrewy/flex-mining>
- **Bulk data:** a public Google Drive folder (see below).

## Access

### Option 1: Signal-theory classification (small, start here)

```python
import pandas as pd
url = ("https://raw.githubusercontent.com/chenandrewy/flex-mining/"
       "main/DataInput/SignalsTheoryChecked.csv")
signals = pd.read_csv(url)   # signalname, Authors, Year, Journal, theory, …
```

### Option 2: Full data-mined returns (large, Google Drive)

```python
# pip install gdown
import gdown
gdown.download_folder(
    "https://drive.google.com/drive/folders/1SZe_aF4ZNvK4ZRx2jQUE1j19KQvBaqWr",
    output="data/flex-mining/", remaining_ok=True)
```

Key files: `DataMinedLongShortReturnsEW.csv`,
`DataMinedLongShortReturnsVW.csv` (~30K strategies, monthly).

## Gotchas (the ones that bite pipelines)

The reason to read this page rather than the repo. The small GitHub CSV was
fetched live on the date above (200, real columns); the bulk returns live in
the linked Drive folder.

- **The bulk data is on Google Drive; use `gdown`.** Plain `requests` on a
  Drive folder URL does not work. `pip install gdown` and use
  `download_folder`. This is the most common failure here.
- **It's large (~500 MB+).** Download once to `data/flex-mining/` and cache.
  Load big CSVs with `usecols=` / chunking or you will OOM.
- **Start from the small file.** `SignalsTheoryChecked.csv` (~65 KB, straight
  from GitHub) is enough to scope an analysis before pulling 30K strategies.
- **EW vs VW are different files.** State which; they tell different stories.
- **It's a benchmark distribution, not a strategy list.** The point is to
  compare a published signal's return against the *distribution* of mined
  returns (is it in the right tail?), not to trade the 30K.
- **Drive folder IDs can rotate.** If `download_folder` fails, check the
  GitHub README for the current link before assuming the data moved.

## What's inside

| File / folder | Description |
|---|---|
| `DataMinedLongShortReturnsEW.csv` | EW long-short returns, ~30K strategies |
| `DataMinedLongShortReturnsVW.csv` | VW long-short returns, ~30K strategies |
| `DataInput/SignalsTheoryChecked.csv` | Published signals: theory vs. empirical |
| `Risk-vs/` | Risk vs. mispricing decomposition outputs |

## Standard operations

- **Benchmark published vs. mined:** is a published signal in the right tail
  of the mined return distribution?
- **Sufficiency:** how much cross-sectional variation does the published set
  capture vs. the full mined set?
- **Pre-publication:** do soon-to-be-published patterns differ before vs.
  after publication?
- **Theory value:** do theory-motivated signals beat empirical ones,
  conditional on mined performance?
- **Always state** EW vs VW, sample period, and any filters.

## Citation

*Chen, A. Y., J. Lopez-Lira, and T. Zimmermann. "Peer-reviewed theory does not
help predict the cross-section of stock returns." Data and code:
https://github.com/chenandrewy/flex-mining, accessed YYYY-MM-DD.*


==============================================================================
# Flow of Funds: Financial Accounts of the United States (Z.1)
# https://instituteforautomatedresearch.org/wiki/datasets/flow-of-funds/
# How to pull the Federal Reserve's Z.1 Financial Accounts (sector balance sheets and flows for the whole US economy) for free with no key, how to read the per-table CSV package and series-code grammar, and the levels-vs-flows gotchas that bite pipelines.
# Verified 2026-06-09 · tested with live no-key pull of z1_csv_files.zip (586 per-table CSVs, 2025:Q4 release; read household balance-sheet table b101, quarterly 1945:Q4 onward)
# Tags: macro, time-series, free, no-api-key, federal-reserve, data:flow-of-funds
==============================================================================

The **Financial Accounts of the United States** (Federal Reserve **Z.1**
statistical release, long known as the **Flow of Funds Accounts**) are the
macro-accounting backbone for the US economy: quarterly balance sheets and
transaction flows for every sector (households and nonprofits, nonfinancial
business, government, each financial sector, and the rest of the world) that
sum to a consistent whole. It is the standard source for household net worth,
sector leverage, credit growth, and who-holds-what. This page is the distilled
access recipe.

- **Cost:** free, no paywall, no key.
- **Coverage:** quarterly, **1945:Q4** onward; released about 10 weeks after
  quarter end, with an earlier `preview` of the upcoming release.
- **Format:** a bulk CSV package (`z1_csv_files.zip`, about 7.5 MB, **586**
  per-table CSVs plus a data dictionary), per-series via the Data Download
  Program, or the same series on FRED.
- **Home:** <https://www.federalreserve.gov/releases/z1/>

## Access

### No key, bulk CSV package (recommended for pipelines)

The download path is **stamped with the release date**, which changes every
quarter; resolve the latest from the release page, or use the stable `preview`
path for the upcoming quarter.

```
https://www.federalreserve.gov/releases/z1/20260319/z1_csv_files.zip   # 2025:Q4 release
https://www.federalreserve.gov/releases/z1/preview/z1_csv_files.zip     # next-quarter preview
```

```python
import io, zipfile, requests, pandas as pd

url = "https://www.federalreserve.gov/releases/z1/20260319/z1_csv_files.zip"
z = zipfile.ZipFile(io.BytesIO(requests.get(url).content))
# Each table is its own CSV under csv/; b101 is households & nonprofits.
hh = pd.read_csv(z.open("csv/b101.csv"))
```

### Data Download Program (per series)

For a handful of series rather than whole tables, the DDP at
`https://www.federalreserve.gov/datadownload/` (release `Z1`) emits CSV/XML for
a selected set; series are addressed by mnemonic (e.g. `FL152000005.Q`).

## Gotchas (the ones that bite pipelines)

The reason to read this page rather than the Z.1 PDF. Verified against the live
package on the date above.

- **586 separate CSVs, one per table.** The zip is not one tidy file: you pick
  the table (`b101` household balance sheet, `b103` nonfinancial corporate,
  `l108` and friends, etc.). A `data_dictionary/` folder in the package maps
  table and series codes.
- **Series-code grammar.** Columns are mnemonics like `FL152000005.Q`: a
  two-letter prefix (`FL` = level / amounts outstanding, `LM` = level at market
  value, `FU` = unadjusted flow, `FA` = flow at a seasonally adjusted annual
  rate, `FR` = revaluation, `FV` = other volume change), then a six-or-more
  digit sector-plus-instrument code, then a `.Q` / `.A` frequency suffix. Decode
  with the Fed's series structure (the `data_dictionary/` folder spells out each
  series); do not guess from the number.
- **Levels vs flows are different tables.** `b` and `l` tables are **levels**
  (stocks); `f` and `fu` tables are **flows** (transactions). Do not difference a
  level series and call it a flow.
- **Change in a level is not the flow.** A level change = transactions (`f`) +
  revaluations (`FR`) + other volume changes. Use the `f`-series when you mean
  actual flows; price changes are netted out there.
- **Flows are seasonally adjusted annual rates.** The `f`-tables report SAAR;
  levels are not annualized. Read the table header before scaling.
- **Households and nonprofits are lumped.** The headline household table (sector
  15) combines households with nonprofit organizations; a nonprofit-only split
  exists in separate tables if you need households alone.
- **The bulk URL date moves.** The dated path (`/20260319/`) is only good for
  one release. Pin it for reproducibility, or script resolving the latest date
  from the release page.

## Tables you actually need

| Table | Sector / content | Type |
|---|---|---|
| `b101` | Households and nonprofits balance sheet (net worth) | levels |
| `b103` | Nonfinancial corporate business balance sheet | levels |
| `l108` | Households credit-market instruments | levels |
| `f101` | Households flows | flows |
| `s` tables | Sector income and saving statements | mixed |
| `data_dictionary/` | Table and series-code descriptions | reference |

## Standard operations

- **Household net worth / wealth:** read `b101`; state real vs nominal and the
  quarter, since the series is revised.
- **Credit and leverage:** combine the relevant `l` level table with the `f`
  flow table; never infer flows by differencing levels.
- **Sector consistency:** the accounts are designed to sum across sectors; use
  that as a sanity check rather than mixing in unrelated sources.
- **Pin the release.** Record the dated release path (or "preview") and your
  access date; every quarter revises history.

## Citation

Board of Governors of the Federal Reserve System. *Financial Accounts of the
United States* (Z.1 statistical release). Washington: Federal Reserve Board.
Cite the specific release date and your access date.


==============================================================================
# Forbes executive compensation surveys
# https://instituteforautomatedresearch.org/wiki/datasets/forbes-executive-compensation/
# The Forbes annual executive compensation surveys (roughly 1970-1992) are the pre-ExecuComp source of U.S. CEO pay. No maintained machine-readable file exists; researchers reconstruct figures from archived print issues or reuse compiled tables. The modern successor is Compustat ExecuComp (1992, licensed).
# Tags: executive-compensation, corporate-governance, historical, no-stable-source, data:forbes-executive-compensation
==============================================================================

:::note[Historical print surveys, no standing machine-readable source]
The Forbes CEO-pay surveys exist only as compilations from print Forbes issues. There is no maintained downloadable dataset, no archive endpoint, and nothing to pull here. The page carries no provenance badge: it documents the dataset and how researchers have reconstructed it, but no data was exercised here.
:::

**The Forbes annual executive compensation surveys** (roughly 1970 to 1992) are the standard pre-ExecuComp source of U.S. CEO pay. Forbes published annual surveys of CEO salary, bonus, and in later years option gains for large U.S. companies. The data exist only as figures in print magazine issues; researchers reconstruct a machine-readable panel by hand-keying from archived issues or by reusing compiled tables from prior studies, most notably Gibbons and Murphy (1992). Used in, for example, [Borgschulte, Guenzel, Liu, and Malmendier (2025)](/wiki/papers/jf/2025/borgschulte-ceo-stress-aging-death-2025/), where the Forbes surveys 1975-1991 (via the Gibbons and Murphy (1992) compilation) form the mortality analysis sample of 2,720 CEOs at 1,501 firms.

- **Cost:** no purchase price for the compiled data (it is an academic hand-collection); accessing original print issues requires library access or archive subscriptions.
- **Machine-readable source:** none maintained; hand-keyed by researchers from print issues.
- **Modern successor:** Compustat ExecuComp (1992 onward, licensed via WRDS).
- **Coverage:** roughly 1970 to 1992, annual; large U.S. companies, not a complete universe.
- **Reference sources:** EPI CEO pay series (https://epi.org/publication/ceo-pay); Gabaix and Landier (2008) (https://pages.stern.nyu.edu/~xgabaix/papers/CEO_EJ.pdf).

## Access

There is no download endpoint or authoritative machine-readable file. To use the Forbes surveys, researchers typically:

1. Obtain the compiled dataset from a prior study. The most common source is Gibbons and Murphy (1992), which covers 1974-1991; Gibbons and Murphy (1990) covers an earlier window.
2. Hand-key additional years from archived print Forbes issues. Some years are digitized on ProQuest Business Collection or similar subscription databases; major research libraries hold print archives.
3. Splice in ExecuComp for 1992 onward (licensed, via WRDS) and apply a concordance to align definitions across the pre- and post-1992 samples.

No official digital archive of the compiled data exists. Researchers who need the data typically contact the authors of studies that built it, subject to data-sharing arrangements.

## Gotchas (the ones that bite pipelines)

- **No authoritative machine-readable file exists.** Each paper that uses pre-1992 CEO pay either hand-keys the data anew or inherits another researcher's compilation, including that compilation's coverage gaps and definitional choices.
- **Forbes pay definitions vary year to year and differ from ExecuComp.** Forbes reported salary plus bonus in early years and later included estimated option gains using varying methodologies. ExecuComp reports TDC1 (total compensation including option grants at grant-date value) and TDC2 (including options exercised). The series are not directly comparable without a definitional bridge.
- **Coverage is incomplete and biased toward larger firms.** Forbes surveyed a subset of large companies; the cross-section is not a fixed universe and changes year to year. Survivorship and selection into the survey introduce bias in the cross-section.
- **Option valuations before 1992 are crude.** Pre-1992 there was no required proxy-statement disclosure of option grant details, so option valuations in the Forbes surveys are approximate or absent depending on the year.
- **Splicing with ExecuComp requires a concordance.** Companies must be matched by name or GVKEY across the two sources; unmatched observations are a common data-quality problem in pre- and post-1992 panels.
- **The 1992 break is a hard discontinuity.** Proxy-statement disclosure rules changed with SEC regulations effective 1993 (for fiscal year 1992), making ExecuComp the post-1992 standard. Pre- and post-1992 samples should be analyzed with care when pooled.

## Reference

| Field | Value |
|-------|-------|
| Source | Print Forbes annual surveys, approximately 1970-1992 |
| Machine-readable | None maintained; hand-compiled by researchers |
| Coverage | Large U.S. companies; survey composition varies year to year |
| Modern successor | Compustat ExecuComp, 1992 onward (licensed, via WRDS) |
| Key compiled source | Gibbons and Murphy (1992): covers 1974-1991 |
| Key required | Not applicable |

## Citation

Cite the specific Forbes survey year(s) and issue, or the compiled secondary source used (e.g., Gibbons and Murphy 1992 with full citation), and note the definition of pay used (salary plus bonus, with or without option gains). If spliced with ExecuComp for 1992 onward, note both sources and the concordance applied. For the Gibbons and Murphy compilation: Gibbons, Robert, and Kevin J. Murphy. "Optimal Incentive Contracts in the Presence of Career Concerns: Theory and Evidence." *Journal of Political Economy* 100, no. 3 (1992): 468-505.


==============================================================================
# DOL Form 5500: ERISA pension & welfare plan filings
# https://instituteforautomatedresearch.org/wiki/datasets/form-5500/
# How to pull DOL Form 5500 plan filings and schedules for free: the apex-redirect-hangs-urllib trap, the ACK_ID-not-EIN join key, and the "Latest is unstable" trap, for automated pipelines.
# Verified 2026-05-16 · tested with live www.askebsa.dol.gov fetch (F_5500_2022, 30 MB) + apex-redirect gotcha confirmed
# Tags: pensions, filings, panel-data, free, no-api-key, dol, data:form-5500
==============================================================================

**Form 5500** (US DOL / EBSA) is the free annual ERISA filing for every
US pension and welfare plan with ≥100 participants. Schedule H gives
plan-level asset breakdowns, including **mutual-fund holdings**, making it
the go-to free source for retirement, household-finance, DC/401(k), and
pension-as-shareholder research. It is what the
[ZeroPaper](https://github.com/alejandroll10/zeropaper) pipeline uses for
retirement and labor-finance work.

- **Cost:** free, no auth.
- **Coverage:** annual from 1999; stable schedule structure from 2009 (post
  EFAST2). Plans <100 participants file 5500-SF (separate).
- **Home:** <https://www.dol.gov/agencies/ebsa/about-ebsa/our-activities/public-disclosure/foia/form-5500-datasets>

## Access

Per-year, per-schedule ZIPs (no auth):

```
https://www.askebsa.dol.gov/FOIA%20Files/{year}/Latest/F_{NAME}_{year}_Latest.zip
NAME ∈ {5500, SCH_A, SCH_C, SCH_D, SCH_G, SCH_H, SCH_I, SCH_R, SCH_MB, SCH_SB}
```

```python
import io, zipfile, requests, pandas as pd

def get_5500(year, name="5500"):
    url = (f"https://www.askebsa.dol.gov/FOIA%20Files/{year}"
           f"/Latest/F_{name}_{year}_Latest.zip")
    z = zipfile.ZipFile(io.BytesIO(
        requests.get(url, stream=True, timeout=120).content))
    return pd.read_csv(z.open(z.namelist()[0]), low_memory=False)

main22 = get_5500(2022)            # ~243K plan filings
sch_h  = get_5500(2022, "SCH_H")   # plan financials / asset breakdown
joined = main22.merge(sch_h, on="ACK_ID")   # filing-level key (see gotchas)
```

## Gotchas (the ones that bite pipelines)

The reason to read this page rather than the DOL site. Verified live on the
date above (the 2022 main file is a ~30 MB ZIP; the apex-redirect below was
reproduced).

- **Hit `www.askebsa.dol.gov`, not the apex.** The apex
  `askebsa.dol.gov` issues a 301 whose `Location` contains a **literal
  space** (`FOIA Files`, not `FOIA%20Files`); confirmed live. That malformed
  header hangs `urllib`. Always request the `www.` host directly.
- **Join schedules on `ACK_ID`, not `EIN`.** `ACK_ID` is the DOL filing ID and
  the correct key to attach a schedule to one filing. `(EIN, PN)` identifies a
  *plan across years*; use that for a plan-year panel, not for joining
  schedules.
- **"Latest" is a moving target.** DOL revises files as late filings arrive.
  Row counts change month to month; snapshot the cache and report the access
  date, or your results aren't reproducible.
- **Slow source.** ~50 KB/s on `urllib`; use `requests` streaming (~2.5 MB/s).
  Expect a multi-minute first download per (year, schedule). Cache to
  `data/form_5500/`.
- **Big in memory.** Schedule H is ~52 MB CSV/year; multi-year panels reach
  hundreds of MB; use a column subset (`usecols=`).
- **Schedule structure stabilized in 2009.** Pre-2009 column names differ;
  don't assume a 2022 schema for a 2005 file.

## Key columns

| Column | Meaning |
|---|---|
| `ACK_ID` | DOL filing ID, the join key for schedules |
| `EIN`, `PN` | Sponsor EIN + plan number; `(EIN, PN)` = plan across years |
| `TOT_ASSETS_EOY_AMT` | Total plan assets, end of year (Sch H) |
| `INT_REG_INVST_CO_EOY_AMT` | **Mutual-fund holdings** (registered inv. cos.) |
| `INT_COMMON_TR_EOY_AMT` | Common collective trusts (DC substitute) |
| `EMPLR_CONTRIB_*_AMT` / `PARTCP_CONTRIB_*_AMT` | Employer / participant contributions |
| `PARTCP_LOANS_*_AMT` | Participant loans outstanding |

Pair `BOY`/`EOY` columns to construct flows.

## Standard operations

- **Plan-year panel:** stack Schedule H across years, key on `(EIN, PN, year)`.
- **Mutual-fund exposure share:**
  `INT_REG_INVST_CO_EOY_AMT / TOT_ASSETS_EOY_AMT`.
- **Implied flows:** `EOY − BOY·(1+r_t)` with a benchmarked plan return.
- **Sponsor link:** match sponsor EIN to Compustat for firm characteristics.
- **Always state** schedule, year, and access date (DOL revises "Latest").

## Citation

*U.S. Department of Labor, Employee Benefits Security Administration, Form 5500
[Schedule, plan year], public-use research files;
https://www.dol.gov/agencies/ebsa/.../form-5500-datasets, accessed
YYYY-MM-DD.*


==============================================================================
# SEC Form ADV (via IAPD): investment-adviser registration
# https://instituteforautomatedresearch.org/wiki/datasets/form-adv/
# How to pull investment-adviser registration data from the SEC for free via IAPD: the firm report, the search API, and the bulk structured feed, plus why Form ADV lives outside EDGAR and the gotchas that bite fund-classification pipelines.
# Verified 2026-05-29 · tested with live IAPD firm ADV report PDF fetch (CRD 105631) + SEC bulk Form ADV filing-data zip reachable (HTTP 200)
# Tags: filings, free, no-api-key, sec, cross-section, data:form-adv
==============================================================================

**Form ADV** is the registration form every SEC- or state-registered
investment adviser must file, and it is **public and free** through the SEC's
**IAPD** (Investment Adviser Public Disclosure) system. It is the source for
what an adviser *is*: assets under management, client and employee counts,
private-fund details, ownership, disciplinary history, and a narrative
brochure of strategies and fees. Papers that classify funds or advisers
(hedge fund vs mutual fund adviser, private-fund flags) read Form ADV; the
[Kwan, Liu & Matthies](/wiki/papers/jf/2026/kwan-liu-matthies-2026/) attention paper
uses it for fund classification.

The one thing to know first: **Form ADV is not on EDGAR.** Advisers file it
through the IARD/CRD system, not EDGAR, so it has its own identifiers (CRD
numbers, `801-`/`802-` file numbers) and its own access points. Looking for it
under a CIK on `data.sec.gov` returns nothing.

- **Cost:** free, no paywall, no key.
- **Auth:** none. A descriptive `User-Agent` header is courteous and avoids
  throttling, same as EDGAR.
- **Coverage:** all SEC-registered advisers, exempt reporting advisers (ERAs,
  partial form), and state-registered advisers.
- **Home:** <https://adviserinfo.sec.gov/> ·
  **Bulk data:** <https://www.sec.gov/foia-services/frequently-requested-documents/form-adv-data>

## Access

### Option 1: IAPD firm report (one adviser, PDF)

Every registered firm has a public report addressed by its **CRD number**:

```
https://reports.adviserinfo.sec.gov/reports/ADV/{CRD}/PDF/{CRD}.pdf
```

```python
import requests
headers = {"User-Agent": "Your Name your@email.edu"}
crd = 105631   # Bridgewater Associates
pdf = requests.get(
    f"https://reports.adviserinfo.sec.gov/reports/ADV/{crd}/PDF/{crd}.pdf",
    headers=headers,
)   # 200, application/pdf: the full Form ADV Parts 1 and 2
```

This is the human-readable filing (Part 1 data + Part 2 brochure). Confirmed
live on the verified date: a multi-megabyte ADV PDF returned for CRD 105631.

### Option 2: Search API (resolve a name to a CRD)

CRD is the join key, so most pipelines start by resolving a firm name:

```
https://api.adviserinfo.sec.gov/search/firm?query={name}&hl=true&nrows=12&start=0&wt=json
```

Returns matching firms with CRD, SEC number, and location. There is a parallel
`/search/individual` endpoint for adviser representatives.

### Option 3: Bulk structured data (the whole population)

For panel work, the SEC publishes the **structured Part 1 data** (the
check-box and numeric fields, including Schedule D private-fund rows) as
downloadable archives, plus monthly Part 2 brochure dumps, on the FOIA page:

```
https://www.sec.gov/files/adv-filing-data-20111105-20241231-part1.zip   # ~700 MB
https://www.sec.gov/files/adv-filing-data-20001019-20111104.zip
```

The complete-filing-data archives give you every adviser's Part 1 fields over
time, which is what you want for classification at scale rather than scraping
PDFs one CRD at a time. Confirmed reachable on the verified date (HTTP 200,
~700 MB zip).

## Gotchas (the ones that bite pipelines)

The reason to read this page rather than the SEC instructions. Verified
against live IAPD and the bulk feed on the date above.

- **Not on EDGAR; CRD, not CIK.** Form ADV uses CRD/IARD identifiers and
  `801-`/`802-` SEC file numbers. There is no CIK and no `data.sec.gov`
  endpoint. Join to EDGAR-based holdings by name or a hand-built CRD↔CIK map,
  not by a shared key.
- **Self-reported and as-of the latest amendment.** The IAPD firm report shows
  the *current* filing. Advisers must file an annual updating amendment within
  90 days of fiscal year-end, but for a point-in-time history you need the
  bulk archive or dated compilation reports, not the live report.
- **AUM is regulated AUM (RAUM), gross.** Part 1 Item 5.F reports *regulated*
  assets under management: gross of leverage, including uncalled capital
  commitments for private-fund advisers. It is not net AUM and is not directly
  comparable to a 13F dollar value.
- **Three registration regimes, different coverage.** SEC-registered advisers
  (generally >$100M RAUM), **state-registered** advisers (smaller, file with
  states), and **exempt reporting advisers** (ERAs, e.g. some private-fund and
  VC advisers) who file only a truncated Form ADV. Do not assume the
  SEC-registered set is the whole universe.
- **Private funds live in Schedule D Section 7.B.1:** one row per reported
  private fund, with a self-classified fund type (hedge fund, private equity,
  etc.). The type is the adviser's own label; treat it as a self-report.
- **Part 2 is prose, not fields.** The brochure (Part 2A) and supplement
  (Part 2B) are narrative PDFs. Strategy, fee, and conflict classification
  means parsing text, not reading a tagged field.
- **Bulk filenames embed date ranges.** The archive filenames carry the
  period they cover (e.g. `...20111105-20241231-part1.zip`), so the URL goes
  stale as the SEC adds new periods. Re-read the FOIA page for the current
  filenames rather than hard-coding a URL.
- **One adviser can advise many funds.** A fund-classification join is
  adviser→funds, often many-to-many; a single ADV does not map one-to-one to a
  single mutual fund. Cross-reference fund-level filings (e.g. EDGAR
  [N-1A](/wiki/datasets/edgar/#form-n-1a-open-end-fund-registration)) for the
  fund side.

## What's in each part

| Part | Form | Contents |
|------|------|----------|
| 1A | structured | RAUM, client/employee counts, custody, ownership, disciplinary, Schedule D (incl. private funds) |
| 1B | structured | State-registration items |
| 2A | brochure (PDF) | Narrative: advisory business, fees, strategies, conflicts |
| 2B | supplement (PDF) | Background of individual advisory personnel |
| 3 | Form CRS | Relationship summary for retail clients |

## Standard operations

- **Classify advisers/funds:** resolve name to CRD (Option 2), then read
  Schedule D and Item 5/Item 7 from the bulk Part 1 data (Option 3) rather
  than scraping PDFs; the structured fields are the classifiable signal.
- **Build a panel:** use the dated bulk archives for point-in-time fields;
  the live IAPD report is current-only and not reproducible as history.
- **Join to holdings or returns:** map CRD to the EDGAR CIK (by name) to link
  an adviser's Form ADV profile to its 13F /
[N-1A](/wiki/datasets/edgar/#form-n-1a-open-end-fund-registration) filings;
there is no shared
  key, so document the crosswalk.
- **Always record the CRD and the as-of date** of the filing you read; ADV is
  amended continuously, so an undated pull is not reproducible.

## Citation

Cite the form, adviser, and source, e.g.: *Bridgewater Associates, LP, Form
ADV, CRD No. 105631, U.S. Securities and Exchange Commission, Investment
Adviser Public Disclosure (IAPD); https://adviserinfo.sec.gov/, accessed
YYYY-MM-DD.* For bulk data, cite the SEC Form ADV complete filing-data
archive and the archive date range.


==============================================================================
# FRED: Federal Reserve Economic Data
# https://instituteforautomatedresearch.org/wiki/datasets/fred/
# How to pull macro and financial time series from FRED for free, including the no-API-key fallback, the series you actually need for finance and macro calibration, and the gotchas that bite automated pipelines.
# Verified 2026-05-16 · tested with live no-key CSV fetch (GDP, USREC, SP500)
# Tags: macro, time-series, free, no-api-key, federal-reserve, data:fred
==============================================================================

**FRED** (Federal Reserve Economic Data, St. Louis Fed) is the single most
useful free source for macro and financial time series: ~800,000 series
covering output, prices, rates, spreads, the cross-section of Treasury yields,
and recession indicators, with a clean API and, crucially, a **no-authentication CSV
fallback**. It is the default calibration source the
[ZeroPaper](https://github.com/alejandroll10/zeropaper) pipeline reaches for
when a model needs macro moments. This page is the distilled access recipe.

- **Cost:** free, no paywall.
- **API key:** free, optional (a fallback works without one).
- **Coverage:** US macro/financial aggregates, many international series.
- **Home:** <https://fred.stlouisfed.org/> ·
  **API docs:** <https://fred.stlouisfed.org/docs/api/fred/>

## Access

### Option 1: No API key (fallback, zero setup)

Any series has a direct CSV endpoint that needs no authentication:

```
https://fred.stlouisfed.org/graph/fredgraph.csv?id=GDP
```

```python
import pandas as pd
gdp = pd.read_csv(
    "https://fred.stlouisfed.org/graph/fredgraph.csv?id=GDP",
    parse_dates=["observation_date"], index_col="observation_date",
)
```

This works for most series and is the right default for a pipeline that
shouldn't depend on a key being present. Use it unless you need search,
metadata, vintages, or bulk pulls.

### Option 2: `fredapi` (preferred when a key is available)

Get a free key at
<https://fred.stlouisfed.org/docs/api/api_key.html> and store it in the
environment (e.g. `.env` as `FRED_API_KEY=...`); never hard-code it.

```python
# pip install fredapi python-dotenv
import os
from dotenv import load_dotenv
from fredapi import Fred

load_dotenv()
fred = Fred(api_key=os.environ["FRED_API_KEY"])
gdp = fred.get_series("GDP")
```

### Option 3: Direct REST API

```
https://api.stlouisfed.org/fred/series/observations?series_id=GDP&api_key={KEY}&file_type=json
```

Useful for series search (`/fred/series/search`), release calendars, and
ALFRED vintage (real-time) data that `fredapi` doesn't expose as conveniently.

## Gotchas (the ones that bite pipelines)

The reason to read this page rather than the FRED docs. Verified against live
data on the date above.

- **`SP500` is price-only and ~10 years deep.** The series begins ~10 years
  back (confirmed: first observation 2016-05-16) and is an *index level*, not
  total return. For asset-pricing work use the
  [Ken French](/wiki/datasets/) market series or CRSP instead, never `SP500`
  for long-horizon return studies.
- **Revisions.** `GDP`/`GDPC1` are revised for years. If your result depends
  on what was *known at the time*, use ALFRED vintages, not the latest series.
- **Mixed frequencies.** Don't silently merge daily and monthly series;
  resample deliberately and document the convention.
- **Discontinued series.** Some IDs stop updating or are superseded; check the
  last observation date before trusting a "current" value.
- **Rate limits.** The keyed API rate-limits bulk pulls, so batch and cache;
  the CSV fallback is fine for low-volume use.
- **Units & seasonal adjustment.** Many series are indices or seasonally
  adjusted (`SA`) variants. Read the series page; don't assume levels or NSA.
- **CSV column name.** The no-key endpoint returns `observation_date` as the
  date column (not `DATE`); parse it explicitly as shown above.

## Series you actually need

FRED has ~800,000 series; for finance and macro calibration the recurring set
is small. Search the site or the API for anything else.

| Series ID | Description | Frequency |
|-----------|-------------|-----------|
| `GDP` | Nominal GDP | Quarterly |
| `GDPC1` | Real GDP | Quarterly |
| `CPIAUCSL` | CPI, all urban consumers | Monthly |
| `PCEPILFE` | Core PCE inflation | Monthly |
| `FEDFUNDS` | Effective fed funds rate | Monthly |
| `GS10` | 10-year Treasury yield | Monthly |
| `TB3MS` | 3-month T-bill rate | Monthly |
| `BAA10Y` | Baa corporate – 10yr Treasury spread | Monthly |
| `UNRATE` | Unemployment rate | Monthly |
| `PCE` | Personal consumption expenditures | Monthly |
| `VIXCLS` | CBOE VIX | Daily |
| `SP500` | S&P 500 index (price-only, ~10yr, see gotchas) | Daily |
| `USREC` | NBER recession indicator (0/1) | Monthly |

## Standard operations

- **Moments:** report mean, std, and autocorrelation of *growth rates* (log
  differences), not levels, when calibrating to a stationary model.
- **Business-cycle stats:** HP-filter or band-pass the cyclical component;
  use `USREC` for recession dating.
- **Term structure / spreads:** combine yield series (`GS10`, `TB3MS`,
  `BAA10Y`) rather than hunting for a pre-computed spread.
- **Real vs. nominal:** deflate with `CPIAUCSL` or the PCE deflator; be
  explicit about which.
- **Always state the sample period and frequency** when reporting any moment:
  FRED series are revised and extended, so an unstated window is not
  reproducible.

## Citation

Cite the series and provider, e.g.: *U.S. Bureau of Economic Analysis, Gross
Domestic Product [GDP], retrieved from FRED, Federal Reserve Bank of St.
Louis; https://fred.stlouisfed.org/series/GDP, accessed YYYY-MM-DD.* Each
series page lists its original source and the exact suggested citation.


==============================================================================
# GSW: Gurkaynak-Sack-Wright Treasury yield curve
# https://instituteforautomatedresearch.org/wiki/datasets/gsw-yields/
# How to pull the Federal Reserve staff's daily fitted US Treasury zero-coupon yield curve (Svensson model, 1961 to present) for free with no key, which mnemonic is which, and the header and compounding gotchas that bite pipelines.
# Verified 2026-06-09 · tested with live no-key CSV pull of feds200628.csv (16958 rows, 1961-06-14 to 2026-05-29; SVENY/SVENPY/SVENF columns + Svensson BETA/TAU parameters)
# Tags: macro, time-series, free, no-api-key, federal-reserve, data:gsw-yields
==============================================================================

The **Gurkaynak-Sack-Wright (GSW)** dataset is the Federal Reserve Board
staff's daily fitted **US Treasury zero-coupon yield curve**, estimated with
the Svensson (1994) six-parameter forward-rate model and updated continuously.
It is the standard source for continuously-compounded zero yields, par yields,
and forward rates at maturities of 1 to 30 years, going back to 1961. This page
is the distilled access recipe.

- **Cost:** free, no paywall, no key.
- **Coverage:** daily (business days), 1961-06-14 to present; longer maturities
  populate only as the sample lengthens. Short maturities start 1961-06-14, but
  the 15-year point begins 1971-11-15 and the 30-year point only 1985-11-25, so
  a deep-history long-maturity series is shorter than the headline 1961 start.
- **Format:** one CSV (`feds200628.csv`), about 16 MB, latest date at the bottom.
- **Caveat:** a staff research product, *not* an official statistical release;
  the whole history can be revised when the curve is re-estimated.
- **Paper / home:** FEDS working paper 2006-28 ·
  <https://www.federalreserve.gov/data/nominal-yield-curve.htm>

## Access

### No key, direct CSV (the only access path needed)

```
https://www.federalreserve.gov/data/yield-curve-tables/feds200628.csv
```

```python
import pandas as pd

url = "https://www.federalreserve.gov/data/yield-curve-tables/feds200628.csv"
# The real header (Date,BETA0,...) is on line 10; skip the 9-line preamble.
gsw = pd.read_csv(url, skiprows=9, parse_dates=["Date"], index_col="Date")

gsw["SVENY10"].dropna().tail()   # 10-year zero-coupon yield, percent, cont. comp.
term = gsw["SVENY10"] - gsw["SVENY02"]   # 10y-2y term spread
```

The companion releases use the same layout: the **TIPS / real** curve is
`feds200805.csv` and the **inflation-compensation** series ship alongside it.
Same `skiprows` trick applies.

## Gotchas (the ones that bite pipelines)

The reason to read this page rather than the raw CSV. Verified against the live
file on the date above.

- **Nine-line preamble.** The file opens with a disclaimer note, a blank line,
  a small legend table mapping each series to its mnemonic, and another blank
  line. The actual column header (`Date,BETA0,...`) is **line 10**, so read with
  `skiprows=9` (or detect the `Date` header). A naive `read_csv` mis-parses the
  legend as data.
- **Mnemonics are not obvious.** `SVENYxx` = continuously-compounded
  zero-coupon yield at `xx` years; `SVENPYxx` = par yield (coupon-equivalent);
  `SVENFxx` = instantaneous forward rate; `SVENnFmm` = the `n`-year forward
  rate `mm` years ahead (e.g. `SVEN1F09` is the one-year rate nine years out).
  All in **percent**.
- **Continuously compounded, not bond-equivalent.** `SVENY` zero yields are
  continuously compounded. To compare with a coupon Treasury quote use the
  `SVENPY` par yields or convert; mixing conventions silently biases spreads.
- **Blanks are missing, not zero.** Early in the sample the curve is fit only to
  about 7 years, so the 10-to-30-year columns are empty (NaN) for older dates.
  `dropna()` per maturity; never fill blanks with 0.
- **The Svensson parameters are in the file.** `BETA0`-`BETA3`, `TAU1`, `TAU2`
  are the fitted parameters; you can reconstruct the yield at any maturity from
  them, but the 1-to-30-year grid is already tabulated, so you rarely need to.
- **Staff product, revised without notice.** The note at the top says so. For
  point-in-time / vintage work, archive the file with its download date; the
  series you pulled today may not match a pull next month.
- **Business days only.** No weekend or holiday rows; align to a trading
  calendar before merging with other daily series.

## Mnemonics you actually need

| Mnemonic | Meaning | Maturities | Unit |
|---|---|---|---|
| `SVENYxx` | Zero-coupon yield (continuously compounded) | 01-30 | percent |
| `SVENPYxx` | Par yield (coupon-equivalent) | 01-30 | percent |
| `SVENFxx` | Instantaneous forward rate | 01-30 | percent |
| `SVENnFmm` | `n`-year forward, `mm` years ahead | `1F01`, `1F04`, `1F09` only | percent |
| `BETA0..3`, `TAU1`, `TAU2` | Svensson curve parameters | per day | mixed |

## Standard operations

- **Term spreads:** build from the zero curve (`SVENY10 - SVENY02`), not from
  coupon quotes, so the maturities are exact.
- **Excess bond returns / Cochrane-Piazzesi:** use the forward-rate columns
  (`SVENFxx`, `SVENnFmm`) directly rather than differencing yields.
- **Monthly studies:** resample to month-end (last business day); state the
  convention, since the file is daily.
- **Always record the access date.** Because the whole history is revised, an
  unstated download date is not reproducible.

## Citation

Gurkaynak, Refet S., Brian Sack, and Jonathan H. Wright. 2007. "The U.S.
Treasury Yield Curve: 1961 to the Present." *Journal of Monetary Economics*
54(8): 2291-2304. Data maintained and updated by the Board of Governors of the
Federal Reserve System; cite the release and your access date.


==============================================================================
# Gyourko-Mayer-Sinai Superstar Cities database
# https://instituteforautomatedresearch.org/wiki/datasets/gyourko-mayer-sinai/
# How to reach the Superstar Cities long-run MSA house-price database (Gyourko, Mayer & Sinai), plus the gotchas: the data sits behind a free openICPSR sign-in while only the AEA appendix PDF is open, it is decadal MSA panels in Stata format, and the public file ends at the published vintage.
# Verified 2026-06-22 · tested with confirmed the replication package (openICPSR project 114839, doi 10.3886/E114839V1) resolves and is live, and fetched the open AEA online appendix PDF (HTTP 200, application/pdf, 1.48 MB); the data files sit behind a free openICPSR sign-in and were not pulled here
# Tags: housing, prices, real-estate, urban, panel-data, free, no-api-key, academic, data:gyourko-mayer-sinai
==============================================================================

**The Superstar Cities database** is the long-run, MSA-level house-price data
assembled by Joseph Gyourko, Christopher Mayer and Todd Sinai for their paper
"Superstar Cities." It tracks real house-price growth across US metropolitan
areas back to 1950, the basis for ranking "superstar" MSAs (those with the
highest long-run real price appreciation) against the rest. It is a standard free
source for cross-MSA long-run housing-return work. Used in, for example,
[Amaral et al. (2025)](/wiki/papers/jf/2025/amaral-superstar-returns-spatial-heterogeneity-2025/),
which extends the Gyourko-Mayer-Sinai MSA database to 2018 via the American
Community Survey for a 316-MSA cross-section.

- **Cost:** free (data behind a no-cost openICPSR sign-in; the appendix PDF is
  fully open).
- **Replication package:** openICPSR project 114839, deposited with the AEA,
  distributed by ICPSR (2019).
- **Coverage:** decadal MSA-level house prices and related series, reaching back
  to 1950; the public file ends at the published vintage.
- **Paper:** *American Economic Journal: Economic Policy* 5(4), 2013, 167-199.

## Access

The replication data is on openICPSR; the journal also posts an open appendix
PDF (documentation, not the panel).

```text
# Replication package (data behind a free openICPSR / ICPSR sign-in; Google
# login works). Scripted requests get HTTP 403, so a human signs in once:
#   https://doi.org/10.3886/E114839V1
#   -> https://www.openicpsr.org/openicpsr/project/114839/version/V1/view

# Open AEA online appendix PDF (documentation only, ~1.48 MB), no login:
#   https://www.aeaweb.org/content/file?id=20254
```

The data files are typically Stata `.dta` plus Excel and a readme; the MSA-by-year
panels are inside the openICPSR deposit, not in the appendix PDF.

## Gotchas (the ones that bite pipelines)

- **openICPSR is a login wall for automation.** Scripted access returns HTTP 403;
  a human must sign in once (free; Google login works) and download. There is no
  anonymous direct-file URL.
- **The open PDF is not the data.** The AEA `content/file?id=20254` appendix is
  documentation; do not mistake it for the MSA panel. The Wharton and Columbia
  faculty pages host the paper PDF, also not the data.
- **Stata-format deposit.** Expect `.dta` files plus a readme, not a single tidy
  CSV; read the readme for variable definitions and the MSA coding.
- **Decadal, long-run frequency.** The series are decadal MSA house prices back to
  1950, not annual; align carefully if joining to annual data.
- **The public file ends at the published vintage.** Later studies (for example
  Amaral et al. 2025) extend it to 2018 via the ACS themselves; the deposit does
  not include that extension, so you build it or use the authors' replication.
- **Do not confuse versions.** "Superstar Cities" is the AEJ: Economic Policy 2013
  article; an earlier NBER working-paper version circulated from 2006. The
  replication package is the AEA deposit (project 114839).

## Reference

| Field | Value |
|-------|-------|
| Authors | Joseph Gyourko, Christopher Mayer, Todd Sinai |
| Replication host | openICPSR project 114839, doi `10.3886/E114839V1` |
| Open appendix | AEA `aeaweb.org/content/file?id=20254` (PDF, documentation) |
| Format | Stata `.dta` plus Excel and readme |
| Coverage | Decadal MSA house prices, back to 1950, to the published vintage |
| Access | Free openICPSR sign-in (no fee); scripted access blocked |
| Key required | No (but sign-in) |

## Citation

Cite Gyourko, Joseph, Christopher Mayer, and Todd Sinai, "Superstar Cities,"
*American Economic Journal: Economic Policy* 5(4), 2013, 167-199, together with
the replication package (openICPSR project 114839, doi 10.3886/E114839V1) and the
retrieval date. If you use an extension of the series (for example to 2018 via the
ACS), describe that construction separately.


==============================================================================
# HCRIS (Medicare Hospital Cost Reports, CMS)
# https://instituteforautomatedresearch.org/wiki/datasets/hcris/
# How to pull hospital cost-report data from CMS HCRIS, including the no-key bulk download, the flat-file worksheet layout, and the gotchas that bite pipelines (form versions, alpha/numeric files, fiscal-year boundaries).
# Verified 2026-06-09 · tested with live no-key download of the HCRIS hospital cost-report archive (downloads.cms.gov/Files/hcris/HOSP10FY2021.zip, ~137 MB)
# Tags: healthcare, hospitals, filings, microdata, free, no-api-key, data:hcris
==============================================================================

**HCRIS** (the Healthcare Cost Report Information System) holds the annual cost
reports that Medicare-certified institutional providers file with the Centers
for Medicare & Medicaid Services (CMS). The hospital report (Form CMS-2552) is
the most-used: facility-level financials, charges, costs, beds, discharges,
uncompensated and charity care, and Medicare utilization, for nearly every US
hospital. The data is free and public. Used in, for example,
[Lewellen](/wiki/papers/jf/2025/lewellen-women-charge-evidence-hospitals-2025/),
where the 2011-2018 hospital reports supply Charity Care and Uninsured Discounts
spending.

- **Cost:** free, public.
- **API key:** none required for the bulk files.
- **Coverage:** institutional Medicare cost reports; the hospital form
  (CMS-2552) for essentially all US hospitals, by provider fiscal year.
- **Home:** <https://www.cms.gov/data-research/statistics-trends-and-reports/cost-reports>

## Access

CMS publishes the data as bulk archives, one per provider type and fiscal year,
under `downloads.cms.gov/Files/hcris/`. The hospital form (CMS-2552-10) files
are named `HOSP10FY<year>.zip`:

```bash
# Hospital cost reports (CMS-2552-10), FY2021 archive, no key
curl -sL -o HOSP10FY2021.zip \
  "https://downloads.cms.gov/Files/hcris/HOSP10FY2021.zip"
unzip HOSP10FY2021.zip
```

Each archive holds the standard HCRIS flat files for that year:

- `*_RPT.CSV` - report-level records (one row per cost report: provider number,
  fiscal-year begin/end dates, status, the report record number that joins the
  others).
- `*_NMRC.CSV` - the numeric data lines (worksheet, line, column, numeric
  value).
- `*_ALPHA.CSV` - the alphanumeric data lines (text-valued cells).
- Plus roster and supporting files.

A single reported figure (say, charity care) is a cell located by
**worksheet + line + column** in the numeric file, joined to the report record
in `RPT`. There is no wide "one row per hospital" table; you pivot the long
worksheet cells yourself using the cost-report form layout.

### Load in Python

```python
import pandas as pd

rpt = pd.read_csv("HOSP10_2021_RPT.CSV", header=None, dtype=str)
nmrc = pd.read_csv("HOSP10_2021_NMRC.CSV", header=None, dtype=str)
# NMRC columns: record number, worksheet code, line, column, value.
# Locate a cell by (worksheet, line, column) from the CMS-2552-10 form,
# then join to RPT on the report record number to get provider + fiscal year.
```

Confirm the column order against the HCRIS documentation for the vintage you
pulled; the flat files ship without header rows.

## Gotchas (the ones that bite pipelines)

- **Two form versions coexist: CMS-2552-96 and CMS-2552-10.** Hospitals moved
  to the `-10` form for cost-reporting periods beginning on or after 1 May 2010.
  Worksheet, line, and column coordinates differ between the two forms; a line
  number is meaningless without knowing which form produced it. The `HOSP10`
  files are the `-10` form; older years use the `-96` layout.
- **Data lives in long worksheet cells, not columns.** A variable is a
  (worksheet, line, column) triple in the numeric file. Build an explicit
  mapping from the form before extracting; the wrong line silently returns a
  different concept.
- **Provider fiscal years are not calendar years and vary in length.** Each
  report covers that hospital's own fiscal year, and short or long periods occur
  (mergers, fiscal-year changes). Align on the begin/end dates in `RPT`, not on
  a calendar-year label.
- **One provider can have multiple reports for a period.** Amended, reopened,
  or superseded reports appear; use the report status and dates in `RPT` to keep
  the right one rather than summing duplicates.
- **The most recent fiscal years are incomplete.** Reports arrive and settle
  with a lag, so the latest year's archive keeps growing and is partial when
  first posted. Treat recent vintages as not-yet-final.
- **Alphanumeric versus numeric files hold different cells.** Text-valued
  fields are in `ALPHA`, numeric in `NMRC`. A field you cannot find in one is
  often in the other.

## Provider types

| File prefix | Form | Provider type |
|-------------|------|---------------|
| `HOSP10` | CMS-2552-10 | Hospitals (2010 form onward) |
| `HOSP` | CMS-2552-96 | Hospitals (1996 form) |
| `SNF10` | CMS-2540-10 | Skilled nursing facilities |
| `HHA10` | CMS-1728-94/-20 | Home health agencies |
| `HOSPICE` | CMS-1984-14 | Hospices |

Match the file prefix and form version before pulling line numbers; layouts are
form-specific.

## Citation

Cite CMS, HCRIS, the form and provider type, the fiscal years, the URL, and the
access date, for example: *Centers for Medicare & Medicaid Services, Healthcare
Cost Report Information System, Hospital Form CMS-2552-10, FY2021, retrieved
from https://downloads.cms.gov/Files/hcris/, accessed YYYY-MM-DD.* Record the
form version and the worksheet/line/column coordinates used so the extract is
reproducible.


==============================================================================
# HMDA: Home Mortgage Disclosure Act loan data
# https://instituteforautomatedresearch.org/wiki/datasets/hmda/
# How to pull HMDA mortgage application and origination records for free: the CFPB data-browser CSV and aggregations API, the 2018 schema break, the privacy-binned public file, the string-ranged DTI trap, and the action_taken and sentinel-code gotchas.
# Verified 2026-06-09 · tested with live ffiec.cfpb.gov/v2/data-browser-api fetch (DC 2022 loan-level CSV + aggregations)
# Tags: mortgages, housing, filings, panel-data, free, no-api-key, cfpb, data:hmda
==============================================================================

**HMDA** (Home Mortgage Disclosure Act) data is the near-census of US mortgage
activity: most home-purchase, refinance, and home-improvement applications and
originations, reported loan by loan with borrower demographics, geography, and
(since 2018) pricing and underwriting fields. It is the standard free source for
mortgage-market, fair-lending, redlining, and household-finance research. This
page is the access recipe for the CFPB data browser and its API.

- **Cost:** free, no auth, no API key.
- **Coverage:** nationwide loan-level. Modern expanded schema from 2018; older
  vintages back to the 1990s on a different, smaller schema.
- **Granularity:** one row per application/loan (the Loan/Application Register,
  LAR), privacy-protected in the public release.
- **Home:** <https://ffiec.cfpb.gov/> ·
  **Data browser:** <https://ffiec.cfpb.gov/data-browser/>

## Access

### Option 1: Data-browser CSV (loan-level)

A no-key endpoint returns filtered loan-level rows as CSV. Filter by year,
geography, and any LAR field:

```python
import pandas as pd

url = ("https://ffiec.cfpb.gov/v2/data-browser-api/view/csv"
       "?states=DC&years=2022&actions_taken=1&loan_purposes=1")
lar = pd.read_csv(url, low_memory=False)   # originated home-purchase loans, DC 2022
```

The same query against `/view/aggregations` returns counts and dollar sums
without pulling rows, which is the right call for tabulations:

```
https://ffiec.cfpb.gov/v2/data-browser-api/view/aggregations?states=DC&years=2022&actions_taken=1
# -> {"aggregations":[{"count":14277,"sum":9.96e9,"actions_taken":"1"}], ...}
```

### Option 2: Snapshot national loan-level files

For a full-year national pull, download the annual snapshot dataset rather than
paging the browser API by state:
<https://ffiec.cfpb.gov/data-publication/snapshot-national-loan-level-dataset/2022>
(swap the year). These are large ZIPs of the same privacy-protected LAR plus the
transmittal sheet (institution names and IDs).

## Gotchas (the ones that bite pipelines)

The reason to read this page rather than the CFPB site. Verified live on the
date above (DC 2022 loan-level CSV and the aggregations endpoint were pulled).

- **The 2018 schema break.** Expanded HMDA added many fields in 2018
  (`interest_rate`, `loan_to_value_ratio`, `debt_to_income_ratio`,
  `property_value`, `applicant_credit_score_type`, `loan_term`, fees). Pre-2018
  vintages lack them and key lenders differently. Do not assume the modern
  schema for an older year.
- **The public file is privacy-protected, not the raw record.** The released LAR
  bins and suppresses fields: ages are bucketed (`35-44`), debt-to-income is a
  range string, free-text and some exact values are withheld, and small cells
  are masked. The unredacted register is confidential (the FFIEC version);
  treat the public file as a privacy-protected extract.
- **`debt_to_income_ratio` is a string with ranges.** Values like `50%-60%`,
  `<20%`, or `>60%` (and `NA`) sit in the column, so it does not parse as a
  number. Map the buckets explicitly; do not `astype(float)`.
- **Sentinel codes, not nulls.** Missing/not-applicable values use codes:
  `9999` (e.g. `co-applicant_age` when there is no co-applicant), `8888`,
  `NA`, and `Exempt` (small-filer partial exemption). Filter these before
  computing statistics rather than treating them as real magnitudes.
- **`action_taken` controls what a row means.** `1` originated, `2` approved not
  accepted, `3` denied, `4` withdrawn, `5` closed incomplete, `6` purchased
  loan, `7` preapproval denied, `8` preapproval approved but not accepted.
  Filter `action_taken=1` for originations; include
  `1-5` for the application funnel; **purchased loans (`6`) are not new
  originations** and double-count if you also have the originating lender.
- **Use the `derived_` fields for demographics.** `derived_race`,
  `derived_ethnicity`, `derived_sex`, and `derived_loan_product_type` are
  CFPB-computed roll-ups of the raw multi-field encodings
  (`applicant_race-1..5`, etc.); the raw fields are messier to aggregate.
- **Lender identity changed in 2018.** Post-2018 the lender is the `lei` (Legal
  Entity Identifier); pre-2018 it was `respondent_id` + `agency_code`. Join to
  the transmittal sheet / panel for institution names.
- **Geography is FIPS.** `state_code` (2-letter), `county_code` (5-digit FIPS),
  and `census_tract` (11-digit) are the geo keys; tract-level fields
  (`tract_minority_population_percent`, `tract_to_msa_income_percentage`) are
  prejoined for redlining work.

## Key columns

| Column | Meaning |
|---|---|
| `activity_year` | Reporting year |
| `lei` | Lender Legal Entity Identifier (post-2018) |
| `action_taken` | Disposition code (see gotchas) |
| `loan_amount` | Loan amount (dollars; binned for some products) |
| `loan_purpose` | 1 purchase, 2 home improvement, 31/32 refi, 4 other, 5 N/A |
| `derived_race` / `derived_ethnicity` / `derived_sex` | CFPB demographic roll-ups |
| `interest_rate`, `rate_spread`, `loan_to_value_ratio` | Pricing/underwriting (2018+) |
| `debt_to_income_ratio` | Range string, e.g. `50%-60%` (see gotchas) |
| `census_tract`, `county_code`, `state_code` | FIPS geography |
| `denial_reason-1..4` | Denial reasons when `action_taken=3` |

## Standard operations

- **Origination vs application analysis:** filter `action_taken=1` for
  originations; keep `1-5` (drop `6` purchases) for the application funnel and
  denial-rate work.
- **Denial-rate / fair-lending:** model `action_taken in (1,3)` on
  `derived_race`/`derived_ethnicity` with tract and lender controls.
- **Pricing:** use `interest_rate` / `rate_spread` only for 2018+ years.
- **Aggregations first:** use `/view/aggregations` for counts and sums; pull
  loan-level CSV only when you need row detail (large states/years are big).
- **Always state** the year(s), the geography filter, and the `action_taken`
  set, since a count is meaningless without them.

## Citation

*Consumer Financial Protection Bureau / Federal Financial Institutions
Examination Council, Home Mortgage Disclosure Act (HMDA) data; retrieved from
the HMDA data browser (https://ffiec.cfpb.gov/data-browser/), accessed
YYYY-MM-DD.*


==============================================================================
# Health and Retirement Study (HRS)
# https://instituteforautomatedresearch.org/wiki/datasets/hrs/
# The HRS is a biennial U.S. panel of older households covering health, income, wealth, retirement, and expectations. It is free academic data behind a registration and data-use agreement; the portal blocked automated requests from this session, so the download was not exercised here.
# Tags: household, retirement, aging, expectations, survey, panel, free, academic, data:hrs
==============================================================================

:::note[Free academic data, registration-walled, not exercised here]
The Health and Retirement Study is **free** for research, but access requires a
free account and acceptance of a data-use agreement at the University of
Michigan HRS portal. The portal (`hrsdata.isr.umich.edu`) returned automated
requests with a 403 from this session, so the end-to-end download was **not**
run here. The page therefore carries **no provenance badge**: it documents the
dataset and the access path, but the pull is unverified under the institute's
Verified discipline.
:::

**The Health and Retirement Study (HRS)** is a biennial longitudinal survey of
U.S. households over age 50, run by the University of Michigan Institute for
Social Research with sponsorship from the National Institute on Aging and the
Social Security Administration. It follows the same respondents and their
spouses over time on health, cognition, employment and retirement, income and
wealth, pensions and Social Security, and expectations (including subjective
probabilities about inflation, recessions, and stock returns). It is used in,
for example,
[Ke](/wiki/papers/jf/2025/ke-intrahousehold-disagreement-macroeconomic-expectations-2025/),
where the macroeconomic expectations of both spouses and household portfolio
outcomes (stock market participation and equity share) come from the HRS panel.

- **Cost:** free for research; registration and a data-use agreement required.
- **Access:** free account at the Michigan HRS data portal, then bulk download.
- **Coverage:** U.S. households age 50 and over, biennial since 1992, with
  refresher cohorts added over time; individual and couple-level records.
- **Home:** <https://hrs.isr.umich.edu> ·
  **Data portal:** <https://hrsdata.isr.umich.edu>

For a household wealth survey with no registration step, see the
[SCF](/wiki/datasets/scf/), which the same retirement-and-wealth work often
pairs with the HRS.

## Access

- **Register, then download.** Create a free account at the HRS data portal,
  accept the data-use agreement, and download the data products in bulk. The
  raw HRS files are released wave by wave.
- **RAND HRS Longitudinal File.** Most analysts use the RAND HRS Longitudinal
  File rather than the raw wave files: it is a cleaned, harmonized,
  cross-wave extract with consistent variable names and derived measures, which
  removes most of the wave-to-wave reconciliation work.
- **Not exercised here.** The portal returned a 403 to automated requests from
  this session, so confirm the current download path in a browser. Access is
  gated by the account and agreement, not by payment.

## Gotchas (the ones that bite pipelines)

These are the failure modes to expect; the download itself was not run here.

- **Registration and a data-use agreement gate every download.** There is no
  anonymous bulk endpoint; an automated pull needs an authenticated session
  under the agreement, so a pipeline cannot fetch it keyless the way it fetches
  SCF or FRED.
- **RAND HRS and raw HRS are different products.** The RAND file harmonizes and
  renames variables across waves; the raw wave files do not. Mixing the two in
  one script breaks on variable names and codings. Pick one.
- **It is biennial and respondent-linked.** Waves are roughly two years apart,
  and individuals are linked to spouses and across waves by household and person
  identifiers. Couple-level analysis requires joining both spouses' records
  within a wave; getting the link keys wrong silently drops or duplicates
  couples.
- **Complex survey design.** Estimates require the supplied person- and
  household-level weights (and the sampling design for standard errors). An
  unweighted mean is biased toward the sample design.
- **An aging panel attrits, partly through mortality.** Respondents leave the
  panel as they die or drop out, and refresher cohorts enter, so the composition
  shifts over time. Account for attrition and entry when building a balanced
  panel.
- **Detailed pension and Social Security modules are separate.** Rollover and
  pension-disposition events need the detailed-pension supplements, not the core
  panel; do not assume the headline file carries them.

## Citation

Cite the Health and Retirement Study (University of Michigan, sponsored by the
National Institute on Aging and the Social Security Administration), the
specific product and waves used (for example the RAND HRS Longitudinal File and
its version), the variables and sample, and the access date. Follow the data-use
agreement's required acknowledgment language.


==============================================================================
# IMF International Financial Statistics (IFS)
# https://instituteforautomatedresearch.org/wiki/datasets/imf-ifs/
# How to pull cross-country macro and external-sector series from the IMF, including the no-key DataMapper API, the SDMX data portal, and the gotchas that bite pipelines (database moves, units, missing-period gaps).
# Verified 2026-06-09 · tested with live no-key pull of the IMF DataMapper API (api/v1/BCA_NGDPD, current account balance as percent of GDP, all countries)
# Tags: macro, external-sector, cross-country, free, no-api-key, time-series, data:imf-ifs
==============================================================================

**IMF International Financial Statistics (IFS)** is a cross-country database of
macroeconomic and external-sector series published by the International Monetary
Fund: balance of payments, international reserves, exchange rates, monetary and
financial accounts, prices, and national accounts, for most member countries.
The data is free and public. Used in, for example,
[Kremens, Martin & Varela](/wiki/papers/jf/2025/kremens-long-horizon-exchange-rate-2025/),
where it supplies each country's current account balance and capital inflows,
both scaled by GDP.

- **Cost:** free, public.
- **API key:** none required for the DataMapper API or the SDMX data portal.
- **Coverage:** annual, quarterly, and monthly series across most IMF member
  countries; external-sector and macro series, many from the 1940s-1960s
  onward depending on country and indicator.
- **Home:** <https://www.imf.org/en/Data>

## Access

Two no-key paths reach IMF data. The simplest is the **DataMapper API**, which
returns a single indicator across all countries as JSON:

```bash
# Current account balance, percent of GDP, all countries, no key
curl -sL \
  "https://www.imf.org/external/datamapper/api/v1/BCA_NGDPD?periods=2019" \
  -o bca_ngdpd.json
# List every available indicator and its code/units:
curl -sL "https://www.imf.org/external/datamapper/api/v1/indicators" -o indicators.json
```

The JSON is keyed `values -> <indicator> -> <ISO3 country> -> <year> -> value`.
The indicator catalogue (`/indicators`) gives each code's label and unit, so
resolve the code you need there first (for example `BCA_NGDPD` is current
account balance as a percent of GDP).

For the full IFS database, series-level dimensions, and quarterly or monthly
frequencies, use the **IMF data portal** (SDMX REST) at
<https://data.imf.org>. Browse the dataset, select the country, indicator, and
frequency dimensions, and either download the CSV or call the SDMX endpoint.
The DataMapper draws a curated subset of indicators across IMF databases; the
SDMX portal exposes the IFS dataflow in full.

### Load in Python

```python
import requests, pandas as pd

j = requests.get(
    "https://www.imf.org/external/datamapper/api/v1/BCA_NGDPD",
    params={"periods": ",".join(str(y) for y in range(2010, 2020))},
    timeout=60,
).json()

# values -> indicator -> country (ISO3) -> {year: value}
rows = [
    {"iso3": c, "year": int(y), "bca_ngdpd": v}
    for c, series in j["values"]["BCA_NGDPD"].items()
    for y, v in series.items()
]
df = pd.DataFrame(rows)
```

## Gotchas (the ones that bite pipelines)

- **The IMF data portal moved, and the legacy SDMX endpoint is being retired.**
  The old `dataservices.imf.org/REST/SDMX_JSON.svc/` JSON service is unreliable
  and on the way out; the current portal is `data.imf.org`. Pin your access
  path and re-check it, rather than hardcoding the legacy host.
- **Indicator codes are not interchangeable across IMF databases.** The same
  concept (say, current account) appears under different codes in IFS, BOP,
  and WEO, in different units (level in USD, percent of GDP, domestic
  currency). Resolve the exact code and unit from `/indicators` before pulling,
  and do not assume two codes that read alike are the same series.
- **DataMapper is a curated subset, not all of IFS.** If an indicator you need
  is missing from `/indicators`, it may still exist in the full IFS dataflow on
  the SDMX portal. Absence from DataMapper is not absence from IFS.
- **Country coverage and start dates vary by indicator.** A series can begin
  decades later for one country than another, and some country-years are simply
  missing. Do not read a gap as a zero; carry it as missing.
- **Vintages get revised.** External-sector and national-accounts figures are
  revised across releases. Record the access date, and for reproducibility note
  the database and the period you pulled, since a later pull can differ.
- **BPM5 versus BPM6 balance-of-payments definitions.** External-sector series
  shifted from the fifth to the sixth Balance of Payments Manual; the
  definitions and signs are not identical. Check which manual a series uses
  before splicing old and new vintages.

## Selected IFS / external-sector indicators

| Indicator | DataMapper code | Unit |
|-----------|-----------------|------|
| Current account balance, percent of GDP | `BCA_NGDPD` | percent of GDP |
| Current account balance, USD | `BCA` | billions USD |
| Gross national savings, percent of GDP | `NGSD_NGDP` | percent of GDP |
| Total investment, percent of GDP | `NID_NGDP` | percent of GDP |

Confirm the exact code, label, and unit against the live `/indicators`
catalogue before use; the set above is illustrative, not the full IFS.

## Citation

Cite the IMF, the database (International Financial Statistics), the indicator
and country, the URL, and the access date, for example: *International Monetary
Fund, International Financial Statistics, current account balance (percent of
GDP), retrieved from https://www.imf.org/en/Data, accessed YYYY-MM-DD.* Record
the indicator code and frequency so the pull is reproducible.


==============================================================================
# IRS Form 990 (Nonprofit Returns)
# https://instituteforautomatedresearch.org/wiki/datasets/irs-form-990/
# How to pull IRS Form 990 nonprofit information returns free with no key via the e-file index on apps.irs.gov, including officer compensation, board composition, and organization financials, plus the gotchas that bite pipelines.
# Verified 2026-06-09 · tested with live no-key download of the IRS annual e-file index (apps.irs.gov/pub/epostcard/990/xml/2018/index_2018.csv)
# Tags: nonprofits, tax-filings, governance, compensation, free, no-api-key, data:irs-form-990
==============================================================================

**IRS Form 990 (Nonprofit Returns)** is the annual information return that
U.S. tax-exempt organizations must file with the Internal Revenue Service.
Returns cover organization financials (total revenue, expenses, and assets),
the names, titles, and compensation of officers, directors, and key employees,
and governance or board information. Since 2016 the IRS has released
machine-readable versions of e-filed returns as XML files, indexed by year.
The data is free, public, and needs no API key. Used in, for example,
[Lewellen](/wiki/papers/jf/2025/lewellen-women-charge-evidence-hospitals-2025/),
which uses 19 years of Form 990 filings for U.S. nonprofit hospitals to measure
hospital financials, CEO and officer names, titles and salaries, and board
composition.

- **Cost:** free, public.
- **API key:** none required.
- **Coverage:** e-filed returns from tax-exempt organizations (primarily
  501(c) entities), indexed from 2011 onward; coverage grows after the
  e-file mandate phased in more fully around 2016.
- **Home:** <https://apps.irs.gov/pub/epostcard/990/xml/>

## Access

Annual index files and per-filing XML documents are hosted on the IRS public
server at `https://apps.irs.gov/pub/epostcard/990/xml/`. Each year has a CSV
index and a corresponding directory of XML files. Download the 2018 index
directly, no authentication:

```bash
# Annual e-file index for 2018, ~58 MB, no key
curl -sL -o index_2018.csv \
  "https://apps.irs.gov/pub/epostcard/990/xml/2018/index_2018.csv"
```

The index CSV has columns: `RETURN_ID`, `FILING_TYPE`, `EIN`, `TAX_PERIOD`,
`SUB_DATE`, `TAXPAYER_NAME`, `RETURN_TYPE`, `DLN`, `OBJECT_ID`. Each row's
`OBJECT_ID` locates that filing's XML document under the same year directory.

**Note on the AWS mirror:** an AWS Registry of Open Data bucket named
`irs-form-990` historically mirrored these filings, but when checked in this
session the S3 bucket listing returned no objects (`s3://irs-form-990/` lists
empty). Use the `apps.irs.gov` index path above as the working no-key route.

### Load in Python

```python
import pandas as pd
import requests

# Read the index; use dtype=str to avoid silent int casting of EINs/IDs
index = pd.read_csv("index_2018.csv", dtype=str)

# Keep only full Form 990 filers (exclude 990-EZ, 990-PF, 990-T, etc.)
returns_990 = index[index["RETURN_TYPE"] == "990"].copy()

# Fetch one filing's XML by OBJECT_ID
obj_id = returns_990.iloc[0]["OBJECT_ID"]
url = f"https://apps.irs.gov/pub/epostcard/990/xml/2018/{obj_id}_public.xml"
response = requests.get(url)
# response.text holds the XML; parse with lxml or xml.etree.ElementTree
```

Adjust the year in the index URL and the `xml/<year>/` path together to pull
other vintages. Inspect the XML namespace and schema version in the document
root before extracting fields, as the layout differs by return type and filing
year.

## Gotchas (the ones that bite pipelines)

- **Only e-filed returns are in the XML set.** Paper-filed returns, which
  smaller organizations and older filings use, are not in the machine-readable
  index. Coverage is incomplete before the e-file mandate phased in, so the
  pre-2016 years underrepresent smaller nonprofits.
- **The XML schema changed across versions and form types.** Form 990, 990-EZ,
  and 990-PF each report different schedules, and the field paths within a
  given form type also vary across filing-year schema versions. Parse
  defensively: check the schema version element in the document root and map
  fields accordingly rather than assuming a fixed XPath.
- **TAX_PERIOD is not SUB_DATE.** The `TAX_PERIOD` column is the fiscal period
  the return covers; `SUB_DATE` is when the return was submitted. Many
  nonprofits use non-calendar fiscal years, so do not align filings on
  submission date when building time series of organizational activity.
- **One EIN can appear multiple times in an index.** An organization may file
  an original return and one or more amended returns, or returns for multiple
  tax periods, all in the same annual index file. Deduplicate deliberately,
  choosing amended-over-original or latest-filing logic as appropriate.
- **Financial fields are self-reported and definitions vary.** Revenue,
  expense, and asset figures come from the organization's own accounting.
  Line-item definitions can differ across filers; reconcile schedules before
  comparing figures across organizations.
- **990 and 990-PF are not the same form.** Public charities file Form 990
  (or 990-EZ below a revenue threshold); private foundations file Form 990-PF.
  The two forms report different schedules and have different compensation
  disclosure structures. Do not pool them without harmonizing fields.

## Return-type families

| Return type | Who files | Notes |
|-------------|-----------|-------|
| 990 | Public charities and other 501(c) organizations above the 990-EZ threshold | Full form; most detail on compensation and governance |
| 990-EZ | Smaller public charities (generally gross receipts under $200 000 and total assets under $500 000) | Abbreviated form; fewer schedule requirements |
| 990-PF | Private foundations | Different schedules; required regardless of size |
| 990-T | Organizations with unrelated business income | Filed in addition to the primary return; separate filing |
| 990-N (e-Postcard) | Very small organizations (gross receipts normally $50 000 or less) | Not in the XML index; submitted separately via a web form |

## Citation

Cite the IRS, the form type, the index year, the retrieval URL, and the access
date, for example: *Internal Revenue Service, Form 990 Series Returns
(machine-readable e-file index and XML filings), 2018 index, retrieved from
https://apps.irs.gov/pub/epostcard/990/xml/2018/index_2018.csv, accessed
2026-06-09.* Record the return types included and the tax-period range covered
so the pull is reproducible.


==============================================================================
# JST Macrohistory Database
# https://instituteforautomatedresearch.org/wiki/datasets/jst-macrohistory/
# How to pull the Jorda-Schularick-Taylor long-run macro-financial panel (18 advanced economies, annual, 1870 onward) for free as a single spreadsheet, what the key series mean, and the subset-citation and crisis-dummy gotchas that bite pipelines.
# Verified 2026-06-09 · tested with live no-key pull of JSTdatasetR6.xlsx (release R6; 2718 rows, 18 countries, 1870-2020, 59 series incl. crisisJST, eq_tr, housing_tr, debtgdp)
# Tags: macro, banks, panel-data, time-series, free, no-api-key, academic, data:jst-macrohistory
==============================================================================

The **Jorda-Schularick-Taylor (JST) Macrohistory Database** is a long-run
cross-country macro-financial panel: **18 advanced economies, annual, 1870 to
2020** (release R6), assembled in one spreadsheet. It is the standard source for long-run
work on credit, banking crises, house prices, returns on the major asset
classes, and business-cycle facts. This page is the distilled access recipe.

- **Cost:** free, no paywall, no key (academic, freely distributed).
- **Coverage:** 18 advanced economies, annual, **1870-2020** in the current
  release (R6); 59 variables per country-year.
- **Format:** a single Excel file (`JSTdatasetR6.xlsx`, about 1.4 MB, one
  `Sheet1`) or the same as Stata (`JSTdatasetR6.dta`).
- **Home:** <https://www.macrohistory.net/database/>

## Access

### No key, direct download

The download links carry a cache-busting token (`?t=...`); the host serves the
current release at a stable file name. Pull the page once to grab the live link,
or use the path below.

```
https://www.macrohistory.net/app/download/9834512569/JSTdatasetR6.xlsx
https://www.macrohistory.net/app/download/9834512469/JSTdatasetR6.dta
```

```python
import pandas as pd

jst = pd.read_excel("JSTdatasetR6.xlsx")   # one sheet, tidy country-year panel
jst = jst.set_index(["country", "year"]).sort_index()
jst.loc["USA", "debtgdp"].dropna().tail()  # private credit / GDP
```

The panel is already tidy: one row per `country` x `year`, with `iso` and `ifs`
country codes, so it merges cleanly onto other country panels.

## Gotchas (the ones that bite pipelines)

The reason to read this page rather than the workbook. Verified against the live
R6 file on the date above.

- **Cite the right subset, not just the headline paper.** The base panel cites
  Jorda, Schularick and Taylor (2017). The **asset-return** columns (`eq_tr`,
  `housing_tr`, `bond_tr`, `bill_rate`, and the `_interp` variants) require
  citing the 2019 *QJE* "Rate of Return on Everything" paper, and the bank
  balance-sheet ratios (`lev`, `ltd`, `noncore`) require the 2021 *RES* paper.
  Using a column without its source citation is the most common JST mistake.
- **Two crisis dummies, and they differ.** `crisisJST` is the current
  systemic-banking-crisis indicator; `crisisJST_old` is the earlier coding.
  Pick one deliberately and say which; results on crisis prediction move with
  the choice.
- **Real vs nominal is encoded in the name.** `rgdpmad`/`rgdpbarro` are real
  (different source vintages, Maddison vs Barro), `gdp` is nominal, `hpnom` is a
  *nominal* house-price index. Deflate with `cpi` yourself when you need real
  series; do not assume.
- **Returns are total returns where marked `_tr`.** `eq_tr` and `housing_tr` are
  total returns (capital gain plus yield); `eq_capgain` / `housing_capgain` are
  price-only. The `_ipolated` / `_interp` columns fill gaps by interpolation,
  which you may want to exclude.
- **Release versions are not interchangeable.** The file name carries the
  release (R6); country coverage, series, and back-revisions change across
  releases. Pin the release number in your replication, not just "JST".
- **Coverage is unbalanced early.** Many series start well after 1870 for
  several countries (wars, data gaps); the panel is not rectangular. Drop NaN per
  series rather than assuming a balanced panel.

## Series you actually need

| Column | Meaning |
|---|---|
| `rgdpmad`, `gdp` | Real GDP (Maddison), nominal GDP |
| `cpi` | Consumer price index (deflator) |
| `tloans`, `tmort`, `thh`, `tbus` | Total / mortgage / household / business loans |
| `debtgdp` | Private credit to GDP |
| `hpnom` | Nominal house price index |
| `stir`, `ltrate` | Short-term and long-term interest rates |
| `eq_tr`, `housing_tr`, `bond_tr`, `bill_rate` | Total returns on equity, housing, bonds, bills |
| `crisisJST` | Systemic banking-crisis dummy |
| `lev`, `ltd`, `noncore` | Bank leverage, loan-to-deposit, noncore funding |

## Standard operations

- **Credit and crises:** the canonical use is predicting crises or output from
  `debtgdp` growth; choose one crisis dummy and report it.
- **Returns on everything:** the `_tr` columns give comparable long-run real
  returns across asset classes once deflated by `cpi`; cite the 2019 QJE paper.
- **Cross-country panels:** index by `country` x `year`; use `iso` for merges.
- **Pin the release** (R6) and access date for reproducibility.

## Citation

Jorda, Oscar, Moritz Schularick, and Alan M. Taylor. 2017. "Macrofinancial
History and the New Business Cycle Facts." In *NBER Macroeconomics Annual 2016*,
volume 31, edited by Martin Eichenbaum and Jonathan A. Parker. Chicago:
University of Chicago Press. Add the 2019 *Quarterly Journal of Economics*
"Rate of Return on Everything" citation when using the return series, and the
2021 *Review of Economic Studies* citation when using the bank balance-sheet
ratios.


==============================================================================
# Ken French Data Library: factors & test portfolios
# https://instituteforautomatedresearch.org/wiki/datasets/ken-french/
# How to pull Fama-French factors, momentum, and sorted test portfolios for free: the percent-not-decimal trap, the header-rows trap, and the monthly/annual-in-one-file trap, for automated pipelines.
# Verified 2026-05-16 · tested with live CSV-zip fetch (F-F_Research_Data_Factors, 200, monthly-fresh)
# Tags: asset-pricing, factors, time-series, free, no-api-key, academic, data:ken-french
==============================================================================

The **Ken French Data Library** (Tuck/Dartmouth) is the canonical free source
for asset-pricing factors and test assets: FF3/FF5/FF6 factors, momentum,
industry portfolios, and 100+ characteristic-sorted portfolio sets. No
authentication, updated monthly. It is what the
[ZeroPaper](https://github.com/alejandroll10/zeropaper) pipeline uses to
compute alphas and build test assets. This page is the distilled recipe.

- **Cost:** free, no key, no auth.
- **Cadence:** updated monthly (the zip I fetched was dated within the month).
- **Coverage:** US + international factors, sorts, industry portfolios.
- **Home:** <https://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html>

## Access

### Option 1: `pandas-datareader` (preferred)

```python
# pip install pandas-datareader
import pandas_datareader.data as web

ff3 = web.DataReader("F-F_Research_Data_Factors", "famafrench", start="1963")
ff3[0].head()   # dict of DataFrames; [0] = monthly: Mkt-RF, SMB, HML, RF
```

### Option 2: Direct CSV zip (no package)

```python
import pandas as pd, zipfile, io, requests

url = ("https://mba.tuck.dartmouth.edu/pages/faculty/ken.french/"
       "ftp/F-F_Research_Data_Factors_CSV.zip")
z = zipfile.ZipFile(io.BytesIO(requests.get(url).content))
df = pd.read_csv(z.open(z.namelist()[0]), skiprows=3)
```

## Gotchas (the ones that bite pipelines)

The reason to read this page rather than the library site. Verified against a
live download on the date above.

- **Returns are in PERCENT, not decimal.** `Mkt-RF = 1.23` means 1.23%, not
  123%. Divide by 100 before compounding or regressing. This is the single
  most common error against this data.
- **CSV has a multi-line text header.** Numeric data starts after a preamble;
  `skiprows=3` for the factors file, but it varies by dataset; never assume,
  inspect.
- **One file often holds multiple tables.** Monthly *and* annual tables are
  stacked in the same CSV, separated by blank lines and a second header.
  Read only to the first blank line, or split deliberately; a naive
  `read_csv` silently concatenates them.
- **Dates are `YYYYMM` integers** for monthly files (e.g. `202401`), not
  parseable dates. Convert explicitly.
- **Value- vs equal-weighted.** Many portfolio sets ship both; the file/column
  must be stated or the result isn't reproducible.
- **`pandas-datareader` returns a dict**, not a DataFrame; `[0]` is monthly,
  `[1]` annual, with a `'DESCR'` key documenting the build.

## Key datasets

| Dataset | Contents |
|---|---|
| `F-F_Research_Data_Factors` | Mkt-RF, SMB, HML, RF (monthly/annual) |
| `F-F_Research_Data_5_Factors_2x3` | + RMW, CMA |
| `F-F_Momentum_Factor` | MOM |
| `25_Portfolios_5x5` | Size × B/M test assets |
| `100_Portfolios_10x10` | Size × B/M, fine grid |
| `6_Portfolios_2x3` | The sorts used to build the factors |

100+ more on the site (industry, international, single-sort characteristics).

## Standard operations

- **Factor models:** regress excess returns on FF3/FF5/FF6.
- **GRS test:** joint test that a set of alphas is zero (use the matching
  test-portfolio set).
- **Sharpe ratios:** mean/std of factor returns, annualized appropriately.
- **Fama-MacBeth:** cross-sectional regressions on portfolio test assets.
- **Always report** sample period and value- vs equal-weighting.

## Citation

*Fama, E. F., and K. R. French. Data from the Kenneth R. French Data Library,
Tuck School of Business, Dartmouth College;
https://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html,
accessed YYYY-MM-DD.* Cite the originating papers (Fama-French 1993, 2015;
Carhart 1997) as appropriate.


==============================================================================
# Florida-UCLA-LoPucki Bankruptcy Research Database (BRD)
# https://instituteforautomatedresearch.org/wiki/datasets/lopucki-brd/
# Case-level records for large U.S. public-company bankruptcies (Chapter 11/7, assets >= $100 million in 1980 dollars) from 1980 through the December 2022 final update, distributed free via a research-use click-through agreement at lopucki.law.ufl.edu; the database is frozen and will not be updated further.
# Verified 2026-06-22 · tested with live no-login download of the Cases table ZIP from lopucki.law.ufl.edu/download_cases_table.php (HTTP 200, application/zip, 2,938,544 bytes) containing the Florida-UCLA-LoPucki Bankruptcy Research Database 1-12-2023 CSV and XLSX, Protocols.pdf, and Stata import/label .do files
# Tags: bankruptcy, corporate-finance, distress, event-data, free, no-api-key, academic, data:lopucki-brd
==============================================================================

**The Florida-UCLA-LoPucki Bankruptcy Research Database (BRD)** is a case-level
record of large U.S. public-company bankruptcy filings, assembled by Lynn M.
LoPucki and now preserved by the University of Florida Levin College of Law (the
site formerly resided at UCLA; lopucki.law.ucla.edu now redirects). Each row is one bankruptcy case. Covered fields include filing,
confirmation, and emergence dates; the federal court, district, and judge;
whether the case was prepackaged or free-fall; the outcome (emerged, acquired,
or liquidated); whether the firm refiled (Chapter 22 = second filing, Chapter
33 = third); CEO turnover around filing; and professional fees. Used in, for
example,
[Griffin, Nini & Smith (2026)](/wiki/papers/jf/2026/griffin-loan-covenant-violations-decline-2026/),
where BRD filing dates support false-negative classification of covenant violations
near bankruptcy.

- **Cost:** free.
- **API key:** none. The data downloads as a ZIP after a click-through form.
- **Coverage:** Chapter 11 (or Chapter 7) cases where the filing firm had assets
  of at least $100 million in 1980 dollars, as reported in the fiscal year before
  filing. Cases from 1980 onward.
- **Final update:** January 12, 2023 (frozen; no further updates planned).
- **Home:** <https://lopucki.law.ufl.edu>

## Access

The Cases table is served via a click-through research-use agreement at the form
page. After checking the agreement box and submitting, the server delivers the
ZIP directly. There is no terms-free direct URL and no API.

```bash
# Complete the form at:
#   https://lopucki.law.ufl.edu/download_cases_table.php
# After the POST accepting the research-use terms, the server returns (verified 2026-06-22):
#   HTTP 200, Content-Type: application/zip, 2,938,544 bytes
#
# Automated pipelines must handle the form POST carrying the agreement checkbox.
# A manually cached copy avoids repeating the click-through.
```

The ZIP contains:
- `Florida-UCLA-LoPucki Bankruptcy Research Database 1-12-2023.csv` (the data)
- `Florida-UCLA-LoPucki Bankruptcy Research Database 1-12-2023.xlsx` (same data)
- `Protocols.pdf` (field definitions, inclusion rules, coding conventions)
- `BRD Import.do` (Stata import script)
- `BRD Labels.do` (Stata variable-label and value-coding script)

A free sample (`/sample_download.php`) and a user manual are available on the
site without the click-through, useful for inspecting field structure before
agreeing to the terms. A separate "Purchase options" page (`/buy_cases_table.php`)
covers additional or derived products; the core Cases table itself is free.

Read `Protocols.pdf` before using the data. It defines the $100M inclusion
threshold, the coding for each field, and conventions for missing or recoded values.

## Gotchas (the ones that bite pipelines)

- **Frozen database.** The site states the BRD will not be updated after the
  December 2022 update (final file date January 12, 2023). No case filed after
  that date will appear. It is now a historical resource.
- **Click-through agreement required; redistribution is restricted.** The terms
  prohibit disclosing the database or its data except, for the purpose of the
  user's own research, to a co-researcher or research assistant who has accepted
  the same conditions. An automated pipeline must handle the form POST with the
  agreement checkbox; there is no direct no-click URL.
- **One row per case, not per firm.** A firm with two bankruptcy filings (Chapter
  22) appears as two rows. Key on the case identifier, not the company name, for
  all joins.
- **Small bankruptcies absent by design.** The $100M-in-1980-dollars threshold
  excludes small filings entirely. The BRD is not a census of all Chapter 11s.
  For completeness checks or work requiring smaller-firm coverage, pair with
  PACER court records or EDGAR filings.
- **CSV and XLSX contain the same data.** Choose one format; do not merge both.
- **Categorical fields require Protocols.pdf for coding.** The Stata .do files
  apply variable labels and value codings; the raw CSV contains numeric codes
  for outcome, filing type, refiling status, and similar categorical columns.
  Without Protocols.pdf the coding for these fields is ambiguous.

## Reference

| Field | Value |
|-------|-------|
| Host | University of Florida Levin College of Law: `lopucki.law.ufl.edu` |
| Download | Click-through research-use form at `/download_cases_table.php` |
| ZIP contents | CSV, XLSX (same data), Protocols.pdf, BRD Import.do, BRD Labels.do |
| Scope | U.S. public companies, Ch. 11/7, assets >= $100M in 1980 dollars, cases 1980-2022 |
| Unit of observation | One row per bankruptcy case |
| Final update | January 12, 2023 (frozen) |
| Key required | No (research-use click-through required) |

## Citation

Cite Lynn M. LoPucki, "Bankruptcy Research Database," available at
`lopucki.law.ufl.edu` (University of Florida Levin College of Law), with the
specific version (January 12, 2023 update) and the access date. Many papers
also cite associated articles by LoPucki and coauthors that document the
coverage decisions and methodology; consult Protocols.pdf for the preferred
citation form posted on the site.


==============================================================================
# Maryland Judiciary Case Search
# https://instituteforautomatedresearch.org/wiki/datasets/maryland-judiciary/
# Maryland's free public court record lookup, why automated access is prohibited (CAPTCHA added March 2022, HTTP 403 to bots), how bulk access via a Public Information Act request works, and the data-use restrictions on individual records.
# Tags: courts, legal, criminal-justice, event-data, no-api-key, view-only, data:maryland-judiciary
==============================================================================

:::note[Free to view one case at a time, automated access prohibited]
Maryland Judiciary Case Search is free to view in a browser but the terms forbid bulk or automated access. A CAPTCHA blocks scraping, and automated requests returned HTTP 403 from this session. The page carries no provenance badge: it documents the dataset and access path, but the end-to-end pull was not exercised here.
:::

**Maryland Judiciary Case Search** (https://casesearch.courts.state.md.us/) is the public web interface for Maryland state court records, covering criminal and civil cases. It provides individual case records including charges, hearing dates, dispositions, bail decisions, and party information. Used in, for example, [Slutzky and Xu (2025)](/wiki/papers/rfs/2025/slutzky-financial-consequences-pretrial-detention-2025/), where 1.08 million criminal cases (2000-2016) with commissioner IDs, release decisions, bail types, and charge categories form the core treatment and control data, and 386,938 civil judgment lien filings (2000-2020) form one outcome variable.

- **Cost:** free to view individual cases in a browser.
- **Automation:** prohibited by terms; a CAPTCHA (added March 2022) blocks scraping; automated requests return HTTP 403.
- **Bulk access:** requires a formal Public Information Act (PIA) request to the Administrative Office of the Courts; no self-serve bulk export.
- **Key required:** no.
- **Content:** criminal and civil state court records (charges, hearings, dispositions, bail types, party information, judgment liens).
- **Home (browser only):** https://casesearch.courts.state.md.us/

## Access

Individual records are freely accessible by case number or party name through the browser interface. No account or key is needed for a single lookup.

Legitimate bulk records are available only through a formal Public Information Act (PIA) request directed to the Administrative Office of the Courts' PIA representative. The AOC controls what fields and date ranges are released and may impose fees for large extracts. The notice and terms are at https://www.mdcourts.gov/casesearch2/notice and the Case Search FAQ.

## Gotchas (the ones that bite pipelines)

- **CAPTCHA and HTTP 403 block all scraping.** A CAPTCHA (TuringTestPage) was added in March 2022 specifically to stop automated scraping; direct HTTP requests return 403 without reaching any data.
- **Terms restrict use, including for background checks.** The Judiciary's notice states the information should not be used to perform background checks on an individual, and the Judiciary assumes no responsibility for use of the data. Read the terms before building any application on this data.
- **Bulk access is through a formal PIA process, not a download.** There is no self-serve bulk export; you must contact the AOC, specify scope, and await approval and possible fee assessment. Lead time is not fixed.
- **Records are not stable over time.** Records can be expunged or sealed after extraction, so a panel assembled at two different dates will differ. Document the extraction date carefully.
- **Name-based search produces false matches.** Common names return multiple hits; case number is the reliable key for linking records.
- **Criminal and civil records are on the same system.** The same interface covers district court criminal cases, circuit court criminal cases, and civil cases (including judgment liens); be precise about which filing type you are extracting.

## Reference

| Field | Value |
|-------|-------|
| Host | https://casesearch.courts.state.md.us/ |
| Access | Free browser view, one case at a time |
| Automation | Prohibited; CAPTCHA-gated (March 2022 onward); HTTP 403 to bots |
| Bulk | Formal PIA request to the Administrative Office of the Courts |
| Key required | No |
| Content | Criminal and civil state court records, judgment liens |
| Data stability | Not stable: records may be expunged or sealed after extraction |

## Citation

Cite Maryland Judiciary Case Search by the case number(s) and retrieval date. If bulk data was obtained via a PIA request, note the request date, the responding office (AOC), and the date range and fields provided.


==============================================================================
# Form N-MFP (Money Market Fund Holdings)
# https://instituteforautomatedresearch.org/wiki/datasets/n-mfp/
# How to pull SEC Form N-MFP monthly money market fund portfolio holdings free with no key via EDGAR full-text search and the Archives endpoint, plus the gotchas around schema versioning, multi-series filers, and the User-Agent requirement.
# Verified 2026-06-09 · tested with live no-key EDGAR full-text search for N-MFP2 (efts.sec.gov, 1,017 hits in Q1 2024) plus pull of a real filing's primary_doc.xml (CIK 356476)
# Tags: money-market-funds, sec-filings, holdings, short-term-funding, free, no-api-key, data:n-mfp
==============================================================================

**Form N-MFP** is a monthly SEC filing that every U.S. registered money market
fund must submit. Each filing discloses fund-level assets under management, fund
category (e.g., Government, Prime, Tax-Exempt), and a security-level schedule of
portfolio holdings: issuer name, CUSIP, maturity date, face value, amortized cost
or market value, and yield. Filings are public on EDGAR as structured XML, free,
and require no API key (SEC asks for a descriptive User-Agent header). Used in,
for example, [Anderson, Du & Schlusche](/wiki/papers/jf/2025/anderson-arbitrage-capital-global-banks-2025/),
where month-end N-MFP holdings and per-fund AUM construct the Bartik shares and
funding-shift measures.

- **Cost:** free, public.
- **API key:** none required. Descriptive `User-Agent` header required by SEC
  fair-access policy.
- **Coverage:** monthly, each month-end snapshot. All U.S. registered money
  market funds. Current form versions are N-MFP2 and N-MFP3 for more recent
  vintages.
- **Home:** <https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&type=N-MFP>

For a commercial alternative with pre-parsed, time-series data, see the
[Crane Data Money Fund Intelligence](/wiki/commercial/crane-mmf/) page; N-MFP is
the underlying public SEC source for those same filings.

## Access

There are two complementary EDGAR endpoints.

**Full-text search across filers** (covers 2001 onward, paginates):

```bash
# All N-MFP2 filings in Q1 2024, no key; replace the User-Agent with your own
curl -s 'https://efts.sec.gov/LATEST/search-index?q=&forms=N-MFP2&startdt=2024-01-01&enddt=2024-03-31' \
  -A "Your Name your-email@example.com"
```

The JSON response contains a `hits.hits` array. Each entry's `_id` field looks
like `"0001145549-24-015515:primary_doc.xml"`. The `_source.display_names` array
holds a string such as `"Fidelity Investments (CIK 0000356476)"` from which you
can parse the CIK; the `cik` field in the JSON may be empty, so parse
`display_names` defensively.

**Pull one filing's XML** (build the archive URL from CIK and accession number
with dashes removed):

```bash
# CIK 356476, accession 0001145549-24-015515 -> dashes removed in the path
curl -s 'https://www.sec.gov/Archives/edgar/data/356476/000114554924015515/primary_doc.xml' \
  -A "Your Name your-email@example.com"
```

The XML contains `<seriesId>`, `<moneyMarketFundCategory>` (e.g., `Exempt
Government`), `<totalValueOtherAssets>`, and repeated `<nameOfIssuer>` rows for
each holding (e.g., `FEDERAL FARM CREDIT BANK`, `FEDERAL HOME LOAN BANK`).

### Load in Python

```python
import json, re, urllib.request, xml.etree.ElementTree as ET

UA = {"User-Agent": "Your Name your-email@example.com"}

# 1. Query the full-text search for a date window
url = (
    "https://efts.sec.gov/LATEST/search-index?q=&forms=N-MFP2"
    "&startdt=2024-01-01&enddt=2024-01-31"
)
req = urllib.request.Request(url, headers=UA)
hits = json.loads(urllib.request.urlopen(req).read())["hits"]["hits"]

for hit in hits:
    filing_id = hit["_id"]  # e.g. "0001145549-24-015515:primary_doc.xml"
    accession_raw = filing_id.split(":")[0]            # "0001145549-24-015515"
    accession_nodash = accession_raw.replace("-", "")  # "000114554924015515"

    # Parse CIK from display_names if the cik field is empty
    names = hit["_source"].get("display_names", [])
    cik_match = re.search(r"CIK\s+(\d+)", " ".join(names))
    if not cik_match:
        continue
    cik = cik_match.group(1).lstrip("0")  # drop leading zeros for URL

    # 2. Fetch the primary XML
    xml_url = (
        f"https://www.sec.gov/Archives/edgar/data/{cik}"
        f"/{accession_nodash}/primary_doc.xml"
    )
    req2 = urllib.request.Request(xml_url, headers=UA)
    tree = ET.parse(urllib.request.urlopen(req2))
    root = tree.getroot()

    # Strip namespace for convenience
    ns = {"n": root.tag.split("}")[0].lstrip("{")} if "}" in root.tag else {}

    # Fund-level fields
    series_id = root.findtext(".//seriesId") or root.findtext(".//n:seriesId", namespaces=ns)
    category  = root.findtext(".//moneyMarketFundCategory") or \
                root.findtext(".//n:moneyMarketFundCategory", namespaces=ns)

    # Holding rows
    for issuer_el in root.iter("nameOfIssuer"):
        print(series_id, category, issuer_el.text)
```

Inspect the tag names in the XML before production use; they vary across form
versions (see Gotchas).

## Gotchas (the ones that bite pipelines)

- **Form version determines XML schema.** The form changed from N-MFP to N-MFP2
  to N-MFP3, and the XML tag names and structure differ across versions. A parser
  keyed to one version breaks on another; branch on the form type returned in the
  search results before parsing.
- **One CIK, multiple series.** A single filing is submitted by a fund family
  (one CIK) and can cover multiple series, each representing a separate fund.
  Key on `<seriesId>`, not just CIK; aggregating at the CIK level mixes funds.
- **Month-end snapshot only.** N-MFP captures holdings as of the last business
  day of the month. Holdings between month-ends are not observed.
- **Full-text search pagination.** The `efts.sec.gov` endpoint paginates results
  and covers filings from 2001 onward. Use narrow date windows and page through
  results rather than pulling a single large query that hits the cap.
- **User-Agent is required.** The SEC fair-access policy requires a descriptive
  `User-Agent` header (name and contact email). Requests without it are
  throttled or blocked.
- **CIK may be embedded in `display_names`, not in a dedicated field.** The
  `cik` field in the full-text search JSON can be empty. Parse the CIK out of
  the `display_names` string and strip leading zeros when building the archive
  URL path.
- **Value convention varies by tag.** Holdings can be reported at amortized cost
  or market value depending on the schema field. Do not assume one convention;
  read the tag name and map it to the fund's stated valuation method.

## Form versions

| Form | Approx. period | Notes |
|------|---------------|-------|
| N-MFP | 2010-2016 | Original structured XML filing; initial schema |
| N-MFP2 | 2016-2023 | Revised schema following 2014 money market reform; tag names changed |
| N-MFP3 | 2024 onward | Further schema revision; check SEC release notes for tag-level changes |

## Key XML fields (N-MFP2 and N-MFP3)

| Tag | Level | Description |
|-----|-------|-------------|
| `seriesId` | Fund | Series identifier for this fund within the filing |
| `moneyMarketFundCategory` | Fund | Fund type: Government, Prime, Tax-Exempt, etc. |
| `totalValueOtherAssets` | Fund | Total assets in dollars |
| `nameOfIssuer` | Holding | Issuer name for each portfolio security |
| `cusip` | Holding | CUSIP of the security, if available |
| `finalMaturityDate` | Holding | Stated final maturity date |
| `percentageOfMoneyMarketFundNetAssets` | Holding | Holding as a share of net assets |

Tag names may differ slightly in N-MFP (the original version); verify against
the filing you are parsing.

## Citation

Cite the SEC, the form type and version, the filer and period, the accession
number or full archive URL, and the access date, for example: *U.S. Securities
and Exchange Commission, Form N-MFP2, [Fund Family Name] (CIK [NNNNNN]),
filing for [Month Year], accession no. [XXXXXXXXXX-XX-XXXXXX], retrieved from
https://www.sec.gov/Archives/edgar/data/, accessed YYYY-MM-DD.* When citing the
EDGAR full-text search endpoint, note the form type, date range, and the
`efts.sec.gov` URL so the pull is reproducible.


==============================================================================
# NBER-CES Manufacturing Industry Database
# https://instituteforautomatedresearch.org/wiki/datasets/nber-ces/
# Annual U.S. manufacturing industry panel (output, employment, capital, materials, price deflators, and TFP) from the NBER and the Census Bureau's Center for Economic Studies, with the no-key download recipe and the gotchas that bite pipelines.
# Verified 2026-06-09 · tested with live no-key CSV pull of NBER-CES (data.nber.org, nberces5818v1_n2012.csv, NAICS panel 1958-2018)
# Tags: manufacturing, productivity, industry-data, macro, free, no-api-key, academic, panel-data, data:nber-ces
==============================================================================

**NBER-CES Manufacturing Industry Database** is an annual panel of U.S.
manufacturing industries, a joint project of the NBER and the U.S. Census
Bureau's Center for Economic Studies (CES). Each row is one industry in one
year and carries output, employment, payroll, capital, materials, energy,
inventories, price deflators, and total factor productivity. It is maintained
by Randy A. Becker, Wayne B. Gray, and Jordan Marvakov (and predecessors). It
is free and public; no API key. Used in, for example,
[Grigoris & Segal](/wiki/papers/jf/2026/grigoris-investment-upstream-downstream-uncertainty-2026/),
which uses it to validate the link between input price uncertainty and supplier
return volatility (an upstream/downstream price-correlation check).

- **Cost:** free, public.
- **API key:** none required.
- **Coverage:** U.S. manufacturing industries, annual. The NAICS-2012 vintage
  pulled here (`nberces5818v1`) covers 1958 to 2018. A separate SIC-based
  vintage also exists.
- **Home:** <https://www.nber.org/research/data/nber-ces-manufacturing-industry-database>

## Access

Files are served as plain CSV from the NBER data directory with no
authentication.

### Step 1: download the CSV

```bash
# Download the NAICS-2012 vintage (1958-2018), no key
curl -sL -o nberces5818v1_n2012.csv \
  "https://data.nber.org/nberces/nberces5818v1/nberces5818v1_n2012.csv"
```

The file is about 5 MB. The directory <https://data.nber.org/nberces/> lists
each vintage and, within a vintage, both the NAICS version and the SIC version
plus a data dictionary.

### Step 2: load in Python

```python
import pandas as pd

df = pd.read_csv(
    "nberces5818v1_n2012.csv",
    dtype={"naics": str},   # keep industry codes as strings
)
df["year"] = df["year"].astype(int)
```

Parse `naics` as a string so industry codes keep their structure and join
cleanly to other NAICS-coded data; parse `year` as an integer.

## Gotchas (the ones that bite pipelines)

- **NAICS and SIC vintages are not directly comparable.** Industry definitions
  differ between the two bases. Pick one basis and stay on it for a given study;
  do not mix rows from a NAICS file with rows from a SIC file.
- **Industry definitions change across NAICS revisions.** A given code can map
  to a different industry in different vintages or revision years. Cross-walk
  codes before pooling long panels across revisions.
- **Deflators are price indices, not dollars.** The `pi*` series are price
  indices (deflators); deflate the nominal series yourself. Confirm the base
  year and the dollar units (the data dictionary describes nominal dollar
  variables in millions, per the data dictionary) in the documentation rather
  than assuming.
- **Two TFP definitions, each as a rate and a level.** TFP comes in a 4-factor
  variant (`dtfp4`, `tfp4`) and a 5-factor variant (`dtfp5`, `tfp5`), and each
  appears both as a growth rate (`dtfp*`) and as an index level (`tfp*`). Do not
  mix the 4-factor and 5-factor definitions, and do not mix the growth rate with
  the index.
- **This vintage ends in 2018.** `nberces5818v1` is not updated annually, so do
  not expect recent years. Check the directory for a newer vintage before
  pinning one.
- **`naics` reads as an integer by default.** `read_csv` will infer it as an
  integer, which discards any leading zeros and breaks joins to other
  NAICS-coded data. Parse it as a string as shown above.

## Columns

Header row of `nberces5818v1_n2012.csv`, verbatim:

| Column | Meaning |
|--------|---------|
| `naics` | NAICS industry code |
| `year` | Year |
| `emp` | Employment (thousands) |
| `pay` | Total payroll |
| `prode` | Production-worker employment |
| `prodh` | Production-worker hours |
| `prodw` | Production-worker wages |
| `vship` | Value of shipments |
| `matcost` | Cost of materials |
| `vadd` | Value added |
| `invest` | Capital investment |
| `invent` | End-of-year inventories |
| `energy` | Cost of electricity and fuels |
| `cap` | Total real capital stock |
| `equip` | Real equipment capital |
| `plant` | Real structures/plant capital |
| `piship` | Price deflator for shipments |
| `pimat` | Price deflator for materials |
| `piinv` | Price deflator for investment |
| `pien` | Price deflator for energy |
| `dtfp5` | 5-factor TFP growth |
| `tfp5` | 5-factor TFP index |
| `dtfp4` | 4-factor TFP growth |
| `tfp4` | 4-factor TFP index |

Nominal dollar variables are in millions of nominal dollars per the data
dictionary; confirm units and the deflator base year in the documentation for
your vintage.

## Citation

Cite the database, the specific vintage and basis, the maintainers, the NBER
home URL, and the access date, for example: *Becker, Randy A., Wayne B. Gray,
and Jordan Marvakov, NBER-CES Manufacturing Industry Database (vintage
nberces5818v1, NAICS-2012 basis, 1958-2018), National Bureau of Economic
Research,
https://www.nber.org/research/data/nber-ces-manufacturing-industry-database,
accessed YYYY-MM-DD.* For reproducibility, pin the exact vintage filename
(`nberces5818v1_n2012.csv`) so the pull can be repeated.


==============================================================================
# NBER Business Cycle Dates
# https://instituteforautomatedresearch.org/wiki/datasets/nber-cycles/
# How to pull the NBER U.S. business cycle peak and trough reference dates as JSON with no API key, plus the gotchas that bite pipelines (announcement lag, day-component conventions, committee judgment versus the GDP rule).
# Verified 2026-06-09 · tested with live no-key JSON fetch of NBER business cycle reference dates (data.nber.org/data/cycles/business_cycle_dates.json)
# Tags: macro, business-cycle, recession, time-series, free, no-api-key, data:nber-cycles
==============================================================================

**NBER Business Cycle Dates** are the official peak and trough dates for U.S.
business cycles, determined by the Business Cycle Dating Committee of the
National Bureau of Economic Research (NBER). A peak marks the end of an
expansion and the start of a contraction; a trough marks the end of a
contraction and the start of an expansion. The data is free, public, and needs
no API key. Used in, for example,
[Coimbra, Gomes, Michaelides & Shen](/wiki/papers/jf/2026/coimbra-pension-plans-asset-pricing-2026/),
where NBER business cycle frequencies calibrate the recession/expansion Markov
chain for the productivity shock process.

- **Cost:** free, public.
- **API key:** none required.
- **Coverage:** all U.S. business cycle turning points determined by the NBER
  committee. The earliest trough in the JSON is December 1854; the earliest
  peak is June 1857. The latest entry reflects the most recently dated cycle.
- **Home:** <https://www.nber.org/research/business-cycle-dating>

## Access

The JSON endpoint returns an array of objects, each with a `"peak"` and a
`"trough"` field as ISO date strings (e.g. `{"peak":"1857-06-01","trough":"1858-12-01"}`).
The earliest entry has an empty peak and a trough of `1854-12-01`, reflecting
that the committee dates the first trough but not the preceding peak. Fetch
directly, no authentication:

```bash
# No key required; returns a JSON array of peak/trough objects
curl -sL "https://data.nber.org/data/cycles/business_cycle_dates.json"
```

Note: the `.csv` variant of that path returns a 404. The JSON endpoint above
is the confirmed no-key working path.

### Load in Python

```python
import pandas as pd

url = "https://data.nber.org/data/cycles/business_cycle_dates.json"
df = pd.read_json(url)

# Convert to datetime; replace empty strings with NaT
df["peak"]   = pd.to_datetime(df["peak"],   errors="coerce")
df["trough"] = pd.to_datetime(df["trough"], errors="coerce")

# Build a monthly recession indicator (1 = contraction, 0 = expansion)
# Expand each peak-trough pair across a monthly date range, then merge
# into a full calendar index.
```

Once `peak` and `trough` are datetime columns, you can expand each row across
a monthly `pd.date_range` to build a 0/1 recession-indicator series aligned
to any monthly panel dataset.

## Gotchas (the ones that bite pipelines)

- **Long announcement lag.** The committee dates a turning point only after
  sufficient data accumulate, often several months to over a year after the
  event. The latest cycle in the JSON may be undated in real time. Do not
  treat the endpoint as a real-time or near-real-time signal.
- **Day component is a convention, not a measurement.** Peaks and troughs are
  identified to the month; the day is set to the first of the month by
  convention. Do not treat the day as meaningful. Work at monthly frequency or
  strip the day when converting to period labels.
- **Monthly cycle dates are distinct from the NBER quarterly dates.** The NBER
  also publishes quarterly turning points (used in some calibration contexts).
  They are not the same series. Confirm which frequency your source references.
- **Determined by committee judgment, not a fixed rule.** The NBER recession
  definition looks at depth, duration, and diffusion across indicators. It is
  not triggered by two consecutive quarters of negative GDP growth. Do not
  substitute one for the other in code or in writing.
- **FRED USREC is a derived artifact.** The FRED series `USREC` (and related
  `USRECM`, `USRECD`) is constructed from these dates but is a different object
  (monthly 0/1). If you use FRED, cite FRED and its construction note. If you
  use the NBER JSON, cite the NBER endpoint. Do not treat them as
  interchangeable in a citation.
- **First entry has a missing peak.** The array element for the 1854 trough
  has an empty string for `"peak"`. Parse with `errors="coerce"` so it becomes
  `NaT` rather than raising.
- **No versioning in the URL.** The endpoint is updated in place when the
  committee issues new dates. Record the access date and the full array in your
  data snapshot so the pull is reproducible.

## Reference

| Field | Value |
|-------|-------|
| URL | `https://data.nber.org/data/cycles/business_cycle_dates.json` |
| Format | JSON array of `{peak, trough}` ISO date strings |
| Frequency | One row per business cycle |
| Earliest trough | December 1854 |
| Earliest peak | June 1857 |
| Update trigger | Committee announcement (irregular, long lag) |
| Key required | No |

## Citation

Cite the NBER, the Business Cycle Dating Committee, the retrieval URL, and the
access date, for example: *National Bureau of Economic Research, Business Cycle
Dating Committee, "US Business Cycle Expansions and Contractions," retrieved
from https://data.nber.org/data/cycles/business_cycle_dates.json, accessed
YYYY-MM-DD.* Record the access date and archive the JSON array at time of
download; the endpoint is updated in place with no version indicator.


==============================================================================
# NBER Working Papers
# https://instituteforautomatedresearch.org/wiki/datasets/nber-working-papers/
# How to pull NBER working paper metadata and full-text PDFs with no API key: the undocumented listing API, the predictable PDF path, and the gotchas that bite pipelines (copyright/redistribution, gated subset, undocumented API, displaydate strings, working-paper numbering).
# Verified 2026-06-23 · tested with live no-key fetch of the NBER listing API (35,851 results) and a full-text PDF (w31010, 4.1 MB application/pdf)
# Tags: working-papers, full-text, metadata, academic, free, no-api-key, data:nber-working-papers
==============================================================================

**NBER Working Papers** are the pre-publication paper series of the National
Bureau of Economic Research, the main circulation channel for new US economics
and finance research. Both the working paper metadata (title, authors, date,
abstract) and the full-text PDFs are reachable over plain HTTP with no API key.
The metadata comes from an undocumented JSON listing endpoint; the PDFs sit at a
predictable path keyed by the working paper number.

- **Cost:** free to fetch; PDFs are copyrighted (see Gotchas).
- **API key:** none required.
- **Coverage:** the full working paper series. The listing API reported 35,851
  working papers as of the verification date, back to w0001.
- **Home:** <https://www.nber.org/papers>

## Access

### Metadata: the listing API

An undocumented JSON endpoint backs the site's paper listing. It paginates and
returns the total count plus an array of paper records:

```bash
# No key. page is 1-based; perPage caps the page size.
curl -sL "https://www.nber.org/api/v1/working_page_listing/contentType/working_paper/_/_/search?page=1&perPage=50"
```

Each result object carries `title`, `authors` (HTML anchor strings),
`displaydate` (e.g. `"June 2026"`), `abstract`, and `url` (e.g.
`/papers/w35366`). The working paper number is the last path segment of `url`,
not a separate numeric field. Walk `page` from 1 until you have collected
`totalResults` records.

### Full text: the PDF path

Given a working paper number `wNNNNN`, the PDF is at a predictable path:

```bash
# No key. Returns application/pdf.
curl -sL "https://www.nber.org/system/files/working_papers/w31010/w31010.pdf" -o w31010.pdf
```

The human landing page for the same paper is
`https://www.nber.org/papers/w31010` (HTTP 200), useful for the abstract,
JEL codes, and citation block.

### Load metadata in Python

```python
import requests

base = "https://www.nber.org/api/v1/working_page_listing/contentType/working_paper/_/_/search"
rows, page = [], 1
while True:
    r = requests.get(base, params={"page": page, "perPage": 100}, timeout=30).json()
    rows.extend(r["results"])
    if len(rows) >= r["totalResults"] or not r["results"]:
        break
    page += 1

# Working paper number is the tail of the url field
for x in rows[:3]:
    wp = x["url"].rsplit("/", 1)[-1]   # e.g. "w35366"
    print(wp, x["displaydate"], x["title"])
```

## Gotchas (the ones that bite pipelines)

- **The PDFs are copyrighted; fetchable is not redistributable.** Free download
  for research use does not grant the right to rehost or bulk-redistribute the
  PDFs. Treat them as inputs to read and cite, not as a corpus to republish.
  Store locally, cite the NBER, and respect the series' terms.
- **A subset of PDFs is access-gated.** Some downloads (notably very recent
  papers) are served free only to specific groups (subscribers, US journalists,
  residents of low-income countries) and otherwise sit behind a paywall. The
  predictable PDF path works for the open majority; do not assume every
  `wNNNNN.pdf` returns 200. Check the status code and fall back to the landing
  page metadata when a PDF is gated.
- **The listing API is undocumented.** The `/api/v1/working_page_listing/...`
  endpoint is not a published, versioned API. The path, query shape, and field
  names can change without notice. Pin nothing to it that you cannot quickly
  repair, and record the response shape with your snapshot.
- **Working paper number lives in the URL, not a field.** The result objects do
  not expose a clean numeric `wpnum` (it was null in testing). Parse the number
  from the tail of `url` (`/papers/w35366` -> `w35366`).
- **`displaydate` is a string, not a date.** Dates come as `"June 2026"`, not
  ISO. Parse to month resolution; do not expect a day component.
- **Working papers are not peer-reviewed and are revised in place.** A `wNNNNN`
  number is the working paper, which may differ from the final journal version.
  The PDF can be updated without a new number. Record the access date; do not
  treat the WP as the version of record for a published paper.
- **Rate-limit your crawl.** Walking 35k+ records and PDFs is a real load. Page
  politely, cache, and avoid hammering the host. `robots.txt` allows `/papers/`
  and `/system/files/` but disallows the admin and UI search paths.
- **Author field is HTML.** `authors` entries are anchor tag strings
  (`<a href="/people/...">Name</a>`), not plain names. Strip tags before use.

## Reference

| Field | Value |
|-------|-------|
| Metadata endpoint | `https://www.nber.org/api/v1/working_page_listing/contentType/working_paper/_/_/search?page=N&perPage=K` |
| PDF path | `https://www.nber.org/system/files/working_papers/wNNNNN/wNNNNN.pdf` |
| Landing page | `https://www.nber.org/papers/wNNNNN` |
| Metadata format | JSON (`totalResults` + `results[]`) |
| Total working papers | 35,851 (as of 2026-06-23) |
| Key required | No |
| PDF license | Copyrighted; free for research use, not for redistribution |

## Citation

Cite the individual working paper, not the series. Use the NBER's own citation
block from the landing page, for example: *Author(s), "Title," NBER Working
Paper No. NNNNN, National Bureau of Economic Research, Month Year.* Record the
access date and note the working paper number; the PDF may be revised in place
without a new number, and the final published version (if any) may differ.


==============================================================================
# NIPA: National Income and Product Accounts (BEA)
# https://instituteforautomatedresearch.org/wiki/datasets/nipa/
# How to pull the US National Income and Product Accounts (GDP and its components) from the BEA for free, the no-key static-file fallback as well as the API, and the units, revision, and table-vs-series gotchas that bite pipelines.
# Verified 2026-06-09 · tested with live no-key pull of BEA NipaDataA.txt (annual, from 1929) and NipaDataQ.txt (quarterly, from 1947Q1); format SeriesCode,Period,Value
# Tags: macro, time-series, free, no-api-key, bea, data:nipa
==============================================================================

The **National Income and Product Accounts (NIPA)** are the US Bureau of
Economic Analysis's official accounts for **GDP and its components**: output,
consumption, investment, government, trade, income, saving, and the price
deflators. They are the primary source for US macro aggregates at higher detail
than FRED's headline series. This page is the distilled access recipe.

- **Cost:** free, no paywall.
- **API key:** free, optional (a no-key static-file fallback exists).
- **Coverage:** annual from **1929**, quarterly from **1947:Q1**; revised
  regularly, with annual and comprehensive revisions.
- **Home:** <https://www.bea.gov/itable/national-gdp-and-personal-income> ·
  **API docs:** <https://apps.bea.gov/api/signup/>

## Access

### Option 1: No API key (static all-series files)

BEA publishes the entire NIPA series set as flat text, no key required:

```
https://apps.bea.gov/national/Release/TXT/NipaDataA.txt   # annual (~12 MB)
https://apps.bea.gov/national/Release/TXT/NipaDataQ.txt   # quarterly (~35 MB)
```

```python
import pandas as pd

url = "https://apps.bea.gov/national/Release/TXT/NipaDataQ.txt"
# Values carry thousands separators inside quotes, e.g. "244,142".
nipa = pd.read_csv(url, thousands=",")
nipa.columns = ["SeriesCode", "Period", "Value"]   # header is %SeriesCode,Period,Value
gdp = nipa[nipa.SeriesCode == "A191RC"]             # nominal GDP, current dollars
```

This is the right default for a pipeline that should not depend on a key.

### Option 2: BEA API (preferred for specific tables)

Get a free key at <https://apps.bea.gov/api/signup/> and store it in the
environment; never hard-code it. The `NIPA` dataset endpoint takes a
`TableName` (e.g. `T10101` for the percent-change-in-real-GDP table) and a
frequency.

```python
import os, requests
key = os.environ["BEA_API_KEY"]
r = requests.get("https://apps.bea.gov/api/data", params={
    "UserID": key, "method": "GetData", "datasetname": "NIPA",
    "TableName": "T10105", "Frequency": "Q", "Year": "ALL", "ResultFormat": "JSON",
})
```

NIPA aggregates are also mirrored on [FRED](/wiki/datasets/fred/) (e.g. `GDP`,
`GDPC1`, `PCE`), which is simpler when you only need a few headline series.

## Gotchas (the ones that bite pipelines)

The reason to read this page rather than the BEA docs. Verified against the live
static files on the date above.

- **Series code, not table, in the flat files.** `NipaData{A,Q}.txt` are long
  format (`SeriesCode,Period,Value`), keyed by BEA **series codes** (e.g.
  `A191RC` nominal GDP, `A191RX` real GDP, `A001RC` national income). The
  human-friendly *tables* (T10101 etc.) only come through the API or iTable;
  map codes to tables with the BEA series register.
- **Values carry thousands commas, quoted.** A value is written `"244,142"`.
  Parse with `thousands=","` or it imports as text.
- **Nominal vs real vs index suffix.** The trailing letters encode the unit:
  `...RC` current dollars (nominal), `...RX` chained real dollars, `...RG` /
  `...RL` index or percent forms. Do not mix codes across unit conventions.
- **Real series are chained, not additive.** Chained-dollar components do **not**
  sum to the chained aggregate (the chaining residual). For decompositions use
  contributions-to-growth tables, not arithmetic on real levels.
- **Revisions are large and routine.** NIPA has monthly third estimates, annual
  revisions, and comprehensive revisions that can reweight history. For
  point-in-time work use the BEA vintage / "real-time" archive, not the latest
  file.
- **Quarterly often reported as annualized rates.** Many quarterly growth and
  flow series are seasonally adjusted annual rates; check the table definition
  before comparing to a level.
- **Period format differs by frequency.** Annual `Period` is a bare year (e.g.
  `1929`); quarterly appends the quarter (`1947Q1`). Parse per file, do not
  assume one format.

## Series you actually need

| Series code | Description | Unit |
|---|---|---|
| `A191RC` | Gross Domestic Product | current $ |
| `A191RX` | Real GDP (chained) | chained $ |
| `A001RC` | National income | current $ |
| `DPCERC` | Personal consumption expenditures | current $ |
| `A006RC` | Gross private domestic investment | current $ |
| `A822RC` | Government consumption and investment | current $ |
| `DPCERG` | PCE price index | index |

## Standard operations

- **Growth rates:** compute on the chained real series (`...RX`); for component
  contributions use BEA's contributions tables, not differences of real levels.
- **Deflators:** divide nominal by real, or use the published price indexes
  (`DPCERG` for PCE).
- **Point-in-time:** use BEA vintage data when the result depends on what was
  known at the time; the latest file embeds all later revisions.
- **State sample and frequency** with any moment; the files are revised and
  extended.

## Citation

U.S. Bureau of Economic Analysis. *National Income and Product Accounts.*
Suitland, MD: BEA. Cite the specific table or series code and your access date,
e.g. *Bureau of Economic Analysis, Gross Domestic Product [A191RC], accessed
YYYY-MM-DD.*


==============================================================================
# NMLS Consumer Access
# https://instituteforautomatedresearch.org/wiki/datasets/nmls/
# NMLS Consumer Access is a free per-record lookup for licensed mortgage loan originators and companies, but the Terms of Use forbid bulk or automated copying and there is no free bulk feed. Paid NMLS B2B Access is the only legitimate programmatic path for panel data.
# Tags: mortgages, licensing, regulatory, no-api-key, view-only, data:nmls
==============================================================================

:::note[Free single-record lookup, bulk and automated access prohibited]
NMLS Consumer Access is free per-record in a browser but the Terms of Use forbid bulk or automated copying. Automated requests returned HTTP 403 (TuringTestPage CAPTCHA gate) from this session. There is no free bulk feed. The page carries no provenance badge: it documents the dataset and access path, but the end-to-end pull was not exercised here.
:::

**NMLS Consumer Access** (https://www.nmlsconsumeraccess.org/) is the public-facing portal of the Nationwide Multistate Licensing System, administered by the Conference of State Bank Supervisors (CSBS). It provides free single-record lookup of licensed mortgage loan originators (MLOs) and companies: name, employer, work address, license status and history, and career history across states. Used in, for example, [Frame et al. (2025)](/wiki/papers/jf/2025/frame-impact-minority-representation-mortgage-2025/), where NMLS links loan officer names, employers, and work addresses to HMDA applications and serves as the source for BIFSG race imputation over 2012-2019 across 255,277 loan officers.

- **Cost:** free for individual lookups; paid for programmatic/bulk access.
- **Automation:** prohibited by Terms of Use; CAPTCHA-gated.
- **Bulk and API access:** paid NMLS B2B Access (contracted through CSBS) or third-party vendors (e.g., Comergence NMLS Verification API); no free bulk download or open API.
- **Key required:** no for the free single-record site; contract required for B2B.
- **Content:** MLO name, employer, work address, license status, career history across states.
- **Home:** https://www.nmlsconsumeraccess.org/

## Access

Individual records are freely accessible by name, NMLS ID, or company. No account or key is needed for a single lookup.

Programmatic access to NMLS data at scale requires paid NMLS B2B Access (contracted through CSBS) or a third-party data vendor. These are commercial licensing arrangements and are not self-serve.

Researchers who need a panel of MLO records (as in Frame et al. 2025) typically obtain a negotiated data access agreement. The HMDA-NMLS link used in that paper was built by the Federal Reserve under a confidential data arrangement, not by scraping the Consumer Access site.

## Gotchas (the ones that bite pipelines)

- **Bulk and automated access is contractually prohibited and CAPTCHA-gated.** The Terms of Use state: "You agree that you will not use any process to monitor or copy NMLS Consumer Access information in bulk, or to make voluminous, excessive or repetitive requests for information," and that you will not use any device or routine to bypass volume-request protections, or duplicate, download, publish, or distribute the information for any purpose other than as expressly permitted.
- **The free site is one record at a time only.** There is no batch export, no pagination endpoint returning multiple records, and no downloadable file.
- **The only legitimate programmatic path is paid B2B.** If you need a large panel of MLO records, budget for the NMLS B2B licensing fee or a third-party vendor contract.
- **BIFSG imputation requires name and address data from this source.** Researchers who apply BIFSG to infer loan officer race (as in Frame et al. 2025) need the name and work-address fields. Obtaining those at scale requires the B2B route or a data access agreement with a regulated party that holds the data.
- **License records change over time.** Renewals, state moves, and employer changes update the record; snapshots are time-stamped at extraction. A panel covering multiple years requires either repeated snapshots or B2B access to the change history.
- **NMLS ID is the stable join key.** Name-based search returns false matches for common names; use the NMLS ID as the primary key when linking to HMDA or other mortgage data.

## Reference

| Field | Value |
|-------|-------|
| Host | https://www.nmlsconsumeraccess.org/ (Conference of State Bank Supervisors) |
| Access | Free single-record lookup (no account, no key) |
| Automation | Prohibited by ToU; CAPTCHA-gated |
| Bulk and API | Paid NMLS B2B Access or third-party vendor (e.g., Comergence) |
| Key required | No for free lookup; contract for B2B |
| Content | MLO name, employer, work address, license status, career history |
| Stable key | NMLS ID |

## Citation

Cite NMLS Consumer Access (Conference of State Bank Supervisors), the NMLS IDs or retrieval date range, and the method used (single-record lookups or B2B access). Note that bulk or programmatic use requires paid B2B access.


==============================================================================
# NOAA hurricane track data (HURDAT2 / NHC)
# https://instituteforautomatedresearch.org/wiki/datasets/noaa-hurricane/
# NOAA National Hurricane Center best-track tropical-cyclone data (HURDAT2): 6-hourly positions, winds, pressure, and landfalls, with the no-key text-file download recipe and the gotchas that bite pipelines.
# Verified 2026-06-09 · tested with live no-key pull of the Atlantic HURDAT2 best-track file (nhc.noaa.gov, hurdat2-1851-2023)
# Tags: climate, natural-disasters, weather, free, no-api-key, event-data, data:noaa-hurricane
==============================================================================

**HURDAT2** (HURricane DATabase, version 2) is the post-season "best track"
archive maintained by NOAA's National Hurricane Center (NHC). For every
tropical and subtropical cyclone it records 6-hourly entries (plus extra
entries at landfall and intensity peaks) with maximum sustained wind, central
pressure, and the 34/50/64-knot wind radii by quadrant. There is a separate
Atlantic file (since 1851) and a separate Northeast and North-Central Pacific
file (since 1949). Used in, for example,
[Kruttli, Roth Tran & Watugala](/wiki/papers/jf/2025/kruttli-pricing-poseidon-extreme-weather-2025/),
which identifies 37 hurricane landfalls (1996-2019), eye location at 6-hour
intervals, and (separately) wind-speed probability forecasts and seasonal
outlooks.

- **Cost:** free, public domain.
- **API key:** none required.
- **Coverage:** Atlantic from 1851, Northeast and North-Central Pacific from
  1949. Each file is reissued after every annual reanalysis.
- **Home:** <https://www.nhc.noaa.gov/data/> (HURDAT2 section); the Atlantic
  and Pacific files are under <https://www.nhc.noaa.gov/data/hurdat/>

The paper also uses real-time NHC forecast and advisory products (wind-speed
probability forecasts and seasonal hurricane outlooks). Those operational
products are distinct from the post-season HURDAT2 best track; see the gotchas.

## Access

The directory <https://www.nhc.noaa.gov/data/hurdat/> lists both the Atlantic
and the Pacific files along with their issue dates. The dated filename changes
with each annual reanalysis, so list the directory and pick the current file
rather than hardcoding a name.

```bash
# Atlantic best-track file (1851-2023 issue), no key.
# The dated filename changes with each annual reanalysis; list the directory first.
curl -sL "https://www.nhc.noaa.gov/data/hurdat/hurdat2-1851-2023-051124.txt" \
  -o hurdat2_atlantic.txt
```

The file is comma-delimited text with two line types. A storm header line has
three fields (storm ID, name, and the number of data lines that follow); each
following line is one observation. Parse it as a state machine, not as a flat
CSV:

```python
import numpy as np
import pandas as pd

def parse_coord(s):
    s = s.strip()
    val = float(s[:-1])
    if s[-1] in ("S", "W"):
        val = -val
    return val

rows = []
with open("hurdat2_atlantic.txt") as f:
    storm_id = storm_name = None
    for line in f:
        fields = [p.strip() for p in line.split(",")]
        # A header line has exactly three populated fields; data lines have many more.
        if len(fields) == 4 and fields[3] == "":
            storm_id, storm_name = fields[0], fields[1]
            continue
        if len(fields) < 8:
            continue
        rows.append({
            "storm_id": storm_id,
            "storm_name": storm_name,
            "date": fields[0],
            "time": fields[1],
            "record_id": fields[2],        # "L" marks a landfall, else blank
            "status": fields[3],           # HU, TS, TD, ...
            "lat": parse_coord(fields[4]),
            "lon": parse_coord(fields[5]),
            "wind_kt": int(fields[6]),
            "pressure_mb": int(fields[7]),
        })

df = pd.DataFrame(rows)
# Missing values are coded -999; convert before any aggregation.
df = df.replace(-999, np.nan)
```

The first storm header in the Atlantic file is `AL011851, UNNAMED`. To work
basin-wide, repeat the pull for the Pacific file from the same directory.

## HURDAT2 line format

| Position | Header line | Data line |
|----------|-------------|-----------|
| 1 | Storm ID (basin letters + cyclone number + 4-digit year, e.g. `AL011851`) | Date `YYYYMMDD` |
| 2 | Storm name (or `UNNAMED`) | Time `HHMM` UTC |
| 3 | Number of data lines that follow | Record identifier (`L` = landfall, else blank) |
| 4 | | System status (`HU` hurricane, `TS` tropical storm, `TD` tropical depression, others) |
| 5 | | Latitude, e.g. `28.0N` |
| 6 | | Longitude, e.g. `94.8W` |
| 7 | | Maximum sustained wind (knots) |
| 8 | | Minimum central pressure (millibars) |
| 9+ | | 34/50/64-knot wind radii in the four quadrants |

Missing values are coded `-999`.

## Gotchas (the ones that bite pipelines)

- **The filename encodes the version and issue date.** A name like
  `hurdat2-1851-2023-051124.txt` carries the issue date (`051124`) and changes
  with each annual reanalysis. Do not hardcode it; list the directory and pick
  the current file.
- **Two line types in one file.** Storm header lines and observation lines have
  different field counts. Parse with a state machine or a two-pass approach that
  tracks the current storm, not as a flat CSV.
- **Missing values are `-999`, not blank.** Convert them to NaN before
  computing means, maxima, or any aggregate, or the numbers will be wrong.
- **Coordinates are strings with hemisphere suffixes.** Values like `28.0N` and
  `94.8W` must be converted to signed decimals before any distance calculation.
- **Units.** Winds are in knots and pressure in millibars. Convert if you need
  meters per second, miles per hour, or hectopascals.
- **Atlantic and Pacific are separate files.** A basin-wide analysis must read
  both; neither file alone covers all storms.
- **Best track is a post-season reanalysis.** It differs from the operational
  advisories issued in real time. For an identification that depends on what
  investors or agents knew at the time (as in the citing paper's real-time
  forecast use), the best track is not the real-time information set. Keep
  best-track and operational products distinct.
- **One storm can make several landfalls.** The `L` record identifier marks a
  landfall, and a single storm can carry more than one such entry. Do not assume
  one landfall per storm.

## Citation

Cite NOAA NHC, the HURDAT2 database, the basin file and its issue date, the NHC
URL, and the access date, for example: *NOAA National Hurricane Center,
HURDAT2 Atlantic best-track file (hurdat2-1851-2023-051124), retrieved from
https://www.nhc.noaa.gov/data/hurdat/, accessed YYYY-MM-DD.* The issue date is
in the filename; record it so the pull is reproducible.


==============================================================================
# NSMO: National Survey of Mortgage Originations
# https://instituteforautomatedresearch.org/wiki/datasets/nsmo/
# How to access the NSMO public-use file from FHFA, covering borrower shopping behavior, mortgage knowledge, and satisfaction linked to administrative credit and servicing data, plus the gotchas that bite pipelines working with survey-weighted microdata.
# Verified 2026-06-09 · tested with NSMO public-use-file page live (fhfa.gov/data/nsmo/puf); the file sits behind a terms-of-use click-through and was not pulled here
# Tags: mortgage, survey, free, cfpb, microdata, data:nsmo
==============================================================================

**NSMO** (the National Survey of Mortgage Originations) is a recurring survey
conducted jointly by the FHFA and the CFPB, and a component of the National
Mortgage Database (NMDB). It draws a sample of newly originated, closed-end
first-lien mortgages and asks borrowers about the shopping and search process,
expectations, mortgage knowledge, and satisfaction, then links the survey
responses to administrative credit-record and loan-servicing data in the NMDB.
FHFA releases a public-use file (PUF). It is used in, for example,
[Bhutta, Hizmo & Ringo](/wiki/papers/jf/2025/bhutta-much-racial-bias-affect-2025/)
for borrower satisfaction and service quality (N=35,162), and
[Bhutta, Fuster & Hizmo](/wiki/papers/jf/2026/bhutta-mortgage-overpayment-borrower-sophistication-2026/)
for borrower sophistication, shopping behavior, and knowledge, merged with
administrative credit and servicing data.

- **Cost:** free.
- **Access:** public-use file behind a terms-of-use click-through (no account or API key needed).
- **Coverage:** closed-end first-lien originations; survey waves from 2013 onward.
- **Home:** <https://www.fhfa.gov/data/nsmo> · **PUF:** <https://www.fhfa.gov/data/nsmo/puf>

## Access

The PUF is downloadable from the FHFA NSMO PUF page
(<https://www.fhfa.gov/data/nsmo/puf>) after accepting a terms-of-use
agreement. No account and no API key are required. A codebook is also
available on the same page.

The NSMO PUF page was confirmed live during this session, but the
end-to-end file download was not exercised here (the file sits behind the
terms-of-use click-through). The page's verified status is therefore
"Source reachable," not "Fetched."

For researchers with FHFA or CFPB affiliation, a restricted internal
NMDB/NSMO file exists with un-masked, uncoarsened data. The papers linked above
use that restricted version, not the PUF. See
<https://www.fhfa.gov/data/nmdb> for access information on the restricted
file.

## Gotchas (the ones that bite pipelines)

- **Survey weights are not optional.** NSMO is a stratified sample of
  originations, not a census. Unweighted statistics are biased. Apply the
  provided survey weights for any population-level inference about borrowers
  or originations.
- **PUF masking does not reproduce restricted-file results.** The public-use
  file is coarsened relative to the internal restricted file: values are
  top-coded or binned, geography is suppressed or rounded, and some
  variables are omitted entirely. Results from PUF analysis will not
  exactly match papers (such as those above) that were built on the
  restricted NMDB/NSMO data.
- **Self-report and non-response bias.** The sophistication, knowledge, and
  satisfaction questions are self-reported. Non-response is non-random (more
  engaged or satisfied borrowers may respond at different rates), and recall
  bias affects questions about the origination process answered after closing.
- **Originated loans only.** Coverage is restricted to closed originations:
  denied applicants and non-applicants are not in scope. The dataset cannot
  speak to credit access at the extensive margin.
- **Sample period depends on release vintage.** Waves accumulate over time,
  so the usable sample period depends on which PUF release vintage you
  download. Check the release notes and state the vintage explicitly in any
  replication package.

## Citation

Cite the FHFA/CFPB National Survey of Mortgage Originations, stating whether
the public-use file or the restricted NMDB file was used, the wave(s)
included, and that survey weights were applied. Example form: *Federal
Housing Finance Agency and Consumer Financial Protection Bureau, National
Survey of Mortgage Originations, Public-Use File, waves YYYY-YYYY, retrieved
from https://www.fhfa.gov/data/nsmo/puf, accessed YYYY-MM-DD; survey weights
applied.*


==============================================================================
# Open Source Asset Pricing (Chen-Zimmermann)
# https://instituteforautomatedresearch.org/wiki/datasets/open-source-asset-pricing/
# How to pull 212 firm-level anomaly signals and pre-built long-short portfolio returns for free: the list-not-string trap, the 1.6 GB bulk trap, and the CRSP-merge-already-done point, for automated pipelines.
# Verified 2026-05-16 · tested with openassetpricing.com + GitHub reachable (pkg/data per docs, not bulk-fetched)
# Tags: asset-pricing, anomalies, equities, panel-data, free, academic, data:open-source-asset-pricing
==============================================================================

**Open Source Asset Pricing** (Chen & Zimmermann 2022, *RFS*) is the free,
reproducible cross-sectional anomaly dataset: **212 firm-level signals** plus
pre-computed long-short and decile portfolio returns, with the CRSP/Compustat
merge already done. It is what the
[ZeroPaper](https://github.com/alejandroll10/zeropaper) pipeline uses to test
whether a model's mechanism maps to a known anomaly. This page is the
distilled recipe.

- **Cost:** free, no auth.
- **Coverage:** 212 signals; original-paper and decile/quintile portfolios.
- **Site:** <https://www.openassetpricing.com> ·
  **Code:** <https://github.com/OpenSourceAP/CrossSection>

## Access

### Option 1: `openassetpricing` package (preferred)

```python
# pip install openassetpricing
from openassetpricing import OpenAP
ap = OpenAP()

# Firm-level signals: predictor arg MUST be a list (see gotchas)
sig = ap.dl_signal("pandas", ["BM", "Mom12m", "AssetGrowth"])
#   → permno, yyyymm, BM, Mom12m, AssetGrowth

port = ap.dl_port("op", "pandas", ["BM"])      # original-paper portfolios
ap.list_port()                                  # op, deciles_ew/vw, quintiles_*
docs = ap.dl_signal_doc("pandas")               # all 212 with paper refs
```

### Option 2: Direct download (fallback)

A ~1.6 GB zipped wide CSV of all predictors is at
<https://www.openassetpricing.com/data/>. Download once, cache under `data/`,
never re-pull.

## Gotchas (the ones that bite pipelines)

The reason to read this page rather than the repo README. Site and GitHub
confirmed reachable on the date above; the package/data interface is as
documented (not bulk-fetched here; the full set is ~1.6 GB).

- **Predictor arguments must be LISTS, not strings.** `["BM"]`, never `"BM"`.
  Passing a string is the #1 failure and the error is not obvious.
- **The bulk download is ~1.6 GB.** `dl_all_signals` / the direct CSV will
  blow memory and time if pulled naively. Request only the signals you need;
  cache aggressively.
- **The CRSP/Compustat merge is already done.** Do not re-merge; signals are
  delivered at `permno × yyyymm`. Re-merging double-counts and misaligns.
- **`yyyymm` is an integer**, not a date; convert before joining to returns.
- **Releases are versioned and periodic.** State the release you used
  (latest tags have been updated as recently as late 2025); results drift
  across releases.
- **Pre-built portfolios beat hand-rolled.** Use `dl_port` deciles for
  long-short spreads rather than re-sorting; it matches the paper's
  methodology and avoids look-ahead in the sort.

## Key signals (10 of 212)

| Signal | Description | Category |
|---|---|---|
| `BM` | Book-to-market | Value |
| `Mom12m` | 12-month momentum (skip last month) | Momentum |
| `AssetGrowth` | Asset growth | Investment |
| `GP` | Gross profitability | Profitability |
| `EP` | Earnings-to-price | Value |
| `Beta` | CAPM beta | Risk |
| `IdioVol` | Idiosyncratic volatility | Risk |
| `Accruals` | Operating accruals | Quality |
| `SUE` | Standardized unexpected earnings | Earnings |
| `ShareIss1Y` | Net share issuance, 1yr | Issuance |

`ap.dl_signal_doc('pandas')` for the full 212 with references.

## Standard operations

- **Long-short spread:** decile 10 − decile 1 from `dl_port("deciles_vw", …)`.
- **Alpha:** regress the long-short series on FF5 (see
  [Ken French](/wiki/datasets/ken-french/)).
- **Signal-zoo test:** does your model's mechanism map onto an existing
  anomaly, or is it genuinely new?
- **Always state** the signals, release/version, sample period, and weighting.

## Citation

*Chen, A. Y., and T. Zimmermann (2022). "Open Source Cross-Sectional Asset
Pricing." Review of Financial Studies. Data from
https://www.openassetpricing.com, release [tag], accessed YYYY-MM-DD.*


==============================================================================
# PACER: federal court records (incl. bankruptcy)
# https://instituteforautomatedresearch.org/wiki/datasets/pacer-bankruptcy/
# How PACER (Public Access to Court Electronic Records) works, what it costs ($0.10/page, $3.00/document cap, $30/quarter waiver), and why automated bulk retrieval is metered rather than free. Covers bankruptcy filings, case-level dockets, and how RECAP partially mirrors paid content.
# Tags: bankruptcy, courts, legal, filings, event-data, no-api-key, paid, data:pacer-bankruptcy
==============================================================================

:::note[Open to anyone, but metered by per-page fees]
PACER is open to any registered user but charges $0.10 per page, so it is not a free bulk dataset and was not pulled here. The metered structure makes large automated pulls expensive and unpredictable. This page carries no provenance badge: it documents the dataset and access path, but the end-to-end pull was not exercised here.
:::

**PACER** (Public Access to Court Electronic Records), at https://pacer.uscourts.gov, is the federal courts' public electronic access system for case dockets and filed documents across all 94 federal district and bankruptcy courts. For bankruptcy research, it covers case-level filings (chapter 7, chapter 11, chapter 13, and others), filing dates, debtor addresses, and docket entries across all federal bankruptcy courts. Used in, for example, [Slutzky and Xu (2025)](/wiki/papers/rfs/2025/slutzky-financial-consequences-pretrial-detention-2025/), where PACER bankruptcy filings from 2000 to 2011 (318,000 filings in Maryland) form the outcome variable.

- **Cost:** $0.10 per page; $3.00 cap per document; fees waived if total court-record spending in a quarter is $30 or less (per the PACER pricing page, https://pacer.uscourts.gov/pacer-pricing-how-fees-work). The waiver means most low-volume users pay nothing in a given quarter.
- **API key:** no API key required; account registration with billing is required.
- **Bulk access:** no free bulk feed; every page is metered at $0.10; there is no single bulk download endpoint.
- **Coverage:** all 94 federal courts (district, bankruptcy, appellate, Court of International Trade); case dockets and documents.
- **Home:** https://pacer.uscourts.gov

## Access

Access requires a PACER account with billing information on file. Registration is free. Retrieval is per-case, per-docket, and per-document: navigate to the court's PACER site, search by case number or party name, and download docket sheets and documents. Each page is billed at $0.10; the docket sheet and each attached document are billed separately.

Bulk or automated retrieval is technically permitted under PACER terms but every page is metered: a scraping run across thousands of cases can accumulate substantial charges. Legislation to eliminate PACER fees has been introduced in Congress several times but none has become law as of 2026, so fees still apply.

The RECAP project (Free Law Project, https://free.law/recap/) mirrors documents that other users have already downloaded and paid for, making them freely available. Coverage is incomplete and is not authoritative, but it is a partial free complement.

## Gotchas (the ones that bite pipelines)

- **Per-page metering makes large pulls expensive and unpredictable.** Page counts per document vary; a 50-page brief costs $3.00 (capped), but a long docket sheet with many entries may run higher before the per-document cap applies.
- **The $30/quarter waiver resets quarterly.** A pipeline that batches downloads at the start of each quarter gets one free tranche; subsequent pulls in the same quarter are billed. Track cumulative quarterly spend.
- **Docket sheets and attached documents are billed separately.** Fetching a docket sheet to find document numbers does not include the documents; each document download incurs its own per-page charge (capped at $3.00 per document).
- **No single bulk download.** You crawl dockets one at a time; there is no bulk export endpoint.
- **RECAP is partial.** The RECAP mirror covers only documents others have already downloaded; new cases, sealed documents, and unpopular dockets are absent. Do not rely on RECAP for a complete case-level panel.
- **Do not assume PACER will become free.** Bills to eliminate PACER fees have been introduced repeatedly but none has been enacted; plan budgets around the current fee schedule.
- **Court-specific access points.** Each federal court has its own PACER site (e.g., `ecf.mdb.uscourts.gov` for the District of Maryland); the central search at `pcl.uscourts.gov` covers all courts but links out to per-court portals for document downloads.

## Reference

| Field | Value |
|-------|-------|
| Host | https://pacer.uscourts.gov (central); per-court ECF sites for documents |
| Fee | $0.10/page; $3.00/document cap; waived if total court-record spend is $30 or less in the quarter |
| Access | Registered account with billing (free to register) |
| Bulk | Metered, no free bulk feed |
| API key | No (account required, no API key) |
| Free alternative | RECAP (partial mirror; not authoritative) |
| Coverage | All 94 federal courts: bankruptcy, district, appellate |

## Citation

Cite PACER by identifying the specific court, case number(s), document(s), and retrieval date. In a methods appendix, note the fee paid or the waiver threshold applied, and whether documents were retrieved directly from PACER or via the RECAP mirror. A representative format: Public Access to Court Electronic Records (PACER), [Court Name], Case No. [X], retrieved [Date]. Fees paid per PACER pricing schedule.


==============================================================================
# Paris historical repeat-rent index
# https://instituteforautomatedresearch.org/wiki/datasets/paris-rents/
# How to reach the long-run Paris rent series from Eichholtz, Korevaar and Lindenthal, plus the honest gotcha that only the 1809-1943 slice is publicly posted (in the shared RFS 2021 workbook), while the deep 1500-1831 repeat-rent index used in some studies remains working-paper-only.
# Verified 2026-06-22 · tested with confirmed the Eichholtz-Korevaar-Lindenthal Paris rent data is reachable on Korevaar's Google Drive (the published 1809-1943 slice downloads as the RFS 2021 headline .xlsx, 29,096 bytes, as corroboration); the keystone Paris 1500-1831 repeat-rent index used by the citing paper is working-paper-only and was not exercised here, so the grade is reachable rather than fetched
# Tags: housing, rents, historical, real-estate, time-series, free, no-api-key, academic, data:paris-rents
==============================================================================

**The Paris historical rent data** is a set of long-run rent series for Paris
compiled from French notarial and archival sources by Piet Eichholtz, Matthijs
Korevaar and Thies Lindenthal. Two distinct products exist, and they differ
sharply in how available they are. The published total-return work includes a
Paris rent and price series for 1809 to 1943, posted openly. A separate,
much deeper repeat-rent index covering 1500 to 1831 is used in some studies but
remains a working paper with no public replication file. Used in, for example,
[Francke (2025)](/wiki/papers/jf/2025/francke-baby-booms-asset-booms-2025/),
which draws on the Paris 1500-1831 repeat-rent series (new rental contracts
only).

- **Cost:** free, no account, no key (the published slice is an Excel file).
- **Publicly available:** the Paris 1809-1943 series, inside the shared RFS 2021
  Amsterdam-and-Paris workbook.
- **Not publicly available:** the Paris 1500-1831 repeat-rent index, which is
  working-paper-only as of this writing.
- **Data page:** Matthijs Korevaar,
  <https://sites.google.com/view/matthijskorevaar/data>

## Access

The only Paris rent data that downloads here is the 1809-1943 slice, which ships
in the same Google Drive workbook as the Amsterdam total-return series (see the
[Amsterdam housing](/wiki/datasets/amsterdam-housing-transactions/) page).

```bash
# RFS 2021 headline series .xlsx (Amsterdam 1900-1979 + Paris 1809-1943);
# 29,096-byte .xlsx, sheets "Sources" and "Headline Series":
curl -sL "https://drive.usercontent.google.com/download?id=1u-cYs7TiB6-rEhcDqCiGkfXsHdP9mWx1&export=download" -o rfs2021_headline.xlsx
```

The deep 1500-1831 Paris repeat-rent index is described in the working paper
"The Housing Affordability Revolution" (predecessor "500 Years of Housing Rents,
Quality and Affordability," SSRN 3418495); no replication package is posted, so
it cannot be fetched here.

## Gotchas (the ones that bite pipelines)

- **The long 1500-1831 index is not public.** The series several studies cite
  (including the one above) is from a working paper with no replication file
  posted. Do not promise a downloadable 1500-1831 Paris panel: it does not exist
  publicly yet.
- **Only the 1809-1943 slice is fetchable.** That is what the green-versus-teal
  grade reflects: the endpoint and the published slice are reachable, but the
  specific long series the citing work used was not exercised here.
- **Shared workbook with Amsterdam.** The published Paris and Amsterdam series
  ship in one Excel file, so cite the same Drive file and mind the link-rot /
  no-DOI caveats noted on the Amsterdam page.
- **New-contract rents, not sitting-tenant rents.** The Paris series is built
  from new rental contracts; these move differently from the rent paid by sitting
  tenants, so do not read it as a tenant cost-of-living series without adjustment.
- **Compiled from archival sources.** Like Amsterdam, the raw French notarial and
  archival records are a separate, harder source; the machine-readable value is
  the compiled series, where one exists.

## Reference

| Field | Value |
|-------|-------|
| Data page | `sites.google.com/view/matthijskorevaar/data` |
| Public series | Paris 1809-1943, in the RFS 2021 headline .xlsx (Drive id `1u-cYs...mWx1`) |
| Not public | Paris 1500-1831 repeat-rent index (working paper, SSRN 3418495) |
| Raw source | French notarial and archival records, separate |
| Format | Multi-sheet Excel (Sources + Headline Series) |
| Key required | No |

## Citation

For the published series, cite Eichholtz, Korevaar, Lindenthal & Tallec (2021),
"The Total Return and Risk to Residential Real Estate," *Review of Financial
Studies* 34(8), 3608-3646, with the Drive file id and retrieval date. For the
1500-1831 series, cite the working paper (Eichholtz, Korevaar & Lindenthal,
"The Housing Affordability Revolution," SSRN 3418495) and note that it is not
yet publicly released.


==============================================================================
# Quarterly Workforce Indicators (QWI)
# https://instituteforautomatedresearch.org/wiki/datasets/qwi-census/
# How to pull Census LEHD Quarterly Workforce Indicators local labor-market statistics free with no key via the LEHD bulk flat files, including the filename scheme, status-flag columns, and the suppression gotchas that bite pipelines.
# Verified 2026-06-09 · tested with live no-key download of a QWI LEHD bulk file (lehd.ces.census.gov/data/qwi/latest_release/us/qwi_us_sa_fa_gn_ns_op_u.csv.gz; 19 MB, 266k rows)
# Tags: labor, employment, regional, firms, free, no-api-key, data:qwi-census
==============================================================================

**Quarterly Workforce Indicators (QWI)** is produced by the U.S. Census Bureau
Longitudinal Employer-Household Dynamics (LEHD) program. It delivers local
labor-market statistics built from linked employer-employee administrative
records: employment levels, hires, separations, job creation and destruction,
turnover, and earnings, broken down by geography (national, state, metro,
county), industry, and worker demographics (sex, age group, race, ethnicity,
education), and by firm age and firm size. The data is free and public. The
LEHD bulk flat-file path used here needs no API key. Used in, for example,
[Johnston-Ross, Ma & Puri](/wiki/papers/jf/2025/johnston-ross-private-equity-financial-stability-2025/),
where county-level QWI startup employment (firm age 0 to 1) and total
employment feed the regional-recovery analysis.

This page documents the LEHD flat-file bulk download path. A separate
[Census Bureau page](/wiki/datasets/census/) covers the api.census.gov API
route (key required) and the Business Dynamics Statistics (BDS) product; this
page does not duplicate BDS material.

- **Cost:** free, public.
- **API key:** none required for the LEHD bulk files. The api.census.gov QWI
  API route requires a free key (returns "Missing Key" without one).
- **Coverage:** quarterly, by state (with national aggregates), from the early
  1990s onward, though state start dates differ.
- **Home:** <https://lehd.ces.census.gov/data/>

## Access

The LEHD flat files live under
<https://lehd.ces.census.gov/data/qwi/latest_release/>. Each subdirectory
corresponds to a geography: `us/` for national aggregates, then one folder per
state FIPS abbreviation (`ak/`, `al/`, ..., `wy/`). Within each folder,
multiple files cover different demographic and firm-characteristic breakdowns.
No authentication is required.

Download the U.S. national, seasonally adjusted, firm-age breakdown file:

```bash
# US national QWI file: seasonally adjusted, by firm age, no key
curl -sL -o qwi_us.csv.gz \
  "https://lehd.ces.census.gov/data/qwi/latest_release/us/qwi_us_sa_fa_gn_ns_op_u.csv.gz"
gzip -dc qwi_us.csv.gz | head
```

The file is approximately 19 MB compressed and contains 266,256 rows.

### Filename scheme

Files follow the pattern:

```
qwi_<geo>_<seasonadj>_<firmchar>_<geo_level>_<raceeth>_<owner>_<ratevar>.csv.gz
```

Confirmed segment codes from the verified file above and LEHD documentation:

| Position | Code | Meaning |
|----------|------|---------|
| `<geo>` | `us`, `ca`, `tx`, ... | Geography (national or 2-letter state abbreviation) |
| `<seasonadj>` | `sa` | Seasonally adjusted |
| `<seasonadj>` | `se` | Seasonally extracted (not seasonally adjusted) |
| `<firmchar>` | `fa` | By firm age |
| `<firmchar>` | `fs` | By firm size |
| `<firmchar>` | `f` | No firm-characteristic detail |
| `<geo_level>` | `gn` | National/state aggregate level |
| `<raceeth>` | `ns` | No race/ethnicity detail |
| `<raceeth>` | `rs` | Race/ethnicity detail included |
| `<owner>` | `op` | Private ownership |

List the directory index to see every available file and confirm the exact
code meanings for your vintage against the LEHD schema documentation at
<https://lehd.ces.census.gov/data/schema/latest/lehd_public_use_schema.pdf>
rather than guessing. Segment codes can change across schema versions.

### Load in Python

```python
import pandas as pd

df = pd.read_csv(
    "qwi_us.csv.gz",
    dtype=str,        # read everything as str first; cast after inspecting
    compression="gzip",
    low_memory=False,
)

# Key identifier columns:
# periodicity, seasonadj, geo_level, geography, ind_level, industry,
# ownercode, sex, agegrp, race, ethnicity, education, firmage, firmsize,
# year, quarter, agg_level

# Then the indicator columns:
# Emp, EmpEnd, HirA, Sep, FrmJbGn, FrmJbLs, EarnBeg, Payroll, ...
# Each has an s-prefixed status/flag twin, e.g. sEmp, sHirA, sSep

# Convert numeric indicators after filtering for acceptable status flags
numeric_indicators = ["Emp", "EmpEnd", "HirA", "Sep",
                       "FrmJbGn", "FrmJbLs", "EarnBeg", "Payroll"]
df[numeric_indicators] = df[numeric_indicators].apply(
    pd.to_numeric, errors="coerce"
)
```

The s-prefixed twin of each indicator (e.g. `sEmp` alongside `Emp`) encodes
a data-quality and suppression status code for that cell. Cells are blank or
hold a sentinel value (e.g. `-1`) where suppressed. These are not zeros and
must not be treated as zeros. Read the status flags and drop or flag
suppressed cells before any analysis.

## Gotchas (the ones that bite pipelines)

- **The api.census.gov QWI API returns "Missing Key" without a free API key.**
  The LEHD flat files used here need no key. If you want zero credentials, use
  the LEHD bulk route; if you use the Census API, register for a free key at
  <https://www.census.gov/data/developers.html>.
- **Suppressed cells are not zeros.** Cells are suppressed or fuzzed for
  confidentiality. Blank cells and sentinel values (e.g. `-1`) in indicator
  columns indicate suppression, not a true zero. Each indicator has an
  s-prefixed status flag column you must read before treating any value as
  valid data.
- **State coverage start dates differ.** States joined the LEHD program in
  different years, so a national panel has ragged left edges by state. A
  series starting in 1990 for some states may start in 1995 or later for
  others. Massachusetts has historically been absent from LEHD.
- **Firm-age and firm-size breakdowns are only in files that carry that
  dimension.** The filename encodes whether firm age or firm size is included.
  Requesting firm-age detail from a file without the `fa` segment returns
  nothing; pick the correct file for your breakdown.
- **QWI counts jobs and worker flows, not firm financials.** Employment,
  hires, separations, and earnings are from administrative records. QWI does
  not contain revenue, assets, or profitability. It is also not the same as
  BDS firm counts (see the [Census Bureau page](/wiki/datasets/census/)).
- **Seasonal adjustment is a file-level choice.** Seasonally adjusted (`sa`)
  and seasonally extracted (`se`) series are in separate files. Do not mix
  them in one panel series.
- **Industry and geography codes follow LEHD conventions.** The `industry`
  column uses NAICS codes at varying levels of aggregation (controlled by
  `ind_level`). The `geography` column uses FIPS codes. Confirm the
  aggregation level in `agg_level` before joining to external sources.

## Key indicators

| Column | Description |
|--------|-------------|
| `Emp` | Beginning-of-quarter employment |
| `EmpEnd` | End-of-quarter employment |
| `HirA` | Total hires (accessions) |
| `Sep` | Total separations |
| `FrmJbGn` | Firm job gains |
| `FrmJbLs` | Firm job losses |
| `EarnBeg` | Average monthly earnings for workers employed at beginning of quarter |
| `Payroll` | Total payroll |

## Geography levels

| Level | Description |
|-------|-------------|
| National | `us/` subdirectory |
| State | One subdirectory per state (e.g. `ca/`, `tx/`) |
| Metro (MSA/CBSA) | Included within state files, `geo_level = M` |
| County | Included within state files, `geo_level = C` |
| WIA/workforce area | Included within state files, `geo_level = W` |

## Citation

*U.S. Census Bureau, Longitudinal Employer-Household Dynamics Program,
Quarterly Workforce Indicators, latest release, retrieved from
https://lehd.ces.census.gov/data/qwi/latest_release/, accessed 2026-06-09.*

Record the schema version, the specific file path, and the seasonal-adjustment
and firm-characteristic segments so the pull is reproducible.


==============================================================================
# SBA 7(a) and 504 loan data (FOIA)
# https://instituteforautomatedresearch.org/wiki/datasets/sba-loans/
# How to download and work with the U.S. Small Business Administration's loan-level FOIA datasets for the 7(a) and 504/CDC programs, including the CKAN portal, direct CSV access, and the gotchas that bite pipelines.
# Verified 2026-06-09 · tested with live no-key CSV pull of the SBA 7(a) FOIA file (data.sba.gov, FY2010-FY2019)
# Tags: small-business, lending, free, no-api-key, loan-data, data:sba-loans
==============================================================================

**SBA 7(a) and 504 loan data (FOIA)** is a set of loan-level files released by
the U.S. Small Business Administration under FOIA. Each record is one approved
loan and includes borrower name and address, lender name and FDIC/NCUA number,
gross approval amount, SBA-guaranteed amount, approval date and fiscal year,
initial interest rate, fixed-or-variable indicator, term in months, NAICS code
and description, project county and state, SBA district office, congressional
district, and jobs supported. The files cover the two main SBA credit programs:
the 7(a) loan-guarantee program and the 504/CDC program. Used in, for example,
[Johnston-Ross, Ma & Puri](/wiki/papers/jf/2025/johnston-ross-private-equity-financial-stability-2025/)
for the number, amount, interest rate, and average size of small-business loans
by county.

- **Cost:** free, public domain.
- **API key:** none required.
- **Coverage:** 7(a) approvals from FY1991 and 504 approvals from FY1991, split
  into FY-range files, refreshed periodically. Each filename encodes an "asof"
  snapshot date.
- **Home:** <https://data.sba.gov/dataset/7-a-504-foia>

## Access

The portal is a CKAN instance. You can query it programmatically to list all
resources, then download any CSV directly without authentication.

### Step 1: list available resources

```bash
curl -s "https://data.sba.gov/api/3/action/package_show?id=7-a-504-foia" \
  | python3 -c "
import json, sys
pkg = json.load(sys.stdin)
for r in pkg['result']['resources']:
    print(r['name'], r['url'])
"
```

This prints each resource name (e.g. `7(a) FY2010-FY2019 (asof 230930)`) and
its direct download URL.

### Step 2: download a CSV file

Pick the URL for the file you need and fetch it directly:

```bash
# Example: 7(a) FY2010-FY2019 snapshot
curl -L -o sba_7a_2010_2019.csv \
  "https://data.sba.gov/dataset/7-a-504-foia/resource/<resource-id>/download/<filename>.csv"
```

Replace `<resource-id>` and `<filename>` with values from the package listing
above; the exact URL is stable for a given snapshot.

### Step 3: load in Python

```python
import pandas as pd

df = pd.read_csv(
    "sba_7a_2010_2019.csv",
    dtype=str,           # read everything as str first; cast after schema alignment
    low_memory=False,
)

# Cast key numeric fields after inspection
df["GrossApproval"] = pd.to_numeric(df["GrossApproval"], errors="coerce")
df["ApprovalDate"]  = pd.to_datetime(df["ApprovalDate"], errors="coerce")
```

A separate data-dictionary XLSX is listed alongside the CSV resources in the
package; download it to map coded fields and confirm column names for the
vintage you are using.

## Gotchas (the ones that bite pipelines)

- **Approvals, not disbursements or outcomes.** Every record means a loan was
  approved. The file carries no information on whether the loan was fully drawn,
  repaid, or defaulted. Default and charge-off information lives in a separate
  SBA purchase/charge-off dataset; join on the loan number if you need it.
- **Multiple FY-range CSVs that must be concatenated; schemas drift.** The data
  is split into files by program and FY range. Column names and coded values
  change across vintages: a field named `NaicsCode` in one file may appear as
  `NAICS` in another. Inspect and align schemas before stacking; do not assume
  a union of columns is safe.
- **Pin the "asof" snapshot date.** Each filename encodes a snapshot date (e.g.
  `asof230930` for the September 30, 2023 cut). Records can be retroactively
  revised or added in later snapshots. To reproduce a prior result, use the
  same snapshot, not the latest file.
- **Geography is recorded at approval.** The project county and state reflect
  where the project was located at the time of approval. There is no update if
  the borrower relocates.
- **NAICS definitions change across years.** NAICS codes are revised on a
  five-year cycle. A code present in FY1997 data may map to a different
  industry description in FY2022 data. Use the NAICS revision year to
  cross-walk codes before aggregating across long panels.
- **7(a) and 504 have different structures; do not pool blindly.** The 7(a)
  program is a direct guarantee to a participating lender. The 504 program
  involves a Certified Development Company (CDC) and a separate third-party
  lender, so the "lender" field means different things. The guaranteed
  percentage and term conventions also differ. Keep the programs separate unless
  you have a specific reason to combine them, and document that choice.

## Citation

Cite the program, the FY-range file(s), and the specific "asof" snapshot date
used, for example: *U.S. Small Business Administration, 7(a) Loan Data
(FOIA), FY2010-FY2019 (as of 2023-09-30), retrieved from
https://data.sba.gov/dataset/7-a-504-foia, accessed YYYY-MM-DD.* The snapshot
date is in the filename; record it so the pull is reproducible.


==============================================================================
# SCF: Survey of Consumer Finances
# https://instituteforautomatedresearch.org/wiki/datasets/scf/
# How to pull the Federal Reserve's Survey of Consumer Finances summary extract (US household wealth, debt, income, and portfolios) for free with no key, why the file has five rows per household, and the weighting and imputation gotchas that bite pipelines.
# Verified 2026-06-09 · tested with live no-key pull of scfp2022excel.zip -> SCFP2022.csv (357 variables, 22975 rows = 5 implicates x ~4595 families; sum of WGT over all rows = 1.313e8, the US household count)
# Tags: macro, cross-section, free, no-api-key, federal-reserve, data:scf
==============================================================================

The **Survey of Consumer Finances (SCF)** is the Federal Reserve Board's
triennial survey of US household balance sheets: assets, debts, net worth,
income, and portfolio composition, with the detail needed to study wealth
distribution and household leverage. The **summary extract public dataset** is
the version most research uses. This page is the distilled access recipe.

- **Cost:** free, no paywall, no key.
- **Coverage:** triennial (1989, 1992, ... 2019, 2022), cross-sectional; the
  2022 summary extract has **357 variables** for about **4,595 families**.
- **Format:** the summary extract as CSV (in `scfp2022excel.zip`), Stata
  (`scfp2022s.zip`), or SAS; plus the larger full public dataset.
- **Home:** <https://www.federalreserve.gov/econres/scfindex.htm>

## Access

### No key, summary extract (recommended starting point)

```
https://www.federalreserve.gov/econres/files/scfp2022excel.zip   # CSV -> SCFP2022.csv
https://www.federalreserve.gov/econres/files/scfp2022s.zip       # Stata
```

```python
import io, zipfile, requests, pandas as pd

url = "https://www.federalreserve.gov/econres/files/scfp2022excel.zip"
z = zipfile.ZipFile(io.BytesIO(requests.get(url).content))
scf = pd.read_csv(z.open("SCFP2022.csv"))

# WGT already nets the five implicates: summing WGT over all rows = US households.
wmean_nw = (scf.NETWORTH * scf.WGT).sum() / scf.WGT.sum()   # weighted mean net worth
```

The summary extract carries the cleaned, ready-to-use aggregates (`NETWORTH`,
`INCOME`, `ASSET`, `DEBT`, and demographic recodes like `AGE`, `EDCL`, `RACE`).
Use the **full public dataset** only when you need a variable the extract does
not expose.

## Gotchas (the ones that bite pipelines)

The reason to read this page rather than the codebook. Verified against the live
2022 extract on the date above.

- **Five rows per household (multiple imputation).** The file has about five
  times as many rows as families (22,975 = 5 x ~4,595): every household appears
  as **five implicates** that fill in missing answers. `Y1` is the implicate
  id, `YY1` the household id. You must use **all five** and not deduplicate, or
  you discard the imputation and bias standard errors.
- **Weights are mandatory, and already net the five implicates.** Nothing is
  representative unweighted. In this summary extract `WGT` is pre-scaled so that
  summing it over all five implicate rows equals the US household count (verified:
  the sum of `WGT` over the full 2022 file is about 131 million). Weight by `WGT`
  directly across all rows; do **not** divide by five again, or totals come out
  5x too small.
- **Proper standard errors need replicate weights.** Point estimates use `WGT`,
  but correct sampling variance needs the separate **replicate-weight** file
  plus the imputation variance across the five implicates; the naive SE on the
  stacked file is wrong.
- **Triennial, and dollars are survey-year.** Surveys are three years apart;
  values are in the survey year's dollars. Deflate to compare across waves, and
  do not treat it as a panel: each wave is a fresh cross-section, not the same
  households.
- **Top wealth is special.** The SCF oversamples wealthy households (via a
  list sample) precisely so top-tail wealth is captured; this is a feature, but
  it means the weights, not raw counts, carry the distribution. Public values
  are also rounded / lightly disclosure-protected.
- **Extract vs full dataset codes differ.** Summary-extract variable names
  (`NETWORTH`, `ASSET`) are constructed aggregates; the full public dataset uses
  raw `Xnnnn` question codes. Do not assume a name carries across the two files.

## Variables you actually need (summary extract)

| Variable | Meaning |
|---|---|
| `YY1`, `Y1` | Household id, implicate id (1-5) |
| `WGT` | Sample weight |
| `NETWORTH` | Household net worth |
| `ASSET`, `DEBT` | Total assets, total debt |
| `INCOME` | Total household income |
| `AGE`, `EDCL`, `RACE` | Age, education class, race/ethnicity of respondent |

## Standard operations

- **Wealth distribution:** compute weighted percentiles of `NETWORTH` with
  `WGT`; report across all five implicates.
- **Means and totals:** weight by `WGT` across all rows; it is already scaled so
  the five implicate copies sum to the population, so neither means nor totals
  need a further division by five.
- **Inference:** use the replicate weights plus the imputation (Rubin) variance
  for standard errors; the macros / packages the Fed documents handle both.
- **Cross-wave:** deflate to a common year and treat each wave as a separate
  cross-section.

## Citation

Board of Governors of the Federal Reserve System. *Survey of Consumer Finances*
(2022). Washington: Federal Reserve Board. Cite the survey wave and your access
date; the Fed's SCF page lists the suggested wording and the codebook.


==============================================================================
# Robert Shiller online data
# https://instituteforautomatedresearch.org/wiki/datasets/shiller-data/
# How to pull Robert Shiller's long-run U.S. stock market (ie_data.xls, CAPE) and home price (Fig3-1.xls) files with no key, plus the gotchas that bite pipelines (the YYYY.MM decimal date, monthly-average prices, the provisional tail, and the changing download link).
# Verified 2026-06-09 · tested with live no-key fetch + parse of ie_data.xls (Data sheet, monthly 1871.01 to 2024.09) and Fig3-1.xls home-price file (annual from 1890); shillerdata.com link and econ.yale.edu mirror both confirmed live
# Tags: market, valuation, housing, macro, time-series, equities, free, no-api-key, data:shiller-data
==============================================================================

**Robert Shiller online data** is the long-run U.S. financial dataset Shiller
maintains alongside *Irrational Exuberance*. Two free files cover most uses: the
stock-market file `ie_data.xls` (monthly S&P Composite price, dividends,
earnings, CPI, the 10-year rate, and the cyclically adjusted price-earnings
ratio, CAPE, back to 1871) and the home-price file `Fig3-1.xls` (a long-run
U.S. real home price index back to 1890). No API key is required. Used in, for
example,
[Li](/wiki/papers/rfs/2025/li-dominate-historical-average-2025/),
which draws CAPE and several Shiller predictors for an out-of-sample equity
premium test, and
[Piazzesi](/wiki/papers/jf/2025/piazzesi-presidential-address-housing-betas-2025/),
which uses the `ie_data.xls` real stock returns and 10-year Treasury series as
cashflow and discount-rate inputs.

- **Cost:** free, public.
- **API key:** none required.
- **Coverage:** `ie_data.xls` monthly from 1871; `Fig3-1.xls` home prices
  annual from 1890 (monthly only in recent decades).
- **Home:** <https://shillerdata.com/> (current host); mirror at
  <http://www.econ.yale.edu/~shiller/data.htm>.

## Access

The download links on shillerdata.com point at a content-delivery blob whose
URL carries a `?ver=` query that changes whenever Shiller re-uploads. The robust
recipe is to scrape the current link off the page; a stable fallback is the
Yale econ mirror, which serves the same file names directly:

```bash
# Stable Yale mirror (same files, direct names)
curl -sL "http://www.econ.yale.edu/~shiller/data/ie_data.xls" -o ie_data.xls
curl -sL "http://www.econ.yale.edu/~shiller/data/Fig3-1.xls"  -o Fig3-1.xls

# Or resolve the current shillerdata.com download link, then fetch it:
curl -sL "https://shillerdata.com/" \
  | grep -oiE 'href="[^"]*ie_data[^"]*"' | head -1
```

### Load in Python

```python
import pandas as pd

# ie_data.xls: the table starts a few rows down; the workbook also has a
# "Disclaimer" sheet, so read the "Data" sheet explicitly with a header offset.
ie = pd.read_excel("ie_data.xls", sheet_name="Data", header=7)

# The Date column is decimal YYYY.MM (e.g. 1871.01 = Jan 1871, 2024.10 = Oct
# 2024). Split it; do NOT treat the fractional part as a fraction of a year.
def ym(x):
    y = int(x)
    m = round((x - y) * 100)
    return pd.Period(freq="M", year=y, month=m)

ie = ie.dropna(subset=["Date"])
ie["period"] = ie["Date"].map(ym)
```

## Gotchas (the ones that bite pipelines)

- **The date is `YYYY.MM`, not a year fraction.** `1871.01` is January 1871 and
  `2024.10` is October 2024. The digits after the decimal are the month number,
  so `1871.1` means October, not "one-tenth into the year." Parse by splitting
  year and month, never by float arithmetic on the year.
- **Prices are monthly averages of daily closes.** The S&P Composite price in
  `ie_data.xls` is the average of that month's daily closing prices, not an
  end-of-month value. Do not merge it against month-end series as if it were a
  point-in-time observation.
- **The tail is provisional and lags the access date.** The most recent months'
  dividends and earnings are estimates, and the file is updated irregularly: a
  pull in mid-2026 ended in September 2024. Do not assume the series runs to the
  month you downloaded it, and flag the estimated tail.
- **Read the `Data` sheet with the right header row.** The workbook opens on a
  notes band and includes a separate `Disclaimer` sheet. The real column header
  for `ie_data.xls` is several rows down (around row 8). Reading the default
  sheet or row 0 returns header noise.
- **Real series and CAPE are precomputed columns.** Real price, real dividend,
  real earnings, CAPE, and the total-return CAPE are computed inside the
  spreadsheet from a specific CPI vintage. Reading the stored values is fine;
  recomputing them with a different CPI series will not match.
- **Home prices are a separate file and mostly annual.** The long-run home price
  index is in `Fig3-1.xls`, not `ie_data.xls`, and is annual back to 1890 with
  monthly detail only in recent decades. Do not expect monthly home prices for
  the nineteenth century.
- **The download URL is not stable.** The shillerdata.com link embeds a
  `?ver=` token that changes on re-upload, so a hardcoded blob URL can go stale.
  Scrape the page for the current link or use the Yale mirror, and record the
  access date.
- **Legacy `.xls` format.** Both files are old-style BIFF `.xls`, not `.xlsx`;
  read them with an engine that handles `.xls`.

## Reference

| File | Contents | Frequency | Start |
|------|----------|-----------|-------|
| `ie_data.xls` (Data sheet) | S&P price, dividends, earnings, CPI, GS10, real series, CAPE | Monthly | 1871-01 |
| `Fig3-1.xls` (Data sheet) | Nominal and real U.S. home price index, building cost, population, long rate | Annual (recent monthly) | 1890 |

| Field | Value |
|-------|-------|
| Current host | `https://shillerdata.com/` |
| Mirror | `http://www.econ.yale.edu/~shiller/data/` |
| Date encoding | Decimal `YYYY.MM` |
| Key required | No |

## Citation

Cite Robert J. Shiller, the file, the host, and the access date, for example:
*Robert J. Shiller, "U.S. Stock Markets 1871-Present and CAPE Ratio"
(`ie_data.xls`), retrieved from https://shillerdata.com/, accessed YYYY-MM-DD.*
For the home price series cite `Fig3-1.xls` from the same source. Shiller's data
underlies *Irrational Exuberance* (Princeton University Press); cite the book if
your text refers to it. Record the access date and archive the file, since both
the data tail and the download URL change over time.


==============================================================================
# SIPP (Survey of Income and Program Participation)
# https://instituteforautomatedresearch.org/wiki/datasets/sipp/
# How to pull SIPP public-use household income and employment microdata from the U.S. Census Bureau with no API key, including the schema JSON for variable definitions, the character-delimited CSV format, and the gotchas that bite longitudinal pipelines.
# Verified 2026-06-09 · tested with live no-key download of the SIPP 2022 public-use schema (www2.census.gov/programs-surveys/sipp/data/datasets/2022/pu2022_schema.json) and confirmed the pu2022_csv.zip data file serves over HTTP
# Tags: household, labor, income, survey, panel, free, no-api-key, data:sipp
==============================================================================

**Survey of Income and Program Participation (SIPP)** is a longitudinal
household survey conducted by the U.S. Census Bureau. The same respondents are
re-interviewed annually, and the interview captures monthly-level information on
labor-force activity and employment transitions, income by source, household and
family composition, and participation in government programs. Public Use (PU)
microdata files are released free with no API key. Used in, for example,
[Choukhmane & de Silva](/wiki/papers/jf/2026/choukhmane-portfolio-choices-risk-preferences-2026/),
where SIPP supplies the labor income process estimation and employment
transition probabilities.

- **Cost:** free, public.
- **API key:** none required.
- **Coverage:** monthly recall for all panel members, re-interviewed annually.
  The redesigned SIPP begins with the 2014 panel, with annual single-year
  panels from 2018 onward; older panels run from 1984 through 2008.
- **Home:** <https://www.census.gov/sipp>

## Access

Panel-year files are organized in per-year subdirectories under
<https://www2.census.gov/programs-surveys/sipp/data/datasets/>. Each year
folder contains a public-use CSV zip, a gzipped CSV, SAS input files, and a
JSON schema with variable definitions and labels. Download the 2022 panel
directly, no authentication:

```bash
# Variable schema (complete pull, 768 KB) -- first variable is SSUID
curl -sL -o pu2022_schema.json \
  "https://www2.census.gov/programs-surveys/sipp/data/datasets/2022/pu2022_schema.json"

# Public-use microdata (large CSV zip)
curl -sL -o pu2022_csv.zip \
  "https://www2.census.gov/programs-surveys/sipp/data/datasets/2022/pu2022_csv.zip"
```

List the top-level directory to see available panel years before downloading:
<https://www2.census.gov/programs-surveys/sipp/data/datasets/>

### Files available per panel year

| File type | Typical filename pattern | Notes |
|-----------|--------------------------|-------|
| CSV zip | `pu<year>_csv.zip` | pipe/character-delimited; parse with schema |
| Gzipped CSV | `pu<year>_csv.gz` | same content, gzip-compressed |
| SAS input files | `pu<year>_sas.zip` | format statements for SAS users |
| Schema JSON | `pu<year>_schema.json` | variable names, labels, value codes |

### Load in Python

```python
import json
import pandas as pd

# 1. Load the schema to inspect variable names and labels
with open("pu2022_schema.json") as f:
    schema = json.load(f)

# schema is a list of variable-definition objects, e.g.:
# [{"name": "SSUID", "label": "Sample Unit Identifier (scrambled)", ...}, ...]
var_names = [v["name"] for v in schema]
print(var_names[:10])

# 2. Read the public-use CSV; the file is pipe/character-delimited
#    Read dtype=str first and cast after inspecting each column
df = pd.read_csv(
    "pu2022_csv.zip",          # pandas reads zip directly
    sep="|",                   # pipe delimiter; confirm with schema if changed
    dtype=str,
    low_memory=False,
)

# 3. Track individuals: combine SSUID (sample unit) + PNUM (person number)
df["person_id"] = df["SSUID"] + "_" + df["PNUM"]

# 4. Apply person weights before computing any means or proportions
df["WPFINWGT"] = pd.to_numeric(df["WPFINWGT"], errors="coerce")
```

Confirm the delimiter and column names against the schema for the specific
panel year you are using; the variable set and coding can differ across panels.

## Gotchas (the ones that bite pipelines)

- **Old and new panels are not comparable.** The 2014 redesign (the 2014 panel,
  then annual single-year panels from 2018 onward) changed the sample design,
  weighting, and reference-period recall relative to the 1984-2008 panels. Do not splice the two generations into a
  single time series without explicit adjustment.
- **Weights are required.** SIPP is a complex survey. Estimates need the
  supplied person or household weights and the set of replicate weights for
  variance estimation. An unweighted mean is biased and not representative of
  the U.S. population.
- **Seam bias inflates transition rates.** Respondents recall month-by-month
  activity since the prior interview. Month-to-month transitions cluster at
  interview-wave boundaries because recall degrades within a reference period;
  measured transition rates are artificially elevated at seams.
- **Longitudinal linking requires two keys.** SSUID identifies the sample
  unit (roughly a household at the time of initial sampling) and PNUM
  identifies the person within it. Both must be combined to track an individual
  across waves. Attrition shrinks the panel over time; account for panel
  non-response in any longitudinal analysis.
- **Income variables are topcoded and imputed.** High incomes are topcoded to
  protect confidentiality, and hot-deck imputation flags are provided for
  missing values. Tail estimates of income distributions and regression
  coefficients on imputed observations are affected; check the imputation flags
  before using income in a model.
- **The data is character-delimited text, not a preformatted table.** The
  public-use file is a flat text file with a pipe delimiter. You must parse it
  against the schema JSON to get variable labels and value codes. Do not assume
  columns are in a fixed position without checking the schema for the specific
  panel year.

## Citation

U.S. Census Bureau, Survey of Income and Program Participation, 2022 Panel,
Public Use Microdata, retrieved from
https://www2.census.gov/programs-surveys/sipp/data/datasets/2022/, accessed
2026-06-09.


==============================================================================
# TNIC (Hoberg-Phillips text-based industries)
# https://instituteforautomatedresearch.org/wiki/datasets/tnic/
# How to pull the Hoberg-Phillips Text-based Network Industry Classifications (TNIC) firm-pair similarity data as no-key bulk files, plus the gotchas that bite pipelines (it is a firm-specific relational network, not a partition; gvkey identifiers; the score is an excess-over-threshold, not a raw cosine).
# Verified 2026-06-22 · tested with live no-key download of tnic3_data.zip from the Hoberg-Phillips Data Library (151 MB zip, tnic3_data.txt = 27,161,831 firm-pair-year rows, year/gvkey1/gvkey2/score, 1988 to 2023)
# Tags: industry-classification, firm-similarity, network, panel, free, no-api-key, academic, data:tnic
==============================================================================

**TNIC** (Text-based Network Industry Classifications) is Gerard Hoberg and
Gordon Phillips's reclassification of US public firms into industries built from
the text of their 10-K product descriptions. Instead of a fixed code like SIC or
NAICS, TNIC gives **each firm its own industry**: the set of other firms whose
product-description text is similar enough to that firm's. The core download is a
firm-pair-year table with a similarity score for every related pair. Used in, for
example,
[Hoberg & Phillips (2025)](/wiki/papers/jf/2025/hoberg-scope-scale-concentration-21st-2025/),
where TNIC similarity drives firm-specific measures of scope and product-market
concentration.

- **Cost:** free, public.
- **API key:** none required (bulk file download from an academic data library).
- **Coverage:** US public firms with a machine-readable 10-K product
  description, annual 1988 to 2023 (TNIC-3 release). Identifier is Compustat
  `gvkey`.
- **Home:** <https://hobergphillips.tuck.dartmouth.edu/industryclass.htm>

## Access

The Hoberg-Phillips Data Library serves each product as a zipped flat file, no
key and no registration. The three main products are TNIC-3 (recommended,
calibrated to be about as granular as three-digit SIC), TNIC-2 (coarser), and
FIC (Fixed Industry Classifications, see below).

```bash
# TNIC-3 firm-pair similarity (151 MB zip -> ~679 MB tab-delimited text):
curl -sL "https://hobergphillips.tuck.dartmouth.edu/idata/tnic3_data.zip" -o tnic3.zip
unzip -p tnic3.zip Readme_tnic3.txt | head   # read the technical notes first
```

The data file `tnic3_data.txt` is tab-delimited with header `year gvkey1 gvkey2
score`: one row per related firm pair per year. A higher score means a nearer
product-market rival.

### Load in Python

```python
import pandas as pd, zipfile

with zipfile.ZipFile("tnic3.zip") as z:
    tnic = pd.read_csv(z.open("tnic3_data.txt"), sep="\t")

# firm-specific industry: every rival of gvkey 1011 in 2023
rivals = tnic[(tnic.gvkey1 == 1011) & (tnic.year == 2023)]
```

## Gotchas (the ones that bite pipelines)

- **It is a relational network, not a partition.** There is no industry code to
  group on. Each firm has its own set of rivals, industries overlap, and
  membership is **not transitive**: A related to B and B related to C does not
  imply A related to C. Do not `groupby(industry)`; work with the firm-pair
  edges, or use FIC (below) when you genuinely need mutually exclusive groups.
- **The identifier is Compustat `gvkey`, not CUSIP/ticker/PERMNO.** Both
  `gvkey1` and `gvkey2` are gvkeys. Link through Compustat to get names, CRSP
  PERMNOs, or CUSIPs; there is no security identifier in the file.
- **`score` is an excess over the inclusion threshold, not a raw cosine.** A pair
  appears only when its pairwise product similarity clears the TNIC threshold,
  and the score is the amount by which it exceeded that threshold, so the
  distribution starts near zero. A score near zero means the pair barely
  qualified. Do not read it as a cosine in [0, 1] or as a probability.
- **Pairs are symmetric and both directions are stored.** `(i, j)` and `(j, i)`
  both appear with the same score, so the file roughly double-counts unordered
  pairs (27.2M rows in the 1988 to 2023 TNIC-3). Deduplicate to `gvkey1 < gvkey2`
  if you want each pair once.
- **Self-pairs are included.** Rows with `gvkey1 == gvkey2` appear (with an empty
  score); drop the diagonal before counting rivals.
- **It is recomputed every year; relationships move.** A firm's rival set in 2010
  differs from 2020 as product descriptions and the firm universe change. Always
  filter on `year`; do not treat a single year as time-invariant.
- **Coverage is 10-K-bound.** Only firms that file a machine-readable 10-K with a
  product-description section are classified. Firms without one (many foreign
  private issuers, firms in non-filing years, very early years) are simply
  absent, not zero.
- **Pick the granularity deliberately.** TNIC-3 is the recommended general-use
  release (about three-digit SIC granularity); TNIC-2 is coarser. They share the
  same column layout but answer different questions.
- **The file is large.** TNIC-3 is 151 MB zipped, about 679 MB and 27.2M rows
  unzipped; TNIC-2 is larger still. Read it in chunks or filter on read.

## FIC vs TNIC

The library also ships **FIC** (Fixed Industry Classifications): a transitive,
mutually exclusive classification with a fixed, preset number of industries,
built from the same text but assigning each firm to a single industry (the
library publishes several granularities). Use FIC when you need a conventional
partition (industry fixed effects, counts per industry); use TNIC when you want
firm-specific rival sets and pairwise similarity.

## Reference

| Field | Value |
|-------|-------|
| Library | `https://hobergphillips.tuck.dartmouth.edu/industryclass.htm` |
| TNIC-3 file | `idata/tnic3_data.zip` -> `tnic3_data.txt` (tab-delimited) |
| Columns | `year`, `gvkey1`, `gvkey2`, `score` |
| Identifier | Compustat `gvkey` |
| Frequency | Annual, 1988 to 2023 (TNIC-3 release) |
| Other products | TNIC-2 (coarser), FIC (fixed transitive classification) |
| Key required | No |

## Citation

The library's readme names the two papers to cite: Gerard Hoberg and Gordon
Phillips, "Text-Based Network Industries and Endogenous Product Differentiation,"
*Journal of Political Economy* 124(5), 2016, 1423-1465; and Gerard Hoberg and
Gordon Phillips, "Product Market Synergies and Competition in Mergers and
Acquisitions: A Text-Based Analysis," *Review of Financial Studies* 23(10), 2010,
3773-3811. Cite the specific release (TNIC-2, TNIC-3, or FIC), the retrieval URL,
and the access date, since the library updates the data over time.


==============================================================================
# Uniswap on-chain data (Ethereum)
# https://instituteforautomatedresearch.org/wiki/datasets/uniswap-blockchain/
# Swap, Mint, and Burn event logs for Uniswap V2/V3 liquidity pools on Ethereum mainnet, pulled directly from the public blockchain via no-key JSON-RPC eth_getLogs; the key operational question is which public RPC endpoints actually serve getLogs on archive blocks without a token. Uniswap V1, used in early studies, has a different architecture and is noted here too.
# Verified 2026-06-22 · tested with no-key eth_getLogs via https://eth.drpc.org returning 32 Uniswap V3 USDC/ETH 0.05% Swap events (pool 0x88e6A0c2dDD26FEEb64F039a2c41296FcB3f5640, first tx 0xd62b5f0ab66c8520683c29a22aa3e275fe3af6cab3dbe342e96feacaf14cb7ae) plus a live eth_blockNumber; V2/V3 Swap topic0 signatures cross-checked
# Tags: defi, blockchain, crypto, market-microstructure, event-data, free, no-api-key, data:uniswap-blockchain
==============================================================================

**Uniswap on-chain data** is the set of Swap, Mint (liquidity added), and Burn
(liquidity removed) events for every Uniswap V2 and V3 pool on Ethereum mainnet,
stored permanently in the public Ethereum blockchain as contract event logs. The
data is retrieved by querying a JSON-RPC endpoint with `eth_getLogs`, filtered
by pool contract address and event topic0 signature. No account and no API key
are required: the chain state is public. Pools are discovered from the Uniswap
factory contracts via their `PairCreated` (V2) or `PoolCreated` (V3) events.
Used in, for example,
[Lehar & Parlour (2025)](/wiki/papers/jf/2025/lehar-decentralized-exchange-uniswap-automated-2025/),
where 95.8 million interactions across 105,098 pools from November 2018 to
December 2022 form the primary dataset.

- **Cost:** free. The Ethereum blockchain is public state.
- **API key:** none required for the working public endpoints listed below;
  several commonly cited endpoints now require a key or block archive requests.
- **Coverage:** Uniswap V2 (launched May 2020, block ~10,000,835) and V3
  (launched May 2021, block ~12,369,621); all pools on Ethereum mainnet.
  Uniswap V1 (launched November 2018) is the earliest version and has a
  different design (see Gotchas); studies of the early period, including Lehar
  & Parlour, cover V1 and V2. Historical pulls require an archive node.
- **V1 factory:** `0xc0a47dFe034B400B47bDaD5FecDa2621de6c4d95`
- **V2 factory:** `0x5C69bEe701ef814a2B6a3EDD4B1652CB9cc5aA6f`
- **V3 factory:** `0x1F98431c8aD98523631AE4a59f267346ea31F984`
- **Home:** <https://uniswap.org>

## Access

Query any JSON-RPC endpoint with `eth_getLogs`, specifying the pool contract
address and the event topic0 hash. The public endpoint `https://eth.drpc.org`
serves archive getLogs with no key (confirmed 2026-06-22). Use small block
windows (20-100 blocks for active pools; see Gotchas).

```bash
# Fetch V3 Swap events for the USDC/ETH 0.05% pool over 100 blocks (drpc, no key)
# Pool: 0x88e6A0c2dDD26FEEb64F039a2c41296FcB3f5640
# V3 Swap topic0: 0xc42079f94a6350d7e6235f29174924f928cc2ac818eb64fed8004e115fbcca67
curl -s https://eth.drpc.org \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "method":  "eth_getLogs",
    "params": [{
      "fromBlock": "0xF42400",
      "toBlock":   "0xF42464",
      "address":   "0x88e6A0c2dDD26FEEb64F039a2c41296FcB3f5640",
      "topics":    ["0xc42079f94a6350d7e6235f29174924f928cc2ac818eb64fed8004e115fbcca67"]
    }],
    "id": 1
  }'
```

Pool addresses are not known in advance; discover them from the factory. To
enumerate all V3 pools, query the V3 factory for `PoolCreated` events using
the same `eth_getLogs` method with the factory address and the `PoolCreated`
topic0. V2 pools are discovered via `PairCreated` on the V2 factory.

### Endpoint status (tested 2026-06-22)

| Endpoint | eth_blockNumber | eth_getLogs (archive) | Notes |
|----------|----------------|----------------------|-------|
| `https://eth.drpc.org` | OK | OK | No key. 10,000-block hard cap; data-volume timeout on busy pools around 200 blocks; 20-100 block windows safe. Recommended primary. |
| `https://1rpc.io/eth` | OK | OK | No key. Returns identical logs. Good failover. |
| `https://ethereum-rpc.publicnode.com` | OK | Blocked | Returns "Archive requests require a personal token" on getLogs. |
| `https://eth-mainnet.public.blastapi.io` | OK | Severely capped | 10-block range cap; impractical for backfill. |
| `https://rpc.ankr.com/eth` | Requires key | Requires key | Key required as of 2026-06. |
| `https://cloudflare-eth.com` | Dead | Dead | Returns -32046. |
| `https://eth.llamarpc.com` | Down | Down | HTTP 521. |

## Gotchas (the ones that bite pipelines)

- **Most well-known public RPCs do not serve archive getLogs.** Several
  endpoints that appear in tutorials return `eth_blockNumber` fine but reject
  `eth_getLogs` on archive blocks with "archive request requires token" or
  fail silently (cloudflare: -32046; llamarpc: 521; ankr: requires key). Test
  the actual getLogs call against a known old block, not just connectivity.
  Hard-code drpc plus 1rpc as the working pair and implement failover.
- **Block-range limits are twofold.** There is a hard block-count cap (drpc:
  10,000 blocks; blastapi: 10 blocks) and a separate data-volume timeout (on
  drpc, a high-activity pool like USDC/ETH 0.05% times out around 200 blocks;
  20-100 blocks is safe). Page in small windows and implement exponential
  back-off on timeouts. A chunker that iterates across windows is required for
  any historical backfill.
- **Uniswap V1 is not V2 or V3, and this page's topic0 hashes and decoding are
  V2/V3 only.** The Lehar & Parlour sample begins at V1 launch (November 2018);
  V1 predates the factory-of-pairs design. Each ERC-20 token has its own Vyper
  exchange contract, discovered from the V1 factory
  (`0xc0a47dFe034B400B47bDaD5FecDa2621de6c4d95`) via its `NewExchange` event.
  V1 trades emit `TokenPurchase` and `EthPurchase` (not `Swap`) and liquidity
  events `AddLiquidity` and `RemoveLiquidity` (not `Mint` / `Burn`). Read the
  event ABI from a V1 exchange contract before decoding; do not reuse the V2/V3
  signatures below for V1 logs.
- **V2 and V3 ABI schemas differ substantially; do not share a decoder.** V2
  Swap has four unsigned amounts (`amount0In`, `amount1In`, `amount0Out`,
  `amount1Out`); price is inferred from the reserve state via the `Sync` event.
  V3 Swap has signed `amount0` and `amount1` (negative = flowing out of the
  pool) plus `sqrtPriceX96`, `liquidity`, and `tick`; price is
  `(sqrtPriceX96 / 2**96)**2` adjusted for the decimal difference between the
  two tokens. V3 Mint and Burn carry `tickLower` and `tickUpper` for the
  concentrated-liquidity range, which V2 does not have. Indexed event parameters
  (sender, owner, recipient) come from `topics[1]` and `topics[2]`, not from
  the `data` field.
- **Historical backfill requires archive access.** V2 launched at block
  ~10,000,835 (May 2020) and V3 at block ~12,369,621 (May 2021). Fetching
  those early blocks requires an archive node. drpc serves archive getLogs
  without a key (confirmed at block 12,380,000). A full backfill across all
  pools is large; discover pools from the factory first, then run per-pool
  windowed scans.
- **Reorgs near the chain tip.** Do not treat the most recent blocks as final.
  Lag at least 64 blocks from the tip, or wait for the finalized tag. Tag each
  event with `blockNumber`, `transactionHash`, and `logIndex` as a composite
  key, and re-fetch the trailing window after each run to self-heal rows from
  reorged blocks.
- **Token amounts are raw integers.** Divide by `10**decimals` per token before
  using values: USDC uses 6 decimals, WETH uses 18. Fetch decimals from the
  token contract's `decimals()` view function or from a maintained token list.
  Always confirm a pool address came from the correct factory (V2 vs V3) before
  trusting its events.

## Reference

### Factory contracts

| Protocol | Factory address | Pool discovery event |
|----------|----------------|----------------------|
| Uniswap V1 | `0xc0a47dFe034B400B47bDaD5FecDa2621de6c4d95` | `NewExchange(address,address)` (per-token Vyper exchange; V1 ABI differs, see Gotchas) |
| Uniswap V2 | `0x5C69bEe701ef814a2B6a3EDD4B1652CB9cc5aA6f` | `PairCreated(address,address,address,uint256)` |
| Uniswap V3 | `0x1F98431c8aD98523631AE4a59f267346ea31F984` | `PoolCreated(address,address,uint24,int24,address)` |

### Event topic0 signatures (keccak256 of the event signature string)

| Protocol | Event | topic0 |
|----------|-------|--------|
| V2 | `Swap(address,uint256,uint256,uint256,uint256,address)` | `0xd78ad95fa46c994b6551d0da85fc275fe613ce37657fb8d5e3d130840159d822` |
| V2 | `Mint(address,uint256,uint256)` | `0x4c209b5fc8ad50758f13e2e1088ba56a560dff690a1c6fef26394f4c03821c4f` |
| V2 | `Burn(address,uint256,uint256,address)` | `0xdccd412f0b1252819cb1fd330b93224ca42612892bb3f4f789976e6d81936496` |
| V3 | `Swap(address,address,int256,int256,uint160,uint128,int24)` | `0xc42079f94a6350d7e6235f29174924f928cc2ac818eb64fed8004e115fbcca67` |
| V3 | `Mint(address,address,int24,int24,uint128,uint256,uint256)` | `0x7a53080ba414158be7ec69b987b5fb7d07dee101fe85488f0853ae16239d0bde` |
| V3 | `Burn(address,int24,int24,uint128,uint256,uint256)` | `0x0c396cd989a39f4459b5fa1aed6a9a8dcdbc45908acfd67e028cd568da98982c` |

### V2 vs V3 decoding summary

| Aspect | V2 | V3 |
|--------|----|----|
| Swap amount fields | Unsigned: `amount0In`, `amount1In`, `amount0Out`, `amount1Out` | Signed: `amount0`, `amount1` (negative = out of pool) |
| Price in Swap | Not present; infer from `Sync(reserve0,reserve1)` | `sqrtPriceX96`; price = `(sqrtPriceX96/2^96)^2` x decimal adjustment |
| Liquidity range | Full range only | Concentrated: `tickLower`, `tickUpper` in Mint/Burn |
| Indexed params in topics | `sender` (Swap, Mint/Burn) | `sender` + `recipient` (Swap); `owner` + ticks (Mint/Burn) |
| Key required | No | No |

## Citation

Cite the Ethereum blockchain as the primary source, specifying the Uniswap
V2 or V3 core contracts (with factory address), the pool address(es) queried
(with token pair and fee tier for V3), the block range covered, the RPC
endpoint used, and the retrieval date. If data were obtained via a derived or
aggregated product (such as The Graph subgraph or Dune Analytics), cite that
layer separately and note the underlying source. Example for a V3 pull:

> Uniswap V3 core contracts (factory `0x1F98431c8aD98523631AE4a59f267346ea31F984`),
> Ethereum mainnet. Swap events fetched via `eth_getLogs` from
> `https://eth.drpc.org`, blocks [from]-[to], retrieved [date].
> Pool: USDC/WETH 0.05% (`0x88e6A0c2dDD26FEEb64F039a2c41296FcB3f5640`).


==============================================================================
# EPA Supply Chain GHG Emission Factors
# https://instituteforautomatedresearch.org/wiki/datasets/us-epa-supply-chain/
# How to pull the EPA's NAICS-level supply-chain greenhouse-gas emission factors from the no-key CSV, the with-margins vs without-margins distinction, and the gotchas that bite pipelines (NAICS vintage, USD year, version).
# Verified 2026-06-09 · tested with live no-key download of EPA Supply Chain GHG Emission Factors v1.3 (SupplyChainGHGEmissionFactors_v1.3.0_NAICS_CO2e_USD2022.csv, 1,016 NAICS codes)
# Tags: esg, emissions, environmental, cross-section, free, no-api-key, data:us-epa-supply-chain
==============================================================================

**The EPA Supply Chain Greenhouse Gas Emission Factors** map each US industry
or commodity (by six-digit NAICS code) to the greenhouse-gas emissions
generated per dollar of its output, expressed as kilograms of carbon-dioxide
equivalent per US dollar. They let you attach an emissions intensity to spending
or revenue by industry without firm-level emissions data. The factors are built
by the Environmental Protection Agency from the USEEIO environmentally-extended
input-output model and are free and public. Used in, for example,
[Allcott, Montanari, Ozaltun & Tan](/wiki/papers/jf/2026/allcott-corporate-social-impact-2026/),
in the corporate-social-impact analysis.

- **Cost:** free, public.
- **API key:** none required.
- **Coverage:** roughly 1,000 commodities at six-digit NAICS; one factor set
  per release (v1.3 uses 2017 NAICS, 2022 USD, AR5 100-year GWP).
- **Home:** <https://www.epa.gov/climateleadership/supply-chain-greenhouse-gas-emission-factors>

## Access

The factor table is a single CSV, no authentication. The CO2e file gives all
gases combined; a by-GHG file breaks them out by gas:

```bash
# Supply Chain GHG factors v1.3, all GHGs combined (CO2e), 2022 USD, no key
curl -sL -o SupplyChainGHGEmissionFactors_v1.3.0_NAICS_CO2e_USD2022.csv \
  "https://pasteur.epa.gov/uploads/10.23719/1531143/SupplyChainGHGEmissionFactors_v1.3.0_NAICS_CO2e_USD2022.csv"
# By individual greenhouse gas:
curl -sL -o SupplyChainGHGEmissionFactors_v1.3.0_NAICS_byGHG_USD2022.csv \
  "https://pasteur.epa.gov/uploads/10.23719/1531143/SupplyChainGHGEmissionFactors_v1.3.0_NAICS_byGHG_USD2022.csv"
```

The CO2e file has one row per six-digit NAICS code with columns: `2017 NAICS
Code`, `2017 NAICS Title`, `GHG`, `Unit`, `Supply Chain Emission Factors without
Margins`, `Margins of Supply Chain Emission Factors`, `Supply Chain Emission
Factors with Margins`, and `Reference USEEIO Code`. The unit is kg CO2e per 2022
USD at purchaser price.

### Load in Python

```python
import pandas as pd

df = pd.read_csv(
    "SupplyChainGHGEmissionFactors_v1.3.0_NAICS_CO2e_USD2022.csv",
    dtype={"2017 NAICS Code": str},
)
# Attach intensity to spend: emissions_kgCO2e = spend_usd * factor_with_margins
factor = df.set_index("2017 NAICS Code")["Supply Chain Emission Factors with Margins"]
```

## Gotchas (the ones that bite pipelines)

- **With-margins versus without-margins are different numbers.** "Without
  margins" is the production emissions per producer-price dollar; "with margins"
  adds the transport, wholesale, and retail margins so the factor applies to a
  **purchaser-price** dollar (what a buyer actually pays). Multiply spend at
  purchaser prices by the with-margins factor; mixing the two double-counts or
  undercounts.
- **The factor's USD year is fixed; deflate your spend to match.** v1.3 factors
  are per **2022** dollar. Spending in another year's dollars must be deflated to
  the factor's USD year before multiplying, or the intensity is off by inflation.
- **NAICS vintage matters for the join.** v1.3 keys on **2017** NAICS codes.
  Firm or transaction data coded to a different NAICS vintage (2012, 2022) needs
  a concordance first; a code can shift industries across vintages.
- **Versions change methodology, not just numbers.** v1.2 (2021 USD) and v1.3
  (2022 GHG data, 2022 USD) differ in the emissions-data year, USD year, and
  minor method. Pin one version across a project and cite it.
- **A few sectors are excluded.** Electricity, government, and households are
  not given factors at the six-digit level. Spend mapped to those NAICS has no
  factor; handle it explicitly rather than letting it drop to zero.
- **It is a spend-based average, not a firm measurement.** The factor is the
  industry-average emissions per dollar from an input-output model, not any one
  firm's footprint. Two firms in the same NAICS get the same intensity per
  dollar regardless of their actual technology.

## Factor columns

| Column | Meaning |
|--------|---------|
| Supply Chain Emission Factors without Margins | kg CO2e per producer-price USD |
| Margins of Supply Chain Emission Factors | the transport/wholesale/retail margin component |
| Supply Chain Emission Factors with Margins | kg CO2e per purchaser-price USD (without + margins) |
| Reference USEEIO Code | the USEEIO model commodity the factor maps to |

## Citation

Cite the EPA, the dataset and version, the NAICS and USD year, the URL, and the
access date, for example: *U.S. Environmental Protection Agency, Supply Chain
Greenhouse Gas Emission Factors v1.3 (2017 NAICS, 2022 USD), retrieved from
https://www.epa.gov/climateleadership/supply-chain-greenhouse-gas-emission-factors,
accessed YYYY-MM-DD.* Record the version, the with/without-margins choice, and
the USD year so the calculation is reproducible.


==============================================================================
# CBOE Volatility Index (VIX)
# https://instituteforautomatedresearch.org/wiki/datasets/vix/
# How to pull the full daily VIX history as a no-key CSV from Cboe, plus the gotchas that bite pipelines (the 1990-2002 backfill vs the original VXO, flat early OHLC, annualized-percentage units, and the family of look-alike vol indices).
# Verified 2026-06-09 · tested with live no-key CSV fetch of the full VIX daily history (cdn.cboe.com VIX_History.csv, 9202 rows, 1990-01-02 to 2026-06-08)
# Tags: market, volatility, options, time-series, equities, free, no-api-key, data:vix
==============================================================================

The **CBOE Volatility Index (VIX)** is Cboe Global Markets' measure of the
market's expectation of 30-day forward volatility of the S&P 500, computed from
a strip of SPX option prices. It is quoted as an annualized percentage (a VIX
of 20 implies about 20% annualized expected volatility). The full daily history
is published free by Cboe with no API key. Used in, for example,
[Stavrakeva, Tang & Sun](/wiki/papers/jf/2026/stavrakeva-dollar-great-recession-2026/),
where log VIX is a dependent-variable proxy for risk aversion in the response
to forward-guidance easings.

- **Cost:** free, public (end-of-day history).
- **API key:** none required.
- **Coverage:** daily, 1990-01-02 to present. Values before 2003 are
  backfilled with the current methodology (see Gotchas).
- **Units:** annualized expected volatility, in percentage points.
- **Home:** <https://www.cboe.com/tradable_products/vix/>

## Access

Cboe serves the entire history as a single no-key CSV from its CDN. Columns are
`DATE,OPEN,HIGH,LOW,CLOSE`, with `DATE` formatted `M/D/YYYY`:

```bash
# No key required; full daily history in one file
curl -sL "https://cdn.cboe.com/api/global/us_indices/daily_prices/VIX_History.csv" -o VIX_History.csv
```

The same CDN path serves the sibling volatility indices by swapping the file
name, for example `VIX9D_History.csv` (9-day), `VIX3M_History.csv` (3-month,
formerly VXV), `VVIX_History.csv` (vol-of-vol), `VXN_History.csv` (Nasdaq-100),
and `VXD_History.csv` (Dow). Confirm you are pulling the index your analysis
names.

### Load in Python

```python
import pandas as pd

url = "https://cdn.cboe.com/api/global/us_indices/daily_prices/VIX_History.csv"
vix = pd.read_csv(url, parse_dates=["DATE"])
vix = vix.rename(columns=str.lower).set_index("date").sort_index()

# The end-of-day level most analyses use is the CLOSE column.
close = vix["close"]
```

## Gotchas (the ones that bite pipelines)

- **The 1990-2002 values are a backfill, not the index that traded then.** Cboe
  redefined VIX in 2003 to a model-free, variance-swap-style formula on a strip
  of SPX options. The pre-2003 portion of the history file is that new formula
  applied retroactively. The index that was actually disseminated in real time
  from 1993 to 2003 was based on at-the-money S&P 100 (OEX) implied volatility
  and now trades under the ticker **VXO**. If your identification depends on
  what investors observed in real time before 2003, you want VXO, not the
  backfilled VIX.
- **Early OHLC is flat.** From 1990-01-02 through the end of 1991 the `OPEN`,
  `HIGH`, `LOW`, and `CLOSE` columns are all equal to the close; only the close
  is meaningful. Distinct intraday open/high/low values begin 1992-01-02. Do
  not compute intraday ranges over the 1990-1991 rows.
- **It is a percentage, not a price.** VIX is annualized expected volatility in
  percentage points and a 30-calendar-day forward measure. Do not treat a level
  change as a return, and do not annualize it again.
- **Many look-alike indices share the path.** VIX9D, VIX3M (VXV), VVIX, VXN,
  VXD, RVX, and the original VXO are distinct series. A wrong file name silently
  returns a different index with the same column layout.
- **No official VIX before 1990.** For longer volatility histories use realized
  volatility or VXO; do not extend the VIX series synthetically without saying
  so.
- **FRED `VIXCLS` is a close-only mirror.** If you pull VIX from FRED, cite FRED
  and note it carries the close only. The canonical primary source with OHLC is
  the Cboe file above.
- **Free history is end-of-day.** Real-time, intraday, and the underlying
  option-level data are licensed Cboe DataShop products. The no-key file is the
  daily history only.

## Reference

| Field | Value |
|-------|-------|
| URL | `https://cdn.cboe.com/api/global/us_indices/daily_prices/VIX_History.csv` |
| Format | CSV: `DATE,OPEN,HIGH,LOW,CLOSE`; `DATE` is `M/D/YYYY` |
| Frequency | Daily |
| Start | 1990-01-02 (current methodology backfilled to here) |
| Units | Annualized expected volatility, percentage points |
| Methodology change | 2003 (SPX strip; pre-2003 real-time index is VXO) |
| Key required | No |

## Citation

Cite Cboe Global Markets, the index name, the retrieval URL, and the access
date, for example: *Cboe Global Markets, "Cboe Volatility Index (VIX),"
historical data retrieved from
https://cdn.cboe.com/api/global/us_indices/daily_prices/VIX_History.csv,
accessed YYYY-MM-DD.* If you use the pre-2003 portion, state in your data
description that those values are the current methodology applied
retroactively, and cite VXO separately if you need the real-time pre-2003
index.


==============================================================================
# Internet Archive Wayback Machine
# https://instituteforautomatedresearch.org/wiki/datasets/wayback-machine/
# How to query the Internet Archive Wayback Machine for historical web-page snapshots from the no-key Availability and CDX APIs, and the gotchas that bite pipelines (coverage is not continuous, a snapshot is a crawl not the live page, rate limits, capture != content change).
# Verified 2026-06-22 · tested with live no-key pull of the Wayback Availability API (archive.org/wayback/available for sec.gov, returned the closest 2010 snapshot)
# Tags: web-archive, alternative-data, text, free, no-api-key, data:wayback-machine
==============================================================================

**The Internet Archive Wayback Machine** is a public archive of historical
snapshots of web pages, captured by crawlers since 1996. For research it is the
standard way to observe a website *as it was* at a past date: page content,
update frequency, whether a site existed, and how it changed over time. The data
is free and public. Used in, for example,
[Hu & Ma](/wiki/papers/jf/2025/hu-persuading-investors-video-based-2025/), where
website update frequency over the three years after a pitch is built from
Wayback Machine captures as a startup-survival measure.

- **Cost:** free, public.
- **API key:** none required (a descriptive `User-Agent` is courteous; the APIs
  are rate-limited).
- **Coverage:** snapshots from 1996 onward; depth varies enormously by site
  (popular sites captured many times a day, obscure ones rarely).
- **Home:** <https://web.archive.org/>

## Access

Two no-key JSON APIs serve the archive programmatically.

**Availability API** answers "is there a snapshot of this URL near this date?":

```bash
# Closest snapshot of sec.gov to 1 Jan 2010
curl -sL "https://archive.org/wayback/available?url=sec.gov&timestamp=20100101"
# -> {"archived_snapshots":{"closest":{"available":true,
#     "url":"http://web.archive.org/web/20100102140201/http://www.sec.gov/", ...}}}
```

**CDX API** lists the full capture history, one row per snapshot, and is what you
want for update-frequency or change-over-time measures:

```bash
# All captures of a URL in 2009, deduplicated to one per day
curl -sL "https://web.archive.org/cdx/search/cdx?url=example.com\
&from=20090101&to=20091231&output=json&collapse=timestamp:8"
```

CDX returns `[urlkey, timestamp, original, mimetype, statuscode, digest,
length]` per capture. The `digest` is a content hash: two consecutive captures
with the same `digest` are byte-identical, which is how you separate a *recrawl*
from an actual *content change*. To fetch the archived page itself, request
`https://web.archive.org/web/<timestamp>/<original-url>` (append `id_` to the
timestamp, e.g. `20100102140201id_`, for the raw capture without the Wayback
toolbar).

### Load in Python

```python
import requests

def captures(url, year):
    cdx = ("https://web.archive.org/cdx/search/cdx?url=" + url +
           f"&from={year}0101&to={year}1231&output=json&collapse=digest")
    rows = requests.get(cdx, timeout=60,
                        headers={"User-Agent": "research (contact@example.org)"}).json()
    return rows[1:]  # first row is the column header
# len(captures(...)) after collapse=digest = number of distinct content versions
```

## Gotchas (the ones that bite pipelines)

- **Coverage is not continuous, and absence is not proof.** Snapshot density
  depends on crawl priority, robots directives, and luck. A gap in captures does
  **not** mean the page did not change or did not exist; it means it was not
  crawled. Treat the archive as an irregular, selection-biased sample, not a
  complete log. Update-frequency measures should be normalized by crawl
  frequency or restricted to well-crawled sites.
- **A capture is a crawl, not a content change.** The raw capture count
  overstates how often a page changed: the same content is recrawled repeatedly.
  Use `collapse=digest` (content hash) or compare the `digest` field across rows
  to count *distinct* versions; counting raw timestamps inflates "activity".
- **Rate limits are real and the APIs return 429.** The CDX endpoint in
  particular throttles aggressively; sustained scraping returns
  `429 Too Many Requests`. Throttle to a few requests per second, back off on
  429, and cache responses. For large jobs the Internet Archive publishes bulk
  CDX/WARC access paths rather than hammering the live API.
- **The snapshot is what the crawler saw, not the live page.** JavaScript-heavy
  pages, lazy-loaded content, and resources blocked at crawl time may be missing
  or rendered differently from the original. The archived HTML can also embed
  the Wayback toolbar; use the `id_` modifier on the timestamp to get the raw
  capture.
- **Timestamps are UTC and 14 digits.** The capture timestamp is
  `YYYYMMDDhhmmss` in UTC. The Availability API's `closest` match can be on
  either side of your target date and arbitrarily far away if coverage is thin;
  always check how far `timestamp` is from what you asked for.
- **`statuscode` matters.** CDX rows include captures of redirects (3xx) and
  errors (4xx/5xx). Filter to `statuscode=200` (and the right `mimetype`) before
  treating a row as a real page view, or `filter=statuscode:200` in the query.

## Endpoints

| Endpoint | Use |
|---|---|
| `archive.org/wayback/available?url=&timestamp=` | Nearest single snapshot to a date (existence check) |
| `web.archive.org/cdx/search/cdx?url=` | Full capture history; `output=json`, `collapse=`, `filter=` |
| `web.archive.org/web/<ts>/<url>` | The archived page itself (`<ts>id_` for the raw capture) |

## Citation

Cite the Internet Archive, the Wayback Machine, the archived URL and capture
timestamp, and the access date, for example: *Internet Archive, Wayback Machine,
snapshot of http://www.sec.gov/ captured 2010-01-02 (https://web.archive.org/web/20100102140201/http://www.sec.gov/),
accessed YYYY-MM-DD.* Record the exact capture timestamp so the snapshot is
reproducible, since the live site will have changed.


==============================================================================
# Zillow research data
# https://instituteforautomatedresearch.org/wiki/datasets/zillow/
# How to pull Zillow Research's free housing metrics (ZHVI, rents, days on market, price cuts) as no-key bulk CSVs, plus the gotchas that bite pipelines (wide format, the filename-is-the-metadata convention, restated history, and RegionID vs FIPS).
# Verified 2026-06-09 · tested with live no-key CSV fetch of the metro ZHVI file (files.zillowstatic.com public_csvs, 896 metros, monthly 2000-01 to 2026-04)
# Tags: housing, real-estate, prices, time-series, panel, free, no-api-key, data:zillow
==============================================================================

**Zillow research data** is the set of housing-market metrics Zillow Research
publishes free: the Zillow Home Value Index (ZHVI, a smoothed measure of typical
home value), the Zillow Observed Rent Index (ZORI), days on market, the share of
listings with a price cut, for-sale inventory, new listings, and median sale and
list prices. Each metric is offered as a no-key bulk CSV; ZHVI spans every
geography level (national, state, metro, county, city, ZIP, neighborhood) while
other metrics cover a subset (check the research-data page for what each
offers). Used in, for
example,
[Amaral, Dohmen, Kohl & Schularick](/wiki/papers/jf/2025/amaral-superstar-returns-spatial-heterogeneity-2025/),
where Zillow days-on-market and asking-price-discount metrics proxy liquidity
across 277 MSAs.

- **Cost:** free, public.
- **API key:** none required (bulk CSV download; there is no official query API).
- **Coverage:** geography from national down to ZIP/neighborhood; most series
  monthly. ZHVI runs from 2000; the rent index (ZORI) starts later (2015).
- **Home:** <https://www.zillow.com/research/data/>

## Access

The research-data page is a directory of download links; the files themselves
live on a static CDN at `files.zillowstatic.com/research/public_csvs/<metric>/`.
Fetch a specific file directly, no key:

```bash
# Metro-level ZHVI, all homes (single-family + condo), middle price tier,
# smoothed and seasonally adjusted, monthly:
curl -sL "https://files.zillowstatic.com/research/public_csvs/zhvi/Metro_zhvi_uc_sfrcondo_tier_0.33_0.67_sm_sa_month.csv" -o zhvi_metro.csv
```

Each file is wide: identifier columns `RegionID, SizeRank, RegionName,
RegionType, StateName` followed by one column per month (header is the
month-end date, e.g. `2024-01-31`).

### Load in Python

```python
import pandas as pd

url = "https://files.zillowstatic.com/research/public_csvs/zhvi/Metro_zhvi_uc_sfrcondo_tier_0.33_0.67_sm_sa_month.csv"
wide = pd.read_csv(url)

id_cols = ["RegionID", "SizeRank", "RegionName", "RegionType", "StateName"]
long = wide.melt(id_vars=id_cols, var_name="date", value_name="zhvi")
long["date"] = pd.to_datetime(long["date"])  # month-end dates
```

## Gotchas (the ones that bite pipelines)

- **ZHVI is a value index, not transaction prices.** It is a smoothed,
  seasonally adjusted measure of the typical home value in a region, not the
  prices of homes that sold. For realized transactions use the median sale price
  files; do not treat ZHVI as a sale-price series.
- **The filename is the metadata.** `sfrcondo` vs `sfr` vs `condo`, the
  `tier_0.33_0.67` (middle third) vs bottom or top tier or bedroom-count cuts,
  `sm` (smoothed) vs raw, and `sa` (seasonally adjusted) vs not are all encoded
  in the file name. Pick the slice deliberately; a different file with the same
  column layout answers a different question.
- **Data are wide; date columns are month-end.** Reshape to long for panel work,
  and read the header dates as month-end timestamps, not month-start.
- **History is restated every release.** Zillow revises the entire back-series
  on each monthly update as methodology and source data change. Two downloads of
  the "same" series from different months will differ. Snapshot the file and
  record the access date.
- **`RegionName` is a label, not a geocode.** Metro rows read like `New York,
  NY`; there is no FIPS or CBSA code in the value file. To merge with Census,
  FHFA, or CoreLogic you must crosswalk `RegionName`/`RegionID` to standard
  geography codes yourself.
- **Composition changes over time.** More metros, counties, and ZIPs enter in
  later years; a region present in 2026 can be missing (NaN) in 2000. Account
  for entry rather than assuming a balanced panel.
- **No official API.** Access is bulk CSV only, and the links on the research
  page can change. Pull from the static CSV host and archive the file.
- **ZTRAX is a different product.** The Zillow Transaction and Assessment
  Dataset (ZTRAX), parcel-level deed and assessment microdata, was a separate,
  application-gated academic dataset (and has been discontinued for new
  agreements). It is not part of these free research CSVs.

## Reference

| Field | Value |
|-------|-------|
| Index page | `https://www.zillow.com/research/data/` |
| CSV host | `https://files.zillowstatic.com/research/public_csvs/<metric>/` |
| Format | Wide CSV: id columns + one column per month (month-end header) |
| Geography | ZHVI: national, state, metro, county, city, ZIP, neighborhood; other metrics a subset |
| Frequency | Mostly monthly |
| ZHVI start | 2000; ZORI (rent) from 2015 |
| Key required | No |

## Citation

Cite Zillow, the specific metric and geography, the retrieval URL, and the
access date, for example: *Zillow Research, "Zillow Home Value Index (ZHVI),"
metro-level CSV retrieved from https://www.zillow.com/research/data/, accessed
YYYY-MM-DD.* Because Zillow restates the full history each release, record the
access date and archive the downloaded file so the vintage is reproducible.


==============================================================================
# Distilled literature
# https://instituteforautomatedresearch.org/wiki/papers/
# Machine-readable distillations of research papers (core results, datasets used, theory tested) so you can see what a paper found without reading all of it.
==============================================================================

import PapersIndex from '../../../components/PapersIndex.astro';

These pages distil each paper to the part you usually need (**its core
results, the datasets it used, and the theory it tested**), with exact
source locators. Read the full paper to replicate or extend it, not to find
out what it found.

Each distillation states its own **provenance honestly**: which paper, that
the source is peer-reviewed or not, that extraction was done by an LLM and
*not* human-verified, and whether the result has been reproduced (almost
never, yet; we say so rather than imply otherwise). Where the source licence
permits, the verbatim PDF is also mirrored, machine-accessible, in the
[Open Library](/library), because the publisher often is not.

## Distilled papers

The table below is generated at build time from each page's own metadata, so a
new distillation is one new file under `papers/` with nothing else to update.

<PapersIndex />

> A distillation is **not** a substitute for the paper under peer review or
> in citation. It is a map. Re-expressed factual results are not
> copyrightable; the synthesis is the Institute's, the findings are the
> authors'. Read the source (mirrored or via DOI) before building on it.


==============================================================================
# The Reversal Interest Rate: Abadi, Brunnermeier & Koby (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/abadi-et-al-reversal-interest-rate-2023/
# Distilled: This paper theoretically characterizes the reversal interest rate, the policy rate below which further monetary easing becomes contractionary for bank lending. In a calibrated New Keynesian model with imperfectly competitive banks and net-worth constraints, the reversal rate is approximately -0.9 percent for aggregate investment and -1.4 percent for bank lending, calibrated to the euro area. American Economic Review 2023, paywalled. Six core results with source locators, the model equations, and the calibration method.
# Tags: paper-summary, macro, monetary-policy, banking, interest-rates, new-keynesian
==============================================================================

**What this is.** The paper's core results, the model it builds on (an infinite-horizon New Keynesian economy with imperfectly competitive banks and net-worth-constrained lending), and the key equations behind the reversal interest rate mechanism: enough to know what it found and how, without reading all 37 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1257/aer.20190150).

## TL;DR

The paper introduces and characterizes the reversal interest rate: the rate below which further monetary easing becomes contractionary for bank lending. The mechanism runs through two competing channels. An interest rate cut initially raises banks' capital gains on long-term bond holdings (the **capital gains channel**), boosting net worth and expanding lending. Simultaneously, lower rates compress banks' net interest income (the **NII channel**), reducing net worth over time. When banks face a binding capital constraint on lending, the NII effect eventually dominates, causing further cuts to shrink, rather than expand, credit supply. In a New Keynesian model calibrated to the euro area, the reversal rate is approximately -0.9 percent for aggregate investment and -1.4 percent for bank lending on impact. Extending rate cuts for prolonged periods ("low-for-long" policies) makes the NII losses cumulate, so policies that initially stimulate lending can eventually turn contractionary. The model also dampens the forward guidance puzzle: in the calibrated economy, an 8-quarter promise to hold rates at -1 percent generates investment and output peaks roughly half as large as in a standard New Keynesian model without bank capital constraints, because agents anticipate the long-run decline in lending. Contemporaneous models by Eggertsson et al. (2019) and Ulate (2021) study related rate-reversal mechanisms through deposit rate floors and bank profitability, but differ in channel and quantitative implications. The role of intermediary net worth in credit supply builds on Brunnermeier and Sannikov (2014); the bank optimization block follows the monopolistic banking firm of Klein (1971).

## Core results

Magnitudes are as reported; `\*` is not applicable (no hypothesis tests in this calibration paper). Locators point to the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Bank lending reversal rate on impact is approximately **-1.4 percent** in the euro area calibration | Figure 2 left, p. 2107; §III.C, p. 2106 | Below -1.4%, a further 10bp Taylor rule cut reduces bank lending at t=0; above -1.4%, the same cut expands lending |
| R2 | Aggregate investment reversal rate on impact is approximately **-0.9 percent**; higher than for bank lending because investment is forward-looking | Figure 2 right, p. 2107; §III.C, pp. 2106-2107 | Below -0.9%, a further 10bp cut reduces aggregate investment on impact; the investment reversal rate lies above the bank-lending reversal rate |
| R3 | A 10bp cut near the reversal rate reduces NII by approximately **6bp** (as a fraction of assets); near the steady state the same cut reduces NII by only about **1bp** | Figure 7 left, p. 2112; §IV.A, p. 2112 | NII effect is approximately 6x stronger at the reversal rate than near the steady state; consistent with Borio, Gambacorta, and Hoffmann (2017) |
| R4 | More deposit-dependent banks' total loans grow roughly **2pp less** than less deposit-dependent banks after a 30bp ECB rate cut in negative territory | Figure 7 right, p. 2112; §IV.A, pp. 2112-2113 | Model prediction lies within 1 SD of Heider, Saidi, and Schepens (2019) DiD estimate of -2pp for banks with 15pp higher deposit-to-asset ratio |
| R5 | **Forward guidance** (8 quarters at -1%) generates investment and output peaks roughly **half as large** as in a frictionless model without bank net-worth constraints | Figure 6, p. 2111; §III.E, pp. 2110-2111 | Frictionless economy reaches investment peak of about 0.87 log points; benchmark model reaches about 0.5 log points; investment reverses by the 8th quarter in the benchmark |
| R6 | The reversal rate for aggregate investment ranges from approximately **-1.8 to -0.5 percent** across the parameter sensitivity analysis, always exists within the calibrated range | Figure 8, p. 2115; §IV.B, pp. 2114-2116 | -1.8% when capital constraints are tight (3.5bp loan rate sensitivity to 25bp capitalization rise); -0.5% when constraints are loose (10.5bp sensitivity); benchmark -0.9% is near center |

**Overall (paper's conclusion).** The reversal interest rate exists under empirically grounded conditions on bank loan and deposit demand: net interest income falls when rates are sufficiently low, and bank lending is constrained by net worth. In the calibrated New Keynesian model for the euro area, the reversal rate lies close to -1 percent for aggregate investment. Its level depends primarily on the tightness of banks' capital constraints and the economy's reliance on bank credit. "Low-for-long" rate environments can initially appear stimulative while eventually becoming contractionary, and this dynamic dampens the power of forward guidance relative to standard New Keynesian models.

## Theory / model

The model is an infinite-horizon New Keynesian economy in discrete time, $$t \in \{0,1,2,\ldots\}$$. The main agents are: households; bank-dependent intermediate goods firms (fraction $$\xi = 0.998$$); non-bank-dependent intermediate goods firms (fraction $$1-\xi$$); a continuum of banks $$j \in [0,1]$$; capital goods producers; monopolistic retailers; and a central bank (Section I, pp. 2087-2092). Two key frictions distinguish the model from standard New Keynesian frameworks: banks have market power in setting deposit and loan rates, and bank lending is constrained by net worth.

**Households.** A household's lifetime utility is (equation 1, p. 2088):

$$
\sum_{t=0}^{\infty} \beta^t \left[ u(C_t, C_{t-1}, H_t) + \zeta \Phi(\mathcal{L}_t) \right] \tag{1}
$$

where $$C_t$$ is consumption, $$H_t$$ hours worked, $$\beta \in (0,1)$$ the discount factor, $$\mathcal{L}_t = \mathcal{L}(D_t, M_t)$$ aggregates real deposits $$D_t$$ and cash $$M_t$$, and $$\Phi$$ is utility from liquid savings with a satiation point $$\mathcal{L}^*$$. The parameter $$\zeta > 0$$ scales liquid-asset demand. The flow utility uses habit formation:

$$
u(C_t, C_{t-1}, H_t) = \frac{(C_t - hC_{t-1})^{1-\sigma}}{1-\sigma} - \chi \frac{H_t^{1+\varphi}}{1+\varphi}
$$

with $$h = 0.62$$, $$\sigma = 1$$ (IES), and $$\varphi = 2$$ (inverse Frisch elasticity; Table 1, p. 2102). Cash earns zero net return, so the presence of cash as a substitute for deposits implies that deposit demand becomes unresponsive to the policy rate when rates are sufficiently low (Lemma 1, p. 2096). This floors the bank's ability to pass rate cuts through to depositors, compressing margins.

**Bank balance sheet and net worth.** Banks extend loans $$L_t$$ and hold long-term government bonds $$B_t^L$$ at price $$Q_t^B$$, financed by deposits $$D_t$$ and net worth $$N_t$$. The balance sheet is (equation 4, p. 2089):

$$
L_t + Q_t^B B_t^L = D_t + N_t \tag{4}
$$

Long-term bonds mature with probability $$1/\tau$$ each period (expected maturity $$\tau$$; calibrated to 13.6 quarters to match average euro area bank bond holdings, p. 2104). The real bond price satisfies (equation 11, p. 2092):

$$
Q_t^B = \frac{1}{\tau} \sum_{s=0}^{\infty} \left(\prod_{r=0}^{s} \frac{1}{1+i_{t+r}}\right) \left(1 - \frac{1}{\tau}\right)^s, \qquad Q^{B*} = \frac{1}{1 + \tau i^*} \tag{11}
$$

Banks pay a fixed fraction $$\gamma$$ of net worth as dividends each period, so net worth accumulates according to (equation 5, p. 2090):

$$
N_{t+1} = (1-\gamma)\left[\frac{1+i_t}{1+\pi_{t+1}}Q_t^B B_t^L + \frac{1+i_t^L}{1+\pi_{t+1}}L_t(i_t^L) - \frac{1+i_t^D}{1+\pi_{t+1}}D_t(i_t^D, i_t) - \Psi^L(N_t, L_t) - \Psi^D(Q_t^B B_t^L, D_t)\right] \tag{5}
$$

where $$\Psi^L(N_t, L_t)$$ is the lending cost (homogeneous of degree one, decreasing in $$N_t$$, increasing in $$L_t$$) and $$\Psi^D$$ is the liquidity cost of deposit issuance. Lending is subject to a capital constraint $$L_t \leq \psi^L N_t$$: when this binds, credit supply is fully determined by net worth.

**Capital gains and NII channels.** An interest rate cut at $$t=0$$ affects bank net worth through two channels. Define $$N_t(N_0, i)$$ as net worth at time $$t$$ when the initial net worth is $$N_0$$ and the policy rate is cut to $$i$$. The total effect decomposes as (equation 17, p. 2097):

$$
\frac{dN_t}{di} = \underbrace{\frac{\partial N_t}{\partial N_0}\frac{dN_0}{di}}_{\text{Capital gains channel}} + \underbrace{\frac{\partial N_t}{\partial i}}_{\text{NII channel}} \tag{17}
$$

The capital gains channel captures the revaluation of existing bond holdings ($$dN_0/di < 0$$: a cut raises $$Q_0^B$$, boosting initial net worth). Lemma 3 (p. 2097) establishes that this channel weakens as the horizon $$t \to \infty$$ or as initial bond holdings $$B^{L*} \to 0$$: once long-term bonds mature, the initial capital gain no longer propagates. The NII channel captures the persistent compression of profit margins. Lemma 2 (p. 2097) shows that for $$i_t \leq \underline{i}$$, a further cut unambiguously reduces NII since deposit demand is unresponsive to the policy rate (cash dominates bonds as outside option). Lemma 4 (p. 2097) shows the NII channel depresses net worth at all future dates when rates are sufficiently low.

**Characterization of the reversal rate.** The reversal interest rate is defined formally in Definition 1 (p. 2093): the *time-t reversal rate* $$i_t^{RR}$$ is the highest interest rate such that bank lending $$L_t(i)$$ is increasing in $$i$$ for all $$i < i_t^{RR}$$. Proposition 1 (p. 2095) characterizes it: $$i$$ is $$i_t^{RR}$$ if (i) the capital constraint binds at all $$i' \leq i$$ and (ii) bank net worth is increasing in the interest rate, $$dN_t(i)/di > 0$$, for all $$i' < i$$. Proposition 2 (p. 2098) guarantees existence when initial bond holdings are small enough; Proposition 3 (p. 2098) guarantees existence when the horizon $$t$$ is long enough. Proposition 4 (p. 2098) establishes that $$i_{t+1}^{RR} \geq i_t^{RR}$$: the reversal rate is weakly increasing in the horizon, so an initially stimulative cut can become contractionary at longer horizons. Proposition 5 (p. 2099) formalizes the low-for-long result: for rates held below $$i^*$$ for long enough ($$T > \bar{T}$$), bank lending eventually falls below its steady-state level $$L^*$$.

## Method

**Solution method.** The quantitative model is solved using Dynare's full nonlinear algorithm rather than perturbation methods, since the model's binding capital constraints create large nonlinearities and non-monotonicities near the reversal rate (§III.A, p. 2101). The economy begins at its steady state; an unanticipated monetary policy shock is announced at $$t=0$$, and the economy evolves deterministically thereafter under perfect foresight.

**Bank's problem.** Banks maximize discounted dividends (equation 6, p. 2090):

$$
\max_{B_t^L,\, i_t^L,\, i_t^D} \sum_{t=0}^{\infty} \beta^t \Lambda_t \gamma N_t \quad \text{subject to (4) and (5)} \tag{6}
$$

where $$\Lambda_t$$ is the household's marginal utility of consumption. In the analytical section (Section II), holding loan and deposit demand curves at their steady-state values, the bank's problem reduces to maximizing net interest income period by period (equation 13, p. 2094):

$$
\text{NII}(N_t, i_t) = \max_{B_t^L,\, i_t^L,\, i_t^D}\; i_t Q_t^B B_t^L + i_t^L L^*(i_t^L) - i_t^D D^*(i_t^D, i_t) \tag{13}
$$

subject to the capital constraint $$L_t \leq \psi^L N_t$$ and liquidity constraint $$Q_t^B B_t^L \geq \psi^D D^*(i_t^D, i_t)$$.

**Optimal loan and deposit rates.** First-order conditions give (equations 14-15, p. 2094):

$$
i_t^L = i_t + \frac{1}{\varepsilon_t^L} + \lambda_t + \mu_t \tag{14}
$$

$$
i_t^D = i_t - \frac{1}{\varepsilon_t^D} + (1-\psi^D)\mu_t \tag{15}
$$

where $$\varepsilon_t^L > 0$$ and $$\varepsilon_t^D < 0$$ are the semielasticities of loan and deposit demand, $$\lambda_t \geq 0$$ is the Lagrange multiplier on the capital constraint, and $$\mu_t \geq 0$$ is the multiplier on the liquidity constraint. When constraints are slack, rates are set at the standard monopolistic markup/markdown over the policy rate. When the capital constraint binds, the loan spread widens, raising borrowing costs for bank-dependent firms and reducing their capital demand.

**Calibration.** The model is calibrated to the euro area; one period equals one quarter. Conventional DSGE parameters (Table 1, p. 2102) are set to standard values: Rotemberg price adjustment cost $$\theta = 70.7$$, capital share $$\alpha = 0.36$$, Taylor rule persistence $$\rho^{mp} = 0.93$$, and inflation coefficient $$\phi^\pi = 2.74$$ (from Coenen et al. 2019). Banking-sector parameters (Table 2, p. 2114) are calibrated to match euro area bank balance sheet moments: fraction of bank-dependent firms $$\xi = 0.998$$ (SME fraction in Eurostat); relative productivity $$A^b/A^{nb} = 0.43$$ (matching a 55.8% SME output share); loan elasticity $$\varepsilon^L = 200$$ (targeting a 2% loan spread from Freriks and Kakes 2021); deposit elasticity $$\varepsilon^D = -275$$ (targeting a 1% deposit spread from ECB MIR data); leverage cost $$\kappa^L = 0.017$$ (targeting 7bp loan rate increase per 25bp capitalization rise, from Macroeconomic Assessment Group 2010); bond maturity $$\tau = 13.6$$ quarters (matching average bank bond holding maturity from Hoffmann et al. 2019); and steady-state net worth-to-loan ratio $$N^*/L^* = 0.155$$ (from Altavilla, Boucinha, and Peydro 2018).

The monetary policy rule in the quantitative model is a Taylor rule with inertia (equation 19, p. 2105):

$$
\frac{1+i_t}{1+i^*} = \left(\frac{1+i_{t-1}}{1+i^*}\right)^{\rho^{mp}} \left(\frac{1+\pi_t}{1+\pi^*}\right)^{\phi^\pi(1-\rho^{mp})} \exp(\epsilon_t^{mp}) \tag{19}
$$

The benchmark results consider a single monetary shock $$\epsilon_0^{mp}$$ at $$t=0$$.

**Capital goods producers.** Capital goods producers solve (equation 8, p. 2091):

$$
\max_{I_{t+1}^z} \sum_{t=0}^{\infty} \beta^t \Lambda_t \left\{ Q_t^{K,z} I_t^z \left[1 - \Xi\!\left(\frac{I_{t+1}^z}{I_t^z}\right)\right] - I_{t+1}^z \right\} \tag{8}
$$

where $$\Xi(\cdot)$$ is a convex adjustment cost function (quadratic, $$\kappa^I = 5$$, implying an investment elasticity of 0.2 with respect to $$Q_t^{K,z}$$). This forward-looking structure explains why the reversal rate for aggregate investment (-0.9%) lies above the reversal rate for bank lending (-1.4%): investment internalizes the future deterioration of bank profits and declines before lending actually reverses.

## Empirical specifications

The quantitative analysis does not use panel regressions. Instead the paper compares **full nonlinear impulse response functions (IRFs)** across initial interest rate levels, using the Dynare solution (§III.A, p. 2101).

**Main results construction (R1, R2).** For each initial policy rate $$i \in \{-1.5\%, -1\%, 0\%, 1\%, 2\%\}$$, the model first computes the economy's response to a Taylor rule shock that reduces the time-0 rate to $$i$$, then adds a marginal 10bp Taylor rule innovation and computes a second IRF. The marginal IRF plotted in Figure 2 (p. 2107) is the difference between these two responses. The reversal rate is where this difference changes sign: at approximately $$i = -1.4\%$$ for bank lending and $$i = -0.9\%$$ for aggregate investment.

**NII validation (R3).** The left panel of Figure 7 (p. 2112) plots the model's marginal NII response (as a fraction of steady-state assets) to a 10bp cut as a function of the initial policy rate, overlaid on the empirical estimate from Borio, Gambacorta, and Hoffmann (2017). The model predicts approximately -1bp near the steady state and approximately -6bp near the reversal rate, consistent in sign and order of magnitude with that study.

**DiD comparison (R4).** To compare model predictions with micro evidence, the paper introduces hypothetical banks calibrated to a deposit-to-asset ratio 15 percentage points below the benchmark. These banks coexist in equilibrium but do not affect aggregate dynamics. Their marginal loan rates and lending quantities are computed for a -30bp Taylor rule shock. The right panel of Figure 7 (p. 2112) plots the log difference in lending at impact as a function of the initial policy rate, comparing the model with the DiD estimate from Heider, Saidi, and Schepens (2019), who use euro area syndicated loan data over June 2014 to December 2015. The model prediction of approximately -2pp is within one standard deviation of the Heider, Saidi, and Schepens (2019) estimate.

**Forward guidance (R5).** The central bank commits to holding the policy rate at -1 percent for $$T = 8$$ quarters, then reverts to the Taylor rule. The impulse responses are compared across the benchmark model and a frictionless version in which banks face no leverage costs ($$\kappa^L = 0$$). The frictionless economy mimics the dynamics of standard New Keynesian models: investment and output peaks are approximately twice those in the benchmark. In the benchmark, the anticipation of future bank-profit deterioration dampens the initial stimulus, and investment reverses by quarter 8 (Figure 6, p. 2111). Eggertsson and Woodford (2003) derive the standard forward guidance channel; the reversal rate mechanism here attenuates it.

**Sensitivity analysis (R6).** Figure 8 (p. 2115) varies $$\kappa^L$$ (the leverage cost parameter, reported as the response of loan rates to a 25bp capitalization target increase) between 3.5bp and 10.5bp and the bank-dependent output share $$A^b/A^{nb}$$ between 0.45 and 0.65. Across this range, the on-impact reversal rate for aggregate investment lies between -1.8 and -0.5 percent, remaining within a narrow band around the benchmark estimate of -0.9 percent.

## Datasets used

This is a theoretical and calibration paper. The quantitative results come from a Dynare solution of the calibrated structural model rather than direct econometric estimation. Calibration targets are drawn from published statistics and the literature:

| Source | Role in paper | Wiki page |
|---|---|---|
| ECB MFI statistical data | Deposit-to-GDP ratio target for $$\mathcal{L}^*$$ (satiation point); equity issuance-to-asset ratio target for $$\hat{N}$$ | No page yet |
| Eurostat (Structural Business Statistics) | Fraction of SMEs in the firm universe ($$\xi = 0.998$$) and SME share of output (55.8%) | No page yet |
| Altavilla, Boucinha, and Peydro (2018) | Tier-1 capitalization ratio target: 15.5% (steady-state $$N^*/L^*$$) | No page yet |
| Hoffmann et al. (2019) | Average bank bond holding maturity: 3.4 years ($$\tau$$) and loan-to-bond ratio ($$L^*/B^{L*}$$) | No page yet |
| ECB MIR database (Freriks and Kakes 2021) | Loan spread of 2% and deposit spread of 1% in steady state | No page yet |

Sample: euro area, calibrated to the period surrounding the ECB's introduction of negative interest rates in 2014. Model period: one quarter.

## When to read the full paper

Use the [original](https://doi.org/10.1257/aer.20190150) if you are: characterizing the effective lower bound on monetary policy when banking frictions are present; building or extending New Keynesian DSGE models with imperfectly competitive banks and balance-sheet constraints; studying the transmission of negative interest rate policy to bank lending and firm investment; evaluating the "low-for-long" forward guidance puzzle with financial intermediaries; or running the Dynare replication (code at [https://doi.org/10.3886/E188611V1](https://doi.org/10.3886/E188611V1)). The key tables and figures are: Figure 2 (reversal rates by initial rate level), Figure 6 (forward guidance dampening), Figure 7 (NII and DiD validation), and Figure 8 (sensitivity analysis).

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(8), 2023. Paywalled; no CC license. This distillation was extracted by an LLM on 2026-06-24 and is **not human-verified or independently reproduced**.

> Abadi, Joseph, Markus Brunnermeier, and Yann Koby. "The Reversal Interest Rate." *American Economic Review* 113, no. 8 (August 2023): 2084-2120. DOI: 10.1257/aer.20190150. Extract-only; all rights reserved by the American Economic Association.


==============================================================================
# Optimal Monetary Policy According to HANK: Acharya, Challe & Dogra (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/acharya-et-al-optimal-monetary-policy-according-2023/
# Distilled: In an analytically tractable HANK model with idiosyncratic income risk, optimal monetary policy places roughly twice as much weight on output stabilization relative to inflation as in RANK (calibrated Upsilon = 1.76 vs 1), adds the level of output to the target criterion (calibrated delta = 0.6), and tolerates inflation to cushion output declines after aggregate shocks. American Economic Review 2023, paywalled. Six core results with source locators, the CARA-normal HANK model, the LQ planning problem, and the HANK target criterion equations.
# Tags: paper-summary, macro, monetary-policy, inequality, hank, new-keynesian, peer-reviewed, unreplicated
==============================================================================

**What this is.** The paper's core results, the CARA-normal HANK model, the welfare-based LQ planning problem, and the HANK target criterion with its defining equations: enough to understand what it found and how, without reading all 42 pages. To replicate or extend the analysis, read the full source at [doi.org/10.1257/aer.20200239](https://doi.org/10.1257/aer.20200239).

## TL;DR

The paper derives optimal monetary policy analytically in a heterogeneous-agent New Keynesian (HANK) model where households face uninsurable idiosyncratic income risk. CARA preferences and normally distributed idiosyncratic shocks (from Acharya and Dogra (2020)) make the model analytically tractable: the economy aggregates linearly, and the planner's welfare can be expressed as a function of aggregate output and a scalar measure of consumption inequality. The key finding is that the HANK planner's loss function differs from the representative-agent (RANK) loss function of Galí (2015) and Woodford (2003) in two ways. First, the planner puts more weight on stabilizing economic activity relative to inflation (calibrated Upsilon = 1.76 vs 1 in RANK). Second, the planner also cares about stabilizing the *level* of output (not just the output gap), because output stabilization reduces consumption risk when income risk is countercyclical (calibrated delta = 0.6 < 1). Following productivity or markup shocks, the HANK planner therefore cushions output declines by accepting positive inflation on impact, in contrast to RANK's divine coincidence (zero inflation and zero output gap following productivity shocks). The analysis builds on the second-order LQ approximation approach of Benigno and Woodford (2005) to handle the inefficient HANK steady state. McKay and Wolf (2022) use a related LQ approach but their planner is not Pareto optimal.

## Core results

Magnitudes are as reported; \* = 5%, \*\* = 1%. Locators point into the source PDF. R3 and R4 magnitudes are approximate figure readings (Figures 4-5, pp. 1768, 1771). R5 is a qualitative irrelevance result.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **HANK loss function puts weight Upsilon(Omega) > 1 on output activity**, scaling up the weight on output relative to inflation compared to RANK | Proposition 3, eq. (33), Figure 3, p. 1763-1765 | In calibration: Upsilon = 1.76; relative weight on price stability roughly halved relative to RANK (epsilon/(kappa \* Upsilon) vs epsilon/kappa) |
| R2 | **HANK target criterion introduces output level**, weighted (1-delta) against the output gap (weighted delta), reducing weight on the price level relative to RANK | Proposition 4, eq. (36), p. 1766 | In calibration: delta = 0.6 (vs 1 in RANK); roughly equal weight on output level stabilization and output gap stabilization |
| R3 | **Following a fall in productivity, HANK planner prevents output from falling to the flexible-price level**, accepting positive inflation on impact; RANK achieves divine coincidence (zero inflation, zero output gap) | Proposition 5, Figure 4, pp. 1767-1768 | HANK: output gap approx. +0.2 pp and inflation approx. +0.025 pp at date 0; both reverse after period T. RANK: output gap = 0, pi = 0 for all t |
| R4 | **Following a positive markup shock, HANK allows a larger inflation increase and a smaller output decline than RANK**, because a large output decline increases consumption inequality | Proposition 6, Figure 5, pp. 1769-1771 | HANK: output approx. -0.2 pp vs RANK -0.3 pp at impact; inflation approx. +0.04 pp vs RANK +0.01 pp at impact |
| R5 | **Acyclical income risk (Omega = 0): HANK optimal policy = RANK optimal policy**, even though inequality exists, because its evolution does not depend on output | Lemma 2, p. 1766 | When Omega = 0: Upsilon = delta = 1; identical target criteria and optimal {yhat, pi} paths in HANK and RANK |
| R6 | **Unequally distributed profits add dividend stabilization to HANK loss function**; when only 10% of households hold dividends (stockholder fraction eta^d = 0.1), the planner raises output approx. 0.1 pp in response to a positive markup shock | Proposition 7, Figure 6, pp. 1773-1775 | K(eta^d) is increasing in profit concentration; eta^d = 0.1: output raised approx. +0.1 pp above equal-distribution baseline |

**Overall (paper's conclusion).** HANK differs from RANK because monetary policy can stabilize consumption inequality. When income risk is countercyclical (the empirically relevant case), the HANK planner puts some weight on stabilizing the level of output in addition to the output gap, tolerates higher inflation after adverse shocks, and implements interest rates that fall by less than in RANK. Extensions to unequal profit distribution and initial wealth inequality (URE channel via Auclert (2019)) reinforce the same qualitative conclusion: both motives lead optimal monetary policy to put more weight on output stabilization relative to RANK.

## Theory / model

The model is a Bewley-Huggett economy with a New Keynesian production side. Households follow a perpetual-youth (Blanchard-Yaari) life-cycle with constant per-period survival probability $$\vartheta$$. Population is normalized to 1. The date $$s$$ problem of household $$i$$ born at date $$s$$ is (p. 1746, eq. 1):

$$
\max_{\{c_t^s(i),\, l_t^s(i),\, a_{t+1}^s(i)\}} E_s \sum_{t=s}^\infty (\beta\vartheta)^{t-s}\, u\!\left(c_t^s(i),\, l_t^s(i);\, \xi_t^s(i)\right) \tag{1}
$$

subject to a budget constraint (simplified, eq. 5, p. 1748):

$$
c_t(i) + w_t l_t(i) + q_t a_{t+1}(i) = w_t \xi_t^s(i) + (1-\tau_t^a) a_t(i) + D_t - T_t, \tag{5}
$$

where $$w_t$$ is the posttax real wage, $$q_t = \vartheta/R_t$$ is the bond price, $$a_t(i)$$ is real actuarial bond holdings, $$D_t$$ are dividends, and $$T_t$$ are lump-sum taxes. Each household faces i.i.d. idiosyncratic disutility-of-labor shocks $$\xi_t^s(i) \sim N(\bar{\xi},\sigma_t^2)$$. Agents have CARA preferences (p. 1747, eq. 4):

$$
u(c, l; \xi) = -\frac{1}{\gamma} e^{-\gamma c} - \rho\, e^{\frac{1}{\rho}(l - \xi)}, \tag{4}
$$

with coefficient of absolute risk aversion $$\gamma$$ and Frisch labor supply parameter $$1/\rho$$. CARA utility with normal shocks enables linear aggregation: equilibrium individual consumption and labor supply are linear in demeaned cash-on-hand (Proposition 1, eqs. 15-16, p. 1750):

$$
c_t^s(i) = \mathcal{C}_t + \mu_t\, x_t^s(i), \qquad l_t^s(i) = \rho\ln w_t - \gamma\rho\, c_t^s(i) + \xi_t^s(i), \tag{15-16}
$$

where $$x_t^s(i) = (1-\tau_t^a)a_t^s(i) + w_t[\xi_t^s(i)-\bar{\xi}]$$ is demeaned cash-on-hand and $$\mu_t$$ is the marginal propensity to consume (MPC) out of cash-on-hand. Aggregate consumption $$\mathcal{C}_t$$ evolves as (eq. 17):

$$
\mathcal{C}_t = -\frac{1}{\gamma}\ln\beta R_t + \mathcal{C}_{t+1} - \frac{\gamma\mu_{t+1}^2 w_{t+1}^2 \sigma_{t+1}^2}{2}, \tag{17}
$$

where the last term is a precautionary savings motive absent in RANK. The MPC satisfies (eq. 18):

$$
\mu_t^{-1} = 1 + \gamma\rho w_t + \frac{\vartheta}{R_t}\,\mu_{t+1}^{-1}. \tag{18}
$$

Lower real interest rates reduce $$\mu_t$$, facilitating self-insurance (the self-insurance channel). Intermediate goods producers face Rotemberg quadratic price adjustment costs; the goods market clears at $$y_t = c_t$$. The nonlinear IS equation is (eq. 21, p. 1751):

$$
y_t = y_{t+1} - \frac{1}{\gamma}\ln\beta\!\left(\frac{1+i_t}{\Pi_{t+1}}\right) - \frac{\gamma}{2}\,\mu_{t+1}^2 w_{t+1}^2 \sigma_{t+1}^2. \tag{21}
$$

The Phillips curve is standard (linearized, eq. 30, p. 1760):

$$
\pi_t = \beta\pi_{t+1} + \kappa(\hat{y}_t - \hat{y}_t^e) + \frac{\varepsilon}{\Psi}\,\hat{\varepsilon}_t, \tag{30}
$$

where $$\kappa = \frac{\varepsilon}{\Psi}\frac{1+\gamma\rho}{\rho/y}$$ and $$\hat{y}_t^e = \frac{1+\rho/y}{1+\gamma\rho}\hat{z}_t$$ is the flexible-price (productively efficient) level of output.

**Welfare decomposition.** The social welfare function is the sum of average household lifetime utilities. By Proposition 2 (eq. 26, p. 1754), the period $$t$$ felicity can be written as:

$$
U_t = u(c_t, n_t;\bar{\xi}) \times \Sigma_t, \tag{26}
$$

where $$\Sigma_t \geq 1$$ is the welfare cost of consumption inequality: it equals 1 under complete markets and exceeds 1 whenever consumption dispersion is positive (higher $$\Sigma_t$$ reduces welfare since $$u(\cdot) < 0$$). The dynamics of $$\Sigma_t$$ are (eq. 27, p. 1755):

$$
\ln\Sigma_t = \frac{\gamma^2}{2}\mu_t^2 w_t^2 \sigma_t^2 + \ln(1-\vartheta + \vartheta\Sigma_{t-1}). \tag{27}
$$

Consumption inequality is driven by within-period consumption risk $$\mu_t^2 w_t^2 \sigma_t^2$$ (idiosyncratic variance passed through the MPC) plus the accumulated effect of preexisting wealth inequality inherited from $$\Sigma_{t-1}$$.

**How monetary policy affects inequality.** Linearizing eq. 27 and using assumptions on the cyclicality of $$\sigma_t$$, the two channels by which monetary policy affects consumption risk become explicit (eq. 31, p. 1760):

$$
\hat{\Sigma}_t = \Lambda\hat{\mu}_t - \gamma y(\Theta-1)\hat{y}_t + \beta^{-1}\tilde{\beta}\,\hat{\Sigma}_{t-1},
$$

where $$\Lambda = \gamma^2\mu^2 w^2\sigma^2 > 0$$ and $$\Theta = 1 - \Lambda\varphi/\gamma$$ with $$\varphi = \partial\ln\sigma_t^2/\partial y_t$$ measuring income-risk cyclicality. The first term captures the self-insurance channel (lower $$\mu_t$$ reduces consumption risk); the second captures the income-risk channel (when risk is countercyclical, $$\Theta > 1$$, higher output lowers $$\Sigma_t$$). Lemma 1 (eq. 32, p. 1762) combines both channels into a single sufficient statistic $$\Omega$$, the cyclicality of consumption risk:

$$
\hat{\Sigma}_t = -\gamma y\,\Omega\!\left[\hat{y}_t - \varkappa(\Omega)\hat{y}_t^e\right] + \beta^{-1}\tilde{\beta}\,\hat{\Sigma}_{t-1}, \tag{32}
$$

where $$\Omega = \frac{\Lambda}{1-\Lambda} + \frac{\Theta-1}{1-\Lambda} \geq \Omega^c = \frac{\Lambda}{1-\Lambda} > 0$$ when risk is acyclical or countercyclical. When $$\Omega = 0$$ (acyclical risk and no self-insurance), monetary policy cannot affect consumption risk, and HANK optimal policy coincides with RANK (Lemma 2).

**The HANK planner's LQ problem.** The planning problem is to minimize the second-order approximation to social welfare losses over sequences $$\{\hat{y}_t, \pi_t\}_{t=0}^\infty$$, subject to the linearized Phillips curve (30). Proposition 3 (eq. 33, p. 1763) states:

$$
\min_{\{\hat{y}_t,\,\pi_t\}_{t=0}^\infty} \frac{1}{2}\sum_{t=0}^\infty \beta^t \left\{\Upsilon(\Omega)\left[\hat{y}_t - \delta(\Omega)\hat{y}_t^e\right]^2 + \frac{\varepsilon}{\kappa}\,\pi_t^2\right\}, \tag{33}
$$

where $$\Upsilon(\Omega) > 1$$ and $$\delta(\Omega) \in (0,1)$$ when $$\Omega \geq \Omega^c > 0$$ (acyclical or countercyclical income risk). In RANK ($$\sigma = 0 \Rightarrow \Omega = 0$$), $$\Upsilon = \delta = 1$$ and eq. (33) reduces to the standard RANK problem (eq. 34):

$$
\min_{\{\hat{y}_t,\,\pi_t\}_{t=0}^\infty} \frac{1}{2}\sum_{t=0}^\infty \beta^t \left\{(\hat{y}_t - \hat{y}_t^e)^2 + \frac{\varepsilon}{\kappa}\,\pi_t^2\right\}. \tag{34}
$$

The HANK loss function has two differences from RANK. First, the weight on economic activity is scaled by $$\Upsilon(\Omega) > 1$$, implying a lower relative weight on price stability. Second, the planner targets $$\hat{y}_t - \delta(\Omega)\hat{y}_t^e$$ rather than the output gap $$\hat{y}_t - \hat{y}_t^e$$: since $$\delta(\Omega) < 1$$, the planner aims to close the gap between $$\hat{y}_t$$ and $$\delta(\Omega)\hat{y}_t^e < \hat{y}_t^e$$, placing positive weight on stabilizing the *level* of output as well as the output gap.

**The HANK target criterion** (Proposition 4, eq. 36, p. 1766) for all $$t \geq 0$$:

$$
\left[1 - \delta(\Omega)\right]\hat{y}_t + \delta(\Omega)\!\left(\hat{y}_t - \hat{y}_t^e\right) + \frac{\varepsilon}{\Upsilon(\Omega)}\,\hat{p}_t = 0, \tag{36}
$$

where $$\hat{p}_t$$ is the log price level. In RANK the target criterion is $$(\hat{y}_t - \hat{y}_t^e) + \varepsilon\hat{p}_t = 0$$ (eq. 37), i.e., flexible price level targeting. The HANK criterion places weight $$(1-\delta)$$ on the output level $$\hat{y}_t$$, weight $$\delta$$ on the output gap, and a *lower* weight $$\varepsilon/\Upsilon < \varepsilon$$ on the price level.

## Method

The planner's problem is solved via the linear-quadratic (LQ) approach, following Benigno and Woodford (2005). A naive LQ approach (maximizing a quadratic approximation to welfare subject to linear constraints) does not yield first-order accurate approximations at an inefficient steady state, as is natural in HANK models with uninsurable income risk. To address this, the paper uses a second-order approximation of the constraints to eliminate first-order terms from the welfare approximation, yielding the LQ problem (33) that is first-order accurate. This mirrors Benigno and Woodford (2005)'s approach in RANK, generalized here to an economy where the steady state features consumption inequality ($$\Sigma > 1$$).

The linearized model consists of three main equations (p. 1753): the IS equation (23), MPC recursion (24), and the Phillips curve (25). Log-linearizing around the zero-inflation steady state and using eq. (20) to eliminate wages:

$$
\hat{y}_t = \Theta\hat{y}_{t+1} - \frac{1}{\gamma y}\!\left(\hat{i}_t - \pi_{t+1}\right) - \frac{\Lambda}{\gamma y}\,\hat{\mu}_{t+1}, \tag{23}
$$

$$
\hat{\mu}_t = -\gamma\mu wy\!\left[(1+\gamma\rho)\hat{y}_t - \hat{z}_t\right] + \tilde{\beta}\!\left(\hat{\mu}_{t+1} + \hat{i}_t - \pi_{t+1}\right), \tag{24}
$$

where $$\hat{i}_t = \ln(1+i_t) - \ln R$$, $$\Theta = 1 - \Lambda\varphi/\gamma$$, $$\Lambda = \gamma^2\mu^2 w^2\sigma^2$$, and $$\tilde{\beta} = \vartheta/R$$. These reduce to the standard RANK IS curve when $$\Lambda = 0$$ (no idiosyncratic risk). The Ramsey plan that solves problem (33) can be implemented by the following interest rate rule (eq. 38, p. 1770):

$$
i_t = i_t^\star + \phi\pi_t + \phi_{\text{gap}}(\Delta y_t - \Delta\hat{y}_t^e) + \phi_y\,\Delta y_t, \tag{38}
$$

where $$\phi_{\text{gap}} = \phi\frac{\Upsilon(\Omega)}{\varepsilon}\delta(\Omega)$$ and $$\phi_y = \phi\frac{\Upsilon(\Omega)}{\varepsilon}[1-\delta(\Omega)]$$. The HANK rule reacts more strongly to changes in output growth and the output gap (relative to $$\pi_t$$) than the RANK rule where $$\Upsilon = 1, \delta = 0$$.

## Empirical specifications

The paper's main results are analytical (propositions and closed-form expressions for $$\Upsilon$$ and $$\delta$$); calibration and impulse response functions (IRFs) are illustrative. Parameters are set to match US annual aggregate and micro targets:

- Annual frequency, real interest rate $$r = 4\%$$, steady-state output $$y = 1$$.
- Survival probability $$\vartheta = 0.85$$ (following Nistico 2016 and Farhi-Werning 2019).
- Standard deviation of income in steady state $$w\sigma(1-\gamma\rho\mu w) = 0.5$$, in line with Guvenen, Ozkan, and Song (2014).
- Cyclicality of income risk $$\varphi = -5.76$$, consistent with Storesletten, Telmer, and Yaron (2004) who find the standard deviation of (log) income rises from 0.12 in expansions to 0.21 in recessions.
- Phillips curve slope $$\kappa = 0.1$$, elasticity of substitution $$\varepsilon = 10$$ (10% steady-state markup).
- Coefficient of relative risk aversion $$\gamma c = \gamma$$ and Frisch elasticity $$\rho/y$$ calibrated so that median household CRA = 2 and Frisch = 1/3 in steady state.
- Persistence of shocks: $$\varrho_z = 0.95^4$$ (productivity) and $$\varrho_\varepsilon = 0.9^4$$ (markup), from Bayer, Born, and Luetticke (2020).
- Shock standard deviations: $$\sigma_z = 0.012$$ and $$\sigma_\varepsilon = 0.034$$ (Bayer, Born, and Luetticke 2020).

With these parameters, the baseline calibration yields $$\Omega = \Omega^c + (\Theta-1) > 0$$ (countercyclical consumption risk) and the welfare parameters $$\Upsilon = 1.76$$ and $$\delta = 0.6$$. IRFs are constructed as the one-standard-deviation impulse response to a date-0 productivity or markup shock, holding the initial wealth distribution at its Ramsey steady state. Figures 4 and 5 (pp. 1768-1770) compare HANK optimal policy (blue) to RANK optimal policy (red dashed) and to the non-optimal policy that sets $$\hat{y}_t = \hat{y}_t^e$$ and $$\pi_t = 0$$ (black dotted). Section V calibrates the unequal dividends extension with $$\mathcal{D}_y < 0$$ (negative output elasticity of dividends in the baseline) and stockholder fractions $$\eta^d \in \{0.1, 0.5, 1\}$$ (Figure 6, p. 1775).

## Datasets used

The paper is theoretical with calibrated parameters; no primary dataset is used directly. The calibration matches moments from the following published studies:

| Dataset / Source | Role in paper | Wiki page |
|---|---|---|
| Guvenen, Ozkan, and Song (2014), SIPP administrative records | Target for steady-state standard deviation of income (set to 0.5) | No page yet |
| Storesletten, Telmer, and Yaron (2004), PSID earnings data | Calibration target for cyclicality of income risk (phi = -5.76) | No page yet |
| Bayer, Born, and Luetticke (2020) HANK calibration | Shock persistence and standard deviation parameters | No page yet |

Replication code and data are available at [doi.org/10.3886/E184261V1](https://doi.org/10.3886/E184261V1).

## When to read the full paper

Use the [original](https://doi.org/10.1257/aer.20200239) if you are: deriving the formal welfare approximation and the exact expressions for $$\Upsilon(\Omega)$$ and $$\delta(\Omega)$$ (online Appendices E.2-E.3); studying extensions to hand-to-mouth households (Appendix H), persistent idiosyncratic risk (Appendix I), or demand shocks (Appendix J); checking the proof of divine coincidence breakdown (Proposition 5, Appendix F); or comparing the URE channel under the non-utilitarian planner to the baseline (Propositions 8-9, Appendix D.4). The locators above point to the key propositions and figures.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(7), July 2023. This distillation was extracted by an LLM on 2026-06-24 and is **not human-verified or independently reproduced**. The article is paywalled; this page contains extracted summaries only (extract-only redistribution).

> Acharya, Sushant, Edouard Challe, and Keshav Dogra.
> "Optimal Monetary Policy According to HANK."
> *American Economic Review* 113, no. 7 (July 2023): 1741-1782.
> DOI: 10.1257/aer.20200239.
> Replication data: doi.org/10.3886/E184261V1.
> All rights reserved, American Economic Association.
> This page is a distillation by the Institute for Automated Research (LLM-extracted, not human-verified, not reproduced).


==============================================================================
# Leaving School VA on the Table: Ainsworth, Dehejia, Pop-Eleches & Urquiola (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/ainsworth-et-al-households-leave-school-value-2023/
# Distilled: Romanian households leave roughly one standard deviation of school value added unexploited when choosing high school tracks; both incomplete information and preferences for curricular focus and peer quality contribute, with preferences explaining 83 percent of the gap that would remain after full information correction. An information RCT raises value added by 0.12 SD for low-achieving students (out of 1 SD potential); a rank-ordered logit and counterfactual simulation decompose the residual. American Economic Review 2023, AEA open access. Seven core results with source locators, datasets used, the model, and the method.
# Tags: paper-summary, school-choice, household-finance, information-economics
==============================================================================

**What this is.** The core results, the model, and the method from the paper in condensed form: enough to know what it found and how. To replicate or extend, read the full source at [doi.org/10.1257/aer.20210949](https://doi.org/10.1257/aer.20210949).

## TL;DR

Romanian households leave roughly one standard deviation of academic value added (VA) unexploited when choosing among high school tracks. Two candidate explanations are examined: households may lack information about which schools add the most academic value, or they may have genuine preferences for other school characteristics. The authors conduct a clustered information RCT - distributing VA rankings at baseline survey sessions - and find the treatment raises the VA of assigned tracks mostly for low-achieving students who were rejected by their top baseline choices (0.12 SD for all low-achieving students; 0.18 SD for the ineligible subgroup). A rank-ordered logit estimated on household preference rankings reveals that preferences for curricular focus and peer quality are far stronger than preferences for academic VA. Counterfactual simulations imply that fully correcting information would close only 17-25 percent of the VA gap for low-achieving students; preferences for other school traits account for 83 percent of the remaining gap.

## Core results

Magnitudes and significance are as reported; \* / \*\* / \*\*\* = 10% / 5% / 1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Households choose tracks at the 67th percentile of VA in their feasible sets; the mean potential gain from switching to the highest-VA option is 1 standard deviation | Table 4, Panel B-C, p. 1062 | Mean percentile rank = 67.1 (all students); potential increase = 1.01 SD = 12 pp baccalaureate pass probability (2019 SD) |
| R2 | VA and selectivity are positively correlated for less- and moderately-selective tracks but negatively correlated for the most selective third | Table 3, Figure 1, p. 1060 | Overall correlation = 0.562 (SE 0.005); most selective third: coefficient = -0.243 (SE 0.024) |
| R3 | Households' VA beliefs are substantially inaccurate: scores are off by 1.1 within-town quintiles on average and explain only 17 percent of the variation in true VA | Tables 5-6, pp. 1063-1064 | Mean absolute error = 1.13 quintiles; R-squared on true VA quintile = 0.17 (VA); 0.33 (selectivity) |
| R4 | The information treatment raises the VA of assigned tracks by 0.05 SD for all students (10%) and 0.12 SD for low-achieving students (5%); no significant effect for high-achieving students | Table 7, p. 1066 | All: 0.048 SD (SE 0.025)\*; low-achieving: 0.121 SD (SE 0.049)\*\*; high-achieving: -0.002 SD (insig.) |
| R5 | Among low-achieving students ineligible for their two top baseline choices, the treatment raises VA by 0.18 SD = 2.21 pp baccalaureate probability (1%) | Table 8, p. 1067 | 0.184 SD (SE 0.065)\*\*\* for the ineligible-for-top-two subgroup; near-zero for those admitted to their most-preferred baseline choice |
| R6 | Fully correcting beliefs raises VA by 0.13-0.20 SD for low-achievers (17-25% of potential) and 0.10-0.23 SD for high-achievers (11-24%), across four model specifications | Table 14, p. 1078 | Change in VA: 0.13-0.20 SD (low-achieving); 0.10-0.23 SD (high-achieving); share of potential increase 17-25% (low) and 11-24% (high) |
| R7 | Households have much stronger preferences for curricular focus (beta = 0.93) than for academic VA (beta = 0.34) or peer quality (beta = 0.34); 83 percent of unexploited VA under accurate beliefs is due to preferences | Table 13, col. 1, p. 1074; Table 15, col. 5, p. 1079 | All rank-ordered logit coefficients sig. at 1%; R-squared = 0.33; 83% of VA left on table under accurate beliefs is attributable to preferences for curricular focus, peer quality, and other traits |

**Overall (paper's conclusion).** Both information and preferences play a role in explaining why households leave VA on the table. Providing information raises the VA of tracks for some students, but only for low-achieving students who were rejected by their top choices. Even fully correcting information would leave most of the VA gap in place because households have strong preferences for curricular focus and peer quality that cause them to forego schools with high VA. Simply making VA information available is unlikely to close the gap.

## Theory / model

The paper has no formal equilibrium model; it investigates two candidate explanations empirically. The preference structure is specified as a linear expected utility function over quality scores (p. 1073, eq. 4):

$$
U_{ij} = \sum_{q} \beta_q \cdot s_{ij}^q + \epsilon_{ij} \tag{4}
$$

where $$U_{ij}$$ is household $$i$$'s expected utility from track $$j$$, $$s_{ij}^q$$ is the household's baseline survey score for track $$j$$ on quality dimension $$q$$ (location, peer quality, VA on the baccalaureate, curricular focus, siblings and friends; scale 1 to 5), $$\beta_q$$ is the preference weight, and $$\epsilon_{ij}$$ follows a Type I extreme value distribution. The model assumes households rank tracks by expected utility and that the serial dictatorship mechanism is incentive compatible, so submitted rankings truthfully reveal preferences.

Value added $$V_{jt}$$ is a track-year effect on the probability of passing the baccalaureate exam, estimated via selection-on-observables (Rothstein 2010). The measures are validated against regression discontinuity estimates at track admissions cutoffs following Angrist, Hull, Pathak, and Walters (2017): school-specific RD cutoffs (the minimum transition score for admission) generate quasi-random variation in track attendance, and the paper shows that all VA measures closely match the resulting causal estimates. For cohorts without baccalaureate data (2015-2017, 2019), VA is extended via local linear forests (Athey et al. 2019), which explain almost 80 percent of the variation in true VA out of sample.

The counterfactual comparison asks what track choices would look like if households had accurate VA beliefs. Two predicted VA values are compared for each student:

- $$V_{i,IS}$$: weighted-average VA across the feasible set using the preference model with inaccurate (baseline survey) scores for VA
- $$V_{i,AS}$$: the same but replacing inaccurate VA scores with within-town quintiles of measured VA

The difference $$V_{i,AS} - V_{i,IS}$$ estimates the effect of accurate beliefs, holding preferences constant.

**Identification.** Sections II-III are purely descriptive (selection-on-observables): they document the VA gap and the accuracy of households' beliefs without claiming causal identification. Section IV is identified by the clustered RCT (randomization at the middle school level, within matched pairs). Section V (preference estimation) is identified by within-town variation in quality scores and track choice under the rank-ordered logit.

## Method

Three methodological components produce the paper's results.

**Information experiment.** Middle schools were assigned to treatment or control by a matched-pair clustered randomization. At the end of the baseline survey session, all schools received a flyer with links to government admissions websites; treatment schools additionally received a ranking of the town's high school tracks by VA. The main treatment effect equation (eq. 1, p. 1065) is:

$$
\text{std}(V_i) = \eta_0 + \eta_1 \cdot T_i + \eta_X' \cdot \mathbf{X}_i + \eta_i \tag{1}
$$

where $$\text{std}(V_i)$$ is the standardized VA of student $$i$$'s assigned track, $$T_i$$ is the treatment indicator, and $$\mathbf{X}_i$$ includes the VA of the track to which the student would have been assigned based on the baseline preference ranking (equal to the VA of the highest-ranked feasible track, set to zero if no feasible track was ranked) and an indicator for whether the student ranked a feasible track. Standard errors are clustered by middle school treatment-control pairs (78 clusters).

The treatment effect on the accuracy of VA beliefs is estimated via:

$$
\left|\text{quint}(V_{jt}) - s_{ij,\text{fs}}^V\right| = \eta_0 + \eta_1 \cdot T_i + \eta_X' \cdot \mathbf{X}_{ij} + \eta_{ij} \tag{2}
$$

where $$\text{quint}(V_{jt})$$ is the within-town quintile of measured VA and $$s_{ij,\text{fs}}^V$$ is the follow-up-survey quality score for track $$j$$ (p. 1069, eq. 2). The effect of treatment on the association between preference ranks and VA is estimated via:

$$
\text{ppr}_{ij,\text{fs}} = \bigl(\delta_1 + \delta_2 \cdot T_i\bigr) \cdot \text{pr}(V_{jt}) + \bigl(\delta_{X,1} + \delta_{X,2} \cdot T_i\bigr)' \cdot \mathbf{X}_{ij} + \delta_{ij} \tag{3}
$$

where $$\text{ppr}_{ij,\text{fs}}$$ is the follow-up-survey percentile preference rank of track $$j$$ and $$\text{pr}(V_{jt})$$ is the within-town VA percentile rank (p. 1070, eq. 3). The coefficient of interest is $$\delta_2$$, which measures how treatment changed the association between VA and preference ranks.

**Rank-ordered logit for preference estimation.** Building on Fack, Grenet, and He (2019) for preference estimation in centralized choice mechanisms, and the heterogeneous-beliefs framework of Kapor, Neilson, and Zimmerman (2020), the preference weights $$\beta_q$$ are estimated by maximizing the log-likelihood of the top-two baseline track choices (eq. 5, p. 1073):

$$
\Pr\!\left(r_{i1}, r_{i2} \mid \mathcal{J}_i, s_{ij}^q\right) = \prod_{l=1}^{2} \frac{\exp\!\left(\sum_q \hat{\beta}_q \cdot s_{il}^q\right)}{\sum_{b \in \mathcal{J}_i \setminus \{r_{i,m:m<l}\}} \exp\!\left(\sum_q \hat{\beta}_q \cdot s_{ib}^q\right)} \tag{5}
$$

Here $$\mathcal{J}_i$$ is the set of all tracks in household $$i$$'s town (the feasible choice set) and the product runs over the first two ranked choices. The denominator at stage $$l$$ excludes all tracks already ranked. Missing quality scores are imputed via a random forest. Standard errors are clustered by middle school.

This page's `proposedVocab` stages `rank-ordered-logit` as a `builds-from` term (distinct from the fixed-effects `conditional-logit` in the registry), and `multidimensional-preferences` as a mechanism term for the preference-constraint channel. The method builds on `randomized-survey-experiment`, `panel-regression`, and `regression-discontinuity-design` from the registry.

## Empirical specifications

**VA-selectivity correlation (R2).** OLS of standardized $$V_{jt}$$ on standardized minimum transition score $$\text{MTS}_{jt}$$, weighted by student count, SE clustered by town-year. Table 3 also presents split-sample regressions by tercile of selectivity to capture the nonlinearity in Figure 1 (2,162,736 students, 5,969 town-years, all towns; 424,508 students, 720 town-years, survey towns).

**Household choice patterns (R1, R3).** Table 4 computes for each student the percentile rank of their chosen track among feasible options (by VA and by selectivity), and the standard deviation gain available from switching to the highest-VA option. Table 5 computes the mean absolute error between household quality scores and within-town quintiles of true VA. Table 6 regresses true quintiles on household scores; SE clustered by middle school (17,460 student-tracks, 188 clusters).

**Treatment effects on track VA (R4, R5).** Regression (1) on the experimental sample (2,692 students, 78 clusters). Subgroup regressions split by achievement (above/below national transition-score median) and by eligibility for the two highest-ranked baseline choices. The treatment is intent-to-treat (flyer distribution at the session); 85 percent of students were assigned to a track.

**Treatment effects on beliefs and rankings.** Regression (2) on 1,525 students with follow-up survey scores; regression (3) on 1,533 students. Both cluster SE by treatment-control pairs. Columns in Table 11 (beliefs) and Table 12 (rankings) distinguish tracks by position in the baseline preference ranking, revealing that effects exist only for tracks other than the household's two top baseline choices.

**Discrete choice preference estimates (R7).** Rank-ordered logit (eq. 5) on 1,170 students in experimental middle schools, using baseline survey scores. Column (1) of Table 13 includes five quality dimensions; columns (2)-(5) explore alternative VA dimensions; column (5) includes all five simultaneously. SE clustered by middle school (150 clusters).

**Counterfactual simulation (R6).** For each student in the feasible set, the model computes the weighted-average VA under inaccurate scores ($$V_{i,IS}$$) and accurate VA scores ($$V_{i,AS}$$):

$$
V_{i,IS} = \sum_{j \in \mathcal{J}_i^c} \text{std}(V_{jt}) \cdot \frac{\exp\!\left(\sum_q \hat{\beta}_q \cdot \bar{s}_{ij}^q\right)}{\sum_{k \in \mathcal{J}_i^c} \exp\!\left(\sum_q \hat{\beta}_q \cdot \bar{s}_{ik}^q\right)}
$$

(Section VB, p. 1076; $$\bar{s}_{ij}^q$$ denotes inaccurate VA scores and baseline scores for other dimensions; $$\mathcal{J}_i^c$$ excludes tracks newly created in 2019 for which no beliefs were elicited). Table 14 presents the mean difference $$V_{i,AS} - V_{i,IS}$$ across four model specifications. Table 15 decomposes residual unexploited VA by zeroing out selected $$\hat{\beta}_q$$ to isolate the contribution of preferences for each dimension.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Romanian Ministry of Education administrative microdata (2004-2017, 2019 cohorts) | Track-year VA estimation; student demographics, transition scores, middle school GPA, assigned tracks, baccalaureate performance; 2,162,736 students | no page yet |
| Baseline survey of parents (194 middle schools, 48 towns, 3,898 students, 2019) | Households' quality scores for tracks on eight dimensions, intended preference rankings, student characteristics; primary data collected by the authors | no page yet (primary data, collected by authors) |
| Endline (follow-up) survey of parents (2019) | Final submitted preference rankings; post-intervention quality scores for tracks | no page yet (primary data, collected by authors) |

Sample: baccalaureate outcomes available for 2004-2014 cohorts; machine-learning VA extension covers 2015-2017 and 2019. Survey and experiment conducted in 2019 in 48 towns selected for moderate size (7-28 tracks) and VA forecastability. Experimental sample (after removing schools that withdrew permission): 2,692 assigned students in 170 middle schools, 45 towns.

## When to read the full paper

Read the [source](https://doi.org/10.1257/aer.20210949) if you are: (i) estimating or validating school value added in a serial dictatorship setting, including the RD validation approach following Angrist, Hull, Pathak, and Walters (2017) (Sections I-II); (ii) studying how information provision affects school or college choices and want a rigorous mechanism decomposition separating information from preferences, building on Hastings and Weinstein (2008) (Section IV); (iii) estimating rank-ordered logit preference weights for multidimensional product quality in a centralized choice mechanism, extending Fack, Grenet, and He (2019) (Section V); or (iv) running simulation-based counterfactuals to separate information and preference constraints on household choices, in the spirit of Kapor, Neilson, and Zimmerman (2020). The Tables 13-15 combination (preference estimation, simulation, and decomposition by quality dimension) is particularly citable. The prior question of whether parents value school effectiveness is studied for New York City by Abdulkadiroğlu, Pathak, and Walters (2020); this paper provides complementary evidence for Romania with the additional advantage of observing the full feasible choice set.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(4), April 2023. AEA open access policy applies to 2023 publications; specific Creative Commons variant not confirmed from Crossref (license field returned empty) or PDF in this session. This distillation was extracted by an LLM (claude-sonnet-4-6) on 2026-06-24 and is **not human-verified or independently reproduced**. Replication data: [doi.org/10.3886/E181263V1](https://doi.org/10.3886/E181263V1).

> Ainsworth, Robert, Rajeev Dehejia, Cristian Pop-Eleches, and Miguel Urquiola.
> "Why Do Households Leave School Value Added on the Table? The Roles of Information and Preferences."
> *American Economic Review* 113, no. 4 (April 2023): 1049-1082.
> DOI: 10.1257/aer.20210949.
> Distilled extract only; redistribution of the verbatim PDF not authorized in this session.


==============================================================================
# Mobility and Congestion in Urban India: Akbar, Couture, Duranton & Storeygard (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/akbar-et-al-mobility-congestion-urban-india-2023/
# Distilled: Develops city-level vehicular speed indices decomposable into uncongested speed and a congestion factor, applied to 57 million simulated Google Maps trips in 180 Indian cities. Uncongested speed explains 70 percent of cross-city speed variance; congestion explains only 13 percent, overturning the view that slow Indian cities are primarily congested. American Economic Review 2023, paywalled. Seven core results with source locators, datasets used, the index methodology, and the empirical specifications.
# Tags: paper-summary, urban-economics, transportation, congestion, india, developing-countries
==============================================================================

**What this is.** The paper's core results, the speed-index methodology (equations and decomposition), and the empirical specifications: enough to understand what was found and how, without reading all 29 pages. To replicate or extend, read the original at [doi.org/10.1257/aer.20181662](https://doi.org/10.1257/aer.20181662) or download the replication archive at [doi.org/10.3886/E182681V1](https://doi.org/10.3886/E182681V1).

## TL;DR

The paper develops a methodology to estimate city-level vehicular speed indices from 57 million simulated Google Maps trips in 180 large Indian cities (population above 300,000 as of 2018, data collected June to November 2019). The headline index is exactly decomposable into an uncongested speed component and a congestion factor. Across the 180 cities, uncongested speed explains 70 percent of the cross-city variance in overall speed; the congestion factor explains only 13 percent. This holds even at peak hours (6-8 PM), where uncongested speed still accounts for 57 percent of variance versus 26 percent for congestion. Slower Indian cities are slower at all hours, not primarily because of traffic. Population density is the dominant correlate of lower speed (working mostly through uncongested speed), while more major roads, a gridded network, and street lighting increase uncongested speed. A hill-shaped relationship between city income and speed reflects two opposing forces: higher income raises uncongested speed through better infrastructure but raises congestion in the upper half of the income distribution. Indian city trips are 70 percent slower on average than comparable US trips.

## Core results

Magnitudes are as reported; SE in parentheses. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Uncongested speed explains 70% of cross-city speed variance**; congestion explains only 13% (all trips, 180 cities) | Table 4, p. 1102 | Variance shares: uncongested = 0.701, congestion = 0.126, covariance = -0.086 |
| R2 | **At high-peak hours (6-8 PM), uncongested speed still explains 57% of variance** vs 26% for congestion | Table 4, p. 1102 | Variance shares, high-peak: uncongested = 0.567, congestion = 0.259 |
| R3 | **Population elasticity of city speed = -0.15**; city area elasticity = +0.17 (density is the binding constraint) | Table 5, col. 1, p. 1104 | log population: -0.15 (0.016); log area: 0.17 (0.017); R\*\*2 = 0.47 |
| R4 | **More major roads increase speed via uncongested speed, not congestion** (consistent with the fundamental law) | Table 5, cols. 2, 5, and 8, p. 1104 | log major roads: 0.069 (0.016) on speed; 0.077 (0.016) on uncongested speed; ~0 on congestion factor |
| R5 | **Hill-shaped income-speed relationship**: speed rises with city earnings up to the 8th decile, then falls | Table 5, col. 3, p. 1104 | earnings: 0.028 (0.0088); earnings\*\*2: -0.0021 (0.00053); turning point at 8th decile |
| R6 | **Indian city trips are 70% slower than US city trips** on average; within-country decile spread is larger in India (36%) than the US (25%) | p. 1108 | US mean trip speed ~70% above India mean; US trip elasticity on trip length is 25-50% larger |
| R7 | **Welfare gain from 10% uncongested speed improvement** far exceeds gains from optimal congestion pricing | p. 1102-1103 | Population-weighted average gain ~Rs 1,157 (~$16) per vehicle commuter per year; Rs 2,696 (~$38) in Delhi |

**Overall (paper's conclusion).** Slow Indian cities are slow at all times of day because of low uncongested speed, not primarily because of traffic congestion. Policy interventions that target uncongested speed (road quality, network design, street lighting) generate substantially larger welfare gains than congestion pricing or ride-sharing promotion. Prior welfare estimates of optimal congestion pricing -- for example Kreindler (2018) for Bangalore and Akbar and Duranton (2018) for Bogota -- are below one percent of travel costs, far below uncongested-speed gains; Brownstone and Small (2005) provide the travel-time reliability valuation framework used to contextualize the welfare calculations. The finding challenges popular reports (e.g., Chin et al. 2018) that characterize certain Indian cities as exceptionally congested: Kolkata is in fact among the least congested but is the second slowest city due to its low uncongested speed.

## Theory / model

The paper has no formal structural economic model. It conceptualizes travel as a consumption problem in which travelers select trips, and city-level speed serves as a price index for the cost of a typical trip. The conceptual framework identifies two components of that price: uncongested speed (the inherent ability of the road network to move vehicles in the absence of traffic) and a congestion factor (the additional delay imposed by other vehicles at peak times).

**Key conceptual result (decomposition identity, p. 1095, equation 3).** Because uncongested and congested speed are defined over identical trip lengths, the following exact additive decomposition holds at the city level:

$$
\hat{f}_c^{fe} = \hat{m}_c^{fe} - \hat{s}_c^{fe} \tag{3}
$$

where $$\hat{s}_c^{fe}$$ is the estimated speed index, $$\hat{m}_c^{fe}$$ is the uncongested speed index, and $$\hat{f}_c^{fe}$$ is the congestion factor. All three are city fixed effects from separate OLS regressions run on the same sample with the same covariates; they therefore add up algebraically, enabling exact variance decomposition.

**Identification.** The paper does not claim causal identification. City-speed indices are estimated by conditioning on trip characteristics (length, distance to center, time of day, day of week, weather, trip type, road class), so the city fixed effect captures the component of speed attributable to the city after holding these trip-level factors constant. Cross-city correlates of the indices (population, roads, income, topography) are interpreted as associations, not causal effects.

## Method

The method has two stages: constructing a comparable speed index per city, and decomposing it into uncongested speed and congestion.

**Stage 1: City speed index (pp. 1093-1094).** The naïve approach, a length-weighted average speed for city $$c$$:

$$
S_c^m = \frac{\sum_{i \in c} D_i}{\sum_{i \in c} T_i} \tag{1}
$$

is not comparable across cities because trip length and distance to the city center vary systematically. The paper instead estimates a log-linear regression of trip speed on city fixed effects and a vector of trip characteristics $$\mathbf{X}_i$$ (equation 2, p. 1093):

$$
\log S_i = \alpha \mathbf{X}_i' + s_{c(i)}^{fe} + \epsilon_i \tag{2}
$$

where $$S_i = D_i / T_i$$ is the speed of trip instance $$i$$, $$c(i)$$ is its city, and $$s_c^{fe}$$ is the city fixed effect used as the speed index. The index $$\hat{S}_c^{fe} = \exp(\hat{s}_c^{fe} + \hat{\phi}^2/2)$$ is a predicted speed for a typical comparable trip in city $$c$$. The same regression is re-estimated twice more: once with log uncongested speed ($$S_i^{nt} = D_i / T_i^{nt}$$, where $$T_i^{nt}$$ is GM's no-traffic duration) as the dependent variable to obtain $$\hat{m}_c^{fe}$$, and once with the log congestion delay $$\log T_i - \log T_i^{nt}$$ to obtain $$\hat{f}_c^{fe}$$. The additive identity in equation (3) follows immediately. This builds on the panel-regression tradition and adapts the price-index approach of Couture, Duranton, and Turner (2018) to a developing-country setting. The finding that more major roads raise uncongested speed without reducing the congestion factor is consistent with the fundamental law of road congestion advanced by Duranton and Turner (2011): new road capacity attracts new traffic and leaves congestion unchanged in aggregate.

**Stage 2: Reliability.** Trip-time unreliability is measured as the ratio of the 90th to 50th percentile of the travel-time distribution across different weekday instances of the same trip, after conditioning on city-specific weekday effects (p. 1095). Unreliability is then used as a dependent variable in a variant of equation (2).

## Empirical specifications

**Benchmark trip-level regression (Table 2, p. 1097).** The core specification regresses log trip speed on city fixed effects, day and time of day effects (30-minute bins), log trip length, log distance to city center, and weather controls:

$$
\log S_i = \alpha_1 \log D_i + \alpha_2 \log \text{Dist}_{ci} + \text{time}_{t(i)} + \text{day}_{d(i)} + \text{type}_{k(i)} + \text{weather}_{w(i)} + s_{c(i)}^{fe} + \epsilon_i
$$

OLS, N = 57,103,181 trip instances (all trips) or 41,249,209 (Intents-weighted weekday trips), 180 cities. Standard errors are robust. Column (1) without weather controls yields trip-length elasticity 0.22 (SE 0.0046); adding distance to center (column 3) raises the distance elasticity to 0.078. Columns (5-7) add route-level controls (gradient, road class shares, intersections, turns against traffic, establishment density) to obtain "narrow" city fixed effects that isolate indirect effects.

**City-level correlates regression (Table 5, p. 1104).** The extracted city fixed effects from equation (2) are regressed on city characteristics with 180 city-level observations and a constant:

$$
\hat{s}_c^{fe} = \beta_1 \log \text{Pop}_c + \beta_2 \log \text{Area}_c + \beta_3 \text{Geography}_c + \beta_4 \text{Roads}_c + \beta_5 \text{Earnings}_c + \beta_6 \text{Earnings}_c^2 + u_c
$$

The same specification is estimated three times (columns 1-3 for speed index, 4-6 for uncongested speed, 7-9 for congestion factor). Coefficients in columns 1-3 equal the analogous coefficients in columns 4-6 minus those in columns 7-9 (from the decomposition identity). Geography variables include elevation variance, water length; roads include log major roads, log street lights, and a grid-conformity index (network shape). Robust standard errors throughout; R\*\*2 reaches 0.64 in the full speed-index specification. The same specification is run for US metro areas (online Appendix M) for cross-country comparison.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Google Maps simulated trips (57M+ instances) | Primary speed data: GM's recommended route duration with and without traffic for 2,730,969 unique origin-destination pairs sampled across 180 Indian cities; also 52M US instances for comparison | no page yet |
| OpenStreetMap via OSMnx | Road network classification (motorways, primary, secondary, tertiary, residential), intersection counts, turns against traffic, network shape (grid index) | no page yet |
| India Census 2011 | City population (sum of town/village level), car and motorcycle ownership shares, road inventory (paved/unpaved, street lights), commute mode shares, earnings from National Sample Survey 2011-2012 | no page yet |
| Global Human Settlement Layer (GHSL) | Urban extent and city boundary delineation; built-up area pixels used to define the 180-city sample | no page yet |
| Intents Mobi actual-trip app | Validation: 90,894 actual weekday trips in 89 cities by professional drivers, used to verify GM speed patterns and calibrate time-of-day weighting | no page yet |
| Meteostat | Weather conditions at time of each trip instance (rain, thunderstorms, wind, humidity, temperature), used as controls | no page yet |
| DMSP nightlight satellite data | Proxy for urban extent and street lighting; used alongside OSM street light data | no page yet |

Sample: 180 Indian cities with 2011 census population above 300,000 (after dropping one city with defective boundary data), June 5 to November 13, 2019. Each trip sampled across 21 instances spanning times of day and days of the week.

## When to read the full paper

Read the original at [doi.org/10.1257/aer.20181662](https://doi.org/10.1257/aer.20181662) if you are:
replicating or extending the speed-index methodology to other countries (the online appendices give full details on trip sampling, city boundary construction, and robustness variants including Laspeyres-type indices and discrete-choice models); studying urban transportation policy in developing countries (the paper discusses implications for congestion pricing, ride-sharing, and road investment); comparing Indian and US urban mobility in detail (Appendix M); or examining walking and transit in India (Appendix A). The replication archive at Akbar et al. (2023) (ICPSR, doi:10.3886/E182681V1) provides data and code.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(4), April 2023. This distillation was extracted by an LLM on 2026-06-24 and is **not human-verified or independently reproduced**. The verbatim PDF is under standard AEA copyright (all rights reserved); only this extract-only summary is provided here.

> Akbar, Prottoy, Victor Couture, Gilles Duranton, and Adam Storeygard.
> "Mobility and Congestion in Urban India."
> *American Economic Review* 113, no. 4 (April 2023): 1083-1111.
> DOI: 10.1257/aer.20181662.
> Replication data: doi:10.3886/E182681V1 (ICPSR).


==============================================================================
# Imperfect Financial Markets and Investment Inefficiencies: Albagli, Hellwig & Tsyvinski (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/albagli-et-al-imperfect-financial-markets-investment-2023/
# Distilled: noisy information aggregation in equity markets creates a rent-seeking motive for incumbent shareholders that causes overinvestment in upside risks and underinvestment in downside risks; in general equilibrium an externality through aggregate share prices dampens overinvestment but amplifies underinvestment. AER 2023, paywalled. Six core theoretical results with equation locators, the partial and general equilibrium models with full equations, and the information-feedback extension. LLM-distilled.
# Tags: paper-summary, asset-pricing, equities, corporate-finance, investment, theory, peer-reviewed, unreplicated
==============================================================================

**What this is.** The paper's core theoretical results, the partial and general equilibrium models with their defining equations, and the empirical-relevance discussion: enough to understand the mechanism and all six propositions without reading all 32 pages. To replicate or extend it, read the full article at [doi.org/10.1257/aer.20170725](https://doi.org/10.1257/aer.20170725).

## TL;DR

Incumbent shareholders who sell a fraction of their equity before dividends are realized have a rent-seeking motive: by distorting investment, they can move the market price in their favor. With upside risk in cash flows (positive return asymmetry), they overinvest to inflate expected prices; with downside risk, they underinvest to avoid price deflation. The magnitude of the distortion scales with three parameters: the percentage return wedge between market-implied and fundamental returns ($$\Delta$$), the fraction of shares traded ($$\alpha$$), and the inverse scalability of investment ($$\chi^{-1}$$). In general equilibrium, the shareholders' collective attempts to boost their firms' share prices lower aggregate dividends, creating an externality that dampens overinvestment with upside risk but amplifies underinvestment with downside risk. A corrective tax implements the efficient allocation in both settings.

## Core results

Propositions reference the paper's own numbering; locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Partial equilibrium investment distortion**: overinvestment for upside risk ($$\alpha\Delta > 0$$), underinvestment for downside risk ($$\alpha\Delta < 0$$); log distortion approximately $$\alpha\Delta\chi^{-1}$$ | Proposition 1, pp. 2333-2334 | $$\hat{K}/K^* = (1+\alpha\Delta)^{1/\chi}$$; dividend losses $$V(\hat{K})/V(K^*) = (1+\alpha\Delta)^{1/\chi}(1-\alpha\Delta\chi^{-1})$$ |
| R2 | **Negative expected dividends** arise when upside risk and high scalability combine: the firm overinvests so severely that it destroys value in expectation | Proposition 1(iv), eq. 4, p. 2333 | Condition: $$\alpha\Delta\chi^{-1} > 1$$ |
| R3 | **Partial equilibrium corrective tax** implements efficient investment by offsetting the return wedge | eq. 5, p. 2336 | $$\tau = 1 - \frac{1}{1+\alpha\Delta}$$ |
| R4 | **General equilibrium unique solution**; investment is lower than in partial equilibrium ($$K_{GE} < K_{PE}$$); with upside risk overinvestment is dampened ($$K_{PE} > K_{GE} > K^*$$); with downside risk underinvestment is amplified ($$K_{GE} < K_{PE} < K^*$$) | Proposition 2, p. 2342 | $$K_{GE}/K^* = \bigl(1 + \alpha\Delta \cdot Q/\hat{Q}\,/\,(1-\alpha+\alpha Q/\hat{Q})\bigr)^{1/\chi}$$ |
| R5 | **Limiting behavior** for highly scalable investments ($$\chi \to 0$$): upside distortions are bounded ($$K_{GE}/K^* \to e^\alpha$$); downside distortions are unbounded ($$K_{GE}/K^* \to 0$$ as investment collapses; GE surplus vanishes relative to PE surplus) | Proposition 3, p. 2342 | Upside: $$\lim_{\chi \to 0} V(K_{GE})/V(K^*) = (1-\alpha)e^\alpha < 1$$; Downside: $$\lim_{\chi \to 0} V(K_{GE})/V(K_{PE}) = 0$$ |
| R6 | **Information feedback** (price-contingent investment) merges both mechanisms: investment is positively correlated with prices (excess sensitivity to market signals), and higher investment predicts lower future equity returns | Proposition 4, pp. 2348-2349 | $$\text{cov}(\hat{K}(z), P(z)) > 0$$; $$\hat{K}(z)/K^*(z)$$ increasing in $$z$$; $$\text{cov}\bigl(\hat{K}(z),\,(V(z)-P(z))/P(z)\bigr) < 0$$ |

**Overall (paper's conclusion).** Even small departures from market efficiency can produce large aggregate investment distortions when investments are highly scalable, and these distortions are compounded by the price externality in general equilibrium. The paper provides a rationale for regulating financial risk-taking by publicly traded firms even when equity markets operate near efficiency.

## Theory / model

The model has three stages and is developed first in partial equilibrium (Section I) and then embedded in general equilibrium (Section II).

**Partial equilibrium baseline (§I.A, pp. 2326-2330).** A single firm. At stage 1, incumbent shareholders choose investment $$k \geq 0$$. At stage 2, they sell a fraction $$\alpha \in (0,1]$$ of shares to outside investors. At stage 3, dividends $$\Pi(\theta, k) \equiv R(\theta)k - C(k)$$ are paid to final shareholders, where $$\theta \sim \mathcal{N}(0, \lambda^{-1})$$ is a stochastic fundamental and $$C(k) = k^{1+\chi}/(1+\chi)$$ with $$\chi \geq 0$$. The parameter $$\chi^{-1}$$ captures the scalability of investment. Ex ante efficient investment $$K^*$$ maximizes $$E[\Pi(\theta,k)]$$.

At stage 2, informed investors (mass 1) each observe private signal $$x_i \sim \mathcal{N}(\theta, \beta^{-1})$$; noise traders place a random demand $$\alpha\Phi(u)$$ with $$u \sim \mathcal{N}(0, \delta^{-1})$$ independent of $$\theta$$. In the unique noisy rational expectations equilibrium, the sufficient statistic for the price signal is $$z \equiv \theta + (1/\sqrt{\beta}) \cdot u$$, and the Lemma (p. 2327) gives the market-clearing price (eq. 1):

$$P(z, k) = E\!\left[\Pi(\theta, k) \mid x = z, z\right] \tag{1}$$

The price equals the expected dividend of the marginal informed trader, who observes both private signal $$x = z$$ and the public signal embedded in the price (also $$z$$). This double-conditioning places excess weight on $$z$$ relative to its true precision as a public signal, generating a systematic bias: the market-implied prior is $$\mathcal{N}(0, \hat{\lambda}^{-1})$$ with $$\hat{\lambda}^{-1} > \lambda^{-1}$$, so prices overweight tail realizations of $$\theta$$. Denote $$\hat{E}[\cdot]$$ as the expectation under this market-implied prior.

At stage 1, incumbent shareholders maximize (eq. 2, p. 2329):

$$\max_{k \geq 0} \; E\!\left[\alpha P(z;k) + (1-\alpha)\Pi(\theta,k)\right] \tag{2}$$

$$= \max_{k \geq 0}\left\{ E[\Pi(\theta,k)] + \alpha E\!\left[P(z;k) - \Pi(\theta,k)\right] \right\}$$

The term $$\alpha E[P(z;k) - \Pi(\theta,k)]$$ is the rent accruing to incumbent shareholders from the price bias. In equilibrium, the distorted investment $$\hat{K}$$ satisfies (eq. 3, p. 2330):

$$C'(\hat{K}) = E[R(\theta)] + \alpha\!\left(\hat{E}[R(\theta)] - E[R(\theta)]\right) \tag{3}$$

Defining the return wedge $$\Delta \equiv \hat{E}[R(\theta)]/E[R(\theta)] - 1$$, the investment ratio is (p. 2332):

$$\frac{\hat{K}}{K^*} = (1 + \alpha\Delta)^{1/\chi}$$

When $$R(\cdot)$$ has upside risk (gains exceed losses at fixed distances from the mean), $$\hat{E}[R(\theta)] > E[R(\theta)]$$ so $$\Delta > 0$$ and $$\hat{K} > K^*$$ (overinvestment). When $$R(\cdot)$$ has downside risk, $$\Delta < 0$$ and $$\hat{K} < K^*$$ (underinvestment). For upside risk with high scalability, expected dividends can turn negative when $$\alpha\Delta\chi^{-1} > 1$$ (eq. 4, p. 2333) as the firm pursues negative-NPV overinvestment to capture rent.

The Grossman and Stiglitz (1980) noisy REE framework underpins the price characterization: prices aggregate private information but the market-clearing condition introduces a systematic bias that shareholders exploit through their investment decision.

**General equilibrium (§II, pp. 2337-2345).** A unit measure of firms indexed $$i$$, each with idiosyncratic fundamental $$\theta_i \sim \mathcal{N}(0, \lambda^{-1})$$. Incumbent shareholders sell an endogenous (and symmetric) fraction $$s$$ of shares. Final shareholders invest through mutual funds (acting as noise traders) and hedge funds (acquiring noisy private information about each firm). Let aggregate market value $$T = \int P_i \, di$$ and aggregate dividends $$V = \int \Pi_i \, di$$. With log preferences $$v_I(C^I_2) + u_I(C^I_3) = \alpha\ln C^I_2 + (1-\alpha)\ln C^I_3$$ for incumbent shareholders (ensuring $$s = \alpha$$ is exogenous), the aggregate intertemporal MRS satisfies (eq. 6, p. 2337):

$$\frac{T}{V} = Q^{-1} = \frac{u'_I((1-s)V)}{v'_I(sT)} = u'_F(sV) \tag{6}$$

The GE equity price for firm $$i$$ is (eq. 7, p. 2339):

$$P_i(z_i, k_i) = \frac{1}{\hat{Q}} \, E\!\left[\Pi(\theta_i, k_i) \mid x = z_i, z_i\right] \tag{7}$$

where $$\hat{Q}$$ is the threshold return on equity required by hedge funds in equilibrium. Aggregating across firms and combining with $$E[P_i] = T = VQ^{-1}$$, the equilibrium $$\hat{Q}$$ satisfies (eq. 8, p. 2340):

$$\hat{Q} = Q \cdot \frac{\hat{E}[\Pi(\theta,K)]}{E[\Pi(\theta,K)]} \tag{8}$$

The ratio $$Q/\hat{Q}$$ is the GE wedge that adjusts the PE rent-seeking incentive. Each firm's incumbents maximize (eq. 10, p. 2340):

$$\max_{k_i \geq 0} \left\{ \alpha \frac{Q}{\hat{Q}} \hat{E}[\Pi(\theta_i, k_i)] + (1-\alpha) E[\Pi(\theta_i, k_i)] \right\} \tag{10}$$

The GE investment ratio and the intertemporal wedge jointly satisfy (eqs. 11 and 13, pp. 2340-2341):

$$\frac{K_{GE}}{K^*} = \left(1 + \alpha\Delta \cdot \frac{Q/\hat{Q}}{1-\alpha+\alpha Q/\hat{Q}}\right)^{1/\chi} \tag{11}$$

$$\frac{Q}{\hat{Q}} = \frac{\chi + 1 - (K_{GE}/K^*)^\chi}{(1+\chi)(1+\Delta) - (K_{GE}/K^*)^\chi} \tag{13}$$

With upside risk ($$\Delta > 0$$), overinvestment by all firms lowers aggregate dividends $$V$$ and thus $$Q$$, making $$Q < \hat{Q}$$ (i.e. $$Q/\hat{Q} < 1$$): the GE wedge attenuates the PE overinvestment, so $$K_{PE} > K_{GE} > K^*$$. With downside risk ($$\Delta < 0$$), underinvestment raises $$Q > \hat{Q}$$: the GE wedge amplifies underinvestment, so $$K_{GE} < K_{PE} < K^*$$. This externality arises because individual shareholders do not internalize that their collective rent-seeking reduces aggregate dividends $$V$$, thereby lowering $$Q$$ and ultimately feeding back to distort the intertemporal incentive of all firms.

## Method

The model is solved analytically throughout. The solution strategy builds on `noisy-rational-expectations` for the price characterization and `dynamic-general-equilibrium` for the fixed-point analysis.

**Price characterization.** The noisy REE price (eq. 1) is derived using the Gaussian signal structure: the market-clearing condition with informed and noise traders yields the sufficient statistic $$z = \theta + (1/\sqrt{\beta})\cdot u$$ (Lemma, p. 2327). Compounding normal distributions implies $$E[P(z;k)] = \hat{E}[\Pi(\theta,k)]$$ under a market-implied prior with inflated variance $$\hat{\lambda}^{-1}$$ (p. 2330). This representation holds for general (nonlinear) payoff functions $$R(\cdot)$$, as shown in the companion paper Albagli, Hellwig and Tsyvinski (forthcoming), making the results robust to the specific return functional form.

**Partial equilibrium investment.** The FOC for investment (eq. 3) follows from differentiating eq. 2 and noting $$E[\partial P(z;k)/\partial k] = \hat{E}[R(\theta)]$$. The power cost structure $$C(k) = k^{1+\chi}/(1+\chi)$$ yields the closed-form investment ratio $$\hat{K}/K^* = (1+\alpha\Delta)^{1/\chi}$$. Comparative statics follow from first-order approximations around the efficient level $$K^*$$; the dividend-loss formula $$V(\hat{K})/V(K^*) = (1+\alpha\Delta)^{1/\chi}(1-\alpha\Delta\chi^{-1})$$ is derived by a second-order expansion of $$\ln(V(\hat{K})/V(K^*))$$ around zero (pp. 2332-2333).

**General equilibrium fixed point.** Proposition 2 (existence and uniqueness) is proved by showing that equations (11) and (13) have a unique solution $$(K_{GE}/K^*, Q/\hat{Q})$$ via continuity arguments and monotone comparative statics (Appendix, pp. 2351-2353). The limiting results in Proposition 3 follow by taking $$\chi \to 0$$ and establishing boundary behavior of the ratio using L'Hopital-type arguments.

**Corrective taxes.** In partial equilibrium, the tax $$\tau$$ on payoff $$R(\theta)k$$ shifts the effective return so that the FOC yields $$K^*$$ (eq. 5, p. 2336):

$$\tau = 1 - \frac{1}{1+\alpha\Delta} \tag{5}$$

In general equilibrium, accounting for the intertemporal wedge $$\hat{Q}/Q = 1 + (1+\chi^{-1})\Delta$$ at the efficient level, the GE corrective tax is (eq. 14, p. 2345):

$$\tau = 1 - \frac{1 - \alpha + \alpha Q/\hat{Q}}{1 - \alpha + \alpha (Q/\hat{Q})(1+\Delta)} \tag{14}$$

This adjusts the PE formula by a Pigouvian correction for the aggregate externality through share prices.

## Empirical specifications

Section III (pp. 2345-2350) studies the model's empirical implications. The paper does not conduct original regressions; it shows that the PE model nests the predictions of two empirical literatures and discusses qualitative consistency with existing estimates.

**Information feedback extension (§III.B, pp. 2346-2349).** The PE model is extended to allow price-contingent investment $$K(z)$$: shareholders commit to an investment rule that the market anticipates. With $$\alpha = 1$$ (full share turnover), shareholders choose $$\hat{K}(z)$$ to satisfy $$C'(\hat{K}(z)) = E[R(\theta)\mid x=z,z]$$, so the equilibrium investment function is (p. 2348):

$$\hat{K}(z) = \left[(1 + 1/\chi)\,P(z)\right]^{1/(1+\chi)}$$

Expected equity returns decrease in investment and price:

$$\frac{V(z)}{P(z)} - 1 = \frac{1+\chi}{\chi}\left(\frac{E[R(\theta)\mid z]}{E[R(\theta)\mid x=z,z]} - 1\right)$$

Proposition 4 (pp. 2348-2349) establishes three predictions: (i) investment is increasing in share prices: $$\text{cov}(\hat{K}(z), P(z)) > 0$$; (ii) excess sensitivity relative to fundamentals: $$\hat{K}(z)/K^*(z)$$ is strictly increasing in $$z$$; (iii) higher investment leads to lower future equity returns: $$\text{cov}(\hat{K}(z), (V(z)-P(z))/P(z)) < 0$$.

**Consistency with existing evidence (§III.A and §III.C, pp. 2346-2350).** The model is consistent with three lines of external evidence:

- Diether, Malloy, and Scherbina (2002) find that stocks in the highest earnings-forecast-dispersion quintile earn about 0.62% per month (roughly 7% annualized) lower returns, consistent with overvaluation from upside-risk overinvestment.
- Polk and Sapienza (2009) estimate a positive relation between share overvaluation (proxied by discretionary accruals) and investment after controlling for Tobin's Q, with stronger effects for firms with higher share turnover (the paper's $$\alpha$$) and lower future returns for overinvesting firms: consistent with Proposition 4(i) and 4(iii).
- David, Hopenhayn and Venkateswaran (2016) calibrate a GE model with the same informational friction (but without the rent-seeking motive) and find it responsible for 20-50% of observed dispersion in the marginal revenue product of capital.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Numerical simulations (Figures 1-5) | Synthetic data generated from model parameters to illustrate comparative statics on investment distortions and efficiency losses; no external dataset is used | No external data |

Sample: none. The paper is theoretical; Figures 1-5 use calibrated parameter values (e.g., $$\alpha = 0.5$$, $$\beta = 1$$, $$\lambda = 1$$) without fitting to real data.

## When to read the full paper

Read the [original](https://doi.org/10.1257/aer.20170725) if you are: building on the model (the Appendix at pp. 2351-2353 contains full proofs of Propositions 1-3 and the GE existence-uniqueness argument); studying the optimal tax design in general equilibrium and its Pigouvian correction; extending the information-feedback model to dynamic or multi-period settings; or seeking the working paper version's analysis of financial transaction taxes and additional policy instruments not covered in the published article.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(9), September 2023. This distillation was extracted by an LLM on 2026-06-24 and is **not human-verified or independently reproduced**. The published article is paywalled; an author manuscript is available at [hal.science/hal-04210328v1](https://hal.science/hal-04210328v1). Replication data are deposited at [doi.org/10.3886/E185081V1](https://doi.org/10.3886/E185081V1).

> Albagli, Elias, Christian Hellwig, and Aleh Tsyvinski. "Imperfect Financial Markets and Investment Inefficiencies." *American Economic Review* 113, no. 9 (September 2023): 2323-2354. DOI: 10.1257/aer.20170725. Extract only; not licensed for reproduction.


==============================================================================
# The Economic Origins of Government: Allen, Bertazzini & Heldring (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/allen-et-al-economic-origins-government-2023/
# River shifts in ancient southern Iraq (~2850BCE) caused new state formation, canal construction, tribute payment, and growth of administrative buildings, supporting cooperative over extractive theories of government origins, in a new archeological panel dataset spanning 3900BCE-2700BCE. American Economic Review 2023, open access. Eight core results with source locators, the identification strategy, and regression specifications; LLM-distilled, not human-verified.
# Tags: paper-summary, economic-history, political-economy, state-formation, public-goods
==============================================================================

**What this is.** The paper's core results, the natural experiment it exploits (river shifts in
ancient southern Iraq), the identification strategy, and the main estimating equations: enough
to know what it found and how, without reading all 39 pages. To replicate or extend it, read
the full source at [https://doi.org/10.1257/aer.20201919](https://doi.org/10.1257/aer.20201919).

## TL;DR

The paper uses river shifts in southern Mesopotamia around 2850BCE as a natural experiment to
test two competing theories of government formation. When a river shifts away from a farming
area, direct irrigation becomes impossible and public canals requiring multi-community
coordination are needed. This creates demand for a coordinating government (cooperative theory),
whereas an extraction-based theory (where government forms to collect surplus) predicts states
would be more likely where the river stays. Constructing a new archeological panel dataset of
5x5 km grid cells across southern Iraq for the period 3900BCE-2700BCE, the paper runs a panel
difference-in-differences design comparing grid cells that lost river access to control cells
that did not. A river shifting away increased the probability of state formation by 14 percentage
points, the probability of canal construction by 12 percentage points, the probability of tribute
payment by 21 percentage points, and the number of administrative buildings by 0.44. The result
is entirely driven by new state formation, not expansion of existing states. Text analysis of
5,885 surviving cuneiform tablets shows increased mentions of lineage-leader titles and tribute
after the shift, consistent with the cooperative interpretation. Extractive theories of Olson (1993)
and Carneiro (1970) predict the opposite sign, and the results reject them.

## Core results

Magnitudes and standard errors are as reported; clustered SE in parentheses, Conley (1999) SE in
brackets. Locators point into the source PDF (pp.2507-2545).

| \# | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | River shift away increases probability of **state formation** by 14pp | Table 3, col 1, p.2530 | +0.14 (SE=0.04) [0.03]; sample mean 0.06; treatment-period mean 0.24; pretrend p=0.23 |
| R2 | Effect driven entirely by **new state formation**, not expansion of existing states | Table 3, col 3-4, p.2530 | New state: +0.11 (SE=0.04) [0.03]; existing state: +0.02 (SE=0.02) [0.01] |
| R3 | River shift increases probability of **canal construction** by 12pp | Table 4, col 1, p.2534 | +0.12 (SE=0.03) [0.02]; mean 0.28; pretrend p=0.81 |
| R4 | River shift increases probability of **defensive wall** in nearest city by 11pp | Table 4, col 2, p.2534 | +0.11 (SE=0.04) [0.03]; mean 0.14; pretrend p=0.57 |
| R5 | River shift **doubles** the probability of **tribute payment** being recorded | Table 4, col 3, p.2534 | +0.21 (SE=0.06) [0.10]; mean 0.19; pretrend p=0.20 |
| R6 | River shift increases **administrative buildings** in nearest city by 0.44 | Table 4, col 4, p.2534-2535 | +0.44 (SE=0.15) [0.17]; mean 0.70; pretrend p=0.69 |
| R7 | Effect concentrated in **high population density** areas before the shift | Table 5, Panel A, p.2536 | High: +0.18 (SE=0.05) [0.03]; low: +0.03 (SE=0.02) [0.03]; Chow p=0.06 |
| R8 | Result holds across the **full 7,000-year panel** (5000BCE-1950CE, all 6 river shifts) | Table 6, col 1, p.2539 | Canal: +0.11 (SE=0.02) [0.02]; mean 0.40 |

**Overall (paper's conclusion, p.2540-2541).** Where rivers shifted away, communities formed new
states, built canals and defensive walls, and paid tribute to their governments. These results are
consistent with cooperative (demand-side) theories: states form to solve coordination failures
in public good provision, not to extract. The first states resembled scaled-up versions of the
lineage social structure that preceded them, with government coordinating between extended kinship
groups (lineages) rather than ruling over individuals.

## Theory / model

The paper has no formal mathematical model. Instead it derives two competing sets of testable
hypotheses from the theoretical literature on state formation and maps them onto the sign of the
main difference-in-differences coefficient.

**Cooperative (demand-side) theory.** Government is an organization with a comparative advantage
in providing public goods (Baumol 1952; Samuelson 1954). Problems of externalities and
coordination failure (Olson 1965) prevent private provision. When a river shifts away, arid-land
farming can continue only via public irrigation canals spanning multiple communities. No individual
community can credibly commit to build and maintain such canals alone. This coordination failure
creates demand for a government. Members pay tribute in exchange for canal access and defense. The
framework builds on Acemoglu and Robinson (2000), who model the bargaining between social groups
that determines whether and what form of government emerges. Under this view, states form where
the river shifted away (positive $$\beta_0^{\text{treatment}}$$ in equation 1).

**Extractive (supply-side) theory.** Government is an organization founded by an elite with
coercive power to manage extraction (Carneiro 1970; Olson 1993). Under Olson (1993), "roving
bandits" settle where the expected tax base is largest and find it worthwhile to provide minimal
stability in exchange for an extractive monopoly. Where a river shifts away, the agricultural
tax base collapses: land is no longer productive without irrigation. A predatory ruler would
prefer to locate where the river remains. Under this view, states form where rivers stay, not
where they shift (negative or zero $$\beta_0^{\text{treatment}}$$). Mayshar, Moav, and Pascali
(2022) provide recent empirical support for a related extractive channel in which taxable surplus
(storable crops) predicts state location. The circumscription theory of Carneiro (1970) also fits
this supply-side cluster: states arise where it is hard to escape extraction.

**Empirical mapping.** The sign of the main coefficient distinguishes the two clusters.
A positive and significant $$\beta_0^{\text{treatment}}$$ (R1 = +0.14) is consistent only with
cooperative theories, since extractive theories predict the opposite or no effect. The placebo
test (estimating the effect of a river shifting closer rather than away) yields a consistently
negative coefficient throughout the paper, confirming the cooperative interpretation.
Sánchez De La Sierra (2020) finds extractive state formation in the Congo using a comparable
natural-experiment design; the two settings differ in the nature of the shock and the role of
public goods versus looting.

## Method

The paper applies a standard panel difference-in-differences estimator via OLS, using the first
large river shift in history (around 2850BCE) as a quasi-random shock. It builds on
`difference-in-differences` and `panel-regression`.

**Treatment definition.** A 5x5 km grid cell $$c$$ is defined as on a river in period $$t$$ if its
centroid is within 5 km of the nearest river. Cell $$c$$ is treated in period $$t = 0$$ (Early
Dynastic I, 2900BCE-2700BCE) if it was on the river in period $$t - 1$$ (the Jemdet Nasr period,
3100BCE-2900BCE) and is no longer on a river in period $$t$$. Treatment is time-invariant and
applies to roughly 13 percent of the 1,374 grid cells in the study area (Table 1, p.2517).
Rivers shift by 30-40 km on average, so treated cells lose water access entirely, not marginally.

**Main estimating equation.** The panel DiD model for the main study period is (equation 1, p.2525):

$$
Y_{ct} = \sum_{k=-4}^{0} \beta_k^{\text{treatment}} \times \mathbf{1}\{\text{period}_k\} \times \text{treated}_c + \rho_c + \gamma_t + v_{ct} + \varepsilon_{ct} \tag{1}
$$

where $$Y_{ct}$$ is the outcome for grid cell $$c$$ in period $$t$$; $$\text{treated}_c$$ equals 1 if cell
$$c$$ loses river access at $$k = 0$$; $$\mathbf{1}\{\text{period}_k\}$$ are indicators for each period
$$k$$ relative to treatment; $$\beta_k^{\text{treatment}}$$ are the period-relative-to-treatment
interaction coefficients (normalized to zero at $$k = -1$$, the last pre-period); $$\rho_c$$ are unit
(grid cell) fixed effects; $$\gamma_t$$ are period fixed effects; $$v_{ct}$$ is a vector of period
fixed effects interacted with time-invariant covariates (survey area indicators, average rainfall,
average temperature, pre-shift urban status); and $$\varepsilon_{ct}$$ is clustered at the grid
cell level. Conley (1999) standard errors with a 484 km spatial cutoff are reported in parallel.

The coefficient of interest is $$\beta_0^{\text{treatment}}$$: the treatment effect in the Early
Dynastic I period (the period of first state formation). Pre-period coefficients
$$\beta_{-2}^{\text{treatment}}, \ldots, \beta_{-4}^{\text{treatment}}$$ serve as pretrend tests;
all are indistinguishable from zero (pretrend p-values reported in Tables 3-4, pp.2530, 2534).

**Extended panel estimator.** For all six river shifts across the full 7,000-year panel
(5000BCE-1950CE), a pooled DiD (equation 2, p.2538) is estimated:

$$
Y_{ct} = \beta \cdot \text{treated}_{ct} + \rho_c + \gamma_t + v_{ct} + \varepsilon_{ct} \tag{2}
$$

where $$\text{treated}_{ct}$$ is now time-varying (equals 1 when cell $$c$$ is treated in period $$t$$
for any of the six shifts), and $$\beta$$ captures the average effect across all shifts.

## Empirical specifications

All regressions are OLS. Unit of observation: 5x5 km grid cell. Time series: archeological period
(average 240 years in the main sample). Standard errors clustered at the grid cell level;
Conley (1999) SE with a 484 km cutoff in brackets.

**State formation (R1-R2, Table 3, p.2530).** Equation (1) with two outcome definitions: column 1
uses an indicator for whether a grid cell is part of a city state, defined using administrative
buildings (palaces, temples, ziggurats) and reconstructed territorial borders; column 2 uses only
building presence without borders. Columns 3 and 4 decompose into new state formation (an indicator
equal to 1 if the nearest city gains state status for the first time) and expansion of an existing
state. All columns include grid cell and period fixed effects plus the four time-invariant covariate
interactions. Pretrend p-values (0.23 and 0.24) confirm no differential pre-trends.

**Public good provision and tribute (R3-R6, Table 4, p.2534).** Four separate regressions using
equation (1):

- Column 1: indicator for canal presence within 5 km of cell centroid, reconstructed from Chicago
  Oriental Institute excavation reports (Adams 1965, 1981; Adams and Nissen 1972).
- Column 2: indicator for defensive wall in the nearest city, from Bryce (2009) and Meyers (1997).
- Column 3: indicator for surviving cuneiform tablet in the nearest city (validated as a proxy for
  tribute payment and redistribution via keyword analysis of transliterated texts; Table RA29 in
  Results Appendix).
- Column 4: total count of palaces + temples + ziggurats in the nearest city (measure of state
  capacity and administrative infrastructure).

**Heterogeneous effects (R7, Table 5, p.2536).** Equation (1) estimated separately in two
subsamples split by the median of the spatial lag of pre-treatment settlement density (a proxy
for returns to coordination). The Chow test for coefficient equality has p-value 0.06. A second
split by geographic costs (FAO potential productivity differential between irrigated and rainfed
barley, and river water flow volume) yields similar heterogeneity.

**Extended panel (R8, Table 6, p.2539).** Equation (2) estimated on 27,106 grid cell-period
observations across all 31 archeological periods (5000BCE-1950CE). Columns 2-3 split the sample
into the first two river shifts (pre-state and state formation) and the four subsequent shifts
(within established states). The effect is positive and similar in both sub-panels.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Chicago Oriental Institute archeological surveys (Adams 1965, 1981; Adams and Nissen 1972) | Settlement history, canal network reconstruction, and city locations; basis for the 5x5 km grid panel covering 5000BCE-1950CE | No page yet |
| Cuneiform Digital Library Initiative (CDLI) | 5,885 transliterated cuneiform tablets from the main study period; text analysis of tribute, canal, and lineage-leader term mentions (Section VII, Figure 6) | No page yet |
| Administrative building and state border data (Heinrich 1982, 1984; Meyers 1997; Bryce 2009) | Identification of palaces, temples, and ziggurats; reconstruction of state territorial borders | No page yet |
| River shift and course reconstructions (Cole and Gasche 1998) | Geographic reconstruction of river courses before and after each of six shifts; defines the treatment variable | No page yet |
| FAO soil and agricultural potential data | Potential productivity of irrigated vs. rainfed barley cultivation; used for geographic heterogeneity subsamples (Table 5, Panel B, p.2536) | No page yet |

Sample: southern Iraq, 1,374 grid cells of 5x5 km covering the area between Baghdad and Basra.
Main study: 3900BCE-2700BCE, five archeological periods (~240 yrs each). Extended study: 5000BCE-1950CE,
31 periods. The replication dataset (including all constructed variables) is publicly available at
[https://doi.org/10.3886/184167EV1](https://doi.org/10.3886/184167EV1).

## When to read the full paper

Use the [original paper](https://doi.org/10.1257/aer.20201919) if you are:
testing theories of state formation in other historical or contemporary settings, where the
identification approach (demand-side shock to coordination costs) can be adapted (Section IV);
studying public good provision and tribute across the full 7,000-year Iraqi panel, including
under the Babylonian and Assyrian empires (Section VI and Table 6); examining the internal
organization of the first states via cuneiform tablet text analysis (Section VII and Figure 6,
p.2541); or extending the dataset to other archeological outcome variables using the replication
data. The online Data Appendix (68 pages) describes dataset construction, coding procedures,
and robustness exercises in full detail.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(10), October 2023. Freely available on
the AEA website; no CC licence detected in Crossref metadata; redistribution is extract-only.
This distillation was extracted by an LLM on 2026-06-24 and is **not human-verified or
independently reproduced**.

> Allen, Robert C., Mattia C. Bertazzini, and Leander Heldring.
> "The Economic Origins of Government."
> *American Economic Review* 113, no. 10 (October 2023): 2507-2545.
> DOI: 10.1257/aer.20201919.


==============================================================================
# Birth of a Nation Media Effects: Ang (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/ang-birth-nation-media-racial-2023/
# Distilled: Ang (2023) provides the first causal evidence that D. W. Griffith's 1915 film The Birth of a Nation increased local lynchings and race riots by approximately fourfold, raised second-KKK klavern probability by 66 pp (2SLS), and predicts 85 percent higher hate crime rates per 100k residents a century later. American Economic Review 113(6), 2023, paywalled. Seven core results with source locators, datasets used, the identification design, and estimating equations.
# Tags: paper-summary, political-economy, media-economics, racial-discrimination
==============================================================================

**What this is.** The paper's core results, the identification design, and the estimating equations: enough to understand what was found and how, without reading all 37 pages. To replicate or extend, read the original at [doi.org/10.1257/aer.20201867](https://doi.org/10.1257/aer.20201867).

## TL;DR

Ang (2023) provides the first causal evidence of the social imprint of D. W. Griffith's 1915 film *The Birth of a Nation*, a fictional depiction of the founding of the Ku Klux Klan that reached an estimated 10 million Americans during a five-year road show. Exploiting the film's staggered county-level distribution, the paper finds sharp spikes in lynchings and race riots in the months of each county's screening (approximately fourfold over the county-level baseline). For longer-run effects, theater presence in 1914 instruments for whether a county received the film; the 2SLS estimate shows a 66 percentage-point increase in the probability of a second-KKK klavern forming by 1930 (Fryer and Levitt (2012) provide Klan history and membership benchmarks). Road-show counties remain significantly more likely to contain hate groups and to experience higher hate crime rates a century later, with effects extending beyond anti-Black incidents to other racial, religious, and sexual minorities. Mediation analysis following Dippel, Ferrara, and Heblich (2020) indicates the long-run effects run almost entirely through the historical formation of second-KKK chapters. Esposito et al. (2023) is a companion paper on the film's rhetorical legacy that independently validates the screening data.

## Core results

Magnitudes as reported. Randomization-inference p-values cited for R1 and R2 (the rarity of events makes conventional standard errors imprecise). Locators point into the source PDF.

| # | Result | Locator | Magnitude as reported |
|---|---|---|---|
| R1 | Lynching probability spiked in the month of the film's county arrival | Figure 3 Panel A, p. 1438 | beta_0 = 0.0011 (mean = 0.0003); ~4x monthly base rate; randomization-inference p < 0.01 |
| R2 | Race riot probability spiked in the three months following the film's arrival | Figure 3 Panel B, p. 1438 | beta_0 = 0.00034 (mean = 0.00009); ~4x base rate; randomization-inference p < 0.05 |
| R3 | Theater presence in 1914 strongly predicts screening (first stage) | Table 3 col. 1, p. 1444 | Coefficient = 0.324 (se = 0.024); mean screened = 0.20; Kleibergen-Paap F = 185.34 |
| R4 | Screening raised klavern probability by 66 pp (2SLS main estimate) | Table 4 Panel C col. 1, p. 1446 | 2SLS = 0.662 (se = 0.138); mean = 0.313; p < 0.001; OLS = 0.111 (se = 0.021) |
| R5 | Screening raised klavern intensity (klaverns per 10,000 US-born White males) | Online Appendix Table A.II; p. 1449 | beta_IV = 0.95 (p = 0.036); implies ~1 million additional Klansmen induced |
| R6 | Road-show counties ~90% more likely to host a hate group in 2000-2019 | Table 5 col. 2, p. 1454 | 2SLS = 0.366 (se = 0.100); mean = 0.374; treatment mean 0.60 vs control mean 0.32 |
| R7 | Hate crime rate ~85% higher in road-show counties, 2000-2018; elevated for anti-Black and anti-other-minority crimes | Table 6 cols. 1-3, p. 1455 | 2SLS = 1.177/100k (se = 0.494; mean = 1.376); anti-Black = 0.516 (p = 0.003); anti-other minorities = 0.561 (p = 0.030); anti-White = 0.101 (insignificant) |

**Overall (paper's conclusion).** Screenings of *The Birth of a Nation* triggered near-term racial violence and fueled the nationwide revival of the Ku Klux Klan. That historical catalysis persists a century later: road-show counties experience markedly higher rates of organized White supremacist activity and hate crimes directed at a wide range of minority groups, effects mediated almost entirely by the Klan chapters the film historically seeded.

## Theory / model

The paper proposes no formal economic model. The research design tests three related hypotheses.

*Short-run violence hypothesis.* Exposure to *The Birth of a Nation* increased racial hate in the county in the months of the screening, through some combination of: (i) persuasion and direct belief change about racial hierarchy, (ii) emotional responses that lowered inhibitions for violence, and (iii) public revelation of latent racism that reduced the perceived social cost of acting on racial animus.

*Medium-run KKK formation hypothesis.* Exposure catalyzed the formation of second-KKK chapters that would not have formed otherwise. Proposed channels include: (a) provision of common cultural symbols and imagery (white robes, cross burning, which the second KKK copied directly from the film's fictionalized Klan, unlike the first Reconstruction-era Klan; Section I.C, p. 1432), (b) reduction of coordination costs among individuals seeking to organize around racial ideology, and (c) persuasion of latent adherents into active participants.

*Long-run persistence hypothesis.* The Klan chapters historically seeded by the film transmitted racial hate across generations through social institutions and networks, so areas with early klavern formation exhibit higher hate-group presence and hate crime rates decades after the second KKK's formal dissolution.

Three mechanism channels are discussed (Section V, pp. 1452-1453):

1. **Social norm erosion and coordination.** Public screenings publicly revealed latent racism and unraveled norms that had suppressed discriminatory behavior, facilitating coordination among individuals predisposed to racial animus. This channel is consistent with Bursztyn, Egorov and Fiorin (2020) and with the heterogeneous-effects finding that the film's catalyzing impact was largest in counties with below-median religious participation, where prior social coordination infrastructure was weakest.
2. **Media imitation.** The second KKK adopted the film's specific iconography: white robes, hoods, and cross burning. None of these practices appeared in the Reconstruction-era Klan; all were drawn directly from the film. Full-count census data also show that White parents in road-show counties became more likely to name their sons "Benjamin," after the film's protagonist.
3. **Persuasion and preference change.** Survey evidence from the 1946 Gallup and 1970s ANES shows road-show counties were associated with less favorable attitudes toward African Americans and greater Klan approval, consistent with a shift in racial preferences among viewers.

## Method

The paper applies two identification strategies, linked by the same instrument.

**Event study (short-run racial violence).** The film's five-year road show created staggered county-level variation in when and whether each county received a screening. This variation, combined with the demonstrated role of market factors (population size, theater capacity, urban density) rather than racial animus in determining where the film was shown, supports a parallel-trends assumption for an event study comparing counties before and after the film's arrival. The key estimate is the coefficient at $$\tau = 0$$ (month of arrival) from Equation 1.

**Instrumental variables (long-run effects).** Whether a county received the film is endogenous to local characteristics. The paper instruments for county-level screening using movie theater presence in 1914, the year before the film's release. Theaters predict screenings because the film required elaborate projection equipment and large paying audiences, so distributors prioritized counties with existing cinema infrastructure. Exogeneity is validated by four pieces of evidence: (i) theater presence in 1914 is uncorrelated with pre-period racial violence, Democratic vote shares, and Black population shares, conditional on controls (Figure 4, p. 1441); (ii) in Kansas, where the film was banned statewide, theaters in 1914 do not predict klavern formation (Table 2 Panel A), supporting the exclusion restriction; (iii) only theaters opened before the road show's end (pre-1919) predict future klavern formation, while post-1918 theaters do not (Figure 5, p. 1443); (iv) Oster's delta ranges from 1.7 to 4.8 in matched samples (p. 1451), indicating that selection on unobservables would need to be several times larger than selection on observables to explain the estimates away. For long-run outcomes, the same IV design is re-estimated with hate groups (2000-2019) and hate crime rates (2000-2018) as dependent variables. DellaVigna and La Ferrara (2015) provide the media-economics benchmark establishing that selection into viewership of entertainment media is driven primarily by demand for entertainment rather than pre-existing racial preferences.

This design extends the approach of Yanagizawa-Drott (2014), who identifies the causal effect of Rwanda radio propaganda on genocide participation using a topography-based instrument for signal reception, to the case of popular fictional entertainment media in a democratic setting.

## Empirical specifications

**Event study (Equation 1, p. 1437).** Weekly county panel data, 1913-1922, all US counties:

$$
y_{c,t} = \delta_c + \lambda_{s,t} + \sum_{\tau=-6}^{6} \beta_\tau \, Show_\tau + \epsilon_{c,t} \tag{1}
$$

Here $$y_{c,t}$$ is an indicator for whether a lynching (or race riot) occurred in county $$c$$ at week $$t$$; $$\delta_c$$ are county fixed effects; $$\lambda_{s,t}$$ are state-week fixed effects, absorbing state-wide shocks and differential secular trends; $$Show_\tau$$ are dummies for months relative to the film's first screening in the county ($$\tau = 0$$ = month of arrival), defined as five-week intervals centered on the week of the film's premiere in each county; $$Show_6$$ ($$Show_{-6}$$) equals 1 for periods six or more months after (before) a screening; the omitted category is the last month before arrival; standard errors clustered by state. Sample: 3,104 US counties (Washington DC dropped due to collinearity with state fixed effects). Outcome means: 0.0003 (lynchings) and 0.00009 (race riots) per county-week.

**IV first stage (Equation 3, p. 1440):**

$$
\text{Screened}_c = \delta_s + \gamma \, \text{Theater}_c + X'_c \Lambda + v \tag{3}
$$

where $$\text{Screened}_c$$ is a binary indicator for whether county $$c$$ received the film from 1915 to 1919; $$\text{Theater}_c$$ is a binary indicator for a movie theater in county $$c$$ in 1914; $$\delta_s$$ are state fixed effects; $$X'_c$$ is the vector of demographic, social-capital, media, and racism controls described in Table 3. The preferred fully-controlled specification (col. 5, Table 3) yields a coefficient of 0.219 (se = 0.023) and Kleibergen-Paap F = 87.45, well above the 16.38 maximal-10%-bias benchmark.

**IV second stage (Equation 2, p. 1440):**

$$
KKK_c = \lambda_s + \beta \, \text{Screened}_c + X'_c \Gamma + u \tag{2}
$$

where $$KKK_c$$ is the outcome of interest (klavern indicator by 1930, hate-group indicator, or hate crime rate); $$\beta$$ is the LATE identifying the average causal effect of screening on compliers; standard errors clustered by state throughout. The same specification is applied to hate-group presence (Table 5) and hate crime rates (Table 6) to produce the long-run estimates.

**Controls** (Table 3) include: total population, density, Black population and share, US-born share, and draft-eligible age share (1910 Census); urban share, illiteracy rate, voter turnout 1912, religious organization share (1906 Census of Religious Bodies), occupational income score, and railroad distances to the two nearest major cities; per-capita newspaper circulation and number of media markets in 1912 (Gentzkow, Shapiro, and Sinkinson 2011); historical lynching count 1900-1905, Democratic vote share 1912, Confederate monuments by 1914, and NAACP chapter presence in 1914.

**Robustness** includes: propensity-score matching across three comparison groups; restriction and exclusion of neighboring counties; restriction to counties with at least one digitized local newspaper; alternative instruments (theaters per 1,000 White residents, maximum seating capacity, year of first theater, showings of *Mickey* or *The Million Dollar Mystery*); spatial-correlation-robust standard errors (Conley 1999; Muller and Watson 2021); and placebo tests using screenings of the 1918 comedy *Mickey*, which show no significant effect on racial violence or Klan formation.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Hand-collected newspaper screening data (newspapers.com, newspaperarchive.com, Library of Congress; 6,266 ads, 1914-1919) | Primary treatment variable: county-level first-screening dates; 621 screened counties | No page yet (hand-collected; see `introducesData`) |
| Historical American Lynching Data Collection Project ("Project HAL"); supplemented by Seguin and Rigby (2019) for non-Southern counties | Short-run outcome: lynching indicator, weekly county panel 1913-1922 | No page yet |
| Race riot records from Gilje (1996) and Red Summer Archive (visualizingtheredsummer.com) | Short-run outcome: race riot indicator, weekly county panel 1913-1922 | No page yet |
| Kneebone and Torres (2015) KKK klavern location data | Long-run outcome: second-KKK klavern presence by 1930; intensive margin klaverns per capita | No page yet |
| House Un-American Activities Committee reports, compiled by Mazumder (2018) | Long-run outcome: third-KKK klavern presence, 1960s | No page yet |
| Southern Poverty Law Center Hate Map (2000-2019) | Long-run outcome: active hate group presence by type (KKK, White supremacist, other) | No page yet |
| FBI Uniform Crime Reports hate crimes, compiled by Kaplan (2020) | Long-run outcome: hate crimes per 100k residents by victim group, 2000-2018 | No page yet |
| Theater location data (Klenotic / mappingmovies.unh.edu 1910; cinematreasures.com) | Instrument: county theater presence in 1914 | No page yet |
| 1910 US Census and 1906 Census of Religious Bodies | Demographic, social-capital, and economic controls | No page yet |

Sample: 3,103 US counties, continental United States. Event study runs at weekly frequency, 1913-1922. Cross-sectional IV uses 1910 pre-treatment characteristics and outcomes measured at 1930 (KKK), 1960s (third KKK), 2000-2019 (hate groups), and 2000-2018 (hate crimes).

## When to read the full paper

Read the original if you are: examining the causal effects of entertainment media on political and social outcomes; studying the origins and geographic persistence of KKK formation; applying staggered event-study or IV designs to county-level historical data; or researching the long-run transmission of racially charged organizations and institutions. Table 4 (p. 1446) contains the main IV estimates; Tables 5 and 6 (pp. 1454-1455) contain long-run hate-group and hate-crime results. Section V (mechanisms, pp. 1452-1453) is the most speculative part of the paper. Yanagizawa-Drott (2014) is the closest design precedent.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(6), June 2023. Published by the American Economic Association. No open-access license detected (Crossref metadata, AEA publisher site, and OpenAlex all confirm paywalled). This distillation was extracted by an LLM on 2026-06-24 and is **not human-verified or independently reproduced**. Extract-only: the PDF is not hosted here.

> Ang, Desmond. "The Birth of a Nation: Media and Racial Hate." *American Economic Review* 113, no. 6 (June 2023): 1424-1460. DOI: 10.1257/aer.20201867.


==============================================================================
# Road to Efficiency: Avoyan & Ramos (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/avoyan-ramos-road-efficiency-communication-commitment-2023/
# Distilled: A laboratory experiment shows that a commitment-enhanced pre-play communication institution (asynchronous revision mechanism) achieves 82 percent efficiency in the minimum-effort coordination game, significantly outperforming cheap-talk communication (64 percent) and the no-communication baseline (48 percent); commitment, asynchronicity, and revision frequency are all necessary ingredients. American Economic Review 2023, paywalled. Nine core results with source locators, the game-theoretic model, and the experimental design.
# Tags: paper-summary, game-theory, experimental, coordination, mechanism-design
==============================================================================

**What this is.** The paper's core results, the game-theoretic model of the minimum-effort game and the asynchronous revision mechanism, and the experimental design with its estimating comparisons: enough to understand what it found and how, without reading all 27 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1257/aer.20171014).

## TL;DR

Avoyan and Ramos run a laboratory experiment in the minimum-effort game, a canonical coordination game where groups chronically fail to reach the payoff-efficient outcome. They introduce a pre-play institution: an asynchronous revision mechanism (RM) in which each player posts a prepared effort choice that is publicly observable, but can be revised only when a stochastic revision opportunity is awarded. This incremental commitment makes prepared actions credible in a way that one-shot cheap-talk communication, as in Blume and Ortmann (2007), cannot. RM achieves 82.1 percent efficiency, versus 47.8 percent with no communication (Baseline) and 64.1 percent with one round of cheap-talk messages (S-CT). The efficiency gain requires all three components simultaneously: commitment (removing it via R-CT drops to 67.2%), asynchronicity (S-RM drops to 67.1%), and frequent revision opportunities (I-RM drops to 69.6%). The dynamic behavior within the pre-play phase also matches the theory: early revisions are forward-thinking (upward moves to lead others toward the efficient effort), while late revisions are myopic payoff-improving down-moves. Rich pre-play communication without commitment, as in Deck and Nikiforakis (2012), does not improve on one-shot cheap talk.

## Core results

Magnitudes and significance as reported; `\*\*\*` = 1%. Locators point to the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | RM significantly increases efficiency over Baseline and cheap talk (S-CT) | Figure 2, p. 2370; Result 1, p. 2371 | RM = 82.1%; Baseline = 47.8% (+34 pp, p < 0.01 MWU); S-CT = 64.1% (+18 pp, p < 0.01 MWU) |
| R2 | RM efficiency (82.1%) and initial efficient-effort rate (85.7%) are both significantly below the 100% theoretical prediction | §IV.B, pp. 2371-2372; Result 2, p. 2372 | Both p < 0.01; exact point predictions of Calcagno et al. (2014) rejected |
| R3 | RM performance is invariant to exogenous initial choices (R-RM) and to Van Huyck, Battalio, and Beil (1990) payoff parameters (RM-VHBB) | pp. 2372-2373; Result 3, p. 2373 | R-RM = 77.8% (MWU p = 0.173 vs RM); RM-VHBB = 82.2% (MWU p = 0.447 vs RM) |
| R4 | Reducing revision frequency (I-RM) or making revisions synchronous (S-RM) each reduces efficiency by 12-15 pp | pp. 2373-2374; Result 4, p. 2374 | I-RM = 69.6% (-12.5 pp, p < 0.001 MWU); S-RM = 67.1% (-15 pp, p < 0.001 MWU) |
| R5 | Removing commitment (R-CT) reduces efficiency by 15 pp to cheap-talk levels | p. 2374; Result 5, p. 2374 | R-CT = 67.2% vs RM = 82.1%, p < 0.001 MWU |
| R6 | Early revisions are forward-thinking; late revisions are myopic payoff-improving | Figure 3, p. 2375; Result 6, p. 2376 | First 10 s: 91.6% forward-thinking, 2.9% myopic-down; last 10 s: 12.5% forward-thinking, 84.4% myopic-down |
| R7 | Groups converging to a common effort profile implement it significantly more often under commitment | pp. 2376-2377; Result 7, p. 2377 | 87.6% follow-through in RM vs 51.1% in R-CT when a common profile is reached |
| R8 | Substantial gap between final pre-play message and implemented effort in R-CT causes 10.7% payoff loss | p. 2377; Result 8, p. 2377 | Payoff at 60th-second message = 10.18; payoff-relevant choice = 9.19; p < 0.001 (MWU) for min effort and freq efficient effort |
| R9 | Richer normative cheap-talk messages (R-R-CT) do not improve efficiency over simple cheap talk (R-CT) | p. 2378; Result 9, p. 2378 | R-R-CT = 68.2% vs R-CT = 67.2%; MWU p >= 0.385 |

**Overall (paper's conclusion).** The three key ingredients of the revision mechanism (commitment, asynchronicity, and frequent revisions) are all necessary to achieve 82 percent efficiency. Removing any one reduces efficiency to the level of standard cheap talk. The paper's summary efficiency ordering is (equation (4), p. 2378):

$$
\text{RM} > \text{S-CT} \approx \text{R-CT} \approx \text{I-RM} \approx \text{S-RM} > \text{Baseline}. \tag{4}
$$

## Theory / model

The stage game is a normal-form game $$(I, (E)_{i \in I}, (\pi_i)_{i \in I})$$, where $$I = \{1, \ldots, n\}$$, $$E$$ is a finite effort set common to all players, and $$\pi_i(\mathbf{e})$$ is the payoff to player $$i$$ for profile $$\mathbf{e} \in E^n$$ (p. 2360). The highest effort is $$\bar{e}$$, the lowest $$\underline{e}$$. In the **minimum-effort game** (equation (1), p. 2360):

$$
\pi_i(\mathbf{e}) = \gamma + \alpha \cdot \min_{j \in I} e_j - \beta \cdot e_i, \tag{1}
$$

where $$\alpha > \beta > 0$$. Laboratory parameters: $$\gamma = 0.18$$, $$\alpha = 0.20$$, $$\beta = 0.04$$, $$E = \{1, 2, 3, 4, 5, 6, 7\}$$, groups of $$n = 6$$. Every profile where all players choose the same effort $$e^*$$ is a strict Nash equilibrium, Pareto-ranked by effort level. The efficient profile $$\bar{e} = (7,\ldots,7)$$ is uniquely Pareto dominant, yet laboratory groups routinely play far below it (Baseline efficiency: 48%).

**K-coordination games.** Following Calcagno et al. (2014) (Definition 1, p. 2361), a component game is a $$K$$-coordination game if for any pair of players $$i, j \in I$$ and any profile $$\mathbf{e}$$:

$$
\frac{\pi_j(\bar{e}) - \pi_i(\mathbf{e})}{\pi_i(\bar{e}) - \pi_i(\underline{e})} \leq K \frac{\pi_j(\bar{e}) - \pi_j(\mathbf{e})}{\pi_j(\bar{e}) - \pi_j(\underline{e})}. \tag{2}
$$

The constant $$K$$ measures payoff similarity: $$K = 1$$ is pure coordination. For the minimum-effort game, condition (2) reduces to $$\alpha / (\alpha - \beta) \leq K$$. At the laboratory parameters, $$K = 0.20/(0.20 - 0.04) = 1.25$$.

**The asynchronous revision game.** The pre-play phase uses discrete time $$t \in \{-T, \ldots, -1, 0\}$$. At $$t = -T$$, an initial effort profile is set simultaneously. During $$t < 0$$, at each instant a revision opportunity arrives for the group with probability $$p \in (0,1]$$; if it arrives, it is allocated to one of the $$n$$ players with equal probability. At $$t = 0$$ (the deadline), the most recently posted efforts are implemented. All past events are publicly observable, so the natural solution concept is subgame perfect equilibrium, called a revision equilibrium (p. 2362).

**Main theoretical result** (Proposition 1, p. 2362). In a discrete-time asynchronous revision game with symmetric arrival rate, if the component game is a $$K$$-coordination game with strict Pareto-dominant profile $$\bar{e}$$ and the condition

$$
(n-2)K < (n-1)
$$

holds, then for any $$\varepsilon > 0$$ there exists $$T' > 0$$ such that for all $$T > T'$$, all revision equilibria have $$\mathbf{e}(0) = \bar{e}$$ with probability at least $$1 - \varepsilon$$.

The proof proceeds by induction (following Calcagno et al. (2014), online Appendix A): (i) $$\bar{e}$$ is absorbing once reached; (ii) far enough from the deadline it is optimal to revise up to $$\bar{e}$$ regardless of others' current choices, since the cost of being alone at $$\bar{e}$$ temporarily is small when the deadline is far. The condition $$(n-2)K < (n-1)$$ ensures the induction goes through in finite time. At the laboratory parameters, the condition is satisfied ($$(6-2)(1.25) = 5 = 6-1$$, boundary case), and backward induction in online Appendix B confirms that the unique revision equilibrium prescribes choosing effort 7 from the start. The paper extends the Calcagno et al. (2014) theory from continuous to discrete time and derives two additional numerical insights: if players can choose effort before the pre-play phase, all should choose 7 from the outset; and far from the deadline, revising to $$\bar{e}$$ is dominant regardless of the current profile.

## Method

Sessions ran at the Center for Experimental Social Science (CESS) at New York University (NYU) and the Interdisciplinary Experimental Laboratory (IELAB) at Indiana University (IU), using z-Tree (Fischbacher 2007), from December 2015 through April 2021 (p. 2365). Participants are randomly assigned to groups of six; each session consists of 10 rounds of the minimum-effort game.

**Revision mechanism (RM) treatment.** Each round starts with all six group members simultaneously choosing an integer in $$E = \{1,\ldots,7\}$$. A 60-second countdown then begins. A real-time graph displays each player's currently posted effort, visible to all. At each second a revision opportunity arrives for the group with probability 0.8; if it arrives, it is allocated to one member uniformly at random (probability $$1/6$$). A player can change their cursor selection at any time, but the posted graph value updates only upon receiving a revision opportunity. At $$t = 0$$ only the posted choice matters for payoffs. Each player expects approximately eight revision opportunities per round (p. 2367).

**Nine treatments** (Table 1, p. 2369):

| Treatment | Communication | Commitment | Subjects | Groups |
|---|---|---|---|---|
| Baseline | None | N/A | 48 | 8 |
| Standard cheap talk (S-CT) | One-shot message | None | 48 | 8 |
| Revision mechanism (RM) | Revisions | Gradual | 96 | 16 |
| Random revision (R-RM) | Revisions | Gradual | 48 | 8 |
| RM-VHBB | Revisions | Gradual | 48 | 8 |
| Infrequent RM (I-RM) | Revisions | Abrupt | 48 | 8 |
| Synchronous RM (S-RM) | Revisions | Gradual | 48 | 8 |
| Revision cheap talk (R-CT) | Revisions | None | 96 | 16 |
| Richer R-CT (R-R-CT) | Revisions | None | 48 | 8 |

S-CT follows Blume and Ortmann (2007): before the effort choice, subjects simultaneously send a public number message; subjects then see all messages for 60 seconds before making their payoff-relevant effort. R-CT follows the RM protocol for the pre-play graph but, unlike RM, the choice at the end of the countdown is not payoff-relevant; subjects choose payoffs on a separate screen after. RM-VHBB uses Van Huyck, Battalio, and Beil (1990) payoff parameters ($$\alpha = 0.2$$, $$\beta = 0.1$$, $$\gamma = 0.6$$). I-RM reduces the group revision probability from 0.8 to 0.1. S-RM makes all revisions synchronous (all six group members receive the revision simultaneously at each opportunity). R-RM has initial effort choices assigned randomly from $$E$$.

**Efficiency measure** (equation (3), p. 2369):

$$
\text{Efficiency} = \frac{\text{Actual} - \text{Min}}{\text{Max} - \text{Min}}, \tag{3}
$$

where Actual is the average amount earned, and Min (Max) is the average minimum (maximum) possible earnings. Normalization enables comparison across payoff specifications.

## Empirical specifications

**Main treatment comparisons (R1, R4-R5, R8-R9).** The primary test is a two-sample Mann-Whitney U (MWU) test with the group's round average as the unit of observation (one group = one independent observation, since standard errors are clustered at the group level). Treatment sizes range from 8 groups (80 group-round observations) to 16 groups (160 group-round observations). Five outcome variables are compared: (i) subject payoffs, (ii) minimum effort of the group, (iii) frequency of efficient effort choice (fraction of group members choosing 7), (iv) fraction of fully coordinated groups (all six members choose the same effort), and (v) equilibrium deviation (average distance between a subject's effort and the group minimum).

**OLS regression (R1 core, Table 2 p. 2372).** Payoffs and the four group-level coordination measures are regressed on treatment dummies (baseline = S-CT treatment) and demographic controls, with standard errors clustered at the group level:

$$
y_{ig} = \beta_0 + \beta_1 \cdot \mathbf{1}[\text{Baseline}]_g + \beta_2 \cdot \mathbf{1}[\text{RM}]_g + \mathbf{X}_{ig}' \gamma + \varepsilon_{ig},
$$

where $$y_{ig}$$ is the outcome for subject $$i$$ in group $$g$$, $$\mathbf{X}$$ includes quiz score and demographics, and errors are clustered by group. The RM coefficient $$\hat{\beta}_2 = 0.21$$ (SE = 0.012) in the payoffs regression confirms significantly higher payoffs in RM relative to S-CT (Table 2, p. 2372).

**Exact theoretical predictions (R2).** The fraction of subjects initially choosing effort 7 is tested against 100 percent using the group average as unit. Average initial choice of 7 is 85.7 percent (93.8% in round 10); both are significantly below 100 percent (p < 0.01). RM efficiency of 82.1 percent is similarly tested against 100 percent.

**Dynamic behavior classification (R6).** Using the R-RM treatment (which introduces initial choice variation), each revision move is classified as: forward-thinking (increases effort even though this would decrease payoff if the game ended immediately, because it initiates a chain reaction when there is enough time); myopic-down (decreases effort toward the group minimum, payoff-improving if the game ended immediately); or other. The proportion of each type is plotted by 10-second interval across the 60-second pre-play phase (Figure 3, p. 2375).

**Communication credibility (R7).** All rounds in which a group converges to a homogeneous message profile during the 60-second pre-play are identified. The fraction of such rounds in which the communicated effort profile is also the payoff-relevant outcome (at $$t = 0$$ for RM, at the separate payoff screen for R-CT) is then compared between RM and R-CT. The 87.6% vs 51.1% gap quantifies the credibility gain from commitment (pp. 2376-2377).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Original laboratory data (CESS/NYU and IELAB/Indiana University, 2015-2021) | Primary experimental observations: effort choices, revision timing, group-level coordination outcomes across 9 treatments | no page yet |

Sample: 528 subjects across 9 treatments, 6 per group, 10 rounds each; 88 groups total.

## When to read the full paper

Read the [original](https://doi.org/10.1257/aer.20171014) if you are: designing or studying pre-play communication institutions in coordination games; testing or extending the Calcagno et al. (2014) theoretical framework to new environments; building laboratory experiments for the minimum-effort game (the online Appendix provides full instructions, backward induction numerical solutions, and complete robustness tables); studying the separate roles of commitment, asynchronicity, and revision frequency in coordination; or comparing against richer real-time communication benchmarks as in Deck and Nikiforakis (2012). The replication data are at the ICPSR archive (https://doi.org/10.3886/E185662V1).

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(9), September 2023, pp. 2355-2381. Copyright American Economic Association 2023. No CC license; redistribution is extract-only. This distillation was extracted by an LLM on 2026-06-24 and is **not human-verified or independently reproduced**.

> Avoyan, Ala, and João Ramos. "A Road to Efficiency through Communication and Commitment." *American Economic Review* 113, no. 9 (September 2023): 2355-2381. DOI: 10.1257/aer.20171014.


==============================================================================
# Dividend Taxes and Allocation of Capital (Comment): Bach et al. (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/bach-et-al-dividend-taxes-allocation-capital-2023/
# Distilled: This comment replicates Boissel and Matray (2022) using their own data and code, finding a coding alteration that suppresses differential pre-trends and showing that "size growth" controls are lagged outcome controls; no corrected specification produces convincing evidence that the 2013 French dividend tax increase raised corporate investment. American Economic Review 2023, paywalled. Three core results with source locators, datasets used, and the estimating equations.
# Tags: paper-summary, replication, public-finance, corporate-investment
==============================================================================

**What this is.** The core findings of this replication comment, with the
estimating equations and event-study diagnostics: enough to know what it
found and how, without reading the source. To replicate or extend, read the
original at [doi.org/10.1257/aer.20221432](https://doi.org/10.1257/aer.20221432).
The replication package is at [doi.org/10.3886/E185061V1](https://doi.org/10.3886/E185061V1).

## TL;DR

Using the same data and code as Boissel and Matray (2022), this comment
identifies two problems in BM's analysis of the 2013 French dividend tax
increase on corporate investment. First, a line in the code plotting BM's
Figure 4 divides two pre-reform event-study coefficients (t=-2 and t=-1) by
1.8, visually attenuating the gap between treated and control firms in the
pre-period. Second, BM's "size growth" controls are controls for the pre-reform
average of the outcome variable (investment rate), which mechanically suppresses
apparent pre-trends by creating mean reversion. After correcting either
problem, no specification produces convincing event-study evidence that the
reform raised investment. The paper concludes that one cannot claim the
dividend tax increase had a positive effect on companies' investment (p. 2049).

## Core results

Magnitudes and locators are as reported in the comment; figures are from the
5-page article.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | A line in BM's plotting code divides two pre-reform event-study coefficients (t=-2 and t=-1) by 1.8, reducing the visual evidence of differential pre-trends | p. 2050 | Factor of 1.8 applied to the t=-2 and t=-1 coefficients; original standard errors left untouched; alteration affects two of four pre-reform years |
| R2 | Without the code alteration, the corrected BM specification shows differential pre-trends: confidence intervals for t=-2 and t=-1 exclude zero | Figure 1, pp. 2049-2050 | CI excludes the null for t=-1 and t=-2 (orange curve); BM's original figure places the null at or inside the CI edge for both periods |
| R3 | Removing the size-growth (lagged outcome) controls leaves significant pre-trends and no clear post-reform positive investment effect in any specification | Figure 2, p. 2051 | Three alternative event-study variants all show significant differential pre-trends; capital-level control specification shows pre-trends significant and in the opposite direction |

**Overall (paper's conclusion).** Using BM's own data and code, the comment
shows that the estimation of the investment impact of the French dividend tax
hike is sensitive to the code alteration and to the choice of controls; no
specification provides clear evidence that dividend taxes encourage investment
(pp. 2051-2052).

## Theory / model

This paper has no formal model. The central hypothesis under test is BM's
parallel trends assumption: absent the 2013 reform, investment trends in
treated (high dividend-tax-burden) and control firms would have evolved
identically. The comment is diagnostic: it tests whether this assumption holds
in the corrected data.

The motivating economic claim, attributed to Boissel and Matray (2022), is
that a higher dividend tax reduces the after-tax return to distributing
earnings, encouraging firms to retain profits and invest. This channel contrasts
with prior evidence. Becker, Jacob, and Jacob (2013) find cross-country
evidence that higher dividend taxes hinder investment. Yagan (2015) finds the
US 2003 dividend tax cut had no effect on corporate investment. Because BM's
result contradicts both of these priors, the comment provides strict
re-examination of the methodology.

The parallel trends assumption is the sole identification claim. The comment
assesses it by inspecting pre-period event-study coefficients; a finding of
significant differential pre-trends invalidates the causal interpretation.

## Method

The comment applies BM's own difference-in-differences estimator, building on
`difference-in-differences` and `event-study` primitives (see `buildsFrom`).
Two diagnostic exercises identify the problems:

**Code alteration check.** The released code for BM's Figure 4 first runs a
regression with year-by-year interaction terms and then uses the output to
produce an event-study plot. In this last step, a command divides the
coefficients for t=-2 and t=-1 by 1.8, while leaving the standard errors
unchanged (p. 2050). Removing this command restores the regression output as
produced (event study 1 in Figure 1).

**Lagged-outcome control check.** BM define their "size growth" control as
"a vector of pre-reform annualized size growth quartile-by-year fixed effects"
(BM p. 2896, quoted p. 2050 of the comment). Size growth refers to capital
growth, which is identical to the investment rate (the main outcome). In a
difference-in-differences setting, conditioning on pre-treatment values of
the outcome forces parallel pre-trends mechanically (Daw and Hatfield 2018;
Chabé-Ferret 2017), biasing the post-treatment estimates upward. Removing
these controls reveals differential pre-trends that were hidden.

## Empirical specifications

BM's baseline difference-in-differences estimator, reproduced as equation (1)
in this comment (p. 2050):

$$
Y_{ijct} = \beta \, \text{Treated}_i \times \text{Post}_t
           + \theta_i + \text{SizeGrowthBin}_{it}
           + \delta_{jt} + \gamma_{ct} + \varepsilon_{ijct}
\tag{1}
$$

where $$Y_{ijct}$$ is total investment scaled by capital in 2011 for firm
$$i$$ in industry $$j$$, cohort $$c$$, year $$t$$; $$\text{Treated}_i \times \text{Post}_t$$
is the main difference-in-differences term (dividend-tax-affected firms
after the 2013 reform); $$\theta_i$$ are firm fixed effects;
$$\text{SizeGrowthBin}_{it}$$ are pre-reform annualized size-growth
quartile-by-year fixed effects (the lagged-outcome controls); $$\delta_{jt}$$
and $$\gamma_{ct}$$ are industry-year and cohort-year fixed effects.

The comment runs three alternative event-study specifications around the
2013 reform (Figure 2, p. 2051):

- **Event study 1** (R2): Equation (1) with the code alteration removed.
  Keeps size-growth controls. Differential pre-trends for t=-2 and t=-1 are
  significant (CI excludes zero).
- **Event study 2** (R3, green): Equation (1) without the alteration and
  without size-growth controls. Pre-trends are present but smaller; post-reform
  effect is unclear.
- **Event study 3** (R3, brown): Equation (1) without the alteration and
  without size-growth controls, but with year dummies interacted with quintiles
  of average pre-reform capital level (2009-2012) to correct for pre-trends.
  Differential pre-trends are significant in the opposite direction.

All three corrected specifications are inconsistent with the parallel trends
assumption that BM's causal interpretation requires.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| INSEE and DGFiP SUSE unified accounting files (2007) | Firm-level investment and capital data for the main DiD analysis | no page yet |
| INSEE Liaisons financières entre societes (2007-2011, 2012-2017) | Ownership structure for identifying dividend-tax-treatment status | no page yet |
| INSEE and DGFiP ESANE enterprise results (2008-2017) | Annual firm results covering the event-study window | no page yet |

All data are French administrative microdata accessed via CASD (Centre
d'Acces Securise aux Donnees) with institutional authorization. These are
the same datasets as used by Boissel and Matray (2022). Sample: French
private non-financial firms around the 2013 dividend tax reform; event-study
window from 4 years pre-reform (approximately 2009) to 5 years post-reform
(approximately 2018); annual frequency.

## When to read the full paper

Read the [original](https://doi.org/10.1257/aer.20221432) if you are:
evaluating the credibility of Boissel and Matray (2022)'s claimed positive
investment effect of dividend taxes; studying how lagged-outcome controls
can produce spurious parallel trends in difference-in-differences designs;
or seeking event-study evidence on the investment effects of dividend
taxation in France. The replication package at
[doi.org/10.3886/E185061V1](https://doi.org/10.3886/E185061V1) includes the
corrected code and data documentation. Locators above point to the exact
figures in the 5-page comment.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(7), July 2023.
This distillation was extracted by an LLM on 2026-06-24 and is
**not human-verified or independently reproduced**. Paywalled; extract-only.

> Bach, Laurent, Antoine Bozio, Arthur Guillouzouic, and Clement Malgouyres.
> "Dividend Taxes and the Allocation of Capital: Comment."
> *American Economic Review* 113, no. 7 (July 2023): 2048-2052.
> DOI: 10.1257/aer.20221432.


==============================================================================
# Electronic Food Vouchers: Banerjee, Hanna, Olken, Satriawan & Sumarto (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/banerjee-et-al-electronic-food-vouchers-evidence-2023/
# Distilled: An at-scale RCT across 105 Indonesian districts (3.4 million households) shows that switching from in-kind rice distribution to electronic food vouchers delivered 46 percent more subsidy to targeted poor households and cut poverty by 20 percent for the bottom 15 percent, driven by improved administrative fidelity rather than price-theoretic mechanisms. American Economic Review 2023, paywalled. Eight core results with source locators, the administrative-fidelity bargaining model, and the estimating equation.
# Tags: paper-summary, macro, panel-regression, cross-section, peer-reviewed
==============================================================================

**What this is.** Core results, the administrative-fidelity bargaining model, and
the estimating equation from Banerjee, Hanna, Olken, Satriawan, and Sumarto (2023):
a distilled skeleton for quick orientation. To replicate or extend, read the original
at [https://doi.org/10.1257/aer.20210461](https://doi.org/10.1257/aer.20210461) and
use the replication package at [https://doi.org/10.3886/E167262V1](https://doi.org/10.3886/E167262V1).

## TL;DR

Indonesia's government randomized 105 districts across the transition from its
in-kind rice subsidy program (Rastra: 10 kg free rice per month) to an electronic
voucher program (BPNT: a debit card worth approximately the same value, redeemable
for rice and eggs at private agents). Forty-two districts converted in 2018; 63 were
randomized to convert in 2019. The voucher program improved fidelity to program
design substantially: nearly all voucher recipients received the full entitlement
amount, versus broad distribution of small amounts in the in-kind program. As a
result, targeted (poor) households received 46 percent more subsidy value on net,
poverty rates fell 20 percent for the bottom 15 percent of the distribution, and
rice quality improved substantially. Price-theoretic channels (price effects,
consumption substitution, self-targeting) explain far less of the difference than
the administrative-fidelity mechanism: local village officials who previously
controlled rice distribution could no longer divert benefits once distribution moved
to private bank agents with individually named debit cards.

## Core results

Magnitudes are as reported; locators point into the source PDF. `\*` = 10%,
`\*\*` = 5%, `\*\*\*` = 1% (randomization inference p-values, Young 2019).

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Targeted households (PMT ≤ 30) received 46% more subsidy** per month in voucher districts than in in-kind districts | Table 1, col. 2, p. 529 | Rp 13,496 more/month (SE = Rp 1,909; p < 0.001); in-kind mean = Rp 29,219 |
| R2 | **Among recipients, voucher households received 85% more** per month (conditional on receiving any assistance) | Table 1, col. 7, p. 529 | Rp 31,333 more/month (SE = Rp 3,190; p < 0.001); in-kind recipient mean = Rp 36,931 |
| R3 | **Non-targeted households (PMT > 30) received 28% less** subsidy in voucher areas | Table 1, col. 3, p. 529 | -Rp 2,532/month (SE = Rp 564; p = 0.002); in-kind mean = Rp 9,162 |
| R4 | **Probability of receiving any subsidy fell** in voucher areas: 16% decline for targeted, 49% decline for non-targeted | Table 1, cols. 5-6, p. 529 | -10.5pp for PMT ≤ 30 (p < 0.001); -14.5pp for PMT > 30 (p < 0.001) |
| R5 | **Poverty rate fell 20% for the bottom 15%** in voucher areas | Table 2, col. 5, p. 537 | -4.3pp from a baseline of 21.0% for PMT ≤ 15 (p = 0.028) |
| R6 | **Rice quality rated 32% higher** in voucher areas (recipient households) | Table 1, col. 8, p. 529 | Coefficient = 0.203 on a 0-1 Likert scale (p < 0.001); in-kind mean = 0.630 |
| R7 | **Total egg protein consumption rose ~4.3%** for targeted households; no change in total rice consumption (consistent with Hastings and Shapiro 2018 earmarking evidence) | Table 3, Panel B, cols. 4-5, p. 539 | +9.3 g/month (p = 0.10) for PMT ≤ 30; rice coefficient = -0.411 kg (p = 0.492) |
| R8 | **No overall price effect** on rice; modest 3.5% increase in the most remote areas only | Table 4, cols. 1, 7, p. 541 | Overall: Rp 129 (p = 0.309); above 75th pct travel time to district capital: Rp 334 (p = 0.027) |

**Overall (paper's conclusion).** Switching from an in-kind food program to
electronic vouchers substantially increased the concentration of benefits to the
poor, primarily by removing local officials from the distribution chain and
replacing them with private bank agents who issued individually named debit cards
(Banerjee et al. 2018 context). Price-theoretic mechanisms (consumption flexibility,
supply-side price effects, self-targeting) are present but small relative to the
administrative-fidelity mechanism. The result parallels the administrative gains from
biometric smartcards documented in India by Muralidharan et al. (2016), here at
larger scale and with electronic vouchers rather than smartcard identification. The
voucher program also costs about half as much to administer (2.1 vs. 4.1 percent
of benefits disbursed).

## Theory / model

The paper has no structural model. It posits a simple Nash bargaining framework
(Section II.C, pp. 532-534) to explain why the voucher program produced a point
mass at the full entitlement amount while in-kind transfers produced a diffuse
distribution.

**Setup.** A beneficiary is entitled to transfer $$b$$ from the program. A village
head can impose a penalty $$X_i$$ on beneficiary $$i$$ (e.g., exclusion from community
activities). The village head and beneficiary split the surplus with bargaining
weight $$\alpha$$ for the village head. The beneficiary's net transfer and the village
head's rent are:

$$
\text{Transfer}_i = b - (1-\alpha)\,X_i, \qquad \text{village head rent} = \alpha\,X_i.
$$

There is a fixed cost $$F$$ for the village head to initiate bargaining with
beneficiary $$i$$.

**In-kind program.** The village head must distribute rice regardless, so $$F$$ is
sunk. The village head always bargains, and the distribution of $$X_i$$ across
beneficiaries produces a spread of realized transfer amounts (matching the broad
histogram in Figure 1, Panel A, p. 526).

**Voucher program.** Distribution moves to private bank agents with individually
named debit cards; the village head no longer has a role in the transfer unless
he actively seeks one. Now $$F$$ is not sunk. The village head bargains only
if $$\alpha X_i > F$$:

$$
\text{Transfer}_i = \begin{cases} b & \text{if } \alpha X_i \leq F \\ b - (1-\alpha)\,X_i & \text{if } \alpha X_i > F. \end{cases}
$$

This generates: (i) a point mass at the full entitlement $$b$$ for beneficiaries
where $$\alpha X_i \leq F$$, (ii) a gap just below $$b$$, and (iii) a left tail for
those with large $$X_i$$. The predicted distribution matches Figure 1, Panel A:
in voucher districts, 81 percent of monthly deliveries are exactly the
nominal Rp 110,000 entitlement, versus 24 percent in in-kind districts (p. 525).

**Identification.** The paper exploits budget-constrained random assignment: 105
districts were deemed potentially ready to convert, but the budget allowed
converting only about 42. The government randomized which 42 were treated in 2018
and which 63 were treated in 2019, stratifying by geography. Balance checks across
11 baseline variables show no significant imbalance (joint F-test p = 0.384; online
appendix Table 1, p. 523). The paper estimates intent-to-treat effects since only
3 of the 63 control districts converted early (p. 522).

## Method

The main estimator is OLS on the randomized intent-to-treat design with
double-LASSO-selected controls (Belloni, Chernozhukov, and Hansen 2014). The
method builds on `panel-regression` for the regression structure and `lasso` for
variable selection.

Control variables $$\mathbf{X}_{hvds}$$ are selected from a large candidate set
(UDB household characteristics, village-census covariates, and district
$$\times$$ urban/rural baseline averages from SUSENAS) using a double LASSO
procedure. The LASSO simultaneously selects variables predictive of (i) the
outcome and (ii) treatment assignment. Including the double-LASSO-selected controls
raises precision without affecting consistency (Belloni, Chernozhukov, and Hansen 2014).

Standard errors are clustered at the district (kabupaten) level, which is the
unit of randomization (d). Permutation-based (randomization inference) p-values
are computed using 1,000 permutations of the treatment vector (Young 2019).

## Empirical specifications

**Main estimating equation.** All outcomes are estimated via a single
equation (equation 1, p. 523):

$$
y_{hvds} = \beta_0 + \beta_1\,\text{Voucher}_{ds} + \mathbf{X}_{hvds}'\,\gamma + \alpha_s + \varepsilon_{hvds}, \tag{1}
$$

where $$y_{hvds}$$ is the relevant outcome for household $$h$$ in village $$v$$,
district $$d$$, stratum $$s$$; $$\text{Voucher}_{ds}$$ is an indicator equal to 1 if
district $$d$$ was randomly assigned to receive the voucher program in 2018;
$$\mathbf{X}_{hvds}$$ is the vector of double-LASSO-selected control variables;
$$\alpha_s$$ is a stratum fixed effect; and $$\varepsilon_{hvds}$$ is the error term.
Standard errors are clustered at the district level; randomization-inference p-values
from 1,000 permutations (Young 2019) are reported in brackets in all tables.

**Outcome variables and samples.** The paper estimates equation (1) on several
outcomes: total subsidy received (Rp/month, the sum of Rastra and BPNT values),
an indicator for receiving any subsidy, total food consumption of rice and eggs
(from the separate SUSENAS consumption module), rice quality (a 0-1 Likert
scale), rice price (for non-eligible households to avoid compositional effects),
and the poverty indicator. Results are presented for (i) the full sample, (ii)
households with PMT score ≤ 30 at baseline (the approximate target population,
PMT ≤ 30 being the program eligibility threshold), and (iii) PMT > 30 (those
not targeted). For the poverty analysis (Table 2, p. 537), the sample is further
restricted to PMT ≤ 25, ≤ 20, ≤ 15, ≤ 10, and ≤ 5 to document heterogeneous
poverty effects at different points of the distribution.

**Price specification.** To isolate the general-equilibrium price effect,
equation (1) is estimated with rice price as the outcome for households not in the
UDB (i.e., those ineligible for the programs, whose reported prices are not
affected by selection into which program they receive). Heterogeneity by supply
shock size and geographic isolation is assessed through interaction terms
$$\text{Voucher}_{ds} \times \text{Variable}_{d}$$ (Table 4, cols. 2-7, p. 541).

**Subsidy-fidelity specification.** To examine overall leakage at the district
level, the unit of observation becomes the district: the fraction of intended
subsidy actually received (subsidy received from SUSENAS divided by intended
subsidy, computed from the official number of beneficiaries times the
entitlement amount) is regressed on $$\text{Voucher}_{ds}$$ with district-level
strata fixed effects (Table 5, p. 544; N = 105 districts).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| SUSENAS (Survei Sosial Ekonomi Nasional) | Primary outcome data: subsidy receipt, food consumption, prices, poverty; March 2018 (baseline) and March 2019 (endline) waves | No page yet |
| Unified Targeting Database (UDB) | Household-level PMT scores and baseline characteristics for control selection and heterogeneity analysis; 2015 data merged by the government using national IDs; deidentified version in replication package | No page yet |
| PODES (Potensi Desa) village census | Village-level baseline control variables (roads, infrastructure, remoteness measures); 2018 wave | No page yet |
| Program administrative data | District-level intended subsidy disbursements (number of official beneficiaries times entitlement) for leakage calculations | No page yet |

Sample scope: 105 districts across Indonesia; approximately one-fifth of Indonesia's
population (53 million individuals); 3.4 million targeted beneficiary households.
Primary analysis uses household-level March 2019 SUSENAS (endline), approximately
66,000 households. Merged deidentified replication data available at
[https://doi.org/10.3886/E167262V1](https://doi.org/10.3886/E167262V1).

## When to read the full paper

Use the [original](https://doi.org/10.1257/aer.20210461) if you are:
studying the design and analysis of large-scale RCTs in the presence of
general-equilibrium effects (Muralidharan and Niehaus 2017); evaluating the
relative merits of in-kind vs. voucher / cash transfer programs in settings with
limited administrative capacity; replicating (the ICPSR replication package at
[https://doi.org/10.3886/E167262V1](https://doi.org/10.3886/E167262V1) contains
all code and data); or reading for the price-effects analysis of the transition
(Section III.B; Cunha, De Giorgi, and Jayachandran 2019 predictions tested at
scale). Table 1 (p. 529) gives the full delivery and targeting results; Table 2
(p. 537) the poverty heterogeneity; Table 3 (p. 539) the consumption results.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(2). This distillation was
extracted by an LLM on 2026-06-24 and is **not human-verified or independently
reproduced**. The AEA copyright applies; no CC licence is recorded in Crossref.
Extract-only; the verbatim PDF is not hosted here.

> Banerjee, Abhijit, Rema Hanna, Benjamin A. Olken, Elan Satriawan, and Sudarno
> Sumarto. "Electronic Food Vouchers: Evidence from an At-Scale Experiment in
> Indonesia." *American Economic Review* 113, no. 2 (February 2023): 514-547.
> DOI: 10.1257/aer.20210461. Copyright 2023 American Economic Association.
> Replication data: DOI 10.3886/E167262V1. This page is an extract by the
> Institute for Automated Research: core results and equations summarized.


==============================================================================
# Information, Mobile Communication, and Referral Effects: Barwick, Liu, Patacchini & Wu (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/barwick-et-al-information-mobile-communication-referral-2023/
# Distilled: Using geocoded cellphone records from a Chinese telecom provider matched to administrative firm data, the paper provides the first direct evidence of increased communication between job seekers and their referrers around job changes (inverted U-shape peaking at the switch month), quantifies a referral effect of 0.35 on job location choice (nearly tripling the baseline probability), and shows referral jobs yield higher wages, shorter commutes, and faster firm growth. American Economic Review 2023, paywalled. Eight core results with source locators, datasets used, the identification strategy, and estimating equations.
# Tags: paper-summary, labor-economics, social-networks, information-economics, urban-economics
==============================================================================

**What this is.** The paper's core results, the identification strategy, and the estimating
equations: enough to know what it found and how, without reading all 38 pages. To replicate
or extend, read the full source at the [original](https://doi.org/10.1257/aer.20200187).

## TL;DR

The paper exploits geocoded cellphone records from a major Chinese telecom provider to study
whether social contacts (referrers) transmit job-relevant information to job seekers. It
documents (i) an inverted U-shaped spike in call frequency between switchers and their
referrers in the months before a job change, with no corresponding pattern for non-referrer
friends; (ii) a referral effect of 0.35 on job location choice - having a social contact
working at a location nearly triples the probability of switching there; and (iii) referral
jobs are of higher quality: they pay more, involve shorter commutes, are more likely to be
full-time, and lead to faster firm growth. Effect heterogeneity shows referrals matter
especially when information asymmetry is more severe, as for young workers, rural-to-urban
movers, and sector-changers. Topa (2001) provides the foundational social-interactions
framework; Bayer, Ross, and Topa (2008) introduced the residential-neighbor proxy
approach that this paper extends using direct communication data.

## Core results

Magnitudes and significance are as reported; \*/\*\*/\*\*\* = 10%/5%/1%. Locators point
into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Referral effect on job location choice**: having a friend at location l increases probability of switching there by 0.35 | Table 3, col 2, p. 1186 | Coeff = 0.35 (SE 0.01); mean baseline probability = 0.09; N = 915,251 switcher-location pairs |
| R2 | **Inverted U-shape in referrer call frequency**: calls between switchers and referrers peak at the job switch month; nonreferrer calls are flat throughout | Figure 3, p. 1188 | Referrer-pair coefficient at month 0 approx +8 above baseline; nonreferrer pairs approx 0 throughout; 238,092 referrer-month obs vs 4,759,176 nonreferrer obs |
| R3 | **Referral effect amplified for high information-asymmetry groups**: rural-to-urban movers and sector-changers show substantially larger referral effects | Table 4, cols 5-6, p. 1191 | Friend x rural-to-urban = +0.32 (SE 0.05); Friend x changing sector = +0.21 (SE 0.02); baseline Friend 0.34 (col 5) / 0.32 (col 6) |
| R4 | **Referral wage premium**: referral jobs pay RMB 620 more per year, about 2 percent above the mean wage | Table 7, col 1, p. 1199 | 0.62 thousand RMB (SE 0.31); mean wage approx 31 thousand RMB/year; N = 17,615 |
| R5 | **Referral jobs are more likely to be full-time**: having a referrer at the new workplace raises the probability of a part-time to full-time transition by 1.4 percentage points | Table 7, col 3, p. 1199 | 0.014 (SE 0.007); approx 2% relative increase (p. 1198); N = 19,431 |
| R6 | **Referral jobs have shorter commutes**: referral raises probability of a shorter commute by 9 percentage points | Table 7, col 4, p. 1199 | 0.09 (SE 0.01); approx one-third of job changes involve a shorter commute; N = 29,117 |
| R7 | **Firm benefit: net labor inflow**: firms hiring through referrals gain 63 percent more workers (log net inflow) in the most saturated specification | Table 8, Panel A col 4, p. 1201 | gamma = 0.63 (SE 0.14); R-squared = 0.66 |
| R8 | **Firm benefit: matching rate**: firms hiring through referrals achieve an 84 percent higher job matching rate (log hires over vacancies) | Table 8, Panel B col 8, p. 1201 | gamma = 0.84 (SE 0.27); average matching rate = 1.53 for large firms |

**Overall (paper's conclusion).** Information provided by social contacts mitigates
information asymmetry in labor markets and facilitates better worker-firm matching.
The inverted U-shape in referrer communication around job changes, absent for
non-referrer friends, provides direct evidence that referrers pass job-relevant
information and rules out homophily and sorting as the sole explanation. Both
workers and firms benefit: referred employees earn more, commute less, and are
more likely to hold full-time positions, while firms that hire through referrals
grow faster and fill vacancies at higher rates.

## Theory / model

This paper has no formal economic model. It tests three linked hypotheses derived
from the theoretical literature on information transmission in labor markets
(Topa 2001):

1. **Information channel hypothesis.** Referrers pass job-relevant information to job
   seekers, generating an increase in communication intensity in the months before the
   job change. The prediction is an inverted U-shape in call frequency between referrer
   pairs centered on the event month, with no such spike for non-referrer friends.

2. **Referral effect hypothesis.** Having a social contact working at a given location
   raises the probability of switching there. The coefficient captures both information
   provision (the referrer informs the job seeker of an opening) and endorsement
   (the referrer vouches for the candidate to the employer). Under either channel,
   the referral should increase location choice probability.

3. **Information-asymmetry amplification.** The referral effect should be larger when
   information between workers and firms is more asymmetric: for young workers with
   limited labor market experience, rural-to-urban movers unfamiliar with urban job
   markets, and sector-changers whose skills are less observable. If the mechanism is
   pure preference (working near friends), no such heterogeneity would be predicted.

**Identification strategy.** The key threats are homophily (friends share unobserved
location preferences) and sorting (friends cluster in locations with unobserved job
opportunities). The paper addresses these via:

- **Origin-destination neighborhood-pair fixed effects** (equation 1, p. 1184):
  the referral coefficient beta is identified from within-pair variation, comparing
  job switchers who move between the same old-new neighborhood pair but have
  different social networks. This controls for all aggregate pair-specific attributes
  including industry composition, labor demand, and amenities.

- **Falsification tests using friend type** (Table 3, cols 3-4, p. 1186): friends
  who recently moved away from location l have a coefficient of 0.07 (far below
  the 0.35 for current referrers). Friends who live in the new location's
  neighborhood but do not work there have a coefficient of 0.15. If homophily or
  the specific neighborhood drove the result, these coefficients would be similar
  to the baseline. Current employment at the destination is what matters, consistent
  with information about job openings being the active ingredient.

- **Vacancy restriction** (Table 3, col 2, p. 1186): the baseline sample restricts to
  switchers with at least one alternative location in the same neighborhood offering
  the same occupation and salary range, ruling out the concern that the friend dummy
  proxies for the only available matching job.

- **Homophily controls** (Table 6, p. 1195): adding same-gender, same-age-group,
  same-birth-county, same-housing-price controls and k-means cluster dummies for
  switcher-friend pairs leaves the estimate stable at 0.33-0.34, showing the baseline
  controls adequately capture sorting.

The paper also compares its call-based referral measure with proxy-based approaches
common in the literature: residential neighbors in the spirit of Bayer, Ross, and
Topa (2008) (coefficient 0.21) and same-birth-county coworkers (0.10). The
call-based measure dominates both proxies by a margin that is statistically
significant at the one-percent level (Table 5, col 3-4, p. 1192).

## Method

The paper applies OLS panel regression with fixed effects and an event study. The
builds on `panel-regression` and `event-study` technique primitives. The four
estimating equations are introduced on pp. 1181-1184 and 1197-1200.

**Preliminary correlation (Table 2, p. 1182).** The relationship between
information flow (call volume) and worker flows across neighborhood pairs is
established via OLS with origin and destination fixed effects. Adding call volume
as a regressor raises the R-squared from 0.037 to 0.17, and doubling call volume
is associated with a 16 percent increase in worker flows (inverse-hyperbolic-sine
specification). This motivates using communication intensity as a proxy for
information provision.

**Main referral regression (eq. 1, p. 1183-1184).** Let $$\text{M}_{il} = 1$$ if job
switcher $$i$$ moves to location $$l$$ within the new workplace neighborhood. The
specification restricts individual $$i$$'s choice set to locations within the destination
neighborhood to absorb heterogeneity across neighborhoods:

$$M_{il} = \beta \text{Friend}_{il} + \mathbf{X}_i \mathbf{Z}_l \gamma + \lambda_{\tilde{c},c} + \epsilon_{il} \tag{1}$$

where $$\text{Friend}_{il} = 1$$ if at least one of $$i$$'s social contacts works at location
$$l$$ three months before the job switch; $$\mathbf{X}_i$$ = individual demographics (gender,
age groups, migration status, total social contacts); $$\mathbf{Z}_l$$ = location amenities
(restaurants, roads and parking lots, schools within 500 m radius); $$\lambda_{\tilde{c},c}$$ = old-by-new
neighborhood-pair fixed effects (20,811 total pairs in unrestricted sample; 16,468 in
the baseline vacancy-restricted sample). Standard errors are clustered at the neighborhood-pair
level.

**Event study (p. 1187).** Call frequency between switcher $$i$$ and friend $$j$$ in
month $$t$$ is regressed on event-time dummies interacted with referrer vs non-referrer
status, covering an event window from 11 months before to 9 months after the job switch
(month $$s = -1$$ is the reference category):

$$\text{Freq}_{ijt} = \sum_{s=-11}^{9} \gamma_s \text{Referral}_{ij} \cdot \mathbf{1}\{t = s\} + \sum_{\substack{s=-11 \\ s \neq -1}}^{9} b_s \text{Nonreferral}_{ij} \cdot \mathbf{1}\{t = s\} + \lambda_i + \tau_t + \epsilon_{ijt}$$

where $$\lambda_i$$ are individual fixed effects and $$\tau_t$$ are calendar month fixed effects.
The coefficients $$\{\gamma_s, b_s\}$$ capture changes in call frequency relative to
the individual's own baseline rate of talking to non-referrer friends. Standard errors
are clustered at the individual level.

**Worker outcome regression (eq. 2, p. 1197).** Labor market outcomes of the referral job:

$$Y_{ilr} = \beta \text{Friend}_{ilr} + \mathbf{X}_i \mathbf{Z}_l \gamma + \lambda_c + \alpha_r + \epsilon_{ilr} \tag{2}$$

where $$Y_{ilr}$$ is a labor outcome for worker $$i$$ at new work location $$l$$ in
residential neighborhood $$r$$; $$\lambda_c$$ = new work neighborhood fixed effect;
$$\alpha_r$$ = residential neighborhood fixed effect; controls $$\mathbf{X}_i$$ include
gender, age groups, migration status, and log number of social contacts. Standard
errors are two-way clustered by residential and new work neighborhood.

**Firm performance regression (eq. 3, p. 1200).** Log firm outcomes:

$$Y_i = \gamma \text{Referral}_i + \mathbf{Z}_i \beta + \lambda_c + \epsilon_i \tag{3}$$

where $$\text{Referral}_i = 1$$ if at least one new hire at firm $$i$$ has a social contact
already working there; $$Y_i \in \{\log(\text{net inflow}),\, \log(\text{matching rate}),\,
\log(\text{growth rate})\}$$; $$\lambda_c$$ = neighborhood fixed effect; $$\mathbf{Z}_i$$
includes firm age, 18 industry dummies, SOE dummy, average employees 2010-2015, average
capital stock 2010-2015, previous employment growth rate, share of female workers, share
of migrants, average employee age, average employee housing price, and the firm's referral
network size. Standard errors are clustered at the neighborhood level.

## Empirical specifications

**Baseline specification (Table 3, p. 1186).** Equation (1) estimated on 915,251
switcher-location pairs (restricted to individuals facing at least one alternative same-neighborhood,
same-occupation, same-salary-range opening). Old-by-new neighborhood-pair fixed effects
(16,468 pairs). The mean within-neighborhood switching probability is 0.09. The baseline
referral coefficient is 0.35 (col 2). Columns 3-4 add falsification friend types (moved-away:
0.07; lives-but-not-works: 0.15); column 5 replaces direct friends with friends-of-friends
(coefficient 0.14, confirming these second-degree links carry less job information); column 6
adds extensive local labor market controls (0.33, similar to baseline).

**Effect heterogeneity (Table 4, p. 1191).** Equation (1) augmented with interaction
terms $$\text{Friend}_{il} \times X_i$$, one per column: (i) distance between old and new
workplaces, coefficient 0.002 (SE 0.0004); (ii) distance between home and new workplace,
0.002 (SE 0.0003); (iii) young dummy (ages 25-34), +0.04 (SE 0.01); (iv) rural-to-urban
switch, +0.32 (SE 0.05); (v) sector-change dummy, +0.21 (SE 0.02). For rural-to-urban
movers and sector-changers the point estimates of the total referral effect are 0.66 and
0.53 respectively, substantially above the base estimate.

**Event study (Figure 3, p. 1188).** 238,092 switcher-referrer-month observations and
4,759,176 switcher-nonreferrer-month observations. The referrer-pair call frequency coefficient
rises from near zero at month -9 to a peak of approximately +8 above baseline at month 0,
then remains elevated post-switch as referrers become coworkers. Nonreferrer friends show
coefficients near zero throughout. Falsification event studies (Figure 4, p. 1190) show that
moved-away and lives-at-new-location friends display flat or mildly elevated patterns with
no inverted U-shape, confirming the information spike is specific to current employment
at the destination.

**Comparison with literature proxies (Table 5, p. 1192).** Equation (1) replaces the
call-based Friend dummy with (i) residential neighbor dummy (0.21, SE 0.01) and (ii)
same-birth-county coworker dummy (0.10, SE 0.01). Columns 3 and 4 each include one
proxy alongside the call-based measure. In column 3, the residential neighbor coefficient
falls to 0.18 (SE 0.01) while the call-based Friend (not neighbor) coefficient is 0.25
(SE 0.01). In column 4, the same-birth-county coefficient falls to 0.09 (SE 0.01) while
the call-based Friend (not same birth county) coefficient is 0.35 (SE 0.03). Both proxy
estimates decline when the direct communication measure is included, confirming the
proxies capture genuine but attenuated social interactions, consistent with Granovetter
(1973) and the approach of Gee, Jones, and Burke (2017).

**Worker benefits (Table 7, p. 1199).** Equation (2) estimated separately for five
outcome variables. Sample sizes vary by outcome due to data availability. All columns
include residential and new-work neighborhood fixed effects. Key results: wage = 0.62
thousand RMB (SE 0.31, N=17,615); coworker housing price difference = 0.07 thousand
RMB/m$^2$ (SE 0.04, N=23,323); PT-to-FT = 0.014 (SE 0.007, N=19,431); shorter
commute = 0.09 (SE 0.01, N=29,117); non-SOE to SOE = 0.012 (SE 0.005, N=15,881).

**Firm benefits (Table 8, p. 1201).** Equation (3) estimated on large-firm locations
(firms with more than 100 employees) to reduce spurious worker-firm linking. In the most
saturated specification (column 4/8/12 for each panel), gamma = 0.63 (SE 0.14) for log
net inflow, 0.84 (SE 0.27) for log matching rate, and 0.45 (SE 0.11) for log firm
growth rate. The estimates are stable across specifications with progressively richer
firm and employee controls, arguing against upward bias from fast-growing firms being
more likely to use referrals.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Geocoded cellphone records, Company A (anonymous northern Chinese city) | Main analysis: social network construction, information-flow measures, work and home location histories for 456,000 users, Nov 2016-Oct 2017 | No page yet (proprietary; provider anonymous under data-sharing agreement) |
| Administrative firm-level records (merged by location) | Industry composition, average payroll, number of employees, capital stock; used for firm-performance regressions and location controls | No page yet (Chinese administrative data; not publicly available) |
| Job postings data (unnamed online platform) | Occupation and salary range at each location; used to restrict baseline sample to switchers facing comparable alternative opportunities | No page yet |
| Residential housing price data | Proxy for coworker socioeconomic status; used as a nonwage benefit measure (delta coworker HP) | No page yet |
| China Family Panel Studies (CFPS, 2014) | Descriptive only: national average demographics and job-search method frequencies for comparison with sample (Figure 1, Table 1) | No page yet |
| US Current Population Survey (2014) | Descriptive only: US job-search method frequencies for cross-country comparison (Figure 1) | No page yet |

Sample: November 2016 to October 2017 (12 months). Final analysis sample: 456,000
individuals with stable work locations for at least 45 weeks and at most two work
locations; 38,102 job switchers (8 percent of sample). Social contacts defined as
anyone with at least one call to or from individual i in the three months prior to the
job switch; on average 50 percent of a user's friends are Company A customers.

## When to read the full paper

Use the [original](https://doi.org/10.1257/aer.20200187) if you are: measuring
information transmission in labor markets with mobile phone data or other digital
footprints; designing or evaluating employee referral programs and need evidence
on the heterogeneity of referral effects by worker type; working on urban labor
mobility in developing economies where formal job-search institutions are weak;
extending the event-study design to other communication technologies (WeChat,
messaging apps) or other information channels; or benchmarking referral-effect
magnitudes for structural job-search models. The online appendix (referenced in
the paper and in the replication package at
[10.3886/E183161V1](https://doi.org/10.3886/E183161V1)) contains detailed
robustness tables and the event study for unemployed job-seekers.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(5), May 2023. Replication
data available at [10.3886/E183161V1](https://doi.org/10.3886/E183161V1). This
distillation was extracted by an LLM on 2026-06-24 and is **not human-verified or
independently reproduced**. No CC licence was found in Crossref metadata; standard
AEA copyright applies; extract-only.

> Barwick, Panle Jia, Yanyan Liu, Eleonora Patacchini, and Qi Wu.
> "Information, Mobile Communication, and Referral Effects."
> *American Economic Review* 113, no. 5 (May 2023): 1170-1207.
> DOI: 10.1257/aer.20200187. Copyright 2023 American Economic Association.
> This page is an extract only; it is not a substitute for the original.


==============================================================================
# Alternative Explanation for the Fed Information Effect: Bauer & Swanson (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/bauer-swanson-alternative-explanation-fed-information-2023/
# Distilled: Bauer and Swanson (2023) show that standard "Fed information effect" regressions suffer from omitted variable bias; once economic news controls are added, monetary policy surprise coefficients reverse sign to match standard macroeconomic theory. A "Fed response to news" channel, supported by their own forecaster survey and financial market evidence, explains the data without invoking Fed private information. American Economic Review 2023, AEA copyright. Seven core results with source locators, datasets used, the model (imperfect information about the policy rule), and the method (OLS with news controls, high-frequency event study).
# Tags: paper-summary, monetary-policy, monetary-economics, central-banking
==============================================================================

**What this is.** The paper's core results, the model (imperfect information about the Fed's policy rule), and the method (OLS with economic news controls plus high-frequency financial-market evidence): enough to see what it found and how, without reading all 37 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1257/aer.20201220) and the [replication package](https://doi.org/10.3886/E181661V1).

## TL;DR

Bauer and Swanson (2023) challenge the "Fed information effect" (FIE): the finding that monetary policy tightenings are associated with *upward* revisions in private-sector GDP and employment forecasts, which prior literature interpreted as the Fed revealing positive private information about the economy. They make four main arguments. First, economic news released in the weeks between the Blue Chip survey and the FOMC announcement is an important omitted variable in standard FIE regressions. Second, once that news is controlled for (regressions with full news vectors), the monetary policy surprise coefficients reverse sign and become consistent with standard macroeconomic theory: hawkish surprises reduce GDP forecasts and raise unemployment forecasts. Third, their direct survey of all 52 Blue Chip forecasters (Section III) confirms that forecasters do not revise in the information-effect direction. Fourth, high-frequency stock market and exchange rate responses to FOMC announcements are equally negative for the "most influential" FIE announcements as for all other announcements, with no sign of a positive information channel. The alternative explanation, which they call the "Fed response to news" channel, is that both the Fed and private forecasters respond to the same public economic data, but the Fed responded more strongly than markets anticipated, generating a spurious positive correlation in simple regressions.

## Core results

Magnitudes and significance as reported. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Standard FIE regressions of Blue Chip forecast revisions on monetary policy surprises have very low R² and fragile coefficients with signs opposite to standard macro theory | Table 1, pp. 672-674 | R² = 0.00-0.06 across unemployment, GDP, and CPI inflation and across samples (N=120, 129, 206, 217); estimates sensitive to sample period and variable forecast |
| R2 | Economic news strongly predicts Blue Chip forecast revisions, confirming the omitted variable | Table 2, p. 677 | R² = 0.64 (unemployment), 0.40 (GDP), 0.31 (CPI inflation); unemployment surprise coefficient = 0.308 (se=0.037); S&P500 change coefficient on GDP = 0.620 (se=0.167) |
| R3 | Economic news predicts high-frequency monetary policy surprises, establishing omitted variable bias in FIE regressions | Table 3, pp. 679-680 | R² = 0.12-0.20 across target, path, and NS surprise measures; S&P500 coefficient \approx 0.15 across all three MPS measures |
| R4 | Controlling for economic news eliminates the FIE: MPS coefficients reverse to conventional signs and R² rises to 31-65% | Table 4, p. 682 | Target factor on GDP = -0.241 (se=0.145); target on unemployment = +0.152 (se=0.074); target on inflation = +0.067 (se=0.088); R² rises from 0-6% to 31-65% |
| R5 | Survey of 52 Blue Chip forecasters: no respondent revises GDP forecast upward after a hawkish surprise; 18 of 23 who revise do so conventionally | Table 5, pp. 683-686 | 0/36 upward GDP revision after hawkish surprise; 18/36 downward; 13/36 do not revise at all in response to the funds rate decision |
| R6 | High-frequency stock market and exchange rate responses to FOMC announcements are equally negative for the ten "most influential" FIE observations | Table 7, p. 689 | S&P500: -8.04 (se=1.91) for top-10 FIE observations vs. -7.14 (se=1.84) for remaining 110 obs; difference not significant |
| R7 | Fed Greenbook and Blue Chip forecast RMSEs are essentially identical; no systematic Fed information advantage | Table 8, pp. 690-692 | Unemployment 0-3Q avg RMSE: GB=0.42, BC=0.42; GDP 0-3Q avg: GB=1.64, BC=1.60; Diebold-Mariano p-values mostly above 0.05 |

**Overall (paper's conclusion).** The response of Blue Chip macroeconomic forecast revisions to FOMC announcements can be fully explained by omitted economic news variables, without invoking the Fed information effect. Once news is controlled for, the monetary policy surprise coefficients become consistent with standard macro models and VARs. The "Fed response to news" channel, in which both the Fed and private forecasters respond to the same publicly available economic data, explains all the empirical patterns.

## Theory / model

Section V (pp. 692-696) presents a simple partial-equilibrium model with imperfect information about the Fed's policy rule. It contains no Fed information effect by construction and illustrates the "Fed response to news" channel.

**Output gap** follows an exogenous AR(1) (eq. 10, p. 693):

$$
x_t = \rho_x x_{t-1} + \eta_t, \qquad \eta_t \sim i.i.d.\, N(0, \sigma_\eta^2), \quad \rho_x \in [0,1)
$$

**Monetary policy rule:** the central bank sets the interest rate linearly in the output gap (eq. 11, p. 693):

$$
i_t = a x_t + \varepsilon_t, \qquad \varepsilon_t \sim i.i.d.\, N(0, \sigma_\varepsilon^2)
$$

The parameter $$a > 0$$ denotes the central bank's responsiveness to the output gap and is known to the central bank but NOT to the private sector. The private sector maintains prior beliefs $$a \sim N(\hat{a}_t, \sigma^2_{a_t})$$ updated from history $$\mathcal{H}_{t-1} = \{i_s, x_s, i_{s-1}, x_{s-1}, \ldots\}$$. The core uncertainty is about $$a$$, not about $$x_t$$ (which is observed by all).

**Expected future rates** (after observing $$x_t$$ but before the FOMC announcement; eq. 12, p. 694):

$$
E[i_{t+j} | x_t, \mathcal{H}_{t-1}] = \hat{a}_t \rho_x^j x_t
$$

**Monetary policy surprise** (the gap between announced $$i_t$$ and prior expectation; eq. 13, p. 694):

$$
mps_t \equiv i_t - E[i_t | x_t, \mathcal{H}_{t-1}] = (a - \hat{a}_t) x_t + \varepsilon_t
$$

This shows that $$mps_t$$ is driven both by the pure exogenous shock $$\varepsilon_t$$ and by the private sector's uncertainty about $$a$$, captured by $$(a - \hat{a}_t) x_t$$. If markets have persistently underestimated $$a$$ (i.e., $$\hat{a}_t < a$$), then $$mps_t$$ will be positively correlated with $$x_t$$, exactly as found empirically in Tables 2 and 3. Cieslak (2018) provides direct evidence that financial markets systematically underestimated the Fed's responsiveness to the economy over this period, consistent with $$\hat{a}_t < a$$ persisting for many periods.

**Bayesian belief update** (eq. 14, p. 694): after observing $$mps_t$$, the private sector updates:

$$
E[a | \mathcal{H}_t] = \hat{a}_t + \omega_t \frac{1}{x_t} mps_t, \qquad \omega_t \equiv \frac{x_t^2 \sigma^2_{a_t}}{x_t^2 \sigma^2_{a_t} + \sigma^2_\varepsilon}
$$

**Interest rate forecast revision** (eq. 15, p. 694): the updated belief implies a forecast path revision of:

$$
E[i_{t+j} | \mathcal{H}_t] - E[i_{t+j} | x_t, \mathcal{H}_{t-1}] = \rho_x^j \omega_t \, mps_t
$$

So interest rate path revisions are a positive, horizon-declining function of $$mps_t$$, replicating the empirical pattern of Gurkaynak, Sack, and Swanson (2005b) without any private Fed information.

**Implication for identification.** Even though $$mps_t$$ may be correlated with $$x_t$$ ex post, it can still be used without adjustment to estimate the causal effects of the exogenous monetary policy shock $$\varepsilon_t$$ on asset prices in narrow-window event-study regressions, because interest rate expectations respond only to $$mps_t$$ and not separately to $$\varepsilon_t$$ (p. 695). However, using $$mps_t$$ as an instrument in structural VARs or local projections is problematic because the exogeneity condition is violated: $$mps_t$$ is correlated ex post with structural shocks to $$x_t$$ (p. 695-696).

## Method

The paper applies three empirical designs. This section states the estimating equations; the identifying assumptions are discussed in the Empirical specifications section. The method builds on `panel-regression` and `event-study`.

**Design 1: Blue Chip forecast revision regressions (replication and extension, eqs. 2-3, pp. 672-673).** Following Campbell et al. (2012) (henceforth CEFJ):

$$
\text{BCrev}_t = \alpha + \beta \, \text{target}_t + \gamma \, \text{path}_t + \varepsilon_t \tag{2}
$$

where $$\text{BCrev}_t$$ is the one-month revision in the Blue Chip consensus forecast, averaged over the 1-, 2-, and 3-quarter-ahead horizons; $$\text{target}_t$$ and $$\text{path}_t$$ are GSS high-frequency factors computed from short-maturity federal funds and Eurodollar futures in a 30-minute window around the FOMC announcement. The Nakamura and Steinsson (2018) variant uses a single composite surprise $$\text{mps}_t$$ (first principal component):

$$
\text{BCrev}_t = \phi + \theta \, \text{mps}_t + \eta_t \tag{3}
$$

Standard errors are bootstrapped (50,000 replications) to account for the generated-regressor nature of the GSS factors.

**Design 2: Economic news controls (eqs. 4-6, p. 676).** The corrected specification adds a news vector $$\text{news}_t$$:

$$
\text{BCrev}_t = \alpha + \beta \, \text{target}_t + \gamma \, \text{path}_t + \delta'\text{news}_t + \varepsilon_t \tag{4}
$$

and analogously for the NS version. $$\text{news}_t$$ includes: unemployment surprise, payrolls surprise, GDP surprise, BBK composite business-cycle index, lagged core CPI measures, core CPI surprise, log change in the S&P500, change in yield curve slope, and log change in commodity prices, all pre-dating the FOMC announcement. Economic news also predicts the monetary policy surprises themselves (eq. 7, p. 679):

$$
mps_t = \alpha + \beta'\text{news}_t + \varepsilon_t \tag{7}
$$

**Design 3: Financial market event study (eq. 8, p. 688).** High-frequency stock and exchange rate regressions:

$$
\Delta \log x_t = \phi + \theta \, \text{mps}_t + \eta_t \tag{8}
$$

run separately over the ten most influential FIE observations and the remaining observations.

**Design 4: Greenbook encompassing regressions (eq. 9, p. 691).** Following Romer and Romer (2000):

$$
X_{t+h} = \alpha + \beta \hat{X}^{GB}_{t+h|t} + \gamma \hat{X}^{BC}_{t+h|t} + \varepsilon_{t+h} \tag{9}
$$

where $$X_{t+h}$$ is the realized macro variable, $$\hat{X}^{GB}$$ and $$\hat{X}^{BC}$$ are the Greenbook and Blue Chip forecasts. A coefficient test asks whether either forecast dominates. Hansen-Hodrick standard errors with $$2(h+1)$$ lags for overlapping horizons.

## Empirical specifications

**Section I (Table 1, R1).** Regressions (2) and (3) are run on four samples: CEFJ replication sample (1990-2007, N=129); NS replication sample (1995-2014, N=120); full sample including unscheduled FOMC announcements (1990-2019, N=217); full sample excluding unscheduled (N=206). Three outcomes: unemployment, GDP, CPI inflation. Identification rests on the assumption that the 30-minute monetary policy surprise window is exogenous to monthly forecast revision determinants; the paper shows this assumption is violated by the economic news omitted variable.

**Section II (Tables 2-4, R2-R4).** Full sample, N=217. Table 2 runs eq. (6) with $$\text{BCrev}$$ as outcome; confirms news has R² of 31-64% for forecast revisions. Table 3 runs eq. (7) with MPS as outcome; confirms news has R² of 12-20% for monetary policy surprises. Table 4 runs eqs. (4) and (5) simultaneously, adding all news controls. Identifying assumption: conditional on economic news, the residual variation in the FOMC surprise is uncorrelated with other determinants of forecast revisions (selection on observables). Standard errors bootstrapped (50,000 reps) throughout.

**Section III (Table 5, R5).** Original survey of 52 Blue Chip Economic Indicators forecasting firms, conducted July-August 2019; 36 responses (70% response rate). Each firm was asked how it revises its GDP, unemployment, and CPI forecasts in response to four components of FOMC announcements: (i) the funds rate decision, (ii) the FOMC statement, (iii) the dot plot, and (iv) the SEP forecasts. Survey answers are self-reported and directional (up/down/no change), not quantitative.

**Section IV (Tables 6-8, R6-R7).** The ten "most influential" FIE observations in the NS regression are identified by the change in the regression t-statistic when that observation is excluded. Regression (8) is run separately for these ten and the remaining 110 NS observations (heteroskedasticity-consistent standard errors). Greenbook comparison (eq. 9) uses 1990-2013 (N=192 observations matched on timing), horizons h=0,1,2,3 quarters, plus the 0-3 quarter average.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Blue Chip Economic Indicators survey (Wolters Kluwer) | Monthly consensus forecasts of GDP growth, unemployment, CPI inflation used as $$\text{BCrev}_t$$ dependent variable; monthly revisions in Sections I-III | [Blue Chip Financial Forecasts](/wiki/commercial/blue-chip-forecasts/) (licensed) |
| GSS monetary policy surprise factors (Gurkaynak, Sack, Swanson 2005b) | Target and path surprise factors from federal funds and Eurodollar futures in 30-minute FOMC windows; also NS first-principal-component surprise | No page yet |
| Money Market Services survey | Market expectations of upcoming BLS/BEA data releases; used to compute the "surprise" component of each macro release | No page yet |
| BLS employment report, BEA GDP release, BLS CPI release | Unemployment surprise, payrolls surprise, GDP surprise, core CPI surprise (components of $$\text{news}_t$$) | No page yet |
| BBK composite business cycle index (Brave, Butters, Kelley 2019) | Comprehensive single monthly business activity index; included in $$\text{news}_t$$ | No page yet |
| S&P500, USD/EUR exchange rate, commodity price index | Financial news controls (lagged in $$\text{news}_t$$) and event-study outcomes in Section IV | No page yet |
| Federal Reserve Greenbook forecasts | Fed's internal forecasts of unemployment, GDP, CPI; compared against Blue Chip accuracy in Section IV.B; public after 5-year lag, available through Dec 2013 | No page yet |
| Authors' own survey of Blue Chip forecasters (July 2019) | Hand-collected survey of 52 professional forecasting firms on how they revise forecasts in response to FOMC announcements; Section III; published in the replication package | No page yet |

Sample: monthly FOMC announcement months, January 1990 to June 2019 (N=217 including unscheduled; N=206 excluding). NS subsample: January 1995 to March 2014 (N=120). CEFJ subsample: January 1990 to June 2007 (N=129). Greenbook comparison: 1990-2013 (N=192).

## When to read the full paper

Read the [original](https://doi.org/10.1257/aer.20201220) if you are: (i) using high-frequency monetary policy surprises as instruments in a structural VAR or local projections framework, where the paper's recommendation to purge the "Fed response to news" component before using them as instruments is directly relevant (Section V, pp. 695-696 and the forthcoming companion paper); (ii) studying whether FOMC announcements transmit private central-bank information to private forecasters (the core question); (iii) replicating or extending the Nakamura and Steinsson (2018) or Campbell et al. (2012) results (exact sample construction details and bootstrap procedure are in Sections I-II and online appendices); or (iv) comparing Fed Greenbook and Blue Chip forecast accuracy (Section IV.B, Table 8 panel structure). The [replication package](https://doi.org/10.3886/E181661V1) at ICPSR contains the data and code.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(3), March 2023. AEA copyright; freely accessible on AEAweb.org past the 3-year embargo (elapsed March 2026). No Creative Commons licence; extract-only. This distillation was extracted by an LLM on 2026-06-24 and is **not human-verified or independently reproduced**.

> Bauer, Michael D., and Eric T. Swanson. "An Alternative Explanation for the 'Fed Information Effect'." *American Economic Review* 113, no. 3 (March 2023): 664-700. DOI: [10.1257/aer.20201220](https://doi.org/10.1257/aer.20201220). Copyright 2023 American Economic Association.


==============================================================================
# Nobel Lecture, Banking and Credit: Bernanke (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/bernanke-nobel-lecture-banking-credit-2023/
# Distilled: Ben Bernanke's Nobel Prize lecture synthesizes his career research showing that informational frictions in credit markets interact with borrower and lender net worth to amplify and prolong economic contractions. The lecture documents that banking and credit disruptions were important sources of the Great Depression and the Great Recession of 2007-2009, and introduces the financial accelerator mechanism through which credit conditions propagate business cycles. American Economic Review 2023, copyright The Nobel Foundation 2022, paywalled. Eight core results with source locators, the Appendix model (moral hazard and credit rationing, eqs. 1-9), and the financial accelerator channel.
# Tags: paper-summary, banking, credit-markets, financial-crises, great-depression
==============================================================================

**What this is.** This is the LLM-distilled skeleton of Ben Bernanke's Nobel Prize lecture, a revised version of the talk delivered in Stockholm on December 8, 2022, published in the American Economic Review. Read the original at https://doi.org/10.1257/aer.113.5.1143 to replicate or extend. Quantitative results in the Core results table originate from studies cited in the lecture; the Appendix model (pp. 1165-1167) is the lecture's own original theoretical content.

## TL;DR

Credit markets are permeated by informational frictions: borrowers know more than lenders about their own riskiness and effort. These frictions, identified by Stiglitz and Weiss (1981) and formalized in the principal-agent framework of Ross (1973), create a wedge between the cost of external and internal finance (the external finance premium) that rises when borrower and lender net worth falls. When a financial shock destroys net worth, the premium spikes, credit contracts sharply, and real activity falls and stays depressed long after the initial shock - a mechanism Bernanke and Gertler (1989) embed in a dynamic general equilibrium model and call the financial accelerator.

Bernanke synthesizes evidence across two crises: the Great Depression of the 1930s and the Global Financial Crisis (GFC) and Great Recession of 2007-2009. In both episodes, banking and credit-market disruptions amplified and prolonged contractions beyond what money-supply declines alone can explain, contradicting the strict monetarist reading of Friedman and Schwartz (1963). A simple Appendix model formalizes how limited liability and moral hazard prevent low-net-worth entrepreneurs from obtaining credit, even at higher interest rates.

## Core results

| # | Result | Locator | Magnitude as reported |
|---|--------|---------|----------------------|
| R1 | Bank loan-to-deposit ratio collapsed (Great Depression) | p. 1150 | Fell from 0.85 (1929) to 0.58 (January 1933), even as deposits shrank |
| R2 | Baa-Aaa credit spread widened sharply (Great Depression) | p. 1151 | Rose from approximately 2.5 pp (1929-1930) to approximately 8 pp (mid-1932) |
| R3 | Banking crisis share of Depression manufacturing employment decline | p. 1154 | Approximately 22% of total 1929-1933 manufacturing employment decline due to banking crisis (Lee and Mezzanotti 2014) |
| R4 | Aggregate bank lending contraction 1929-1933 | p. 1155 | Reduced by 15% (Mitchener and Richardson 2019) |
| R5 | Bank credit/GDP declined sharply in the GFC | p. 1159 | Fell 8.5% between Q4 2008 and Q4 2013; did not regain December 2008 nominal level for nearly three years |
| R6 | Employment losses: pre- vs. post-Lehman contrast | p. 1160 | Fewer than 1.2M jobs lost Jan-Aug 2008; 5.4M jobs lost Sep 2008-Apr 2009 |
| R7 | Real GDP growth at the GFC trough | p. 1160 | Annualized -8.7% in Q4 2008, -4.7% in Q1 2009 |
| R8 | Net-worth credit threshold (Appendix model) | p. 1167, eq. (9) | Credit extended iff pR - r(1-w) >= [p/(p-q)]e; entrepreneurs below the endowment threshold receive no loans |

**Overall (paper's conclusion).** When Bernanke began this research in the late 1970s, the prevailing view treated financial markets as a veil with no independent macroeconomic role. Monetarists (Friedman and Schwartz 1963) and real-business-cycle economists both marginalized credit. The research synthesized here established banking and credit markets as central to macroeconomic fluctuations: credit-market disruptions are a quantitatively important cause of economic contractions, operating through balance-sheet channels that persist long after the initial shock.

## Theory / model

The paper's formal theoretical content is the Appendix: "Net Worth, Borrowing, and Investment" (pp. 1164-1167). It presents a stripped-down two-period moral hazard model illustrating why only borrowers with sufficient net worth can obtain credit. The model motivates the inverse relationship between net worth and the external finance premium that underlies the financial accelerator (Bernanke and Gertler 1989, 1990) and the credit channel of monetary policy (Bernanke and Gertler 1995).

**Setup (p. 1165).** There are $$M$$ risk-neutral agents with heterogeneous endowments $$w_i \in [0, 1]$$. A subset of $$N \ll M$$ agents ("entrepreneurs") can invest in risky projects. Each entrepreneur can choose either a "good" project (probability $$p$$ of producing $$R$$ units when effort $$e$$ is exerted) or a "fair" project (probability $$q < p$$ of producing $$R$$, no effort required). Project choice and effort are private information. The condition for the good project to be socially preferred (p. 1165, eq. 1) is:

$$pR - e \geq qR \tag{1}$$

An entrepreneur with endowment $$w_i$$ must borrow $$1 - w_i$$ from competitive financial intermediaries (coalitions of non-entrepreneurs) at opportunity cost $$r$$.

**Optimal contract (p. 1166, eqs. 2-5).** The intermediary maximizes expected profit subject to the entrepreneur's participation constraint (3), incentive-compatibility constraint (4), and limited liability (5):

$$\max\{pR - r(1 - w_i) - C_p^*\} \tag{2}$$

$$C_p^* - e \geq rw_i \tag{3}$$

$$C_p^* \geq C_q^* + e \tag{4}$$

$$C_s, C_u \geq 0 \tag{5}$$

where $$C_p^* = pC_s + (1-p)C_u$$ and $$C_q^* = qC_s + (1-q)C_u$$ are expected entrepreneur payments under the good and fair projects, and $$C_s, C_u$$ are state-contingent payments on success and failure.

**Zero-profit and credit threshold (p. 1166-1167, eqs. 6-9).** Under competitive intermediation, zero profits require:

$$C_p^* = pR - r(1 - w_i) \tag{6}$$

Setting $$C_u = 0$$ (which maximizes the entrepreneur's incentive to choose the good project) yields the incentive constraint:

$$(p - q)(C_s - C_u) \geq e \tag{7}$$

Substituting (6) with $$C_u = 0$$:

$$pC_s = pR - r(1 - w_i) \tag{8}$$

Inserting (8) into the incentive-compatibility constraint and rearranging gives the credit condition (p. 1167, eq. 9):

$$pR - r(1 - w_i) \geq \left[\frac{p}{p-q}\right] e \tag{9}$$

Equation (9) is the key result: only entrepreneurs whose endowment $$w_i$$ is large enough to satisfy this inequality receive loans. Below the threshold, the intermediary cannot simultaneously cover its opportunity cost and give the entrepreneur enough residual stake to deter the fair project. The higher the borrower's net worth $$w_i$$, the more the intermediary can pay on success and the less it needs to screen or monitor - a direct demonstration that access to credit depends on net worth, not just the project's expected return.

## Method

This paper is a Nobel lecture and survey; it introduces no new estimation method. The analytical framework it synthesizes rests on two building blocks: (i) the external finance premium, defined as the all-in cost of external finance net of the safe interest rate, and (ii) the net worth of borrowers and lenders as the key linking variable between financial conditions and economic activity (pp. 1145-1147). The external finance premium framework builds on Stiglitz and Weiss (1981) (credit rationing under asymmetric information) and the principal-agent analysis of Ross (1973).

The financial accelerator (Bernanke and Gertler 1989, 1990; Bernanke, Gertler, and Gilchrist 1999) is the main conceptual tool for business-cycle analysis: endogenous changes in net worth feed back into borrowing costs and investment, amplifying any initial shock. A downturn lowers net worth, raises the external finance premium, reduces borrowing and investment, and further reduces net worth. The credit channel of monetary policy (Bernanke and Gertler 1995) follows from this: unanticipated policy tightenings reduce asset values and net worth, raising the external finance premium and magnifying the total output effect beyond what the safe-rate channel implies.

For the Depression analysis, the empirical methodology in Bernanke (1983) is reduced-form time-series regression of industrial production on bank-failure measures alongside money growth - a descriptive design that documents forecasting power rather than identifying a structural causal effect. Subsequent work applies causal designs: instrumental variables for bank distress (Calomiris and Mason 2003), city-level variation in bank exposure (Lee and Mezzanotti 2014), and interbank contagion networks (Mitchener and Richardson 2019). For the GFC, Bernanke (2018) uses dynamic factor models to compare the forecasting power of lender-stress indicators against borrower-distress indicators.

## Empirical specifications

The lecture surveys empirical designs from prior work rather than reporting original estimation. The key specifications are as follows.

**Great Depression, aggregate time series (Section II, p. 1154).** Bernanke (1983) regressed industrial production on bank-failure variables and money growth in a vector autoregression. Bank-lending variables had significant forecasting power for industrial production over and above monetary aggregates - evidence that credit channels operate independently of money. Identification is descriptive; the challenge (noted in the lecture) is that bank credit and money move together, making it difficult to disentangle their effects without additional variation.

**Great Depression, cross-sectional and international (Section II, pp. 1154-1155).** Calomiris and Mason (2003) use Depression-era state and county data, instrumenting for bank distress using pre-crisis bank-risk predictors, and find that loan supply explains important cross-sectional variation in economic activity. Lee and Mezzanotti (2014) use city-level data to show that industries reliant on external finance were relatively more affected by local bank distress, attributing approximately 22% of manufacturing employment decline 1929-1933 to the banking crisis (p. 1154). Bernanke and James (1991) used a 24-country panel, finding that countries experiencing banking crises had sharper output declines controlling for gold-standard exit timing (Eichengreen and Sachs 1985).

**GFC, timing evidence (Section III, pp. 1159-1160).** The central empirical argument is timing: the recession became "great" only after the Lehman collapse intensified the financial crisis in fall 2008. In the eight months before Lehman (January to August 2008) the US economy lost fewer than 1.2 million jobs; in the eight months after (September 2008 to April 2009) it lost 5.4 million. Real GDP fell at annualized rates of 8.7% in Q4 2008 and 4.7% in Q1 2009, the sharpest post-WWII declines before the pandemic (p. 1160; data from FRED).

**GFC, dynamic factor models (Section III, p. 1159).** Bernanke (2018) used dynamic factor models to show that high-frequency indicators of lender stress are significantly better predictors of output, consumption, employment, and other macro variables than indicators tied to mortgage delinquency and housing. This timing argument supports the lender-distress hypothesis over the borrower-distress hypothesis (Mian and Sufi 2014b).

**Credit channel of monetary policy (Section IV, pp. 1161-1163).** Empirical tests exploit compositional shifts in firms' external finance (Kashyap, Stein, and Wilcox 1993), cross-firm borrowing cyclicality by size using post-1958 data from the Quarterly Financial Report of Manufacturing Firms (Bernanke, Gertler, and Gilchrist 1996), and asset-price responses around FOMC announcements (Gertler and Karadi 2015). Small firms - which face higher and more cyclically sensitive external finance premiums (Bernanke, Gertler, and Gilchrist 1996) - reduce external borrowing more than large firms after monetary tightenings, consistent with the credit channel.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---------|--------------|-----------|
| FRED (Federal Reserve Economic Data) | Bank credit outstanding relative to GDP, mortgage delinquency rates, business formations; cited explicitly pp. 1159, 1161 and footnotes 16, 18 | [FRED](/wiki/datasets/fred/) |
| Aggregate US banking and deposit data (1929-1933) | Loan-to-deposit ratios, bank failure counts, corporate credit spreads; sourced from Calomiris (1993) and other Depression-era historical compilations | no page yet |
| 24-country Depression-era international panel (Bernanke and James 1991) | Cross-country output and banking crisis data used to compare Depression severity by gold-standard status | no page yet |
| Quarterly Financial Report of Manufacturing Firms | Composition of external finance for small vs. large manufacturing firms (post-1958); cited p. 1162 | no page yet |

Sample scope: US aggregate credit and banking data; international Depression-era panel (24 countries); period 1929-2013; mixed frequency (monthly, quarterly, annual).

## When to read the full paper

Read the source if you want: (i) the canonical synthesis of Bernanke's career research on credit markets and macroeconomic fluctuations, written by the economist most responsible for developing the credit-market disruption framework; (ii) the Appendix model (eqs. 1-9, pp. 1165-1167) as a worked example of how limited liability and moral hazard create a net-worth threshold for credit access; (iii) the financial accelerator mechanism and the credit channel of monetary policy as laid out in Section IV; or (iv) a structured map to the broader research program: Bernanke (1983), Bernanke and Gertler (1989, 1990), Bernanke, Gertler, and Gilchrist (1999), and Bernanke (2018) are all directly synthesized here. The Diamond and Dybvig (1983) bank-run model and the Stiglitz and Weiss (1981) credit-rationing framework are also placed in context.

## Attribution and rights

This article is copyright © The Nobel Foundation 2022 and is published in the *American Economic Review* with the permission of the Nobel Foundation. Paywalled; extract-only redistribution. LLM-distilled by claude-sonnet-4-6; not human-verified; not reproduced.

> Bernanke, Ben S. 2023. "Nobel Lecture: Banking, Credit, and Economic Fluctuations." *American Economic Review* 113(5): 1143-1169. https://doi.org/10.1257/aer.113.5.1143


==============================================================================
# Individuals and Organizations as Sources of State Effectiveness: Best, Hjort & Szakonyi (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/best-et-al-individuals-organizations-sources-state-2023/
# Distilled: Using 16 million Russian public procurement transactions (2011-2016), the paper measures that individual bureaucrats and organizations jointly explain 39 percent of the variation in quality-adjusted government procurement prices, with bureaucrats and organizations each accounting for roughly half. Bid preferences favoring domestic suppliers reduce prices when implemented by less effective bureaucrats but raise them when implemented by more effective ones. American Economic Review 2023, paywalled. Seven core results with source locators, datasets, the procurement model, and the variance decomposition method with defining equations.
# Tags: paper-summary, public-economics, bureaucracy, procurement, industrial-policy
==============================================================================

**What this is.** The core results, model, and method of this paper, condensed to know what it found and how, without reading all 47 pages. To replicate or extend, read the full source at [doi.org/10.1257/aer.20191598](https://doi.org/10.1257/aer.20191598) and the replication package at [doi.org/10.3886/E184107V1](https://doi.org/10.3886/E184107V1).

## TL;DR

The paper quantifies how much of the Russian government's procurement performance is attributable to the individuals and organizations who run it. Using administrative data on 16 million off-the-shelf purchases (2011-2016), it estimates that individual bureaucrats and public-sector organizations together explain 39 percent of the variation in quality-adjusted prices paid. The identification strategy exploits the fact that many bureaucrats work with multiple organizations and vice versa, providing thousands of quasi-experiments from bureaucrat-organization switches. The paper is related to the evidence on active vs passive waste in public contracts by Bandiera, Prat and Valletti (2009), which it extends by separately identifying individual and organizational sources of inefficiency. Effective bureaucrats (those who pay lower prices) also achieve better spending quality (fewer delays, renegotiations, and cost overruns), so the multitasking problem is mild. In a second part, the paper introduces bid preferences that give a 15 percent price advantage to domestic suppliers. On average, these preferences achieve their goal of increasing domestic sourcing at near-zero cost, but this average masks large heterogeneity: prices fall substantially for ineffective bureaucrats and rise for effective ones, consistent with a model where participation costs are the key friction.

## Core results

Magnitudes and significance are as reported. `\*` = 5%; `\*\*\*` = 1%.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Bureaucrats and organizations jointly explain **39 percent of the variation** in quality-adjusted procurement prices | Table 2 col 6, p. 2143 | Combined bur+org SD within connected sets = 0.489 log points; SD of log price \| good, month = 1.280; ratio = 38.2% (covariance shrinkage method) |
| R2 | **Bureaucrats account for ~21 percent and organizations for ~26 percent** of price variation separately | Table 2 col 6, p. 2143 | Covariance-shrunk SD: bureaucrats = 0.263 log points, organizations = 0.338; bur-org correlation = 0.311 (positive assortative matching) |
| R3 | **Moving lowest-quartile bureaucrats to the 75th percentile** would reduce procurement costs by 4.6 percent; moving both below Q25 to Q75 saves 13.9 percent | §IVC, pp. 2143-2144 | 4.6% savings from bureaucrats only (Q1 to Q75); 13.9% combined (~US\$10 billion/year, ~0.7% of nonresource GDP) |
| R4 | **Price effectiveness and spending-quality effectiveness are positively correlated** across bureaucrats and organizations | Figure 3, p. 2148 | Correlation = 0.43 for bureaucrats; 0.48 for organizations; low-price buyers also deliver fewer delays and renegotiations |
| R5 | **Bid preferences achieve domestic sourcing at near-zero average cost**: prices are unaffected on average but the probability of a domestic supplier winning rises 14 percent | Table 5, p. 2155 | Price effect: -0.004 (SE 0.010), not significant; domestic winner probability: +0.042 (SE 0.005) in pharmaceuticals |
| R6 | **Bid preferences reduce prices by up to 12 percent for the least effective bureaucrats** but raise prices for the most effective | Table 6 cols 1 and 4, Figure 7, pp. 2156-2158 | Coefficient on Bureaucrat FE x Preferred x PolicyActive: -0.090\*\*\* (SE 0.018) full sample; -0.466\*\*\* (SE 0.090) pharmaceuticals |
| R7 | **Optimal bid penalty varies across bureaucracy types**, from 23 percent for the most effective to 10 percent for the least effective subgroup | Figure 9, pp. 2162-2163 | Equivalent penalty for same price outcome as the 15% rule achieves overall, across ten effectiveness deciles |

**Overall (paper's conclusion).** State effectiveness is embedded in individuals and organizations: bureaucrats and their employers account for a large fraction of procurement price variation, and effectiveness is embodied in individual characteristics (experience, network, auction design quality) rather than in corruption measures. Because raising bureaucratic effectiveness directly is often difficult, a feasible alternative is to tailor policy design to the capacity of the implementing bureaucracy. A "buy local" bid preference regime that is counterproductive for high-state-capacity contexts works well for low-state-capacity contexts.

## Theory / model

Section III presents a stylized model of public procurement in which state effectiveness is modeled as costs imposed on potential suppliers wishing to participate. A pair consisting of a procurement bureaucrat and an end-user organization (jointly, the bureaucracy) purchases an off-the-shelf good through a descending open-outcry auction, approximated as a second-price sealed-bid auction (p. 2133, following Milgrom 2004).

There are two potential suppliers: a foreign firm $$F$$ with higher expected productivity (Pareto parameter $$\delta_F$$) and a local firm $$L$$ (parameter $$\delta_L < \delta_F$$). Both face a common fulfillment cost component $$\log(\bar{\theta}) = \mathbf{X}'\boldsymbol{\beta} + \alpha_\theta + \psi_\theta$$, where $$\mathbf{X}$$ are observable item attributes and $$\alpha_\theta$$ and $$\psi_\theta$$ are bureaucrat and organization fulfillment-cost shifters. The bureaucracy also imposes participation cost $$c_i$$ on each supplier, determined by specification parameters $$\alpha_c$$ and $$\psi_c$$:

$$
c_i = \frac{\bar{\theta}}{1 + \delta_i} - \frac{\bar{\theta}}{1 + \delta_L}\sqrt{1 - \alpha_c - \psi_c}
$$

Suppliers independently decide whether to pay $$c_i$$ and learn their type $$\theta_i$$. In the Nash equilibrium (Proposition 1, p. 2134), entry probabilities satisfy $$q_i = \sqrt{\kappa(1 - \alpha_c - \psi_c)}$$ where $$\kappa = \min\!\left\{\left[\frac{1 + \delta_F + \delta_L}{1 + \delta_F}\right]^2, \frac{1}{1 - \alpha_c - \psi_c}\right\}$$. Expected log prices are (equation 1, p. 2134):

$$
E[\log(p)] = \log(\bar{\theta}) - \frac{q_F q_L}{\delta_F + \delta_L} = \mathbf{X}'\boldsymbol{\beta} - \frac{\kappa}{\delta_F + \delta_L} + \bar{\alpha} + \bar{\psi}, \tag{1}
$$

where $$\bar{\alpha} = \alpha_\theta + \frac{\kappa}{\delta_F + \delta_L}\alpha_c$$ and $$\bar{\psi} = \psi_\theta + \frac{\kappa}{\delta_F + \delta_L}\psi_c$$. Equation (1) delivers two predictions: (i) bureaucracies that impose higher fulfillment costs $$(\alpha_\theta, \psi_\theta)$$ pay higher prices for otherwise identical goods; (ii) bureaucracies that impose higher participation costs $$(\alpha_c, \psi_c)$$ pay higher prices and also attract fewer bidders. Both channels reduce price performance and are captured by the combined effects $$\bar{\alpha}$$ and $$\bar{\psi}$$.

**Bid preferences with heterogeneous effectiveness (Section IIIB, Proposition 2, p. 2135).** A 15 percent bid preference gives local bidder $$L$$ a price advantage: a foreign winner receives only $$\gamma = 0.85$$ times her bid as the contract price. The model predicts three regimes of bureaucratic effectiveness: (i) for very effective bureaucracies with low participation costs $$(\alpha_c + \psi_c \leq \underline{c})$$, both firms still enter, the preference primarily shifts bidding, and prices rise; (ii) for intermediate effectiveness $$(\underline{c} < \alpha_c + \psi_c \leq \bar{c})$$, the foreign firm drops out and the local winner receives the maximum price, raising prices further; (iii) for ineffective bureaucracies $$(\bar{c} < \alpha_c + \psi_c)$$ with the highest participation costs, the local firm's increased willingness to enter outweighs the bidding effect, participation rises and prices fall. The model thus predicts that preferences compress cross-bureaucracy variation in performance: they help the least effective the most and harm the most effective.

## Method

The paper extends the variance decomposition method pioneered by Abowd, Kramarz, and Margolis (1999) from the private sector to public procurement, treating bureaucrats as workers and organizations as firms. The reduced-form model for log unit price of purchase $$i$$ procured by organization $$j$$ through bureaucrat $$b(i,j)$$ is (equation 2, p. 2138):

$$
p_i = \mathbf{X}_i\boldsymbol{\beta} + \tilde{\alpha}_{b(i,j)} + \tilde{\psi}_j + \varepsilon_i, \tag{2}
$$

where $$\mathbf{X}_i$$ includes log quantity, good fixed effects, month fixed effects, and interactions of two-digit HS product categories with years, regions, and lot size. The parameters $$\tilde{\alpha}_b$$ and $$\tilde{\psi}_j$$ are overall effects including connected-set intercepts (equation 3, p. 2139):

$$
p_i = \mathbf{X}_i\boldsymbol{\beta} + \alpha_{b(i,j)} + \psi_j + \gamma_{s(b,j)} + \varepsilon_i, \tag{3}
$$

where $$\gamma_{s(b,j)}$$ is a connected-set fixed effect normalizing $$\alpha$$ and $$\psi$$ to mean zero within each connected set. Identification of bureaucrat and organization effects separately requires observation of bureaucrats working with multiple organizations and organizations working with multiple bureaucrats (the switchers). The data contain 616 connected sets with an average density of 5.2 organizations per bureaucrat and 4.8 bureaucrats per organization.

The variance decomposition (equation 4, p. 2140) attributes total price variation to its sources:

$$
\text{var}(p_i) = \text{var}(\alpha_{b(i,j)}) + \text{var}(\psi_j) + 2\,\text{cov}(\alpha_{b(i,j)}, \psi_j) + 2\,\text{cov}(\alpha_{b(i,j)} + \psi_j,\, \gamma_{s(b,j)} + \mathbf{X}_i\boldsymbol{\beta}) + \text{var}(\gamma_{s(b,j)} + \mathbf{X}_i\boldsymbol{\beta}) + \text{var}(\varepsilon_i). \tag{4}
$$

The paper addresses two finite-sample estimation problems. First, **limited mobility bias**: because the bureaucrat-organization network has only 616 connected sets with bounded within-set mobility, the naive OLS estimates of $$\hat{\alpha}_b$$ and $$\hat{\psi}_j$$ contain large sampling errors that inflate the apparent variance of each component and create a spurious negative covariance. The paper applies two bias corrections: (a) a **split-sample estimator** (following Finkelstein, Gentzkow, and Williams 2016 and Silver 2016) that randomly splits the sample and estimates the two fixed effects on each half, then forms variance components as cross-half covariances; (b) a **covariance-shrinkage estimator** that forms minimum-MSE predictions of the full vector $$(\hat{\alpha}_b, \hat{\psi}_j)$$ by weighting via a matrix $$\Lambda^*$$ accounting for both own-variance shrinkage and cross-component covariance (analogous to the shrinkage in Chetty, Friedman, and Rockoff (2014) but generalized to two dimensions). The covariance-shrinkage method is the preferred specification because it yields a plausible positive bur-org correlation of 0.311, unlike the split-sample or OLS methods which give negative estimates (Table 2, column 6, p. 2143).

Effective bureaucrats also display distinguishable process behaviors (Sections IVD-IVE). A LASSO procedure selecting 30 process predictors from 85 potential variables shows that effective bureaucrats set lower reservation prices, attract larger and more diverse supplier pools, specialize in narrower product ranges, and avoid fiscal-year-end spending rushes, all consistent with the participation-cost channel in the model.

## Empirical specifications

**Event study for causal identification (Section IVA, Figure 1, p. 2137).** To support the causal interpretation of bureaucrat and organization effects, the paper exploits 65,000 events in which organizations switch the bureaucrat they work with. Prices change sharply and in the expected direction when a new bureaucrat takes over, with an 18 percent price decrease when switching from a worst-quartile to a best-quartile bureaucrat. There is no systematic pre-trend, and price changes are symmetric for switches in both directions, consistent with the identifying assumption that drift in effectiveness and switches are uncorrelated.

**Average effect of bid preferences (equation 5, p. 2153; Table 5, p. 2155).** The paper uses a generalized difference-in-differences design, exploiting year-to-year variation in which goods appear on the domestic-preference list:

$$
y_{igt} = \mathbf{X}_{igt}\boldsymbol{\beta} + \mu_g + \lambda_t + \delta\,(\text{Preferenced}_{gt} \times \text{PolicyActive}_t) + \varepsilon_{igt}, \tag{5}
$$

where $$\text{Preferenced}_{gt}$$ indicates that good $$g$$ is on the preference list in year $$t$$ and $$\text{PolicyActive}_t$$ indicates the policy is in effect. Standard errors are clustered by month and good. Because a minimum of one local and one foreign bidder must be present for preferences to apply, $$\hat{\delta}$$ is an intent-to-treat effect. An event-study analog (equation 6, p. 2153) stacks all list-publication events in a window of -3 to +4 months:

$$
p_{igt} = \mathbf{X}_{igt}\boldsymbol{\beta} + \mu_g + \lambda_t + \sum_{s=-3}^{4}\delta_s\,\text{Preferenced}_{gt} \times \mathbf{1}\!\left\{t - \text{ListMonth}_t = s\right\} + \varepsilon_{igt} \tag{6}
$$

Pre-trend coefficients $$\hat{\delta}_s$$ for $$s = -3, -2, -1$$ are indistinguishable from zero (Figure 6, p. 2154), validating parallel trends. The preferred estimate of $$\hat{\delta}$$ is -0.004 (SE 0.010) for log prices, not significantly different from zero (Table 5, column 1), while the probability that an auction is won by a domestic supplier rises by 4.2 percentage points in pharmaceuticals. This near-zero average price effect differs from the positive price effects of domestic bid preferences found by Marion (2007) in US highway procurement, where preferences reduced competition and raised prices by 3.8 percent; the difference is consistent with the lower baseline state capacity in the Russian context reducing the scope for anti-competitive effects.

**Heterogeneous effects by bureaucratic effectiveness (equation 7, p. 2156; Table 6, p. 2157).** To test the model's prediction that the policy effect varies by state effectiveness, the paper interacts the DiD design with the covariance-shrunk bureaucrat and organization effects:

$$
y_{igt} = \mathbf{X}_{igt}\boldsymbol{\beta} + \mu_g + \lambda_t + \theta_b\hat{\alpha}_b + \theta_j\hat{\psi}_j + \delta\,(\text{Preferenced}_{gt} \times \text{PolicyActive}_t)
$$

$$
+ \rho_b\,\text{Preferenced}_{gt}\hat{\alpha}_b + \rho_j\,\text{Preferenced}_{gt}\hat{\psi}_j + \eta_b\,\text{PolicyActive}_t\hat{\alpha}_b + \eta_j\,\text{PolicyActive}_t\hat{\psi}_j
$$

$$
+ \pi_b\,(\text{Preferenced}_{gt} \times \text{PolicyActive}_t)\hat{\alpha}_b + \pi_j\,(\text{Preferenced}_{gt} \times \text{PolicyActive}_t)\hat{\psi}_j + \varepsilon_{igt}, \tag{7}
$$

where $$\hat{\alpha}_b$$ and $$\hat{\psi}_j$$ are the covariance-shrunk estimates from equation (3). The key parameter is $$\hat{\pi}_b$$: a negative estimate means that bureaucrats with higher baseline prices (higher $$\hat{\alpha}_b$$, i.e., less effective) experience larger price declines under preferences. Table 6 reports $$\hat{\pi}_b = -0.090$$ (SE 0.018) in the full sample and $$-0.466$$ (SE 0.090) in the pharmaceuticals subsample, both significant at 1 percent. An analogous decile-level regression (equation 8, p. 2157) shows that the estimated price effect is monotonically decreasing from about -20 percent for the bottom decile of bureaucratic effectiveness to slightly positive for the top decile (Figure 7, p. 2158). Heterogeneity by organization effectiveness $$\hat{\psi}_j$$ is small and imprecisely estimated, consistent with the model's prediction that it is participation costs (driven by bureaucrats) rather than fulfillment costs (shared by both) that generate the heterogeneous policy effects.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Russian federal procurement register (zakupki.gov.ru / EIS) | 16 million auction requests, protocols, and contracts for off-the-shelf goods, 2011-2016; main data for all price and spending-quality outcomes | No page yet |
| LVEMD pharmaceutical list (MinZrav 2016) | Barcode-level classification of drugs for the pharmaceuticals subsample; matches procurement items to active ingredient and manufacturer | No page yet |
| ClearSpending.ru anomalies database | Six spending-quality proxies (delays, cost overruns, contract renegotiations, end-user complaints, cancellations, quality bans) from a civil-society monitoring platform | No page yet |
| Rosstat firm databases | Financial characteristics of winning suppliers (balance sheets, employment) for the process-measurement analysis | No page yet |
| Ruslana (Bureau Van Dijk) | Additional firm financials and company identifiers for matching procurement winners to financial data | No page yet |

Sample period: January 1, 2011 through December 31, 2016. All prices converted at an exchange rate of 43 rubles per US dollar. The analysis sample (connected sets) contains 11,339,187 observations. The pharmaceuticals subsample contains 181,963 observations.

## When to read the full paper

Use the [original](https://doi.org/10.1257/aer.20191598) for:

- **Replication**: the online appendices (Appendices A-H) contain the text classification algorithm, full variance-decomposition derivations, event-study robustness checks, all organization-side correlates, and sensitivity to alternative shrinkage procedures.
- **AKM extension**: Section IVB and Appendix C give the formal derivation of the covariance-shrinkage estimator and proofs that it yields lower-bound estimates of the true bureaucrat and organization variances.
- **Policy design**: Section VD and Figure 9 work out how to calibrate a bid-preference rate to a bureaucracy's effectiveness distribution, the starting point for any government adapting the framework to its own context.
- **Corruption robustness**: Section IIC and online Appendix D show why the quality-adjusted price measure is informative even in the presence of corruption, and provide direct tests that corruption is unlikely to explain the main findings.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(8), August 2023, pp. 2121-2167. DOI: [10.1257/aer.20191598](https://doi.org/10.1257/aer.20191598). Replication data: [doi.org/10.3886/E184107V1](https://doi.org/10.3886/E184107V1).

This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. No open license was found via Crossref; the paper is paywalled. This page contains extracted results and equations only (extract-only redistribution).

> Best, Michael Carlos, Jonas Hjort, and David Szakonyi. "Individuals and Organizations as Sources of State Effectiveness." *American Economic Review* 113, no. 8 (August 2023): 2121-2167. DOI: 10.1257/aer.20191598.


==============================================================================
# Regulation Design in Insurance Markets: Bhaskar, McClellan & Sadler (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/bhaskar-et-al-regulation-design-insurance-markets-2023/
# Distilled: The paper models insurance regulation as a delegation problem and shows a regulator can implement the socially optimal allocation by requiring each firm menu to include at most two latent contracts that are never purchased in equilibrium but deter the firm from misusing its private signal about consumers. American Economic Review 2023, paywalled. Six core results with source locators, the formal model, and the mechanism with equations.
# Tags: paper-summary, insurance, regulation, mechanism-design, information-economics, peer-reviewed, unreplicated
==============================================================================

**What this is.** The paper's core propositions, the formal delegation model it builds on, and the latent-contract mechanism with its defining equations: enough to understand what was proved and how, without reading all 35 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1257/aer.20210710).

## TL;DR

A regulator seeks to restrict the menus of insurance contracts a firm may offer. The firm privately observes a signal about each consumer's type, giving it an informational advantage the regulator lacks. The paper's main result (Theorem 1) shows that despite this asymmetry, the regulator can implement any socially optimal allocation by augmenting each intended menu with at most two "latent contracts": high-coverage expensive options and low-coverage cheap options that are never chosen in equilibrium but deter the firm from offering menus designed for different signals. Under an order condition that higher signals correspond to higher consumer coverage need, these latent contracts can be constructed from the model's primitives. A separate result shows that when the regulator maximizes consumer welfare, more firm information weakly improves welfare under optimal regulation: the regulator can turn the firm's data advantage into a tool rather than a threat.

## Core results

Locators reference the published version. No numerical magnitudes are reported; all results are theoretical propositions.

| # | Result | Locator | Formal claim |
|---|---|---|---|
| R1 | Main theorem: any incentive-compatible, participation-feasible allocation can be implemented with at most two latent contracts per menu | Theorem 1, p. 2560 | There exist contracts {c̄^s, c_s}_{s∈S} such that policy R = {M^s_{a\*} ∪ {c̄^s, c_s}}_{s∈S} implements a\* in perfect Bayesian equilibrium |
| R2 | Noncontractible loss model: optimal allocation gives all agents in category x the same contract, equalizing marginal utility across all types and events | Prop. 1(i), pp. 2561-2562 | c_x = (p\*, t_x\*) where t_x\*(ω) equates the marginal utility of transfers u_z across all (ω, x) pairs |
| R3 | Under supermodularity of utility in transfer and category, transfers are strictly increasing in category and one downward latent contract per menu suffices | Prop. 1(ii), p. 2562 | t_x\*(ω) strictly increasing in x for each ω; policy R = {M_x}_{x∈X} with M_x = {c_x, c̄_x} implements a\* |
| R4 | Consumer welfare is weakly higher under optimal regulation when the firm has more information | Prop. 2, p. 2563 | If S' is more informative than S, an optimal policy under S' attains W ≥ the optimal W under S |
| R5 | Local improvements: any nearby allocation can be implemented with latent contracts proportionally close to existing contracts | Prop. 3, pp. 2564-2565 | There exists K ≥ 1 such that for any Δ > 0 and any allocation â with d(â, a) < Δ satisfying agent incentives and firm participation, there is a policy R̂ with D(R̂, R) ≤ KΔ implementing â |
| R6 | Price cap plus latent contracts: a data-driven price cap on the highest-risk menu, augmented with latent contracts on all other menus, robustly improves on the laissez-faire outcome using only observable market data | Prop. 5, pp. 2568-2569 | For menu M_n (highest signal), price-capped menu M^C_n combined with latent contracts c̄^i on menus M_1,...,M_{n-1} leads the firm to offer M^C_n after signal s_n; highest-risk types purchase full coverage at a lower price p^C < p^{m_n}_n |

**Overall (paper's conclusion).** Regulators can leverage a firm's informational advantage against itself. By requiring firms to include off-path latent contracts in each menu, the regulator implements her preferred allocation without observing the firm's signal, and the mechanism requires each consumer to face at most three contracts. This insight extends to price caps and other standard regulatory tools when combined with targeted latent contracts inferred from market data.

## Theory / model

The framework builds on the monopolistic insurance screening model of Stiglitz (1977), in which a firm offers menus to separate risk types, and on the delegation theory of Holmstrom (1984), who studies a principal restricting the action set of a better-informed agent. Baron and Myerson (1982) study optimal price regulation of a monopolist with privately known costs; the key difference here is that the firm contracts with individual consumers and screens them using its data advantage. Chade and Schlee (2012) characterize optimal insurance allocation under adverse selection; the noncontractible loss model below embeds their canonical formulation as a special case. Galperti (2015) shows commitment devices can screen time-inconsistent agents; latent contracts serve an analogous off-path incentive role here. Brunnermeier, Lamba, and Segura-Rodriguez (2020) study how a firm better informed than consumers designs profit-maximizing menus; this paper adds a regulator who turns that asymmetry into a welfare tool.

The model is a three-stage game (regulator, firm, consumers) with heterogeneous agents and a privately informed intermediary (pp. 2554-2556).

**Types and preferences.** Each agent has a two-dimensional type $$\tau = (x, \theta) \in T = X \times \Theta$$, where $$X \subset \mathbb{N}$$ is the finite set of categories and $$\Theta \subset \Delta(\Omega)$$ is the set of risk types. Categories capture utility differences in the event of a loss (treatment cost, severity of need); risk types determine the probability distribution over verifiable loss events $$\Omega$$. Types are distributed according to $$\mu \in \Delta(X \times \Theta)$$. A contract $$c = (p, t)$$ specifies a premium $$p \in \mathbb{R}_+$$ and a transfer function $$t : \Omega \to \mathbb{R}_+$$. Agent $$\tau$$ buying contract $$c$$ obtains expected utility (p. 2554):

$$
U(\tau, c) := \sum_{\omega \in \Omega} \theta(\omega)\, u(t(\omega) - p,\, \omega,\, x) \tag{1}
$$

where $$u : \mathbb{R} \times \Omega \times X \to \mathbb{R}$$ is strictly increasing and strictly concave in its first argument, with the Inada condition $$\lim_{z \to \underline{z}_{\omega,x}} u_z = \infty$$. The firm's expected profit from selling $$c$$ to type $$\tau$$ is:

$$
\Pi(\tau, c) := p - \sum_{\omega \in \Omega} \theta(\omega)\, t(\omega) \tag{2}
$$

**Information structure.** The firm privately observes a signal $$s \in S \subset \mathbb{N}$$ about each agent's type; the regulator observes neither the signal nor the type, only the aggregate joint distribution of types and signals. Let $$\mu(\cdot \mid s)$$ denote the conditional type distribution given signal $$s$$ and $$\text{supp}(s)$$ the support of $$\mu(\cdot \mid s)$$.

**Regulatory policy and timing.** A regulatory policy $$\mathcal{R}$$ is a set of menus (each menu is a set of contracts). Timing: (i) the regulator chooses $$\mathcal{R}$$; (ii) the firm decides whether to enter, then offers each agent a menu from $$\mathcal{R}$$ based on its signal; (iii) agents choose a contract from the offered menu; (iv) loss events realize and transfers are paid.

**Aggregate welfare.** An allocation $$\mathbf{a} = \{c^s_\tau\}$$ specifies, for each type and signal, a contract. Total consumer welfare and firm expected profit from allocation $$\mathbf{a}$$ are (p. 2556):

$$
W(\mathbf{a}) := \sum_{\tau \in T} \mu(\tau) \sum_{s:\, \tau \in \text{supp}(s)} \Pr(s \mid \tau)\, U(\tau, c^s_\tau) \tag{3}
$$

$$
\pi(\mathbf{a}) := -k + \sum_{\tau \in T} \mu(\tau) \sum_{s:\, \tau \in \text{supp}(s)} \Pr(s \mid \tau)\, \Pi(\tau, c^s_\tau) \tag{4}
$$

The regulator maximizes a social welfare function $$F(\mathbf{a})$$ satisfying $$F(\mathbf{a}) = -\infty$$ whenever $$W(\mathbf{a}) = -\infty$$; a canonical special case is $$F(\mathbf{a}) = \beta W(\mathbf{a}) + (1-\beta)\pi(\mathbf{a})$$ for $$\beta \in (0,1]$$.

**Order condition (Assumption 1, p. 2558).** The main result requires that higher signals correspond to higher coverage need. There exists a loss event $$\omega_1 \neq \omega_0$$ such that $$u(z, \omega, x)$$ is supermodular in $$z$$ and $$x$$, and for every signal $$s$$, a maximal type $$\bar{\tau}^s = (\bar{x}^s, \bar{\theta}^s)$$ and a minimal type $$\underline{\tau}^s = (\underline{x}^s, \underline{\theta}^s)$$ exist in $$\text{supp}(s)$$ such that:

- the maximal type has a weakly higher category and higher likelihood ratio $$\bar{\theta}^s(\omega_1)/\bar{\theta}^s(\omega_0) \geq \theta(\omega_1)/\theta(\omega_0)$$ for all types in the support of $$s$$; and
- for $$s' > s$$, the maximal category $$\bar{x}^{s'} \geq \bar{x}^s$$ and the maximal likelihood ratio in signal $$s'$$ is strictly greater than in signal $$s$$.

This means higher signals are supported on types with weakly higher category and strictly higher maximal risk, capturing that individuals with poorer health or greater need are more likely to generate higher firm signals.

**Relaxed problem.** The regulator first solves a relaxed problem (RP) that ignores firm incentive constraints, treating the firm as if the regulator could directly observe the signal (pp. 2559-2560):

$$
\max_{\mathbf{a}}\; F(\mathbf{a}) \quad \text{s.t.} \quad \pi(\mathbf{a}) \geq 0 \text{ and } U(\tau, c^s_\tau) \geq U(\tau, c)\; \forall\, \tau \in \text{supp}(s),\, c \in M^s_\mathbf{a},\, s \in S \tag{RP}
$$

The firm's incentive constraints to truthfully offer the menu intended for each signal are then the binding constraint for implementation in the actual problem.

**Noncontractible loss model (special case).** For the main application, each verifiable event $$\omega$$ contains a set of noncontractible states $$\hat{\omega}$$; agents in category $$x$$ face the distribution $$\nu_x(\hat{\omega} \mid \omega)$$ over them, suffering loss $$\tilde{\ell}(\hat{\omega})$$. Utility in event $$\omega$$ is (p. 2557):

$$
u(t(\omega) - p,\, \omega,\, x) := \int_{\hat{\omega} \in \omega} v\!\left(e - p + t(\omega) - \tilde{\ell}(\hat{\omega})\right) d\nu_x(\hat{\omega} \mid \omega) \tag{5}
$$

where $$e$$ is initial wealth and $$v$$ is strictly concave. In the canonical Stiglitz (1977) model this collapses to two states (no loss, loss $$\ell$$) with a single category $$x$$.

## Method

The proof of Theorem 1 constructs two families of latent contracts for each signal $$s$$, using the order condition to ensure they deter deviations without distorting on-path allocations.

**Latent contract properties (p. 2560).** For each signal $$s$$, the two contracts $$(\bar{c}^s, \underline{c}^s)$$ added to menu $$M^s_{a^*}$$ satisfy:
1. Every type $$\tau \in \text{supp}(s)$$ weakly prefers the allocated contract $$c^s_{\tau, a^*}$$ to both latent contracts.
2. For any higher signal $$s' > s$$, all types $$\tau \in \text{supp}(s')$$ with $$\theta(\omega_1)/\theta(\omega_0) > \bar{\theta}^s(\omega_1)/\bar{\theta}^s(\omega_0)$$ and $$x \geq \bar{x}^s$$ prefer the downward latent contract $$\bar{c}^s$$ over any contract in $$M^s_{a^*}$$.
3. For any lower signal $$s' < s$$, all types $$\tau \in \text{supp}(s')$$ with $$\theta(\omega_1)/\theta(\omega_0) < \underline{\theta}^s(\omega_1)/\underline{\theta}^s(\omega_0)$$ and $$x \leq \underline{x}^s$$ prefer the upward latent contract $$\underline{c}^s$$ over any contract in $$M^s_{a^*}$$.

**Construction of the downward latent contract (Lemma 1, p. 2574).** Fix signal $$s$$ and let $$(\bar{x}^s, \bar{\theta}^s)$$ be the maximal type in its support. Let $$(p', t')$$ be a reference contract with $$U(\tau, (p', t')) > -\infty$$ for all $$\tau$$. The downward latent contract $$(\bar{p}, \bar{t})$$ is constructed so that the maximal type is indifferent between $$(\bar{p}, \bar{t})$$ and $$(p', t')$$. The indifference condition is (equation A1, p. 2574):

$$
\frac{\bar{\theta}^s(\omega_1)}{\bar{\theta}^s(\omega_0)} \Big[ u\!\left(\hat{t} - \bar{p},\, \omega_1,\, \bar{x}^s\right) - u\!\left(t'(\omega_1) - p',\, \omega_1,\, \bar{x}^s\right)\Big] = u\!\left(t'(\omega_0) - p',\, \omega_0,\, \bar{x}^s\right) - u\!\left(-\bar{p},\, \omega_0,\, \bar{x}^s\right) \tag{A1}
$$

For fixed $$\hat{t} > t'(\omega_1)$$, a unique $$\bar{p} \in [p' - t'(\omega_0),\, -\underline{z}]$$ solving (A1) exists because the right-hand side is finite and decreasing in $$\bar{p}$$, and utility is unbounded above. Raising $$\hat{t}$$ drives the firm's expected profit from selling $$(\bar{p}, \bar{t})$$ toward $$-\infty$$, because $$\lim_{\hat{t} \to \infty} \bar{p}'(\hat{t}) = 0$$ (the marginal increase in $$\bar{p}$$ vanishes, so $$\hat{t} - \bar{p}(\hat{t}) \to \infty$$) while the transfer cost $$\bar{\theta}^s(\omega_1) \hat{t} \to \infty$$ (p. 2574). Choosing $$\hat{t}$$ large enough makes the downward latent contract arbitrarily unprofitable for the firm, so no signal $$s' < s$$ is worth offering after signal $$s$$. By supermodularity, any type in a higher signal's support who could prefer $$\bar{c}^s$$ over the intended contract is exactly the kind of type the regulator wants to attract away from the wrong menu.

An analogous upward latent contract $$(\underline{p}, \underline{t})$$ is constructed with lower payouts in $$\omega_1$$ and a higher transfer in $$\omega_0$$, targeting the minimal type $$\underline{\tau}^s$$, to deter the firm from offering a lower-signal menu after signal $$s$$.

**Equilibrium verification (proof of Theorem 1, pp. 2576-2577).** With $$M^s = M^s_{a^*} \cup \{(\bar{p}^s, \bar{t}^s), (\underline{p}^s, \underline{t}^s)\}$$, consider the strategy profile: the firm offers $$M^s$$ after signal $$s$$; on-path agents choose their allocated contract; if the firm deviates to $$M^{s'}$$, all types who weakly prefer the corresponding latent contract choose it. For $$\hat{t}$$ sufficiently large, the latent contracts are chosen by enough types that the deviation is unprofitable regardless of what other types choose, since the expected loss from latent-contract purchases is unbounded. Hence the firm's optimal response is to follow the regulator's prescribed menu after each signal.

**Data-driven price cap (Proposition 5, pp. 2568-2569).** For the canonical one-category model the regulator can observe three objects from laissez-faire market data: the contracts sold in each menu $$M_i = \{(p_i^1, t_i^1), \ldots, (p_i^{m_i}, t_i^{m_i})\}$$; the fraction $$\lambda(c_i^k)$$ of agents choosing each contract; and the fraction $$\rho(c_i^k)$$ suffering a loss. From these she infers: the loss amount $$\ell = t_n^{m_n}$$ (the transfer paid by the full-coverage contract on the highest-signal menu); the highest-risk types $$\bar{\theta}^{s_i} = \rho(c_i^{m_i})$$; and their menu share $$\lambda(c_n^{m_n})$$. The price-capped menu $$M_n^C$$ charges $$\bar{p}_i = \min\{p_i, p^C\}$$ for each contract. Adding to each other menu a latent contract with premium $$\bar{p}^i = p_i^{m_i} + B^i$$ (surcharge $$B^i \geq 0$$) and add-on transfer $$\bar{t}^i = \ell + A^i$$ ($$A^i \geq 0$$) constructs these purely from observable parameters, implementing the price cap without requiring knowledge of the full type distribution.

## Empirical specifications

This is a pure theory paper. There are no regression specifications or estimated equations. The illustration in Section I (p. 2552, Figure 1) uses calibrated parameters from Handel, Hendel, and Whinston (2015) to visualize the latent contract construction: CARA utility $$u(z) = -e^{-\alpha z}/\alpha$$ with risk-aversion coefficient $$\alpha = 0.0004$$, low risk $$\underline{\theta} = 0.1$$, high risk $$\bar{\theta} = 0.2$$. The figure shows indifference curves and the shaded region of contracts that attract the type $$(H, \bar{\theta})$$ but repel category $$L$$ consumers. These are calibration parameters used only for visual illustration; no data are fitted or tested.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Handel, Hendel, and Whinston (2015) CARA calibration | Parameters for the illustrative Figure 1 only (not data in the usual sense: α, θ, θ̄ are fixed prior estimates) | No page yet |

No empirical data are used in the main theoretical analysis or proofs.

## When to read the full paper

Use the [original](https://doi.org/10.1257/aer.20210710) if you are: designing or evaluating menu-based regulatory policies for insurance or other selection markets; studying delegation theory where an agent uses private data to screen principals; extending the framework to multiple firms, multidimensional types, or settings where regulators cannot impose purchase mandates; or working through the formal proofs and the online Appendix extensions (cream-skimming, constrained regulatory power, moral hazard). The six locators above point to the exact theorem and proposition pages.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(10), 2023. No open-access licence was found in Crossref or OpenAlex metadata (access: closed). This page is a distillation extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. Redistribution: extract-only.

> Bhaskar, Dhruva, Andrew McClellan, and Evan Sadler.
> "Regulation Design in Insurance Markets."
> *American Economic Review* 113, no. 10 (October 2023): 2546-2580.
> DOI: 10.1257/aer.20210710.
> © 2023 American Economic Association.
> Reproduced here in extract form only under fair-use scholarly commentary.


==============================================================================
# When Losses Turn into Loans: Blattner, Farinha & Rebelo (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/blattner-et-al-losses-turn-loans-cost-2023/
# Distilled: Distressed banks respond to ratio-based capital shortfalls by reallocating credit toward borrowers whose loan losses they underreport, using the 2011 EBA capital exercise in Portugal as a natural experiment. The credit misallocation accounts for about 22 percent of Portugal's allocative-efficiency decline in 2012. American Economic Review 2023, paywalled. Seven core results with source locators, datasets used, the identification design, and the defining equations.
# Tags: paper-summary, banking, credit-supply, zombie-lending, capital-requirements
==============================================================================

**What this is.** The paper's core results, the natural-experiment identification design, the loan-loss underreporting algorithm, and the firm-level empirical specifications with their defining equations: enough to know what it found and how, without reading all 42 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1257/aer.20190149).

## TL;DR

Distressed banks respond to ratio-based capital shortfalls not only by cutting overall credit but by distorting the composition of credit supply: they reallocate lending toward firms whose loan losses they have been underreporting, thereby delaying the recognition of those losses and protecting their reported capital ratios. Using the October 2011 European Banking Authority (EBA) capital exercise as a natural experiment affecting a subset of large Portuguese banks, Blattner, Farinha, and Rebelo develop a bunching-based algorithm to measure loan-loss underreporting at the monthly firm-bank level, show that exposed banks increase credit supply to underreported borrowers by about 2 percentage points per quarter while cutting credit to all other firms, and trace this credit reallocation through to a widening of capital and labor wedges that accounts for roughly 22 percent of the decline in aggregate allocative efficiency in Portugal in 2012.

## Core results

Magnitudes and significance are as reported; `\*\*`/`\*\*\*` = 5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Exposed banks increase credit to underreported firms, cut credit to all others**: triple-interaction coefficient positive and significant during EBA, negative and significant for baseline group | Figure 3 Panel A; Table A2 cols. 2-3, p. 1615 | +2 pp quarterly credit growth for underreported firms at exposed banks; -2 pp for all other relationships at exposed banks (each ~4% of 1 SD of credit growth) |
| R2 | **Firm-level credit reallocation is real and not undone by substitution**: total credit rises for underreported firms with high exposure to EBA banks, falls for all others | Figure 5 Panel A, p. 1619 | +16% cumulative credit for underreported firms borrowing entirely from exposed banks; -14% for all other firms relative to base quarter 2011:III |
| R3 | **Credit shock transmits to labor**: IV elasticity of labor w.r.t. credit supply is large and significant | Table 5 Panel B col. 2, p. 1625 | Elasticity = 0.52 (SE 0.094); first-stage F = 111.2 |
| R4 | **Credit shock transmits to capital**: IV elasticity of capital w.r.t. credit supply is significant | Table 5 Panel B col. 4, p. 1625 | Elasticity = 0.14 (SE 0.046); first-stage F = 111.2 |
| R5 | **Total EBA intervention caused large allocative-efficiency loss**: aggregating all firm-level wedge changes explains majority of 2012 AE decline | Table 6 Panel A col. 1, p. 1628 | Total estimated AE effect: -6.59% = 54% of actual -12.24% within-sector AE decline in 2012 |
| R6 | **Credit reallocation to underreported firms alone accounts for ~22% of the AE decline**: reallocation component isolated via simulation | Table 6 Panel B col. 1, p. 1628 | Mean -2.71% AE (range -0.89% to -4.43%); mean = 22% of actual AE decline |
| R7 | **Bunching validity**: underreporting is statistically higher in overdue buckets immediately before a jump in the regulatory deduction rate, confirming strategic behavior | Table B1 Panel A, p. 1636 | Coefficients 0.014-0.451 across collateral types and increment sizes (all positive and significant); placebo using the other collateral type's rate increment yields negative coefficients |

**Overall (paper's conclusion).** Ratio-based capital requirements create distorted lending incentives when banks are already in distress: exposed banks intensify loss underreporting and roll over credit to underreported borrowers to avoid booking additional losses. This credit misallocation prevents inputs from being reallocated to their highest-value uses, widening the dispersion of capital and labor wedges and contributing meaningfully to aggregate productivity decline.

## Theory / model

The paper has no formal equilibrium model; its conceptual framework is that ratio-based capital requirements create distorted incentives for already-distressed banks (tested hypotheses) and that these distortions propagate to real outcomes through credit misallocation.

**Identification.** The EBA in October 2011 announced that a subset of European banks had to meet a 9 percent Core Tier 1 ratio (with an additional sovereign debt buffer) by June 2012 (p. 1610). In Portugal, four banking groups (seven banks) were affected. The capital shortfall was determined by:

$$
\frac{\text{Core Tier 1} - \text{sovereign debt buffer}}{\text{RWA}} \geq 0.09. \tag{EBA threshold}
$$

A bank is defined as exposed if it was subject to the EBA exercise AND had a large capital shortfall (above-median sovereign bond holdings among eligible banks). The control group consists of eligible banks with below-median holdings plus all other commercial banks operating in Portugal. The key assumption is that observed credit reallocation is driven by the supply side (the shock to exposed banks' incentives) rather than differential credit demand from underreported firms. The paper supports this with parallel pre-trends, firm-level liquidity checks, and the result that exposed banks increase credit only to underreported (not to overdue-but-correctly-reported) firms.

**Wedge measurement.** Firm-level distortions are quantified as wedges in the first-order conditions of a Cobb-Douglas production function (equations 4-5, p. 1622):

$$
\alpha_s \frac{Y_{it}}{K_{it}} = (r_t + \Delta_t)(1 + \tau_{it}^K), \tag{4}
$$

$$
\beta_s \frac{Y_{it}}{L_{it}} = w_t(1 + \tau_{it}^L), \tag{5}
$$

where $$\tau_{it}^K$$ and $$\tau_{it}^L$$ are the capital and labor wedges (gaps between marginal revenue products and user costs), $$\alpha_s$$ and $$\beta_s$$ are sector-level capital and labor income shares, $$r_t + \Delta_t$$ is the depreciation-adjusted interest rate, and $$w_t$$ is the wage. Underreported firms have substantially lower marginal revenue products (mean MRPL = 37,300 vs. 48,240 for performing firms; mean MRPK = 47,440 vs. 63,160), indicating they are far from the efficient allocation even before the EBA shock (Table 4, p. 1622).

**Tested hypotheses:** (i) capital-constrained banks increase loss underreporting; (ii) they roll over credit to underreported borrowers to avoid forced loss recognition; (iii) credit reallocation to low-productivity borrowers widens wedge dispersion and lowers allocative efficiency; (iv) the mechanism is loss delay, not risk shifting (no increase in risky lending; see Table 3, p. 1618). Peek and Rosengren (2005) and Caballero, Hoshi, and Kashyap (2008) documented zombie lending in Japan; European reduced-form evidence exists (Schivardi, Sette, and Tabellini 2022 find no TFP effect in Italy) but has not established causality; this paper introduces quasi-experimental variation to establish causality and links the channel to input misallocation.

## Method

**Loan-loss underreporting algorithm.** The key methodological contribution is a Markovian excess-mass algorithm (Section I and Appendix A, pp. 1605-1608) that measures strategic delay in loan-loss reporting using the Portuguese Credit Register (Central de Responsabilidades de Credito), which reports the overdue loan balance in each regulatory deduction-rate bucket at the monthly firm-bank level.

Denote the observed loan balance in overdue bucket $$k$$ in month $$t$$ as $$B_{ib}(t;k)$$. In the absence of misreporting, the balance in bucket $$k$$ at $$t$$ equals the amount that moved up from bucket $$k-1$$ in the previous period. Excess mass is the deviation from this identity. When there are no flows the baseline expression is $$E(t;k) = B(t;k) - B(t-1; k-1)$$ (p. 1606). In general, inflows (new overdue installments) and outflows (repayments, restructurings, write-offs) require adjustment; the paper's equation (1) is:

$$
E(t;k) = \bigl[B(t;k) - IN(t;k)\bigr] - \bigl[B(t-1; k-1) - OUT(t; k-1)\bigr]. \tag{1, p. 1607}
$$

For multi-month buckets the appendix reformulates this using unobserved monthly sub-bucket balances $$C(t;c)$$ (where $$B(t;k) = \sum_{c \in k} C(t;c)$$), giving equation (A3, p. 1632):

$$
E(t;k) = \bigl[C(t;c) - IN(t;c)\bigr] - \bigl[C(t-1;c-1) - OUT(t;c-1)\bigr], \tag{A3}
$$

where $$IN(t;k)$$ denotes new installments falling overdue and $$OUT(t;k-1)$$ denotes repayments and restructurings. Excess mass is set to zero when negative (additional restriction) and adjusted for December window-dressing (Appendix A4). The algorithm is validated by the bunching result: excess mass is significantly higher in buckets immediately before a jump in the mandatory deduction rate, and the effect increases with the size of the regulatory increment (Table B1, p. 1636; equation B1, p. 1635):

$$
\frac{\text{excess mass}_{ibkct}}{\text{overdue loans}_{ibkct}} = \sum_{j=1}^{5} \beta_j \Delta \text{deduction rate}_j + \varphi_b + \theta_t + \mu_t + \epsilon_{ibkct}. \tag{B1}
$$

**Productivity decomposition.** To aggregate firm-level wedge changes to the macro level, the paper follows Osotimehin (2019) and decomposes aggregate TFP growth into technical efficiency (TE), within-sector allocative efficiency ($$\Delta AE_{\text{within}}$$), and between-sector allocative efficiency ($$\Delta AE_{\text{between}}$$) (equation 7, p. 1625):

$$
\Delta \ln TFP \simeq \Delta TE + \Delta AE_{\text{within}} + \Delta AE_{\text{between}}. \tag{7}
$$

Within-sector allocative efficiency in sector $$s$$ depends on the weighted sum of firm-level wedge changes (equations 9-10, pp. 1626-1627); allocative efficiency deteriorates when wedge dispersion across firms grows. The paper focuses on the within-sector component since between-sector allocative efficiency is small in the data (Figure 6, p. 1627).

## Empirical specifications

**Firm-bank DiD (primary credit results, R1).** The estimating specification is a dynamic differences-in-differences with a triple interaction at the firm-bank-quarter level (equation 2, p. 1613):

$$
g_{ibt}^{\text{credit}} = \sum_{\tau=-2}^{5} \beta_\tau^{\text{treat}}(\text{period}_\tau \times \text{exposed}_b)
  + \sum_{\tau=-2}^{5} \beta_\tau^{\text{period}}(\text{period}_\tau \times \text{underreported}_{ib})
$$

$$
+ \sum_{\tau=-2}^{5} \beta_\tau^{\text{treatgroup}}(\text{period}_\tau \times \text{underreported}_{ib} \times \text{exposed}_b)
  + \beta_1^{\text{base}}(\text{underreported}_{ib} \times \text{exposed}_b) + \beta_2^{\text{base}} \text{underreported}_{ib}
$$

$$
+ \alpha_2 \mathbf{X}_{ibt} + \theta_{it} + \varphi_b + \epsilon_{ibt}. \tag{2}
$$

Here $$g_{ibt}^{\text{credit}} = \text{credit}_{ibt}/\text{credit}_{ib,t-1} - 1$$, $$\text{exposed}_b$$ is a bank-level dummy, $$\text{underreported}_{ib}$$ is a dummy for lending relationships with prior loss underreporting, $$\text{period}_\tau$$ groups quarters into three-quarter windows, $$\theta_{it}$$ are firm-quarter fixed effects (absorbing all firm-level credit demand shocks), and $$\varphi_b$$ are bank fixed effects. Standard errors are two-way clustered at the firm and bank level. N = 1,981,219. The coefficients of interest are $$\beta_\tau^{\text{treatgroup}}$$, measuring the treatment effect for underreported firms at exposed banks.

**Firm-level DiD (credit confirmation, R2).** To confirm that firms do not undo the bank-level reallocation through other lenders, a dynamic firm-level specification instruments for the firm-level credit shock using the pre-EBA borrowing share from exposed banks (equation 3, p. 1618):

$$
\Delta \log \text{credit}_{it} = \sum_{t=-5}^{10} \Delta_t^{\text{treatgroup}}(\text{quarter}_t \times \text{treatment}_i \times \text{underreported}_i)
  + \sum_{t=-5}^{10} \Delta_t^{\text{treatment}}(\text{quarter}_t \times \text{treatment}_i)
  + \text{controls} + \alpha_1 \mathbf{X}_{it} + \theta_i + \epsilon_{it}. \tag{3}
$$

N = 1,346,771 firm-quarter observations. Standard errors clustered at the firm level.

**IV for real effects (R3-R4).** To estimate pass-through to inputs, an IV strategy instruments the log change in firm-level credit with the normalized pre-EBA borrowing share from exposed banks interacted with the underreported dummy (equation 6, p. 1623):

$$
\Delta \log y_{is} = \gamma \Delta \log \text{credit}_{is} + \text{controls} + u_{is}, \tag{6}
$$

where $$y_{is}$$ is labor or capital. The instrument is the first-stage version of equation (3). The regression is estimated at annual frequency for 2012. N = 104,499. Standard errors clustered by industry. First-stage F-statistic = 111.2 (above Stock-Yogo 5% maximal-bias criterion). Capital elasticity = 0.14, labor elasticity = 0.52; TFP elasticity is near zero and insignificant (Table 5 Panel B, p. 1625).

**Productivity aggregation (R5-R6).** Firm-level predicted wedge changes are computed as:

$$
\frac{\Delta \hat{\tau}_{it}^X}{1 + \tau_{it,t-1}^X} = \left(\hat{\gamma}_1^X + \hat{\gamma}_2^X \times \text{capital wedge}_{i,t-1}\right) \times \left(\hat{\Delta}^{\text{treatment}} \text{borrowing share}_{is}\right), \tag{12}
$$

for $$X = K, L$$. These firm-level changes are aggregated using equations (9)-(10) to estimate the contribution to within-sector allocative efficiency. The between-sector contribution (equation 11) is small and excluded from the headline calculation.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Portuguese Credit Register (Central de Responsabilidades de Credito) | Monthly firm-bank loan balances by overdue bucket; universe of lending relationships above EUR 50 (2009-2015); primary source for underreporting algorithm and DiD | No page yet |
| Simplified Corporate Information / IES (Informacao Empresarial Simplificada) | Annual mandatory firm census; value added, employment, capital, sales, balance sheet; used for wedge measurement and productivity decomposition (2009-2015) | No page yet |
| Banco de Portugal quarterly bank balance sheet data | Bank-level capital ratios, sovereign bond holdings, liquidity; determines EBA exposure definition (2010-2012) | No page yet |
| EBA capital exercise disclosures | Bank eligibility and sovereign debt buffer used to define the exposed/nonexposed distinction (October 2011) | No page yet |

Sample: quarterly loan data 2009:I to 2014:IV; annual firm data 2009-2015. Firms cover 81% of Portuguese sales and 73% of assets. Underreporting measured on firm-finance loans (36% of banks' corporate portfolio; 73% collateralized). 56% of firms have multiple lending relationships, required for the within-firm firm-bank DiD.

## When to read the full paper

Use the [original](https://doi.org/10.1257/aer.20190149) if you are: (i) replicating the underreporting algorithm or DiD design (appendices A-C give the full algorithm, validity checks, and productivity decomposition details); (ii) studying how ratio-based capital requirements distort credit composition in distress (the paper tests risk-shifting vs. loss-delay mechanisms in Table 3); (iii) interested in the link between credit misallocation and aggregate TFP measurement in a bank-dependent economy; or (iv) extending the Hsieh and Klenow (2009) or Restuccia and Rogerson (2008) wedge framework to a credit-supply channel.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(6), June 2023. Replication data available at [ICPSR E120003V1](https://doi.org/10.3886/E120003V1). This distillation was extracted by an LLM on 2026-06-24 and is **not human-verified or independently reproduced**.

> Blattner, Laura, Luisa Farinha, and Francisca Rebelo. "When Losses Turn into Loans:
> The Cost of Weak Banks." *American Economic Review* 113, no. 6 (June 2023): 1600-1641.
> DOI: 10.1257/aer.20190149. Copyright 2023 American Economic Association.
> Paywalled; this page is an extract-only distillation.


==============================================================================
# Long and Short Run of Trade Elasticities: Boehm, Levchenko & Pandalai-Nayar (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/boehm-et-al-long-short-run-trade-2023/
# Distilled: using MFN tariff variation and local projections, this paper estimates the trade elasticity at every time horizon, finding -0.76 in the short run and approximately -2 in the long run, converging over 7-10 years. Long-run estimates are substantially smaller in absolute value than conventional wisdom, implying welfare gains from trade five to six times larger than standard estimates. AER 2023, paywalled. Six core results with source locators, datasets, the dynamic model, and the MFN instrumental variable.
# Tags: paper-summary, international-trade, trade-elasticity, gravity, local-projections
==============================================================================

**What this is.** This is a machine-distilled skeleton of the paper. Read the original at [doi.org/10.1257/aer.20210225](https://doi.org/10.1257/aer.20210225) to replicate or extend.

## TL;DR

Boehm, Levchenko, and Pandalai-Nayar estimate the trade elasticity at every time horizon from 0 to 10 years. Using variation in MFN (most-favored-nation) tariffs across minor trading partners as a plausibly exogenous instrument, and local projections (Jordà 2005) to trace the full time path, they find an elasticity of -0.76 one year after a tariff shock, converging to approximately -2 after 7-10 years. Conventional log-levels OLS gravity estimates (-3.7 to -7.0) are biased by omitted bilateral taste and trade-cost shocks correlated with tariffs; controlling for bilateral unobservables sharply reduces the estimates. The lower long-run elasticity implies that the welfare-relevant trade elasticity is about -1, and applying the Arkolakis, Costinot, and Rodriguez-Clare (2012) gains-from-trade formula shows welfare gains five to six times larger than under the conventional value of -5.

## Core results

| # | Result | Locator | Magnitude as reported |
|---|---|---|---|
| R1 | Short-run (h=1) trade elasticity, preferred baseline IV | Figure 2, p. 876; Table 3 col 1, p. 885 | -0.76 (se=0.11) |
| R2 | Long-run (h=10) trade elasticity, preferred baseline IV | Figure 2, p. 876; Table 3 col 1, p. 885 | -2.12 (se=0.32) |
| R3 | Log-levels OLS with multilateral resistance FE only | Table 1 cols 1-2, p. 880 | -3.70 (se=0.02) to -6.96 (se=0.05) |
| R4 | Log-levels OLS with bilateral product FE (importer-exporter-HS4) | Table 1 col 3, p. 880; Table 2 col 6, p. 882 | -1.04 (se=0.02) |
| R5 | Welfare gains from trade, ACR formula, welfare-relevant theta=-1 | Figure 7, pp. 900-901 | US: 5.27%; world median (64 countries): 22.9% |
| R6 | Sectoral long-run elasticity range, 11 HS sections | Figure 3, pp. 877-878 | -0.75 to -5 (median years 7-10) |

**Overall.** The preferred IV estimates of the trade elasticity are -0.76 in the short run, falling to about -2 in the long run. It takes 7-10 years for estimates to stabilize. The welfare-relevant elasticity (adding 1 to the tariff-exclusive estimate to account for tariff-inclusive spending) is about -1, implying gains from trade five to six times larger than under the conventional elasticity of -5. Controlling for bilateral unobservables is the single most important factor distinguishing the paper's estimates from conventional ones.

## Theory / model

The paper develops a partial equilibrium (PE) dynamic model of sluggish adjustment to trade cost shocks (§V.A, pp. 890-897), nesting dynamic versions of the Krugman (1980), Melitz (2003), and Arkolakis (2010) models. Trade in period t is:

$$X_t = p_t^* q_t n_t,$$

where $$p_t^*$$ is the exporters' price exclusive of tariffs, $$q_t$$ is quantity per unit mass, and $$n_t$$ is the mass of active exporters. Crucially, $$p_t^*$$ and $$q_t$$ adjust instantaneously to tariff changes, while $$n_t$$ is predetermined by one period: it captures the entry and investment decisions made last period.

The value of exporting $$v_t$$ and the mass dynamics are governed by (eqs. 8-9, p. 891):

$$v_t = \frac{1}{1+r}\,E_t\bigl[\pi_{t+1} + (1-\delta)v_{t+1}\bigr], \tag{8}$$

$$n_t = n_{t-1}(1-\delta) + G(v_{t-1}), \tag{9}$$

where $$\delta$$ is the exogenous exit rate and $$G(\cdot)$$ is an increasing function mapping the value of exporting to new entrants. Solving (8) forward and (9) backward yields analytical expressions for $$v_t$$ and $$n_t$$ in terms of the tariff path.

Define four elasticities (eq. 7, p. 891): $$\eta_{q,p} := \partial\ln q/\partial\ln p^*$$, $$\eta_{q,\tau} := \partial\ln q/\partial\ln\tau$$, $$\eta_{p,\tau} := \partial\ln p^*/\partial\ln\tau$$, $$\eta_{\pi,\tau} := \partial\ln\pi/\partial\ln\tau$$. The short-run trade elasticity, where $$n_t$$ is fixed (eq. 12, p. 892), is:

$$\varepsilon^0 := (1 + \eta_{q,p})\eta_{p,\tau} + \eta_{q,\tau}. \tag{12}$$

The long-run elasticity adds the endogenous adjustment of $$n$$ to its steady-state value (eq. 13, p. 892):

$$\varepsilon := \varepsilon^0 + \chi\eta_{\pi,\tau}, \tag{13}$$

where $$\chi := g(v)v/G(v) > 0$$ captures the elasticity of the mass of exporters with respect to the value of exporting. Because $$\eta_{\pi,\tau} < 0$$ and $$\chi > 0$$, the long-run elasticity is strictly larger in absolute value than the short-run elasticity. In the CES-monopolistic competition version, $$\varepsilon^0 = -\sigma$$ and $$\varepsilon = -\sigma(1 + \chi)$$.

**Proposition 2** (p. 894) establishes that $$\lim_{h\to\infty}\varepsilon^h = \varepsilon$$ as long as the tariff shock is not fully mean-reverting, validating the horizon-10 estimates as long-run estimates.

**Proposition 3** (p. 895) shows that the model delivers the local-projections estimating equation (2) up to first order. The importer-product-time and exporter-product-time fixed effects absorb weighted averages of past, present, and expected future demand and supply shifters: dynamic analogues of the Anderson and van Wincoop (2003) multilateral resistance terms.

The model is calibrated with $$\sigma = 1.1$$, $$\chi = 0.82$$, $$\delta = 0.25$$, $$r = 0.03$$ to match the empirical time path of elasticities (Figure 5, p. 897). Convergence to the long run is geometric at rate $$\delta$$, taking approximately a decade.

## Method

The horizon-h trade elasticity $$\varepsilon^h$$ is estimated by combining **local projections** (Jordà 2005, builds-on) with a **WTO MFN instrumental variable**. The key innovation relative to conventional gravity estimation is that a separate regression is run at each horizon $$h = 0, 1, \ldots, 10$$, tracing the full impulse-response function of trade to tariff shocks without imposing a parametric dynamic model.

The horizon-h trade elasticity is defined (eq. 1, p. 865) as:

$$\varepsilon^h := \frac{\Delta_h \ln X_{i,j,p,t}}{\Delta_h \ln \tau_{i,j,p,t}}, \tag{1}$$

where $$i$$ indexes the importing country, $$j$$ the exporting country, $$p$$ the product, $$t$$ time, and $$\Delta_h x_t := x_{t+h} - x_{t-1}$$ is the h-period change. The long-run elasticity is $$\varepsilon = \lim_{h\to\infty}\varepsilon^h$$.

**Instrument.** To address the endogeneity of tariffs, the paper exploits the WTO's MFN principle: when an importing country changes its applied MFN tariff on a product, all WTO partners trading on MFN terms experience that change. Minor trading partners (not among the top-10 exporters of product p to importer i) are unlikely to have driven the tariff change. The baseline instrument (eq. 5, p. 870) is:

$$\Delta_0\ln\tau^{\text{instr}}_{i,j,p,t} = \mathbf{1}\!\left\{\tau_{i,j,p,t} = \tau^{\text{appliedMFN}}_{i,j,p,t}\right\} \times \mathbf{1}\!\left\{\tau_{i,j,p,t-1} = \tau^{\text{appliedMFN}}_{i,j,p,t-1}\right\} \times \left(\ln\tau^{\text{appliedMFN}}_{i,j,p,t} - \ln\tau^{\text{appliedMFN}}_{i,j,p,t-1}\right), \tag{5}$$

retaining only observations where exporter j is not a top-10 trading partner of importer i (in total trade or in product p). Countries in preferential trade agreements (PTAs) with the importer serve as the control group: their applied tariffs differ from MFN rates so they do not experience the MFN tariff change. The instrument is equivalent to an instrumented difference-in-differences comparing minor MFN partners (treated) to PTA partners (control).

Standard errors are clustered at the country-pair-product level throughout; first-stage F-statistics exceed 10 at all horizons (Online Appendix Table B2).

## Empirical specifications

The combined specification estimated at each horizon h is (eq. 4, p. 867):

$$\Delta_h \ln X_{i,j,p,t} = \beta^h \Delta_h \ln\tau_{i,j,p,t} + \delta^{d,h}_{i,p,t} + \delta^{s,h}_{j,p,t} + \delta^{b,h}_{i,j,p} + u^h_{i,j,p,t}, \tag{4}$$

where $$\delta^{d,h}_{i,p,t}$$ is an importer-HS4-year fixed effect, $$\delta^{s,h}_{j,p,t}$$ is an exporter-HS4-year fixed effect, and $$\delta^{b,h}_{i,j,p}$$ is a source-destination-product fixed effect (absorbing bilateral trends in trade). All specifications include one lag of log changes in tariffs and trade as pretrend controls. When $$\Delta_0\ln\tau^{\text{instr}}_{i,j,p,t}$$ instruments for $$\Delta_h\ln\tau_{i,j,p,t}$$, the IV estimator $$\hat{\beta}^h$$ identifies $$\varepsilon^h$$.

To account for tariff autocorrelation, the paper also runs a complementary local projection of the tariff change (eq. 3, p. 867):

$$\Delta_h \ln\tau_{i,j,p,t} = \beta^h_\tau \Delta_0\ln\tau_{i,j,p,t} + \delta^{d,\tau,h}_{i,p,t} + \delta^{s,\tau,h}_{j,p,t} + \delta^{b,\tau,h}_{i,j,p} + u^{\tau,h}_{i,j,p,t}, \tag{3}$$

with the trade elasticity recovered as $$\varepsilon^h = \hat{\beta}^h_X / \hat{\beta}^h_\tau$$. This is important because tariff changes are autocorrelated in the data: about 80% of the initial shock survives at 5 years and 75% at 10 years (Figure 1 panel A, p. 875). Failing to account for this autocorrelation would cause h-period differences to conflate elasticities across horizons.

**Bias diagnostics (Table 1, p. 880).** A log-levels OLS specification assuming all tariff variation is exogenous, similar to Head and Ries (2001), yields coefficients of -3.70 to -6.96. Conventional static gravity in log-levels without bilateral effects:

$$\ln X_{i,j,p,t} = \beta\ln\tau_{i,j,p,t} + \delta^d_{i,p,t} + \delta^s_{j,p,t} + u_{i,j,p,t}$$

yields $$\hat{\beta} = -6.96$$ (col 2) or $$-3.70$$ (col 1 without multilateral resistance). Adding bilateral product fixed effects $$\delta^b_{i,j,p}$$ in log-levels (Table 2, p. 882) yields $$-1.04$$, demonstrating that controlling for bilateral unobservables is the dominant force pushing estimates toward the paper's preferred IV values.

The 5-year IV baseline is $$-1.24$$ (Table 3 col 1, p. 885) and the 10-year baseline is $$-2.12$$, consistent across a wide range of robustness checks: alternative pretrend lags (Table 3 cols 2-3), alternative clustering (col 5), constant sample (col 6), and extensive margin specifications (cols 7-8).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| BACI (CEPII version of UN COMTRADE) | Trade values and quantities at HS6 level, 183 economies, 1995-2018; main outcome variable | no page yet |
| UN TRAINS (UNCTAD, 1995-2018) | Applied and MFN tariff rates at HS6 level; source of the instrument and the tariff regressor | no page yet |

Sample: 183 economies, 5,000+ HS6 product categories, 1995-2018 (annual). Baseline estimation sample at h=1: approximately 26 million country-pair-product-year observations. Additional gravity covariates (distance, common border, common language, colonial relationship) from the CEPII GeoDist database are used in robustness exercises.

## When to read the full paper

Read the original when you need: (i) first-stage F-statistics and the full robustness matrix across all horizons (Online Appendix Tables B1-B8); (ii) the multicountry multisector general equilibrium extension and GE impulse responses of US imports to tariff shocks (Figure 6, pp. 898-899); (iii) sectoral heterogeneity results broken down by 11 HS sections and benchmarked to Ossa (2015) (Figure 3, pp. 877-878); (iv) proofs of Propositions 1-3 on the impulse-response of firm mass and the micro-foundation of the estimating equation (Online Appendix C); or (v) country-level gains-from-trade quantification at alternative elasticity values (Online Appendix Table B9).

## Attribution and rights

Boehm, Christoph E., Andrei A. Levchenko, and Nitya Pandalai-Nayar. 2023. "The Long and Short (Run) of Trade Elasticities." *American Economic Review* 113(4): 861-905. doi:10.1257/aer.20210225.

Replication code: Boehm, Levchenko, and Pandalai-Nayar (2023). "Replication Data for: The Long and Short (Run) of Trade Elasticities." American Economic Association / ICPSR openICPSR. doi:10.3886/E182781V1.

This page is an LLM-distilled extract (not human-verified, not reproduced). Results are drawn from Figures 1-7 and Tables 1-4 of the published paper. Extract-only redistribution; the original is paywalled via the AEA.


==============================================================================
# Technological Change and Job-Loss Consequences: Braxton & Taska (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/braxton-taska-technological-change-consequences-job-2023/
# Distilled: Using Burning Glass Technologies online vacancy data to measure within-occupation technological change, Braxton and Taska show that workers displaced from more tech-exposed occupations suffer earnings declines over 7 percentage points larger per standard deviation of exposure, are 17 pp more likely to switch occupations, and that a calibrated structural search-and-matching model attributes 45 percent of post-displacement earnings losses to technological change. American Economic Review 2023, paywalled. Six core results with source locators, datasets used, the simple two-period model and the quantitative model with Bellman equations, and the empirical specifications.
# Tags: paper-summary, labor-economics, technological-change, job-loss
==============================================================================

**What this is.** The paper's core results, the two-period model and the
quantitative model with their key equations, and the empirical specifications:
enough to understand what was found and how, without reading all 38 pages. To
replicate or extend, read the full source at the
[original](https://doi.org/10.1257/aer.20210182).

## TL;DR

Braxton and Taska use within-occupation changes in computer and software skill
requirements from the Burning Glass Technologies vacancy database (2007-2017)
to measure technological change. Merging this measure with the Current
Population Survey Displaced Workers Supplement (DWS), they document that
workers displaced from occupations with greater technological change (i) suffer
larger earnings declines, (ii) are more likely to switch to a lower-paying
occupation, and (iii) show earnings losses concentrated entirely among those
who do switch. A calibrated search-and-matching model with an "up-to-the-task"
production function attributes 45.5 percent of average post-displacement
earnings losses to technological change (occupation-specific human capital
accounts for 34.5 percent; moving lower on the wage ladder for 20 percent).

## Core results

Magnitudes and significance are as reported; all regressions use clustered
standard errors at the occupation level (SE in parentheses). Locators point
into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Workers displaced from higher-tech-change occupations suffer **larger earnings declines** | Table 3, Col 1, p.295 | Coefficient on $$\Delta\bar{z}_o$$ = -0.0354 (SE 0.0114); 1 SD higher exposure implies >7 pp larger earnings decline |
| R2 | Higher tech-change exposure raises the **probability of occupation switching** after displacement | Table 4, Col 1, p.297 | Coefficient = 0.0847 (SE 0.0269); 1 SD higher exposure implies 17 pp greater probability of switching |
| R3 | **Earnings losses are concentrated among occupation switchers**; stayers are unaffected | Table 5, Col 3, p.299 | Interaction ($$\Delta\bar{z}_o \times S_{i,o,t}$$) = -0.0494 (SE 0.0180); coefficient for stayers = 0.000532 (SE 0.0131), insignificant |
| R4 | Tech-change exposure does **not raise the probability of displacement** | Table 6, Col 1, p.301 | Coefficient = -0.00125 (SE 0.00320), t-stat = -0.39; economically negligible (0.25 pp vs mean displacement rate of 6.7%) |
| R5 | Occupation **stayers** more exposed to tech change earn **more** over time | Table 7, Col 1, p.302 | Coefficient = 0.00270 (SE 0.000705); 1 SD above mean implies >0.5 pp higher earnings over 12 months |
| R6 | Model decomposition: **tech change accounts for 45.5%** of post-displacement earnings decline | Figure 5, p.313; §V.E | Full model: 7.63% average earnings decline; model without tech change: 4.16%; residual after also removing occ-specific HC (experience): 1.53% |

**Overall (paper's conclusion).** Technological change, measured by changes in
computer and software skill requirements in vacancy postings, explains a large
share of post-displacement earnings losses. The mechanism is occupation
switching: workers whose occupation introduced new technology during their
employment period no longer have the skills to match with newly created jobs
there, and transition to lower-tech, lower-wage occupations. This gives a
new rationale for the human-capital-decline modeling device used by
Ljungqvist and Sargent (1998) and others, and suggests retraining policies
may be a productive complement to unemployment insurance. The earnings losses
after job loss documented by Jacobson, LaLonde and Sullivan (1993) and by
Couch and Placzek (2010) are here decomposed by channel; Huckfeldt (2022)
shows displacement losses concentrate among occupation switchers, and this
paper provides technological change as the mechanism driving the switch;
Davis and von Wachter (2011) document larger earnings losses in recessions,
and the paper suggests tech-change acceleration in recessions may contribute.

## Theory / model

### Simple two-period model (Section I, pp. 283-285)

The paper begins with a two-period model to derive the testable predictions.
Two occupations, L (low-tech) and H (high-tech), use technology levels
$$z_L < z_H$$, where $$z_L = (1-\eta)z_H$$ with $$\eta > 0$$. An "up-to-the-task"
production function defines output of a worker-firm match (p.283):

$$
f(h, z) = \begin{cases} z, & \text{if } h \geq z \\ 0, & \text{otherwise} \end{cases}
$$

Workers have general human capital $$h \sim F(h)$$ and receive share
$$\omega \in (0,1)$$ of output as wages. In period 2, a new technology
$$z'_H = (1+\gamma)z_H$$ is introduced in occupation H (embodied in new matches,
as in Mortensen and Pissarides (1998) and Violante (2002)). Workers employed
in H in period 1 who do not have skills for the new technology (i.e.,
$$z'_H > h \geq z_H$$) must move to occupation L after displacement.

Define $$\pi = \frac{F(z'_H) - F(z_H)}{1 - F(z_H)}$$ as the share of H-employed
workers who lack the skills to use the new technology (p.284). The model
delivers three predictions:

- **Prediction 1**: If $$\pi > \frac{\gamma}{\eta + \gamma}$$, workers displaced from
  H experience larger average earnings losses than workers displaced from L.
- **Prediction 2**: Workers displaced from H are more likely to switch
  occupations (whenever $$\pi > 0$$).
- **Prediction 3**: The larger earnings losses among H-displaced workers are
  concentrated among occupation switchers.

These predictions guide the empirical strategy in Section IV.

### Quantitative model (Section V, pp. 303-312)

The quantitative model extends the simple model to an infinite-horizon,
overlapping-generations search environment with $$K = 10$$ occupations. Time is
discrete. Technology grows at rate $$g > 0$$ per year (calibrated to 1.5 percent,
from the Burning Glass data). Occupation $$k$$ has technology intensity
$$c_k \in [0,1]$$, so the technology level in occupation $$k$$ at time $$j$$ is
$$z_{k,j} = c_k z_j$$. Workers live $$T = 120$$ quarters (30 years).

Workers are heterogeneous in general human capital $$h$$ and occupation-specific
experience $$x \in \{E, N\}$$. The up-to-the-task production function is
(p.308):

$$
f(c_k z, h, x) = \begin{cases} A_x c_k z, & \text{if } A_x h \geq c_k z \\ 0, & \text{otherwise} \end{cases}
$$

where $$A_N = 1$$ (inexperienced) and $$A_E = 1.12$$ (experienced), capturing a
12 percent productivity premium from occupation-specific human capital
(following Kambourov and Manovskii (2009)).

Matching follows a constant-returns-to-scale matching function (p.308):

$$
M(s, v) = \frac{sv}{\left(s^\xi + v^\xi\right)^{1/\xi}}, \quad \xi = 1.6
$$

Technology at an existing match and the worker's general human capital both
evolve stochastically, depreciating at rate $$\mu = 1/(1+g)$$ each period with
probability $$\iota = 0.25$$ per quarter:

$$
Z(z) = z' = \begin{cases} z\mu, & \text{with pr. } \iota \\ z, & \text{with pr. } 1-\iota \end{cases}
\qquad
H(h) = h' = \begin{cases} h\mu, & \text{with pr. } \iota \\ h, & \text{with pr. } 1-\iota \end{cases}
$$

**Bellman equations (pp. 306-307).** For an inexperienced, unemployed worker
of age $$t$$ with human capital $$h$$, the value function satisfies:

$$
U_t^N(h, 0) = b + \beta E\!\left[\hat{U}_{t+1}^N(h', 0)\right], \quad \forall t \leq T
$$

where $$b$$ is the public insurance transfer (calibrated so insurance replaces
41.2 percent of lost earnings, using PSID data from 2001-2013), and
$$\hat{U}_{t+1}^N$$ is the value of search over occupations $$k \in \mathcal{K}$$
and wage piece rates $$\omega \in [0,1]$$. For an inexperienced, employed worker
at a firm using technology $$z$$ in occupation $$k$$ with piece rate $$\omega$$:

$$
W_t^N(h, z, k, \omega) = \omega f(c_k z, h, N)
+ \beta E\!\Bigl\{\delta\hat{U}_{t+1}^N(h', k)
+ (1-\delta)\bigl[\lambda_E \hat{W}_{t+1}^E(h', z', k, \omega)
+ (1-\lambda_E)\hat{W}_{t+1}^N(h', z', k, \omega)\bigr]\Bigr\}
$$

where $$\delta = 0.10$$ per quarter is the exogenous job-destruction rate (from
Shimer (2005)) and $$\lambda_E = 0.05$$ is the quarterly probability of becoming
experienced in the current occupation.

## Method

The empirical approach exploits cross-occupation heterogeneity in the change
in computer and software skill requirements between 2007 and 2017, as
measured in the Burning Glass Technologies vacancy database. Following
Hershbein and Kahn (2018), the tech-change measure $$z_{o,t}$$ for occupation
$$o$$ in year $$t$$ is the share of vacancies in that occupation listing a
computer or software related skill. The paper defines the change as
$$\Delta z_o = z_{o,2017} - z_{o,2007}$$, normalized to mean zero and unit SD.

The identifying assumption (selection-on-observables) is that conditional
on controls including the initial level of computer requirements in 2007, the
change in employment share in the occupation, age, education, gender, tenure
before layoff, and unemployment spell duration, the change in tech requirements
is uncorrelated with potential outcomes. Crucially, Section IV.E (Table 6)
shows that $$\Delta z_o$$ is uncorrelated with the probability of displacement
itself, consistent with the model's structure and validating the approach.

The quantitative model is calibrated to match aggregate labor market moments
for the 2010-2017 period. The model is solved using value function iteration
over a discretized state space, with K = 10 occupations grouped by computer
and software requirements from Burning Glass (the technology intensity
$$c_k$$ is calibrated from smoothed earnings ratios across occupation groups
using CPS data).

## Empirical specifications

### Main earnings and switching regressions (Section IV, pp. 292-299)

The baseline specification (equation 1, p.293) is:

$$
Y_{i,o,t} = \alpha + \beta\Delta\bar{z}_o + \Gamma \mathbf{X}_{i,o,t} + \varepsilon_{i,o,t} \tag{1}
$$

where $$Y_{i,o,t}$$ is the outcome for individual $$i$$ displaced from occupation
$$o$$ in DWS wave $$t$$ (change in log earnings, indicator for occupation
switching, etc.); $$\Delta\bar{z}_o$$ is the occupation-level tech-change measure
normalized to mean zero and unit SD; and $$\mathbf{X}_{i,o,t}$$ is a vector of
controls (age, log unemployment-spell duration, pre-displacement computer
requirements in 2007, tenure before layoff, years of education, gender,
DWS survey year, full-time indicators, change in occupation employment
share). Standard errors clustered at the occupation level (four-digit SOC).

Applied to the sample of 6,742 displaced workers in the DWS (waves
2010, 2012, 2014, 2016, 2018, restricted to ages 25-65, employed before
and after displacement, non-top-coded earnings), specification (1) produces
the earnings (Table 3, R1) and occupation-switching (Table 4, R2) results.

To test whether earnings losses are concentrated among occupation switchers
(Model Prediction 3), the paper estimates (equation 2, p.298):

$$
\Delta\ln(\text{Earn}_{i,o,t}) = \alpha + \gamma S_{i,o,t} + \beta\Delta\bar{z}_o
+ \eta\!\left(\Delta\bar{z}_o \times S_{i,o,t}\right) + \Gamma \mathbf{X}_{i,o,t} + \varepsilon_{i,o,t} \tag{2}
$$

where $$S_{i,o,t}$$ is a dummy equal to one if the individual switches occupations
following displacement. The coefficient $$\eta$$ on the interaction captures
whether the earnings-loss effect of tech change is concentrated among switchers.
Table 5, Column 3 shows $$\hat{\eta} = -0.0494$$ (SE 0.0180) and $$\hat{\beta} =
0.000532$$ (SE 0.0131, insignificant), confirming that occupation stayers are
unaffected by tech change while switchers bear the full cost (R3).

### Displacement probability and occupation stayers (Sections IV.E-F, pp. 300-302)

The same specification (1) is applied to two additional outcomes:
(i) an indicator for being displaced (Table 6), using the full DWS sample of
239,509 individuals (not just those who regained employment), showing that
$$\hat{\beta} = -0.00125$$ (SE 0.00320, t = -0.39), confirming tech change does
not raise displacement probability (R4); and (ii) the 12-month change in log
earnings for occupation stayers in the CPS-ORG (Table 7, N = 150,330),
showing $$\hat{\beta} = 0.00270$$ (SE 0.000705), confirming occupation stayers
benefit from tech change (R5).

### Model decomposition (Section V.E, p.312)

The model is used as a laboratory to decompose earnings losses. Starting from
the full model (average earnings decline 7.63 percent), the paper sets the
technology growth rate to zero ($$g = 0$$) and re-solves the model; the average
decline falls to 4.16 percent, implying tech change accounts for
$$(7.63 - 4.16) / 7.63 = 45.5$$ percent of the total decline. Removing
occupation-specific human capital (setting $$A_E = A_N = 1$$) further reduces
the decline to 1.53 percent, attributing 34.5 percent to occ-specific HC.
The residual 1.53 percent (20 percent share) is attributed to moving lower on
the job ladder.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Burning Glass Technologies vacancy database (2007-2017) | Primary source for the tech-change measure: share of vacancies listing computer or software skills by four-digit SOC occupation per year | No page yet |
| CPS Displaced Workers Supplement (DWS), 2010-2018 waves | Primary displaced-worker sample: earnings and occupation before and after displacement for 6,742 workers | No page yet |
| O\*NET (vintage 15.1, 2005-2010) | Validation of Burning Glass tech-change measure via computer-knowledge ratings by occupation (Figure 1); task content measures (Table 1) | No page yet |
| American Community Survey (ACS), 2007 and 2017 | Employment shares by occupation as controls for demand shifts; calibration of smoothed occupation earnings | No page yet |
| CPS Outgoing Rotation Group (CPS-ORG) | Earnings gains for occupation stayers (Table 7, N = 150,330); nondisplaced comparison group (Table 2) | No page yet |

Sample (displaced workers): ages 25-65, employed both before and at time of
DWS, non-top-coded earnings both before and after, displaced 2007-2017.
Quantitative model calibrated to 2010-2017 at quarterly frequency.

## When to read the full paper

Read the [original](https://doi.org/10.1257/aer.20210182) if you are:
examining the role of technological change in earnings dynamics more broadly
(the paper tests against a range of controls and alternative occupation
definitions); building on the Burning Glass vacancy data for measuring
skill requirements (Section III and online Appendices B-C); extending the
search-and-matching model to study retraining policy (the paper references
a companion paper on optimal retraining subsidies); or replicating the
decomposition exercise (replication data are available at the ICPSR).

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(2), February 2023.
This distillation was extracted by an LLM on 2026-06-25 and is **not
human-verified or independently reproduced**. The AER is paywalled; no
open-access or CC rights were identified. Extract-only.

> Braxton, J. Carter, and Bledi Taska. "Technological Change and the
> Consequences of Job Loss." *American Economic Review* 113, no. 2
> (February 2023): 279-316. DOI: 10.1257/aer.20210182.
> Replication data: https://doi.org/10.3886/E181166V1


==============================================================================
# Behavioral Foundations of Default Effects: Brot-Goldberg, Layton, Vabson & Wang (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/brot-goldberg-et-al-behavioral-foundations-default-effects-2023/
# Distilled: Default rules in Medicare Part D have large, persistent effects on enrollment and drug utilization; beneficiary passivity is insensitive to the value of the default even when following it causes drug consumption losses up to 30 percent. Evidence favors "mental gap" over "frictional" models of default-following, implying that optimal policy should match beneficiaries to their best plans rather than incentivize active choice. AER 2023, paywalled. Seven core results with source locators, datasets used, the theoretical framework, and the empirical specifications.
# Tags: paper-summary, health-insurance, behavioral-economics, default-effects
==============================================================================

**What this is.** The paper's core results, the theoretical framework distinguishing frictional from mental gap models of default-following, and the three natural-experiment designs used to test them: enough to know what was found and how, without reading all 41 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1257/aer.20210013).

## TL;DR

Default rules in the Low-Income Subsidy (LIS) segment of Medicare Part D determine what plan a beneficiary is enrolled in, and most beneficiaries never change their default assignment. Using random assignment of new enrollees to benchmark plans and two regression discontinuity and difference-in-differences designs that exploit premium-subsidy threshold crossings, the authors show that: (i) default changes shift virtually all beneficiaries through passive reassignment, not active choice; (ii) these reassignments reduce annual drug spending by 6.4 percent on average, with losses up to 30 percent for those assigned to poorly fitting plans; and (iii) beneficiaries whose randomly assigned defaults are worst-fitting are only marginally more likely to make an active choice, even as the consumption losses grow dramatically. A variance decomposition reveals that two-thirds of variation in latent attention comes from random transitory shocks within a beneficiary over time, not from permanent friction heterogeneity. This evidence favors "mental gap" models (Handel and Schwartzstein (2018)), where attention is driven by factors unrelated to the stakes, over the "frictional" class of models used in most prior work (Handel (2013); Abaluck and Gruber (2011)). The welfare implication: optimal default policy should paternalistically match beneficiaries to their best plans, not try to incentivize active choice.

## Core results

Magnitudes and significance are as reported. Locators refer to pages in the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Only 16% of new LIS beneficiaries opt out before initial auto-enrollment**; 84% passively follow their randomly assigned default. Even after 5 years, only 45% have ever made an active choice | Figure 1, p. 2727 | Active choice rate stays flat below 50% across all quintiles of default-plan fit, including for beneficiaries assigned to plans covering few of their prior drugs |
| R2 | **A change in default from remaining to reassignment raises plan switching by 96 percentage points**, driven almost entirely by passive reassignment (93.6 pp), not active choice (2.4 pp) | Table 2, p. 2734 | RD estimate beta = 0.960 (SE 0.001); average switching rate for controls at cutoff = 0.006 |
| R3 | **Reassignment default reduces annual drug spending by 6.4 percent** (approximately $213 off a $3,329 control base), a non-trivial consumption loss | Table 3 Panel A, p. 2738 | DiD coefficient: -0.064 (SE 0.004) on log drug spending; effect persists at least 24 months; robust to drug-level and class-level price normalizations |
| R4 | **Worst-fitting default plans reduce drug spending by 12.6 percent**, nearly 3x the 4.3 percent loss for better-fitting defaults; high-fit-variance subsamples face losses up to nearly 30 percent | Table 3 Panel B, p. 2738; Table 6, p. 2747 | Worst-quintile interaction: -0.083 (SE 0.007); top 50% / 25% / 10% variance subsamples face total losses of -24.9% / -27.6% / -29.8% |
| R5 | **Active choice propensity barely responds to the value of the default**: beneficiaries assigned to worst-fitting plans are only about 1.5 pp more likely to make an active choice than those assigned to better plans | Table 5 Panel B, p. 2747; Figure 7, p. 2749 | Worst-quintile differential: 0.015 (SE 0.002); slope of active choice vs drug consumption loss is nearly flat across the full range; ∂a/∂v^d ≈ 0 |
| R6 | **26.5 percent of beneficiaries are "sometimes choosers"** whose plan exits, far exceeding the ~4.3 percent predicted by a constant-friction frictional model | p. 2752 | Prior passive auto-assignees make an active choice 7.7% of the time following plan exit; prior active choosers make an active choice only 25.6% of the time - far below the ~100% predicted by a constant-friction model |
| R7 | **Two-thirds of variation in latent attention is driven by within-beneficiary transitory shocks**, not permanent friction differences: transitory share = 66.6%, permanent heterogeneity = 33.4% | Table 7, p. 2754 | Observable characteristics (age, gender, race, Elixhauser index) explain only 3.2% of total attention variance; unobservable permanent heterogeneity = 30.6%; transitory shocks = 66.2% |

**Overall (paper's conclusion).** Beneficiaries are overwhelmingly passive in their Medicare Part D plan choices, even when that passivity generates large and immediate drug consumption losses. Because active choice is insensitive to the value of the default, the evidence fits mental gap rather than frictional models of default-following. This means that policymakers cannot incentivize their way out of passivity: optimal default design should match beneficiaries to their best outcomes, not shock them into making active choices.

## Theory / model

Section V (pp. 2740-2743) formalizes two classes of default-following models and derives the welfare implications that distinguish them, following the taxonomy of Handel and Schwartzstein (2018).

**Active choice decision.** Beneficiary $$i$$ makes an active choice if and only if $$A(v^*, v^d, c, \ldots) \geq 0$$, where $$v^*$$ is the payoff from the best active choice, $$v^d$$ is the payoff of the default option, and $$c$$ represents the psychic, search, or attention cost of making an active choice (p. 2741). Active choosers receive $$v^* - c$$; passive followers receive $$v^d$$.

**Welfare.** Expected welfare is (p. 2742):

$$
W = \Pr\!\bigl(A \geq 0\bigr) \cdot E\bigl[v^* - c \mid A \geq 0\bigr] + \Pr\!\bigl(A < 0\bigr) \cdot E\bigl[v^d \mid A < 0\bigr]
$$

Differentiating with respect to the default value $$v^d$$ yields the welfare derivative (equation 2, p. 2742):

$$
\frac{\partial W}{\partial v^d} = \underbrace{1 - a(v^*, v^d, \ldots)}_{\substack{\text{benefit to}\\\text{inframarginal}\\\text{passive agents}}} + \underbrace{\frac{\partial a}{\partial v^d}}_{\substack{\text{attention}\\\text{elasticity}}} \times \underbrace{E\!\bigl[v^* - c - v^d \mid A = 0\bigr]}_{\substack{\text{value of choice for}\\\text{marginal agents}}} \tag{2}
$$

where $$a = \Pr(A \geq 0)$$ is the aggregate share making an active choice.

**The frictional vs mental gap distinction.** In "frictional" models, $$\partial a / \partial v^d < 0$$: agents respond to a worse default by making active choices more often (the second term in equation (2) is non-zero), so optimal policy can improve welfare by setting the default so badly that all agents switch - the "shocking default" logic of Carroll et al. (2009) and Bernheim, Fradkin, and Popov (2015). In "mental gap" models, $$\partial a / \partial v^d = 0$$: active choice is driven by random attention or contextual salience, not by material stakes. The second term in equation (2) drops out, leaving $$\partial W / \partial v^d = 1 - a > 0$$: welfare always rises when the default improves, so the optimal default maximizes $$v^d$$ (a "smart" default matching beneficiaries to their best plan, as in Thaler and Sunstein (2003)).

**Structural model of latent attention (Section VII, p. 2752).** To decompose the variation in active choice, the paper models latent attention for beneficiary $$i$$ in year $$t$$ as:

$$
A_{it} = c_i + k_{it}
$$

where $$c_i \sim \mathcal{N}(\mu X_i, \sigma^2)$$ captures permanent, beneficiary-specific drivers of attention (with observable characteristics $$X_i$$) and $$k_{it} \sim \mathcal{N}(0, 1)$$ are i.i.d. transitory shocks across periods. A beneficiary makes an active choice when $$A_{it} \geq 0$$. Estimated by maximum likelihood on sequences of active/passive choices observed for the same beneficiary over multiple plan-exit events (Online Appendix E). The key testable implication is the share of total variance explained by $$k_{it}$$ (transitory) versus $$c_i$$ (permanent): a frictional model with fixed individual frictions predicts the permanent component dominates; the empirical result (two-thirds transitory) falsifies that prediction.

## Method

The paper uses three natural experiments within the LIS segment of Medicare Part D. All exploit the institutional structure of the LIS program, where defaults are quasi-randomly assigned or quasi-randomly changed via premium-subsidy threshold crossings.

**Experiment 1 - New enrollee random assignment (Section II, pp. 2726-2729).** When a beneficiary first qualifies for Medicare at age 65, she is randomly assigned to a benchmark plan in her service region (stratified randomization to avoid insurer gaming). The paper tracks cumulative active choice rates over 60 months post-enrollment using the 20 percent CMS claims sample, using `difference-in-differences`-style comparisons across subgroups (fit quintiles, health status) and cohort-year cells.

**Experiment 2 - Regression discontinuity in plan benchmark status (Section III, pp. 2729-2735).** For continuing LIS beneficiaries enrolled by auto-assignment, the default switches from "remain in your plan" to "be randomly reassigned to a new benchmark plan" if the incumbent plan sets a year-$$t$$ premium bid just above the regional subsidy level (losing benchmark status). This provides quasi-random variation in default rules around the subsidy cutoff. The `regression-discontinuity-design` estimating equation is (p. 2733):

$$
\Pr(\text{Switch Plans})_{it} = \beta \cdot \mathbf{1}\!\bigl\{B_{jt} - S_{rt} > 0\bigr\} + \gamma^{-} (B_{jt} - S_{rt}) \cdot \mathbf{1}\!\bigl\{B_{jt} - S_{rt} \leq 0\bigr\} + \gamma^{+} (B_{jt} - S_{rt}) \cdot \mathbf{1}\!\bigl\{B_{jt} - S_{rt} > 0\bigr\} + \delta X_{it} + \epsilon_{it}
$$

where $$B_{jt}$$ is the year-$$t$$ monthly premium bid for plan $$j$$, $$S_{rt}$$ is the regional LIS subsidy for year $$t$$, and $$X_{it}$$ includes individual controls. Standard errors clustered at incumbent-plan-by-year level; bandwidth restricted to within $6 of the cutoff; region-year fixed effects included. The design is validated by density tests (Figure 2, p. 2731) and covariate balance checks (Figure 3, p. 2732; Table A1).

**Experiment 3 - Stacked difference-in-differences for drug consumption (Section IV, pp. 2736-2740).** To estimate drug consumption effects, the paper stacks all market-year pairs where some plans lose benchmark status, creating a series of experiment-specific `difference-in-differences` comparisons. The pooled estimating equation (equation 1, p. 2736) is:

$$
y_{itd} = \beta\,(\text{BenchmarkLoss}_{id} \times \text{Post}_{td}) + \gamma_{id} + \eta_{hd} + \epsilon_{itd} \tag{1}
$$

where $$y_{itd}$$ is log total allowed prescription drug spending for beneficiary $$i$$ in quarter $$t$$ in experiment-cohort $$d$$, $$\gamma_{id}$$ are individual-by-experiment fixed effects, and $$\eta_{hd}$$ are event-time-by-experiment fixed effects. Standard errors clustered at the beneficiary level. Final stacked DiD sample: 5,574,684 person-by-experiment-quarter observations.

## Empirical specifications

**RD enrollment effects (R1-R2).** Linear probability model on 460,729 beneficiary-year observations from the RD analysis sample. Dependent variable = switched plans (December of year $$t-1$$ to January of year $$t$$). Running variable: incumbent plan bid minus regional subsidy. Bandwidth: within $6 of year-$$t$$ benchmark; "de minimis" plans (bid above cutoff but below cutoff plus $2) excluded. Results decomposed by outcome type: overall switching, active-choice switching, benchmark-plan active switching, and passive-reassignment switching (columns 1-4 of Table 2, p. 2734).

**DiD drug consumption (R3-R4).** Stacked DiD per equation (1), outcome = log drug spending. Panel A of Table 3 (pooled across all reassigned beneficiaries): $$\hat{\beta} = -0.064$$ (SE 0.004). Panel B of Table 3 (interacted with worst-quintile fit indicator): main effect $$\hat{\beta} = -0.043$$ (SE 0.004), worst-quintile interaction $$\hat{\delta} = -0.083$$ (SE 0.007), implying total loss of -12.6% for worst-fitting plans vs -4.3% for others. Effects robust to drug-level price normalization (-5.0%, column 2) and class-level normalization (-2.1%, column 3). Drug-type heterogeneity: high-value drugs -6.2%, chronic drugs -7.8%, non-chronic -2.8% (Table 4, p. 2740). Online Appendix C shows effects persist 24 months.

**Active choice elasticity (R5).** Same stacked DiD as equation (1) augmented per equation (3) (p. 2745) to identify $$\partial a / \partial v^d$$:

$$
y_{itd} = \delta_t(\text{LowestFit}_{id} \times \text{BenchmarkLoss}_{id}) + \beta_t\,\text{BenchmarkLoss}_{id} + \gamma_{id} + \eta_{hd} + \epsilon_{itd} \tag{3}
$$

where $$\text{LowestFit}_{id}$$ indicates that the beneficiary's randomly assigned default plan falls in the bottom quintile of formulary fit for her drug consumption. Identification comes from random assignment of plans to beneficiaries. Outcome = binary active-choice indicator by December. Main effect: $$\hat{\beta} = 0.047$$ (SE 0.001). Worst-quintile differential: $$\hat{\delta} = 0.015$$ (SE 0.002) - a 1.5 pp increase in active choice despite a -8.3 pp additional consumption loss. Slope of active choice vs consumption loss is nearly flat across the full range (Figure 7, p. 2749).

**Structural attention decomposition (R7).** Maximum likelihood estimation of the bivariate normal model $$A_{it} = c_i + k_{it}$$ on the exiting-plans subsample (32,852 beneficiaries from 84 exiting plans, 28.3% of whom had previously made an active enrollment choice). Variance shares bootstrap-estimated with 1,000 draws. Without covariates: permanent heterogeneity 33.4% [31.4%, 35.4%]; transitory 66.6% [64.6%, 68.6%]. With age/gender/race/health covariates: observable permanent variation = 3.2% [2.9%, 3.8%]; unobservable permanent = 30.6%; transitory = 66.2%. These shares rule out frictional models with constant, stable individual frictions, as they predict the permanent component should dominate.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CMS Medicare Beneficiary Summary File (2007-2015) | Beneficiary demographics, plan enrollment by month, LIS and Medicare Advantage eligibility | no page yet |
| CMS Medicaid Personal Summary File (NY and TX, 2004-2010) | Demographics and Medicaid eligibility pre-Medicare enrollment for the NY/TX Medicaid-Medicare linked subsample | no page yet |
| CMS Part D Event Files (2007-2015), 20% sample | Prescription drug claims: NDC code, quantity, days supply, date, cost charged to beneficiary and Medicare | no page yet |
| CMS MAX RX Files (NY and TX, 2004-2010), 20% sample | Medicaid prescription claims pre-Medicare enrollment, used to measure prior drug use and plan fit | no page yet |
| CMS Plan Election Type File (2007-2015) | Tracks whether each enrollment spell began through active choice or default auto-assignment; also records the assigned default plan even when the beneficiary opted out | no page yet |
| CMS Plan Characteristics Data (publicly available) | Plan-level monthly premiums, benchmark status, and formulary coverage by year and service region | no page yet |

Sample overview: CMS administrative data for dual-eligible LIS beneficiaries over 65 enrolled in Medicare Parts A, B, and D. Broad LIS sample: 4,628,704 beneficiary-year observations. National MCR initial enrollment sample: 216,772. NY/TX Medicaid-Medicare linked sample: 14,218. RD analysis sample: 1,989,603 (from Table 1, p. 2726); 460,729 beneficiary-year observations restricted to within $6 of the subsidy cutoff. Stacked DiD: 5,574,684 person-by-experiment-quarter observations.

## When to read the full paper

Read the [original](https://doi.org/10.1257/aer.20210013) if you are: (i) evaluating the welfare economics of default design in health insurance and need the full derivation of the optimal default conditions in Online Appendix D; (ii) replicating the RD or stacked DiD designs for other Part D populations or subgroups; (iii) extending the structural attention decomposition to new settings; (iv) studying the institutional details of LIS premium-benchmark rules to design a natural experiment; or (v) modeling attention as a mixture of permanent frictions and transitory shocks in other contexts. The full paper includes Online Appendix E (MLE procedure), Online Appendix C (24-month persistence), and additional robustness tests for the DiD price normalizations and alternative fit measures not covered by this distillation.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(10), October 2023, pages 2718-2758. DOI: [10.1257/aer.20210013](https://doi.org/10.1257/aer.20210013). Replication data and code: [ICPSR E184423V1](https://doi.org/10.3886/E184423V1). This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. The article is paywalled; redistribution is extract-only.

> Brot-Goldberg, Zarek, Timothy Layton, Boris Vabson, and Adelina Yanyue Wang.
> "The Behavioral Foundations of Default Effects: Theory and Evidence from
> Medicare Part D." *American Economic Review* 113, no. 10 (October 2023):
> 2718-2758. DOI: 10.1257/aer.20210013.


==============================================================================
# Smart Contracts and the Coase Conjecture: Brzustowski, Georgiadis-Harris & Szentes (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/brzustowski-et-al-smart-contracts-coase-conjecture-2023/
# Distilled: A durable-good monopolist with access to general dynamic contracts (smart contracts) earns an equilibrium payoff strictly above the low buyer valuation for any discount factor, refuting the Coase conjecture. American Economic Review 2023, paywalled. Four core theoretical results with source locators, the formal model (incentive-compatible abiding contracts), and the two-lemma proof strategy.
# Tags: paper-summary, game-theory, mechanism-design, dynamic-contracting
==============================================================================

**What this is.** The paper's main theorem, the model, the key definitions, and the two-lemma proof strategy: enough to know what it found and how, without reading all 26 pages. To replicate or extend, read the full source at [doi.org/10.1257/aer.20220357](https://doi.org/10.1257/aer.20220357).

## TL;DR

The Coase conjecture states that a durable-good monopolist who cannot commit to future prices will clear the market arbitrarily quickly as the discount factor approaches one, earning only the low buyer valuation. This paper shows the conjecture fails when the seller has access to general dynamic contracts, analogous to smart contracts used in digital markets. The main result (Theorem 1, p. 1343) is that the seller's largest equilibrium payoff is bounded away from the low valuation by a constant that does not depend on the discount factor. The driving mechanism is information storage: smart contracts can hold buyer information the seller does not possess, and abandoning a contract destroys that information. This creates a credible commitment device that breaks the Coasian logic even though the seller retains discretion to switch contracts each period.

## Core results

All results are theoretical; magnitudes are those reported in the paper's propositions and figures.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Coase conjecture fails with dynamic contracts: seller's equilibrium payoff bounded away from $$v_l$$ for all $$\delta$$ | Theorem 1, p. 1343 | There exists $$\underline{\pi} > v_l$$ such that $$\pi(\mathcal{C},\delta) \geq \underline{\pi}$$ for all $$\delta \in (0,1)$$ |
| R2 | Any $$\delta$$-abiding contract provides a lower bound on the seller's equilibrium payoff | Lemma 1, p. 1345 | For any $$\delta$$-abiding $$d \in \mathcal{D}$$, $$\pi(\mathcal{C},\delta) \geq v(d,\delta)$$ |
| R3 | $$\delta$$-abiding contracts exist for all $$\delta \in (0,1)$$ with payoff strictly above $$v_l$$ | Lemma 2, p. 1347 | For all $$\delta \in (0,1)$$, there exists $$d_\delta \in \mathcal{D}$$ with $$v(d_\delta,\delta) \geq \underline{\pi} > v_l$$ |
| R4 | Posted-price seller payoff (Doval and Skreta (2022)) converges to $$v_l$$ as $$\delta \to 1$$; dynamic-contract payoff stays bounded away | Figure 1, §III, p. 1353 | For $$v_l=1, v_h=3, \mu=0.95$$: simple-and-direct equilibrium payoff exceeds posted-price equilibrium payoff for all $$\delta$$; lower bound $$\underline{\pi} > v_l = 1$$ holds uniformly |

**Overall (paper's conclusion).** The Coase conjecture (first articulated by Coase (1972) and formalized by Stokey (1981) and Gul, Sonnenschein, and Wilson (1986)) reflects not only the seller's limited commitment power but also a restricted contract space (price posting). When the contract space expands to general dynamic contracts, the information stored in the contract can deter the seller from abandoning it, effectively providing commitment not because she is bound but because abandonment is unprofitable.

## Theory / model

**Setup.** There is one seller of a durable, indivisible good and one buyer. The buyer's willingness to pay is binary: high ($$v_h$$) or low ($$v_l$$), with $$v_h > v_l > 0$$. The probability of $$v_h$$ is $$\mu \in (v_l/v_h, 1)$$, so the static monopoly price is $$v_h$$. Time is discrete, indexed by $$0, 1, \ldots$$, and both parties discount at the common factor $$\delta \in (0,1)$$. If trade occurs at time $$T$$ at transfer $$p_t$$ (paid each period), payoffs are (p. 1339):

$$
\delta^T v - \sum_{t=0}^{\infty} \delta^t p_t \quad (\text{buyer}), \qquad \sum_{t=0}^{\infty} \delta^t p_t \quad (\text{seller}).
$$

In the initial period, the seller offers a contract from a set $$\mathcal{C}$$ (the contract space). The contract specifies allocations and transfers for every period when it is active. Each subsequent period, the seller decides whether to deploy the current contract or replace it. Contracts resemble smart contracts: they execute automatically once accepted, can be voided by the seller, and may hold information the seller herself cannot observe.

**Simple and direct contracts.** The paper works primarily with a subset $$\mathcal{D} \subset \mathcal{C}$$ of "simple and direct" contracts. A simple and direct contract asks the buyer to report his valuation once in the initial deployment period and makes no further requests; thereafter, the buyer can only accept or reject. Such a contract $$d$$ deployed for $$\tau$$ consecutive periods induces unconditional trade probability and expected transfer in period $$\tau$$ (p. 1344):

$$
X_\tau(v) = \mathbf{x}_\tau(v) \prod_{t=0}^{\tau-1}\bigl[1 - \mathbf{x}_t(v)\bigr], \qquad P_\tau(v) = \mathbf{p}_\tau(v)\,\mathbf{x}_\tau(v)\prod_{t=0}^{\tau-1}\bigl[1 - \mathbf{x}_t(v)\bigr].
$$

**Incentive compatibility.** Let $$U(v,\hat{v},d,\delta)$$ denote the buyer's expected payoff when his type is $$v$$, he reports $$\hat{v}$$, and the contract $$d$$ is deployed forever (p. 1344):

$$
U(v,\hat{v},d,\delta) = \sup_{T \geq 0} \sum_{t=0}^{T} \delta^t \bigl[X_t(\hat{v})\,v - P_t(\hat{v})\bigr].
$$

**Definition 1** (p. 1344): Contract $$d \in \mathcal{D}$$ is $$\delta$$-incentive compatible if for each $$v \in \{v_l, v_h\}$$,

$$
v \in \arg\max_{\hat{v} \in \{v_l,v_h\}} U(v,\hat{v},d,\delta).
$$

**Seller's payoff.** If the incentive-compatible simple-and-direct contract $$d$$ is actively deployed forever, the seller's payoff is (p. 1345):

$$
v(d,\delta) = \mu \sum_{t=0}^{\infty} \delta^t P_t(v_h) + (1-\mu)\sum_{t=0}^{\infty} \delta^t P_t(v_l).
$$

**Abiding contracts (Definition 2, p. 1345).** The key concept. Contract $$d = (X_\tau, P_\tau)_{\tau=0}^{\infty} \in \mathcal{D}$$ is $$\delta$$-abiding if it is $$\delta$$-incentive compatible and:

- (i) $$\sum_{t=T}^{\infty} \delta^{t-T}\bigl[X_t(v)v - P_t(v)\bigr] \geq 0$$ for all $$v \in \{v_l,v_h\}$$, $$T \geq 0$$, so the buyer's continuation payoff is always nonnegative;
- (ii) $$\mu_t(d) \leq v_l/v_h$$ for all $$t \geq 1$$, so conditional on no trade, the seller becomes pessimistic enough that the static monopoly price drops to $$v_l$$;
- (iii) $$\mu_T(d)\sum_{t=T}^{\infty}\delta^{t-T}P_t(v_h) + [1-\mu_T(d)]\sum_{t=T}^{\infty}\delta^{t-T}P_t(v_l) \geq v_l$$ for all $$T \geq 1$$, so the seller's continuation payoff exceeds $$v_l$$ in every future period.

Here $$\mu_t(d)$$ is the seller's posterior probability that the buyer's valuation is $$v_h$$ in period $$t$$, given that the contract has been actively deployed.

**Theorem 1** (p. 1343): *There exists $$\underline{\pi} > v_l$$ such that for all $$\delta \in (0,1)$$,*

$$
\pi(\mathcal{C},\delta) \geq \underline{\pi}.
$$

## Method

The proof proceeds in two lemmas, proved separately and then combined.

**Lemma 1** (p. 1345): *If $$d \in \mathcal{D}$$ is $$\delta$$-abiding, then $$\pi(\mathcal{C},\delta) \geq v(d,\delta)$$.*

The argument: suppose an equilibrium yields the seller less than $$v(d,\delta)$$. Modify the equilibrium so the seller always deploys $$d$$ and the buyer always accepts. On the equilibrium path, the seller's payoff is exactly $$v(d,\delta)$$. Off the path, the seller cannot profitably deviate in the initial period because any alternative contract gives her at most $$v(d,\delta)$$ (by construction of the modification). In subsequent periods, by condition (iii) of Definition 2, the seller's continuation payoff from $$d$$ exceeds $$v_l$$, which is also an upper bound on what she could get by abandoning $$d$$ (abandonment destroys its information content, leaving only the option to clear the market at $$v_l$$). Conditions (i)-(ii) ensure the buyer has no incentive to reject the contract after the initial period.

**Lemma 2** (p. 1347): *For all $$\delta \in (0,1)$$, there exists $$d_\delta \in \mathcal{D}$$ such that $$v(d_\delta,\delta) \geq \underline{\pi} > v_l$$.*

The construction uses a three-parameter family of simple and direct contracts indexed by $$(\alpha, \beta, p) \in [0,1]^2 \times [v_l, v_h]$$. In period $$\tau = 0$$, if the buyer reports $$v_h$$, trade occurs with probability $$\alpha$$ at price $$p$$; if he reports $$v_l$$, there is no trade. For each subsequent period $$\tau > 0$$, trade occurs with probability $$\beta$$ at a price equal to the buyer's initial report, regardless of type.

The high-type buyer's incentive constraint (binding at optimum, eq. (1), p. 1348) is:

$$
\alpha(v_h - p) \geq \frac{\beta\delta}{1-\delta+\beta\delta}(v_h - v_l). \tag{1}
$$

Conditional on no initial trade, the seller's posterior updates via Bayes' rule (eq. (2), p. 1348):

$$
\tilde{\mu}(\alpha) = \frac{(1-\alpha)\mu}{1-\mu+(1-\alpha)\mu}. \tag{2}
$$

Condition (ii) of Definition 2 requires the static monopoly price after the initial period to be $$v_l$$, i.e., (eq. (3), p. 1349):

$$
v_l \geq \tilde{\mu}(\alpha)\,v_h. \tag{3}
$$

The abiding constraint (4) (p. 1349) requires that the seller's continuation payoff from keeping $$d$$ deployed exceeds her payoff from clearing the market at $$v_l$$ immediately. Setting $$\bar{\beta}(\alpha)$$ to be the $$\beta$$ that binds constraint (4) and $$\bar{p}(\alpha)$$ to be the $$p$$ that binds constraint (1), the seller's payoff from the resulting contract is (eq. (7), p. 1350):

$$
v(\alpha) = \mu\alpha\,\bar{p}(\alpha) + (1-\mu\alpha)\,v_l. \tag{7}
$$

The optimal $$\alpha^*$$ maximizes $$v(\alpha)$$ subject to constraint (3). The paper shows (via the envelope theorem) that $$v(\alpha^*)$$ is strictly larger than $$v_l$$ for all $$\mu \in (v_l/v_h, 1)$$ and does not depend on $$\delta$$ for large enough $$\delta$$. Setting $$\underline{\pi} = \min\{\pi_{\bar{\delta}}, \hat{\pi}\}$$ for a small-$$\delta$$ bound $$\hat{\pi}$$ completes the proof.

**Proof of Theorem 1.** Lemma 2 guarantees a $$\delta$$-abiding contract $$d_\delta$$ with $$v(d_\delta,\delta) \geq \underline{\pi} > v_l$$ for every $$\delta$$. Lemma 1 then implies $$\pi(\mathcal{C},\delta) \geq v(d_\delta,\delta) \geq \underline{\pi}$$.

**Discussion.** The paper builds on the approach of Laffont and Tirole (1988) to combine one-period and multi-period contracts in a dynamic principal-agent setting. It compares its result to the model of Doval and Skreta (2022), where the seller is restricted to one-period contracts and the Coase conjecture holds (the seller's payoff converges to $$v_l$$ as $$\delta \to 1$$). The key difference is that one-period contracts have no information content to lose upon abandonment, so the seller always faces the temptation to clear the market quickly. With general dynamic contracts, information stored in the contract deters abandonment; this is the role played by smart-contract-style information storage.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| No empirical data | Pure theory paper; all results are derived from the formal model | n/a |

## When to read the full paper

Read the source at [doi.org/10.1257/aer.20220357](https://doi.org/10.1257/aer.20220357) if you are: studying the robustness discussions (continuous types, side contracts, interim participation, buyer rejection as endogenous abandonment trigger, Section III, pp. 1353-1356); interested in the mechanism-design methodology for modeling limited commitment via an expanded contract space; or comparing the paper's lower bound with the full-commitment payoff and the posted-price equilibrium (Figure 1, p. 1353). The online Appendix contains existence proofs and the result for $$\mathcal{C} = \mathcal{D}$$.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(5). This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. The journal version is paywalled; an LSE eprint is available at [eprints.lse.ac.uk/117950/](http://eprints.lse.ac.uk/117950/1/).

> Brzustowski, Thomas, Alkis Georgiadis-Harris, and Balázs Szentes. "Smart Contracts and the Coase Conjecture." *American Economic Review* 113, no. 5 (May 2023): 1334-1359. DOI: 10.1257/aer.20220357.


==============================================================================
# A Signal to End Child Marriage: Buchmann, Field, Glennerster, Nazneen & Wang (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/buchmann-et-al-signal-end-child-marriage-2023/
# Distilled: A clustered RCT in rural Bangladesh showed a small conditional financial incentive (cooking oil, ~US$16/year) for adolescent girls to remain unmarried reduced underage marriage by 19 percent and increased schooling, while a traditional empowerment program had no marriage effect and raised dowry. A signaling model explains child marriage persistence as a pooling equilibrium driven by information asymmetry about bride type. American Economic Review 2023, free after 12-month AEA embargo. Seven core results with source locators, the signaling model, and the empirical specifications.
# Tags: paper-summary, development-economics, gender, child-marriage, marriage-markets
==============================================================================

**What this is.** Core results, the signaling model of marriage timing, and
the main empirical specifications from a 44-page paper: enough to know what
it found and how, without reading the full source. To replicate or extend,
read the original at the [DOI](https://doi.org/10.1257/aer.20220720).

## TL;DR

The paper runs a clustered RCT in rural Bangladesh (2007-2017) testing two
policy approaches to child marriage: (i) a conditional financial incentive
(cooking oil, approximately US$16/year) paid to families while their adolescent
daughters remain unmarried, and (ii) a standard adolescent empowerment program
(Kishore Kontha). The incentive is effective at reducing child marriage: girls eligible for two
years are 19 percent less likely to marry underage and complete significantly
more schooling, with no marriage-market penalty (dowry and husband quality are
unchanged). The empowerment program fails to reduce marriage and raises dowry,
consistent with a perverse signaling effect. The paper then develops a
signaling model in which early marriage signals socially conservative bride
type (valued by grooms), and supports this mechanism via spillovers: untreated
less-conservative women near incentive communities also delay marriage, while
conservative women do not.

## Core results

Magnitudes as reported; \*\*\* = 1%, \*\* = 5%, \* = 10%. ITT estimates
from OLS with Huber-White SEs clustered at the community level and union fixed
effects (Tables 2-4). Table 5 uses a proximity-based spillover design within
non-incentive communities.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Incentive reduces underage marriage (< 18) | Table 2, col 2, p. 2661 | -7.4 ppts\*\*\* (age 15, 19% reduction); -4.9 ppts\*\*\* (age 15-17, 17%); control means 38.5% and 29.3% |
| R2 | Incentive increases average age at marriage | Table 2, cols 6-7, p. 2661 | +0.32 years\*\*\* (3.9 months) for age 15; +0.21 years\*\*\* (2.5 months) for age 15-17 |
| R3 | Incentive increases school enrollment and attainment | Table 3, cols 2-4, p. 2664 | +8.6 ppts\*\*\* in school at midline for age 15; +5.0 ppts at endline; control means 48.2% (midline) and 27.8% (endline) for age-15 cohort |
| R4 | No marriage-market penalty from incentive (null result) | Table 4, col 1, p. 2666 | Dowry: US$17.6 (SE 24.3, not sig.); husband education: -0.166 (SE 0.119, not sig.); control mean US$888 |
| R5 | Empowerment raises dowry with no marriage improvement | Table 4, col 1; Table 2, col 1, p. 2666, 2661 | Dowry +US$57.3\*\*\* (6% above control); marriage rate coefficient -0.007 (SE 0.008, not sig.) |
| R6 | Spillovers on untreated nearby women; null for ineligible ages | Table 5, cols 1-2, p. 2678 | Age 15-17 within 500m: -2.9 ppts\* married <18; placebo age 7-14: -0.9 ppts (not sig.) |
| R7 | Spillovers are specific to less socially conservative (nonpreferred-type) women | Table 5, cols 3-4, p. 2678 | Low SC: -14.7 ppts\*\* married <18; high SC: +9.3 ppts (not sig.); control means 36.0% and 47.9% |

**Overall (paper's conclusion).** The pattern is inconsistent with child
marriage reflecting strong cultural preferences for young brides or unraveling
in a thin marriage market. A signaling model in which early marriage signals
socially conservative (preferred) bride type explains why: a small financial
incentive orthogonal to type can shift the equilibrium from universal early
marriage to later marriage for all women who can credibly claim eligibility.
The empowerment program's failure, and its associated dowry increase, is also
predicted by the model: making more women appear nonpreferred strengthens the
incentive for conservative women to signal type by marrying early.

## Theory / model

The model is a two-period marriage-market signaling game (Section IV,
pp. 2666-2675). Women have measure $$|W| = 1$$ and men measure $$|M| > 1$$.
Women are heterogeneous in an unobservable type $$\Theta \in \{\Theta_H, \Theta_L\}$$,
where $$\Theta_H$$ (preferred, socially conservative) has population fraction
$$f \in (0,1)$$ and is privately known. Men are homogeneous and desire $$\Theta_H$$ women.

If a woman marries in period $$t_1$$ she has education $$E_L$$; if she delays to
$$t_2$$ her education rises to $$E_H > E_L$$. The total transferable utility of a
union between a man and woman of type $$\Theta_j$$ and education $$E_j$$ is
$$\mu(\Theta_j, E_j)$$, increasing in both arguments. Two conditions govern
equilibrium (p. 2669):

**(i) Preferred type is first-order in desirability:**

$$
\mu(\Theta_H, E_H) \;>\; \mu(\Theta_H, E_L) \;>\; \mu(\Theta_L, E_H) \;>\; \mu(\Theta_L, E_L) \tag{1}
$$

**(ii) Single-crossing, type and education are substitutes:**

$$
\mu(\Theta_H, E_H) - \mu(\Theta_H, E_L) \;<\; \mu(\Theta_H, E_L) - \mu(\Theta_L, E_H) \;<\; \mu(\Theta_L, E_H) - \mu(\Theta_L, E_L) \tag{2}
$$

Condition (2) implies preferred types gain less from delayed marriage (higher
education) than nonpreferred types. Dowry $$D$$ can be conditioned only on
observable characteristics (marriage timing, education); men compete for
available women and receive their outside options in equilibrium.

**Result 1** (p. 2671, Appendix A): Under a liquidity constraint
$$Y < \omega_M + \mu(\Theta_H, E_H) - 2\mu(\Theta_H, E_L)$$, no separating equilibrium
exists: nonpreferred types cannot afford the dowry $$D|t_2$$ that would credibly
signal preferred type.

**Result 2** (p. 2672, Appendix C): When separation is infeasible, the unique
PBE surviving the Cho-Kreps intuitive criterion is pooling on $$t_1$$ (all women
marry early). Equilibrium dowries are:

$$
D|t_1 = \omega_M - \bigl[f\,\mu(\Theta_H, E_L) + (1-f)\,\mu(\Theta_L, E_L)\bigr], \qquad D|t_2 = \omega_M - \mu(\Theta_L, E_H)
$$

Pooling on $$t_2$$ fails the intuitive criterion because preferred types would
profitably deviate to $$t_1$$ (where grooms know with certainty that early
entrants are preferred type). Pooling on $$t_1$$ survives because neither type
can credibly signal itself by deviating to $$t_2$$. This is the child-marriage
equilibrium of Wahhaj (2018) but generated here by differential returns to
education rather than preference for young brides.

**Corollary 2** (p. 2675): A decrease in $$f$$ (more nonpreferred types)
makes pooling more likely. This is the mechanism by which the empowerment
program can worsen early marriage: by reducing the perceived fraction of
preferred types, it strengthens the signaling incentive for conservative women.

**Result 4** (pp. 2673-2674): If a conditional incentive $$C$$ satisfying

$$
C \;>\; \mu(\Theta_H, E_L) - \bar{\mu}(E_H) - \bigl[\mu(\Theta_H, E_H) - \mu(\Theta_H, E_L)\bigr]
$$

is offered randomly to fraction $$\tau$$ of women (treatment orthogonal to type),
a semiseparating PBE exists and Pareto-dominates the pooling equilibrium.
Untreated preferred types marry at $$t_1$$; all other women delay to $$t_2$$.
Equilibrium dowries under the incentive:

$$
D|t_1 = \omega_M - \mu(\Theta_H, E_L), \qquad D|t_2 = \omega_M - \bar{\mu}(E_H)
$$

where $$\bar{\mu}(E_H) \equiv \frac{\tau f\,\mu(\Theta_H, E_H) + (1-\tau)(1-f)\,\mu(\Theta_L, E_H)}{\tau f + (1-\tau)(1-f)}$$
is the expected match quality when the fraction of preferred types among $$t_2$$
entrants is $$f' < f$$. Crucially, untreated nonpreferred types also delay (the
**spillover**): because treatment status is unobservable, delaying marriage is
no longer a certain signal of being the nonpreferred type (Prediction 4,
p. 2676).

## Method

**Experimental design.** A clustered randomized trial in 460 communities
across five subdistricts in rural Bangladesh, randomized in ratio 1:2:1:2 to
(i) incentive only, (ii) empowerment only, (iii) both, (iv) control.
Randomization was stratified by union (administrative grouping of approximately
10 communities) and community size. All girls aged 15-17 and unmarried at
program start (January 2008) were eligible for the conditional incentive; 92
percent received it at least once. The empowerment program (Kishore Kontha)
reached 93 percent of eligible girls.

The incentive program was implemented January 2008 to August 2010. Main outcome
data come from an endline survey conducted 2016-2017, approximately 4.5 years
after program completion.

**Estimation.** ITT effects are estimated via OLS linear probability models
(`panel-regression`) with Huber-White robust standard errors clustered at the
community level (the unit of randomization). The theoretical mechanism is
formalized in the `signaling-game-pbe` model (Results 1-4 above). Field and
Ambrus (2008) establish the education channel in Bangladesh; Duflo, Dupas, and
Kremer (2015) and Baird, McIntosh, and Ozler (2011) provide comparable evidence
from transfer programs in other contexts.

## Empirical specifications

**Main ITT regression** (Equation 1, p. 2659):

$$
Y_{icu} = \alpha + \beta_1 I_c + \beta_2 E_c + \beta_3 (I_c \times E_c) + \beta_4' \mathbf{X}_{ic} + \varepsilon_{icu} \tag{3}
$$

where $$Y_{icu}$$ is outcome for person $$i$$ in community $$c$$ and union $$u$$;
$$I_c$$ = assignment to the incentive program; $$E_c$$ = assignment to the
empowerment program; $$\mathbf{X}_{ic}$$ includes village population tercile and union
fixed effects, age indicators, household size, older unmarried sister in
household, school enrollment, mother's education, and public transport access
(a proxy for remoteness). The interaction $$I_c \times E_c$$ tests complementarity.
Tables 2, 3, and 4 report $$\hat{\beta}_1$$, $$\hat{\beta}_2$$, $$\hat{\beta}_3$$.
The main analysis sample covers 15,576 women in the parents' survey.

**Spillover specification** (Table 5, p. 2678): Estimated on nonincentive
communities only, comparing women within 500 meters of an incentive community
center to women farther away:

$$
Y_{icu} = \alpha + \gamma \cdot \text{Close}_c + \delta' \mathbf{W}_{ic} + \varepsilon_{icu} \tag{4}
$$

where $$\text{Close}_c = 1$$ if the community center is within 500 meters of the
closest incentive community center. $$\mathbf{W}_{ic}$$ includes union fixed effects,
baseline characteristics, and distances to the closest community center and safe
space (to rule out urbanicity confounds). Placebo test: same regression for
girls aged 7-14 who are observably ineligible. Type heterogeneity: equation (4)
is estimated separately for girls with above-median and below-median social
conservatism index (columns 3-4 of Table 5), testing Prediction 4 of the
signaling model.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Authors' RCT survey data (parents' and young women's surveys) | Main analysis: marriage outcomes, schooling, dowry, husband characteristics; 460 communities, 24,095 girls at baseline, 15,576 in endline parents' sample | No page yet (hand-collected; replication data at https://doi.org/10.3886/E192114V1) |
| Bangladesh Demographic and Health Surveys (DHS, waves 2004-2017) | Descriptive context: trends in marriage age, education, and child mortality 2004-2017 (Figure 1, p. 2646) | No page yet |

Sample scope: rural Bangladesh, five subdistricts (Daulatkhan, Babuganj, Muladi,
Patuakhali Sadar, Bauphal, Bhola Sadar). Program period January 2007 to
September 2017. Main endline wave: 2016-2017.

## When to read the full paper

Read the [original](https://doi.org/10.1257/aer.20220720) if you are: (i)
extending the signaling model to other cultural contexts with bride-type
heterogeneity or dowry; (ii) designing a conditional transfer program for child
marriage reduction; (iii) examining the cost-effectiveness calculation of
US$1,010 NPV per US$1,000 invested (Buchmann et al. 2021 companion paper);
or (iv) testing Predictions 1-4 of the signaling model in a new setting. Tables
2, 3, and 5 contain the core empirical results; Section IV formalizes the
model; Appendix A carries the proofs.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(10), October 2023.
This distillation was extracted by an LLM on 2026-06-25 and is **not
human-verified or independently reproduced**. The article is freely accessible
after the AEA 12-month embargo (expired October 2024) and via NIH PMC; no CC
license applies, so redistribution is extract-only.

> Buchmann, Nina, Erica Field, Rachel Glennerster, Shahana Nazneen, and Xiao Yu Wang.
> "A Signal to End Child Marriage: Theory and Experimental Evidence from Bangladesh."
> *American Economic Review* 113, no. 10 (October 2023): 2645-2688.
> DOI: 10.1257/aer.20220720.


==============================================================================
# Second-Best Fairness: Cappelen, Cappelen & Tungodden (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/cappelen-et-al-second-best-fairness-trade-2023/
# Distilled: Large-scale experimental evidence from 26,500 spectators in the US and Norway on how people trade off false positives against false negatives in second-best fairness decisions. A majority are false negative averse across three economic environments, with substantial heterogeneity by country and political affiliation. American Economic Review 2023, AEA copyright. Six core results with source locators, datasets used, the theoretical model, and the estimation strategy.
# Tags: paper-summary, behavioral-economics, fairness, redistribution
==============================================================================

**What this is.** The core results, theoretical model, and empirical strategy from this paper: enough to understand what was found and how, without reading all 28 pages. To replicate or extend, read the original at the [DOI](https://doi.org/10.1257/aer.20211015).

## TL;DR

The paper examines how people trade off false positives (paying an undeserving individual) against false negatives (not paying a deserving individual) in second-best fairness decisions. Across three large-scale experiments in the United States and Norway (26,500 spectators total), the large majority of spectators are false negative averse: they prefer risking a false positive over a false negative. When the probability of a false claim is 50 percent, 72.4 percent of spectators still choose to pay, consistent with placing higher weight on avoiding a false negative. However, about 20 percent are strongly false positive averse. Americans are more false positive averse and less false negative averse than Norwegians, and right-wing spectators exhibit the same pattern within both countries. These second-best fairness preferences strongly predict policy attitudes on unemployment benefits and income redistribution, above and beyond stated fairness views and altruism.

## Core results

Magnitudes are as reported; `\*\*\*` = 1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Paying falls monotonically with false-claim probability; 72.4% still pay at Pr(f)=0.5**, implying a FN-FP gap of 44.8 pp | Table 3 col 1, p. 2473; Figure 1, p. 2472 | Coeff at 50% false-claim prob: -17.6 pp (SE=0.016, p<0.001); baseline (Pr(f)=0) = 90.0%; at 100%: -79.7 pp; FN-FP difference = 44.8 pp (p<0.001) |
| R2 | **Type-share lower bounds: FP averse 20.3%, FN averse 65.2%** (pooled compensation experiment) | Table 4 upper panel, p. 2474 | FP averse LB 20.3% (SE=0.015); symmetric UB 14.5% (SE=0.037); FN averse LB 65.2% (SE=0.027) |
| R3 | **Majority have highly asymmetric preferences: 20.3% strongly FP averse, 43.5% strongly FN averse** | Figure 2 upper-left panel, p. 2475; text p. 2474 | Strongly FP averse (beta <= 0.25): 20.3%; strongly FN averse (beta >= 0.75): 43.5% of pooled sample |
| R4 | **US spectators are more FP averse and less FN averse than Norwegians** across all experiments | Table 6 right panel, p. 2481 | Strongly FP averse US vs Norway: +11.8 pp (SE=0.017, p<0.001); strongly FN averse: -10.0 pp (SE=0.021, p<0.001) |
| R5 | **Right-wing spectators are less FN averse and more FP averse** than non-right-wing spectators | Table 6 left panel, p. 2481 | FN averse -9.5 pp (SE=0.010, p<0.001); strongly FP averse +6.2 pp (SE=0.019, p<0.001); strongly FN averse -12.1 pp (SE=0.022, p<0.001) |
| R6 | **Second-best fairness preferences predict policy attitudes independently of stated fairness views** | Table 7, p. 2482 | Paying predicts support for generous unemployment benefits: coeff 0.562 (SE=0.025, p<0.001); income inequality: 0.354 (SE=0.025, p<0.001); survives controls for fairness views, efficiency costs, altruism, religiosity |

**Overall (paper's conclusion).** The majority of spectators in both countries are false negative averse in all three experiments. A significant minority is strongly false positive averse. Country and political differences are of similar magnitude: the US-Norway gap mirrors the right-wing/non-right-wing gap within each country. These second-best fairness preferences are strongly predictive of real-world policy attitudes on redistribution and social insurance, suggesting they are a fundamental ingredient in the political economy of welfare institutions.

## Theory / model

The paper proposes a simple expected utility framework to characterize second-best fairness preferences (Section I, pp. 2461-2462). Consider an environment where a spectator must choose a payment $$y$$ for an individual whose claim may or may not be false. Let $$m(f)$$ be the fair payment if the claim is false and $$m(c)$$ the fair payment if it is correct, with $$m(c) > m(f)$$. The probability the claim is false is $$\Pr(f)$$ and the probability it is correct is $$1 - \Pr(f)$$.

In line with Cappelen et al. (2013a), the spectator dislikes any payment that deviates from what is fair. The expected utility of paying $$y$$ is (equation 1, p. 2461):

$$
EU(y) = \Pr(f)\, u\!\left(y - m(f)\right) + \bigl[1 - \Pr(f)\bigr]\, u\!\left(y - m(c)\right), \tag{1}
$$

where $$u(\cdot)$$ is weakly decreasing in the deviation from the fair payment and $$u(0) = 0$$. The spectator faces a binary choice between paying, $$y = m(c)$$, and not paying, $$y = m(f)$$. These yield (equations 2 and 3, p. 2461):

$$
EU\!\left(y = m(c)\right) = \Pr(f)\, u\!\left(m(c) - m(f)\right), \tag{2}
$$

$$
EU\!\left(y = m(f)\right) = \bigl[1 - \Pr(f)\bigr]\, u\!\left(m(f) - m(c)\right). \tag{3}
$$

Since $$|m(c) - m(f)| = |m(f) - m(c)|$$, define $$\beta$$ as the relative weight the spectator places on avoiding a false negative versus a false positive (p. 2462):

$$
\beta = \frac{u\!\left(m(f) - m(c)\right)}{u\!\left(m(c) - m(f)\right) + u\!\left(m(f) - m(c)\right)}.
$$

Three types follow: False Positive Averse ($$\beta < 1/2$$), Symmetric ($$\beta = 1/2$$), False Negative Averse ($$\beta > 1/2$$).

Observation 1 (p. 2462): The spectator is indifferent between paying and not paying when $$\Pr(f) = \beta$$; strictly prefers not to pay when $$\Pr(f) > \beta$$; and strictly prefers to pay when $$\Pr(f) < \beta$$. The individual switching threshold directly reveals $$\beta$$. This also predicts that the choice between the two options should be independent of the size of $$m(c) - m(f)$$ (the payoff size), which Table 5 confirms (the high-stakes treatment changes behavior by -4.3 pp, not significant after multiple-testing correction).

Upper and lower bounds on type shares are derived from $$sp(\Pr(f))$$, the share of spectators paying at each treatment value (Observation 4, p. 2469):

$$
S_U = 2 \times \max\!\left\{0,\, \min\!\bigl(sp(0.25) - sp(0.5),\; sp(0.5) - sp(0.75)\bigr)\right\},
$$

$$
FP_U = 1 - sp(0.5), \quad FN_U = sp(0.5), \quad FP_L = FP_U - 0.5\, S_U, \quad FN_L = FN_U - 0.5\, S_U.
$$

## Method

The estimation strategy is between-subject OLS on the binary payment indicator, applied to the pooled sample and separately for each country (Section III, pp. 2468-2471). The approach builds on `randomized-survey-experiment` (spectators randomly assigned to treatment arms) and `panel-regression` (linear probability model with controls and population weights).

The main specification for treatment effects (equation 4, p. 2468):

$$
e_i = \alpha + \alpha_1 P(0.25)_i + \alpha_2 P(0.5)_i + \alpha_3 P(0.75)_i + \alpha_4 P(1)_i + \gamma \mathbf{X}_i + \varepsilon_i, \tag{4}
$$

where $$e_i$$ is an indicator equal to one if spectator $$i$$ pays, $$P(0.25)_i$$ through $$P(1)_i$$ are treatment indicators for each false-claim probability level (baseline: $$\Pr(f) = 0$$), and $$\mathbf{X}_i$$ is a vector of controls (income, education, gender, age, political ideology). Estimates are population-weighted. Multiple testing corrected via Holm-Bonferroni and Romano-Wolf procedures.

For additional robustness treatments at $$\Pr(f) = 0.5$$ (equation 5, p. 2470):

$$
e_i = \alpha + \alpha_1 M_i + \gamma \mathbf{X}_i + \varepsilon_i, \tag{5}
$$

where $$M_i$$ indicates the specific additional treatment (doubled stakes, nationality framing, or endowment introduction). For the cost treatments (equation 6, p. 2470):

$$
e_i = \alpha + \alpha_1 C(0.1)_i + \alpha_1 C(0.3)_i + \gamma \mathbf{X}_i + \varepsilon_i, \tag{6}
$$

where $$C(0.1)_i$$ and $$C(0.3)_i$$ indicate treatments where the spectator bears a personal cost of US$0.1 or US$0.3 for paying.

For policy attitudes (equation 7, p. 2471):

$$
\text{pol}_i = \alpha + \alpha_1\, \text{pay}_i + \gamma \mathbf{X}_i + \varepsilon_i, \tag{7}
$$

where $$\text{pol}_i$$ is stated support for unemployment benefit generosity or income equalization on a seven-point scale, and $$\text{pay}_i$$ is an indicator for having paid in the experiment.

## Empirical specifications

Three experiments share the same between-subject design. In each, spectators are randomly assigned to one of five treatments where $$\Pr(f) \in \{0, 0.25, 0.5, 0.75, 1\}$$, then decide whether to pay a worker whose claim may be false.

**Compensation experiment** (Section II.A, pp. 2463-2465). Workers are recruited on an international online labor market platform. It is randomly determined whether they are offered work. Those not offered work are entitled to a compensation of US$4 ($$m(c) = 4$$, $$m(f) = 0$$). Spectators are told the false-claim probability for the matched worker and decide whether to pay. The main sample is 5,395 spectators (2,695 US, 2,700 Norway). Additional treatments at $$\Pr(f) = 0.5$$ test: (i) high stakes (US$8 compensation, Panel A Table 5); (ii) nationality framing, where stakes are reported in local currency and workers are implied to be compatriots (Panel A); (iii) three endowment-plus-cost arms (Panel B). All produce null effects, consistent with the theoretical prediction that stake size and in-group salience should not affect the FP-FN trade-off (RESULT 3).

**Earnings experiment** (Section II.B, pp. 2465-2466). Same structure but the claim is for earnings from completing a 15-minute task rather than compensation for not being offered work. Spectators in 5,391 observations (main study, excluding pilot). Treatment effects are tested for differences from the Compensation experiment via interaction terms (Figure 3, Panel A). All interaction effects are small and not robust to multiple-testing correction (RESULT 4), confirming the preferences are not specific to the compensation context.

**Unemployment experiment** (Section II.C, p. 2466). A nonincentivized survey experiment where respondents decide whether to hypothetically pay unemployment benefits to someone with a known false-claim probability. The pattern of results closely matches the incentivized experiments (RESULT 5). Respondents are somewhat more false positive averse in this policy domain, consistent with the policy debate around welfare fraud (Figure 3, Panel B).

Country-level regressions run equation (4) separately for the US and Norway. Pooled cross-country and cross-political-spectrum comparisons use 22,476 observations (all three experiments combined) with interaction terms for Norwegian nationality and for right-wing political affiliation (Table 6). Almås, Cappelen, and Tungodden (2020) provide the comparison framework for cross-country distributive preference differences. Policy attitude regressions (Table 7) pool all experiments and add controls for fairness views and efficiency beliefs; Alesina and Angeletos (2005) is the background reference for the fairness-redistribution link.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Norstat population panel (US and Norway) | Recruitment of 22,500 spectators, quota-matched on age, gender, and geography to be nationally representative | No page yet |
| Amazon Mechanical Turk | Recruitment of 6,250 workers (consequential decisions) and 4,000 pilot spectators from earlier Earnings versions | No page yet |

All primary data are author-generated experimental data. Replication data and code are publicly available at the AEA/ICPSR OpenICPSR archive (https://www.openicpsr.org/openicpsr/188201). Sample: summer 2022 (main study); pilot collected 2019. Workers: 2,250 AMT main study + 4,000 pilot.

## When to read the full paper

Read the [original](https://doi.org/10.1257/aer.20211015) if you are: designing second-best social insurance policies where eligibility is uncertain; studying cross-country or political heterogeneity in redistributive preferences; classifying spectator types (FP vs FN averse) in a behavioral experiment; or extending the framework to judicial or disability contexts. The locators above point to the exact tables and figures for each result.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(9), September 2023. AEA copyright; not yet freely available on AEAweb (3-year embargo expires September 2026). This distillation was extracted by an LLM on 2026-06-24 and is **not human-verified or independently reproduced**. Redistribution is extract-only; the PDF is not hosted here.

> Cappelen, Alexander W., Cornelius Cappelen, and Bertil Tungodden. "Second-Best Fairness: The Trade-Off between False Positives and False Negatives." *American Economic Review* 113, no. 9 (September 2023): 2458-2485. DOI: 10.1257/aer.20211015.


==============================================================================
# Partisanship and Fiscal Policy in Economic Unions: Carlino, Drautzburg, Inman & Zarra (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/carlino-et-al-partisanship-fiscal-policy-economic-2023/
# Distilled: Using a regression discontinuity design on close gubernatorial elections, the paper shows Republican governors spend 0.29 percentage points less (elasticity) per 1 percent increase in federal intergovernmental transfers than Democratic governors, instead reducing debt and cutting taxes with a two-year lag; a calibrated New Keynesian two-state monetary union model implies the IG transfer impact multiplier falls by 0.58 under equal partisan representation relative to an all-Democratic benchmark. American Economic Review 113(3), 2023, paywalled. Eight core results with source locators, the NK model equations, and the RDD specification; LLM-distilled, not human-verified.
# Tags: paper-summary, fiscal-policy, political-economy, fiscal-federalism, macroeconomics
==============================================================================

**What this is.** The core results, the New Keynesian two-state monetary union model, and the regression discontinuity design that identifies partisan differences in state governors' propensity to spend federal intergovernmental (IG) transfers: enough to understand what the paper found and how, without reading all 37 pages. To replicate or extend it, read the full source at [doi:10.1257/aer.20210147](https://doi.org/10.1257/aer.20210147).

## TL;DR

The paper documents that the party of the state governor is a source of heterogeneity in how federal IG transfers affect the macroeconomy. Using a regression discontinuity design on close gubernatorial elections, it finds Republican governors spend 0.29 percent less (elasticity) per 1 percent increase in IG aid than Democratic governors, and instead reduce debt immediately and cut top income tax rates with a two-year lag. A calibrated New Keynesian model of a two-state monetary union with these estimated partisan fiscal rules implies an intertemporal trade-off: the IG transfer impact multiplier is 1.22 when all governors behave like Democrats but falls to 0.64 under equal partisan representation, a difference of 0.58; the long-run discounted multiplier, however, is 0.16 higher under equal partisanship (0.80 vs. 0.64) because delayed Republican tax cuts stimulate future output. The paper extends Besley and Case (2003), who documented partisan spending differences using OLS, to a causal RDD design, and adds the aggregate macroeconomic implications via a calibrated NK model. The partisan differences are only statistically significant in the post-Reagan era and grew with national polarization.

## Core results

Magnitudes are as reported; preferred estimates use the robust RDD estimator (quadratic MOV polynomial, state x party and year x party fixed effects, bandwidth 10 percentage points). Locators point to the published version.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Republican governors spend less per 1% increase in IG aid than Democratic governors (MPS elasticity partisan difference) | Table 2, col. 5, p. 713 | $$\gamma_{r,\text{inc}} = -0.290$$ (SE 0.098); t approximately -3.0 with party FE |
| R2 | Republican governors cut spending more per 1% decrease in IG aid than Democrats (asymmetric response to cuts) | Table 2, col. 5, p. 713 | $$\gamma_{r,\text{cut}} = 0.524$$ (SE 0.239); statistically significant at 5% |
| R3 | Dollar-for-dollar: Democratic MPS near $1.35 per $1 IG increase; Republican partisan difference -$1.58, leaving Republican total near zero | Table 4, col. 2, p. 716 | Democrat: $1.346 (SE 0.600); Rep diff: -$1.576 (SE 0.892); Republican total not different from zero |
| R4 | Republican states immediately lower debt outstanding relative to Democratic states; effect persists for at least 3 years | Figure 7, Panel A, p. 723 | Debt elasticity partisan difference approximately -0.25 per 1% IG increase; statistically significant |
| R5 | Republican top marginal income tax rates are lower, but only with a 2-year lag; no partisan tax difference on impact | Figure 7, Panel B, p. 723 | Log tax-rate elasticity partisan difference approximately -1% after 2 years; zero on impact and year 1 |
| R6 | Democratic-led states show higher GDP growth per dollar of IG aid, with the gap persisting up to 3 years | Figure 8, p. 724 | Partisan GDP growth elasticity (Rep - Dem): approximately -0.334% per 1% IG increase; marginally significant |
| R7 | IG transfer impact multiplier falls by 0.58 under equal partisanship vs. all-Democratic baseline | Table 7, rows 1-2, col. 1-3, p. 731 | Impact multiplier: 1.22 (all Democrats) vs 0.64 (equal partisanship); difference -0.58 (SE 0.33) |
| R8 | Long-run PDV multiplier is 0.16 higher under equal partisanship than all-Democratic, because of delayed Republican tax cuts | Table 7, rows 1-2, col. 1-3, p. 731 | Long-run multiplier: 0.64 (all Democrats) vs 0.80 (equal partisanship); difference +0.16 (SE 0.09) |

**Overall (paper's conclusion).** Partisanship of state governors alters how federal IG transfers translate into macroeconomic activity. More Democratic governors raise the short-run IG impact multiplier; more Republican governors raise the long-run discounted multiplier by channeling aid into delayed tax cuts. The resulting intertemporal trade-off means the optimal partisan composition of governors depends on whether the federal government prioritizes short-run or long-run stimulus. As national polarization increased after 1980, partisan differences in MPS widened and became statistically detectable, implying the aggregate consequences of partisanship for IG policy have grown over time.

## Theory / model

The model is a two-state New Keynesian monetary union with home state size $$n$$ and foreign state size $$1-n$$. The states are symmetric except for their governors' partisan MPS. Each state has constrained households (fraction $$\mu$$, hand-to-mouth) and unconstrained households (fraction $$1-\mu$$, with complete markets and bond holdings). Households in each state have utility (eq. 4, p. 724):

$$
u(C_t, N_t, G_{st,t}) = \frac{1}{1-1/\varepsilon_C} C_t^{1-1/\varepsilon_C} - \kappa_n^i \frac{N_t^{1+1/\varepsilon_N}}{1+1/\varepsilon_N} + v(G_{st,t}) \tag{4}
$$

where $$C_t$$ is aggregate consumption, $$N_t$$ labor supply, $$G_{st,t}$$ state government services, $$\varepsilon_C$$ the elasticity of intertemporal substitution, $$\varepsilon_N$$ the Frisch elasticity, and $$\kappa_n^i$$ a leisure weight that differs by household type $$i \in \{c, u\}$$. The budget constraint for unconstrained agents is (eq. 5, p. 725):

$$
P_t C_t^u + B_t^u \leq (1 - \tau_t^f - \tau_t^{st}) W_t N_t^u + B_{t-1}^u R_{t-1}^n + Pr_t + Tr_t^u \tag{5}
$$

Intermediate goods firms produce using labor only under decreasing returns to scale (eq. 6, p. 725):

$$
y_{h,t}(z) = A_t \times N_t(z)^{1-\alpha} \tag{6}
$$

where $$\alpha \in (0,1)$$ is the fixed-factor share. Firms face Calvo price stickiness: with probability $$\xi$$ a firm cannot reoptimize; calibrated $$\xi = 0.735$$ at annual frequency to match a peak defense spending multiplier of 0.8.

**State government fiscal rules.** Government consumption responds to IG transfers through the partisan MPS $$\psi_{IG}$$ (eq. 7, p. 725):

$$
G_{st,t} = \psi_{IG}\!\left(\frac{IG_t}{P_t} - \bar{IG}\right) + G_{st,t}^x \tag{7}
$$

The Republican home state uses $$\psi_{IG} = 0$$ (calibrated: Republican spending effect not significantly different from zero) while the Democratic foreign state uses $$\psi_{IG}^* = 1.576$$ (robust RDD estimate from Table 4, col. 2). State tax rates adjust smoothly to service debt and to cover expenditure net of IG revenue (eq. 8, p. 726):

$$
\tau_{st,t} = \rho_\tau \tau_{st,t-1} + (1-\rho_\tau)\!\left\{\bar{\tau}_{st} + \psi_{st,b}\!\left[(R_{t-1}^n - 1)\frac{b_{st,t-1}}{\Pi_t} - (\bar{R}^n - 1)\frac{\bar{b}_{st}}{\bar{\Pi}}\right] + \psi_{st,E}\!\left[G_{st,t-1} - \bar{G}_{st} - \frac{IG_{t-1} - \bar{IG}}{P_t}\right]\right\} \tag{8}
$$

Calibrated: persistence $$\rho_\tau = 0.35$$, debt loading $$\psi_{st,b} = 0.99$$, expenditure-net-of-IG loading $$\psi_{st,E} = 0.85$$ (Table 6, p. 729). This yields no tax change in year 1, a fall after year 2, and reversal toward zero by year 3, consistent with the micro estimates in Figure 7.

**Monetary policy.** The common central bank follows a Taylor rule reacting to union-wide inflation and the output gap (eq. 9, p. 726):

$$
R_t^n = \left(\frac{\bar{\Pi}}{\beta}\right)^{\rho_r}\!\!\left[\!\left(\frac{\Pi_t^{agg}}{\bar{\Pi}}\right)^{\psi_{r,\pi}}\!\!\left(\frac{Y_t^{agg}}{Y_t^{flex}}\right)^{\psi_{r,y}}\right]^{1-\rho_r} \tag{9}
$$

where $$\Pi_t^{agg} = n\Pi_t + (1-n)\Pi_t^*$$ and $$Y_t^{agg} = nY_t + (1-n)Y_t^*$$ are population-weighted. Calibrated: $$\rho_r = 0.75$$, $$\psi_{r,\pi} = 1.5$$, $$\psi_{r,y} = 0.5$$. The model builds on Nakamura and Steinsson (2014) by adding heterogeneous state fiscal rules that encode partisan differences. It is linearized around a deterministic steady state and solved numerically with Dynare (Adjemian et al. 2011).

## Method

The identification design exploits the regression discontinuity in gubernatorial elections. Near a 50-50 vote outcome (margin of victory MOV = 0), the party of the winning governor is as good as randomly assigned. This yields a consistent estimate of partisan differences in MPS conditional on IG aid changes, without relying on the assumption that IG aid is itself randomly allocated. The design follows Lee, Moretti, and Butler (2004) and Ferreira and Gyourko (2009), who used close-election RDDs for US House representatives and mayors.

MOV is signed positive (negative) for a Democratic (Republican) winner. Standard errors are two-way clustered by state and year throughout. The bandwidth $$\bar{m}$$ on absolute MOV is chosen by cross-validated MSE minimization using linear MOV controls ($$q = 1$$). The preferred "robust" estimator uses quadratic MOV controls ($$q = 2$$) with the same bandwidth; the preferred bandwidth is 10 percentage points with fixed effects and 11 without. Internal validity is assessed via Table 1 (p. 710), which shows no significant differences in pre-determined covariates across Democratic and Republican winners in close elections. External validity is assessed in Section III: rolling-window OLS shows the same partisan pattern for margins up to 30 percentage points, and candidate ideological differences are uncorrelated with closeness.

## Empirical specifications

**Baseline panel regression (eq. 1, p. 707).** Changes in log state expenditure $$\Delta \ln E_{s,t}$$ in state $$s$$, fiscal year $$t$$, are regressed on log changes in IG aid interacted with the governor's party lagged one year ($$\text{Rep}_{s,t-1} = 1$$ for Republican):

$$
\Delta \ln E_{s,t} = (\gamma_{0,\text{inc}} + \gamma_{r,\text{inc}} \times \text{Rep}_{s,t-1})\Delta \ln IG_{s,t}^{\text{inc}} + (\gamma_{0,\text{cut}} + \gamma_{r,\text{cut}} \times \text{Rep}_{s,t-1})\Delta \ln IG_{s,t}^{\text{cut}} + \mu_0 + \mu_r \times \text{Rep}_{s,t-1} + \text{fixed effects} + e_{s,t} \tag{1}
$$

where $$\Delta \ln IG_{s,t}^{\text{inc}} = \max\{0, \Delta \ln IG_{s,t}\}$$ captures IG increases and $$\Delta \ln IG_{s,t}^{\text{cut}} = \min\{0, \Delta \ln IG_{s,t}\}$$ captures decreases. The Democratic MPS elasticity is $$\gamma_{0,\text{inc}}$$ for increases and $$\gamma_{0,\text{cut}}$$ for decreases; the Republican partisan difference is $$\gamma_{r,\text{inc}}$$ (expected negative: Republicans spend less on increases) and $$\gamma_{r,\text{cut}}$$ (expected positive: Republicans cut more when aid falls). Fixed effects are state x party and year x party in the preferred specification.

**RDD specification (eq. 2, p. 708).** Equation (1) is augmented with a polynomial of order $$q$$ in the MOV and its interactions with IG changes, estimated on observations within bandwidth $$|MOV_{s,t-1}| \leq \bar{m}$$:

$$
\Delta \ln E_{s,t} = (\gamma_{0,\text{inc}} + \gamma_{r,\text{inc}} \times \text{Rep}_{s,t-1})\Delta \ln IG_{s,t}^{\text{inc}} + (\gamma_{0,\text{cut}} + \gamma_{r,\text{cut}} \times \text{Rep}_{s,t-1})\Delta \ln IG_{s,t}^{\text{cut}} $$

$$+ \sum_{\hat{s} \in \{\text{cut},\text{inc}\}}\sum_{p=1}^{q} (\gamma_{0,\hat{s},m,p} + \gamma_{r,\hat{s},m,p} \times \text{Rep}_{s,t-1})\Delta \ln IG_{s,t}^{\hat{s}} \times MOV_{s,t-1}^p $$

$$+ \sum_{p=1}^{q} (\beta_{0,m,p} + \beta_{r,m,p} \times \text{Rep}_{s,t-1})MOV_{s,t-1}^p + \mu_0 + \mu_r \times \text{Rep}_{s,t-1} + \text{fixed effects} + e_{s,t} \tag{2}
$$

The RDD identifies partisan differences $$(\gamma_{r,\text{inc}}, \gamma_{r,\text{cut}})$$ but not the Democratic baseline MPS $$(\gamma_{0,\text{inc}}, \gamma_{0,\text{cut}})$$. The paper benchmarks Republican MPS to zero (consistent with the near-zero robust estimate) and sets Democratic MPS to 1.576 (from dollar-level Table 4, col. 2). Headline results are R1 and R2 above (Table 2, cols. 5 and 2 respectively).

**Debt and tax rate specifications.** Equations (1) and (2) are re-estimated replacing $$\Delta \ln E_{s,t}$$ with the log level of total debt outstanding and the log of the state top marginal income tax rate, estimated at horizons 0-3 years (Figure 7, p. 723). These specifications produce R4 and R5: Republicans lower debt immediately; tax cuts arrive with a two-year lag. Results R4-R5 do not appear in a formal regression table; Figure 7 plots the RDD coefficient path with 68% and 90% pointwise confidence bands.

**State GDP specification.** Equation (2) with cumulative GDP growth as the dependent variable produces R6 (Figure 8, p. 724): a 0.334 percent Democratic advantage in GDP growth per 1 percent increase in IG aid, which persists for up to three years.

**Polarization extension (eq. 3, p. 720).** Interacting partisan MPS differences with a political polarization index $$PPC_{t-1}$$ (normalized; from Azzimonti 2018) shows that partisan MPS differences grew with polarization: the interaction is $$-0.13 \times PPC_{t-1}$$ (SE 0.06) for IG increases. Before 1980, $$PPC_{t-1}$$ averaged 1.1 standard deviations below the mean, implying near-zero partisan differences (consistent with the pre-Reagan evidence). After 1990, $$PPC_{t-1}$$ averaged 0.8 standard deviations above the mean, implying partisan differences of approximately $$-0.25$$ (standard error 0.09).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| US Census Bureau, State and Local Government Finance (State Government Finances series) | Primary source for state expenditure, revenue, debt, and tax data; 1983-2014 | no page yet |
| Federal Assistance Award Data System (FAADS) via National Archives / SAM | Identifies federal aid programs with matching provisions (excluded from main IG measure); 1983-2010 | no page yet |
| Council of State Governments, The Book of the States | Gubernatorial election outcomes, party affiliation, and term data | no page yet |
| US Bureau of Economic Analysis, Regional Economic Accounts | State-level GDP (personal income), used for R6 state GDP growth specifications | no page yet |
| US Bureau of Labor Statistics / Federal Reserve Bank of St. Louis (FRED) | State unemployment rate (used in robustness checks, Table 5); macro data | [FRED](/wiki/datasets/fred/) |
| S&P Global Ratings, History of US State Ratings | State bond downgrade dummies for fiscal-stress robustness checks (Table 5) | no page yet |
| Klarner (2013) State Partisan Balance Data (Harvard Dataverse) | State partisan composition of legislature; used in robustness checks and polarization analysis | no page yet |
| Bonica (2014) DIME Database on Ideology | Ideological scores for gubernatorial candidates; external validity check in Figure 5 | no page yet |

Sample: 48 US states, fiscal years 1983-2014. Alaska, Wyoming, and North Dakota excluded from main sample (large sovereign wealth funds). Close-election RDD subsample: elections with absolute MOV at most 10-11 percentage points.

## When to read the full paper

Read the [original](https://doi.org/10.1257/aer.20210147) if you are: calibrating the partisan heterogeneity in state fiscal responses for a fiscal federalism model (Tables 2 and 4 give the full set of MPS estimates by party, sign, and specification); studying the macroeconomic implications of state partisanship for federal stimulus design (Section V and Table 7 provide the NK model results and counterfactuals); assessing external validity of RDD designs for close gubernatorial elections (Section III, Figures 4-6, Appendix C); comparing IG multiplier estimates across methodologies (Chodorow-Reich (2019) surveys ARRA-era estimates and finds a headline multiplier of 1.7, higher than the estimates here partly because it implicitly assumes the ZLB constraint); or extending the model to incorporate the zero lower bound or other heterogeneous states. The replication data at [doi:10.3886/E177001V1](https://doi.org/10.3886/E177001V1) includes all cleaned datasets and Dynare code.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(3), March 2023. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. The AEA retains copyright; no CC licence was found in Crossref metadata. Reproduction of figures or tables requires AEA permission.

> Carlino, Gerald, Thorsten Drautzburg, Robert Inman, and Nicholas Zarra.
> "Partisanship and Fiscal Policy in Economic Unions: Evidence from US States."
> *American Economic Review* 113, no. 3 (March 2023): 701-737.
> DOI: 10.1257/aer.20210147.
> Replication data: [doi:10.3886/E177001V1](https://doi.org/10.3886/E177001V1).


==============================================================================
# Occupational Exposure to Capital-Embodied Technical Change: Caunedo, Jaume & Keller (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/caunedo-et-al-occupational-exposure-capital-embodied-2023/
# Distilled: Using the first measures of capital-embodied technical change (CETC) at the occupational level, Caunedo, Jaume, and Keller show that CETC accounts for 95% of gross US labor reallocation between 1984 and 2015, with heterogeneous capital-labor substitutability (not the extent of CETC) as the key driver. American Economic Review 2023, AEA standard (free access). Seven core results with source locators, datasets used, the model, and the method.
# Tags: paper-summary, labor-economics, technical-change, wage-inequality
==============================================================================

**What this is.** The paper's core results, the structural model of occupational capital and worker sorting, and the method for measuring CETC at the occupational level, with the defining equations: enough to know what was found and how, without reading the full 44 pages. To replicate or extend, read the original at [doi:10.1257/aer.20211478](https://doi.org/10.1257/aer.20211478).

## TL;DR

Caunedo, Jaume, and Keller construct the first direct measures of capital-embodied technical change (CETC) at the occupational level, covering 24 BEA equipment categories and 327 US occupations from 1984 to 2015, by combining NLP-extracted tool use from the 1977 Dictionary of Occupational Titles and O\*NET with BEA quality-adjusted capital stocks. They also estimate the elasticity of substitution between capital and labor in each one-digit occupation via an instrumental-variables strategy. Embedding these measures in a general equilibrium model of occupational choice (Roy 1951 tradition with Frechet efficiency draws), they find that CETC accounts for 95% of gross US labor reallocation between 1984 and 2015 and 51% of the rise in the college premium. The key driver is not the extent of CETC but heterogeneity in the elasticity of substitution across occupations: without it, CETC would generate less than 10% of the observed high-skill employment shift.

## Core results

Magnitudes are as reported; locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | CETC accounts for **72% of the labor reallocation toward high-skill occupations** (professionals, managers, technicians) between 1984 and 2015 | Table 1, p. 1667 | 7.23 pp of 10.06 pp observed high-skill employment share increase |
| R2 | CETC accounts for **58% of the employment loss in middle-skill occupations** (machine operators, precision production, admin. services, sales, mechanics) | Table 1, p. 1667 | -7.82 pp of -13.58 pp observed middle-skill employment share decline |
| R3 | CETC drives **95% of gross labor reallocation** across all occupations | Table 1, p. 1667 | 2.89 pp of 3.04 pp average absolute employment share change; data gross reallocation = 3.0 pp |
| R4 | CETC accounts for **51% of the rise in the college premium** between 1984 and 2015 | Table 2, p. 1671 | 15.56 pp of 30.58 pp observed college premium increase |
| R5 | CETC **widens the gender wage gap by 17.49 pp**, primarily by raising wages per efficiency unit in mechanics/transportation (male-intensive) and managerial occupations | Table 2, p. 1671 | Without CETC, gender wage gap would have closed by 45.50 pp instead of 28.01 pp |
| R6 | **Occupational IV elasticities of substitution range from 0.65 to 2.18**; aggregate IV elasticity = 0.88, consistent with prior aggregate estimates | Figure 3, p. 1657; App. Table B.III, p. 1681 | Technicians: 0.65 (SE 0.21); mechanics/transp.: 0.73; managers: 0.93; admin. services: 2.18 (SE 0.50); aggregate: 0.88 (SE 0.24) |
| R7 | **Heterogeneity in elasticity of substitution is the primary channel**: imposing a common elasticity (sigma = 0.82) reduces CETC's high-skill employment shift from 7.23 pp to 0.40 pp, less than 10% of the baseline | Table 1, "identical elasticity" column, p. 1667 | 0.40 pp vs 7.23 pp baseline; equalizing CETC paths across occupations changes reallocation by only 0.19 pp |

**Overall (paper's conclusion).** The heterogeneity in the types of capital used across occupations, and consequently in the elasticity of substitution between capital and labor, is the primary channel through which CETC shapes employment reallocation and wage inequality. CETC reallocates employment out of middle-skill occupations (higher capital-labor substitutability) and into high-skill occupations (higher complementarity). Computer-specific CETC alone explains only 10% of the college premium rise; communication equipment and software each explain 12-15%, reinforcing the importance of broad capital measurement relative to prior estimates by Burstein, Morales, and Vogel (2019) who attributed 60% to computers. The routinization mechanism in Autor, Levy, and Murnane (2003) is broadly consistent with the findings, but the substitution channel driven by heterogeneous elasticities, not task content per se, is quantitatively primary. The capital-skill complementarity framework of Krusell, Ohanian, Rios-Rull, and Violante (2000) is extended here to allow heterogeneous substitutability across nine occupation groups.

## Theory / model

The model extends Greenwood, Hercowitz, and Krusell (1997) to include multiple occupations with heterogeneous exposure to CETC, and adopts the Roy (1951) occupational choice framework with Frechet efficiency draws.

**Occupational production.** A representative producer in occupation $$o$$ uses a constant-returns CES technology combining capital $$k_{ot}$$ and labor $$n_{ot}$$ to produce occupational output $$y_{ot}$$ (equation (8), p. 1661):

$$
y_{ot} = \left[ \alpha k_{ot}^{\frac{\sigma_o - 1}{\sigma_o}} + (1-\alpha) n_{ot}^{\frac{\sigma_o - 1}{\sigma_o}} \right]^{\frac{\sigma_o}{\sigma_o - 1}}, \tag{8}
$$

where $$\sigma_o \geq 0$$ is the elasticity of substitution between capital and labor, which differs across occupations. Occupations differ in two dimensions: the technology embodied in capital (CETC) and this elasticity.

**Final good producer.** Final consumption is a CES aggregator of occupational goods (p. 1662):

$$
y_t = \left( \sum_o \omega_{ot}^{1/\rho} y_{ot}^{(\rho-1)/\rho} \right)^{-\rho/(\rho-1)},
$$

where $$\rho$$ is the (absolute) demand elasticity for occupational output and $$\omega_{ot}$$ is an occupational demand shifter capturing offshoring and structural change forces.

**Capital producer.** Each unit of occupational capital is produced from the final good at a rate of transformation $$q_{ot}$$, so the user cost satisfies $$\lambda^k_{ot} = 1/q_{ot}$$. CETC in occupation $$o$$ is the decline in the user cost of occupational capital relative to consumption: a rise in $$q_{ot}$$ is the capital-embodied improvement.

**Worker occupational choice.** The economy has $$H$$ labor groups (defined by age, gender, education). Worker $$i$$ of type $$h$$ draws efficiency units $$n_{oht}(i)$$ from a Frechet distribution with scale $$T_{oht}$$ and shape $$\theta$$. Worker $$i$$ of type $$h$$ chooses the occupation that maximizes wages:

$$
o^*_h(i) = \arg\max_o \{ w_{oht}(i) \}, \tag{13}
$$

where $$w_{oht}(i) = n_{oht}(i) \lambda^n_{ot}$$ is compensation and $$\lambda^n_{ot}$$ is the wage per efficiency unit (endogenously equated across workers in equilibrium). The Frechet property delivers a closed-form occupational allocation (equation (21), p. 1678):

$$
\pi_{oht} = \frac{T_{oht} (\lambda^n_{ot})^\theta}{\sum_{o'} T_{o'ht} (\lambda^n_{o't})^\theta}, \tag{21}
$$

with labor supply elasticity $$\eta_{n\lambda^n_o} = \theta - 1 = 0.30$$ (using $$\theta = 1.30$$ estimated from Mincerian wage residuals).

**Equilibrium wages.** From the zero-profit condition of the occupational producer, the wage per efficiency unit satisfies (equation (18), p. 1677):

$$
\lambda^n_{ot} = \left[ \left(\frac{1}{1-\alpha}\right)^{\sigma_o} (\lambda^y_{ot})^{1-\sigma_o} - \left(\frac{\alpha}{1-\alpha}\right)^{\sigma_o} (\lambda^k_{ot})^{1-\sigma_o} \right]^{\frac{1}{1-\sigma_o}}. \tag{18}
$$

## Method

The method has two parts: measuring occupational CETC from newly constructed data, and estimating the capital-labor elasticity via instrumental variables. It builds on `instrumental-variables`, `panel-regression`, `text-classification`, and `roy-occupational-sorting`.

**Occupational capital stocks and CETC.** The paper covers all 24 BEA equipment and software categories. Quality-adjusted stocks for each category $$j$$ are initialized in 1984 using nominal stocks as the base and then iterated forward:

$$
k_{ot} = k_{ot-1} e^{\gamma^k_{ot}}, \quad k_{o,1984} = \sum_j \lambda^k_{j,1984} k_{oj,1984}, \tag{1}
$$

where $$\gamma^k_{ot} = \sum_j \omega_{ojt} \gamma^k_{ojt}$$ is the expenditure-share-weighted average growth rate of the equipment categories used in the occupation. The user cost of capital for equipment $$j$$ follows the Jorgenson (1963) no-arbitrage condition (p. 1647):

$$
\lambda^k_{jt} = \frac{p^k_{jt-1}}{\lambda^c_{t-1}} \left[ R - (1-\bar\delta_{jt}) \frac{p^k_{jt}/\lambda^c_t}{p^k_{jt-1}/\lambda^c_{t-1}} \right],
$$

where $$p^k_j$$ is the quality-adjusted price, $$\lambda^c$$ is the price of consumption, $$R = 1.02$$ is the gross return on a safe asset, and $$\bar\delta_{jt}$$ is the average physical depreciation. Occupational CETC is then the implied user cost of occupational capital (equation (2), p. 1647):

$$
\lambda^k_{ot} = \frac{\sum_j \lambda^k_{jt} k_{ojt}}{k_{ot}}. \tag{2}
$$

**Occupational capital requirements.** The capital requirement index assigns the fraction of each equipment category's aggregate services to each occupation, using the tools reported by workers in that occupation (equation (3), p. 1649):

$$
\text{req}_{ojt} = \frac{\tau_{ojt} l_{ot}}{\sum_o \tau_{ojt} l_{ot}}, \tag{3}
$$

where $$\tau_{ojt}$$ is the count of tools from category $$j$$ used by occupation $$o$$ at time $$t$$, and $$l_{ot}$$ is full-time-equivalent workers. The tool data for 2015 come from O\*NET; for 1984, NLP string matching is applied to the 1977 Dictionary of Occupational Titles (DOT) to extract the same tool taxonomy, and then linearly interpolated between the two years.

## Empirical specifications

**Elasticity of substitution (Section II.A).** The structural equation for the capital-labor ratio is estimated as a time-series regression for each one-digit occupation (equation (5), p. 1655):

$$
\ln\!\left(\frac{k_{ot}}{\tilde n_{ot}}\right) = \beta_{1o} + \beta_{2o} t + \beta_{3o} \ln\!\left(\frac{\tilde\lambda^n_{ot}}{\lambda^k_{ot}}\right) + \varepsilon_{ot}, \tag{5}
$$

where $$k_{ot}/\tilde n_{ot}$$ is the observed capital-labor ratio (labor adjusted for efficiency via observable demographics), $$\tilde\lambda^n_{ot}/\lambda^k_{ot}$$ is the ratio of the measured labor price to the capital user cost, $$\beta_{3o}$$ identifies $$\sigma_o$$, and $$\beta_{2o}$$ captures the rate of factor-augmenting technical change. The OLS estimate is biased because relative factor prices are endogenous to capital-labor ratios. Instruments exploit exogenous labor supply shifts: (i) 16-year lagged live births interacted with 1984 occupation-education shares $$sh^e_{oe,1984}$$, and (ii) aggregate trade shocks for occupations with weak first-stage F-statistics (mechanics/transportation and low-skill services). All regression series span 1984 to 2015 (32 annual observations per occupation).

**Workers' exposure to CETC (Section II.B).** Under constant returns and competitive markets, the cross-price elasticity of occupational labor demand with respect to the user cost of capital (equation (7), p. 1659) is:

$$
-\frac{d\ln(n_o)}{d\ln(\lambda^k_o)} = \frac{\eta_{n\lambda^n}(\rho - \sigma_o) \frac{\lambda^k_o k_o}{\lambda^y_o y_o}}{\rho + \eta_{n\lambda^n} + (\sigma_o - \rho) \frac{\lambda^k_o k_o}{\lambda^y_o y_o}}, \tag{7}
$$

where $$\sigma_o$$ is the elasticity of substitution (Section II.A), $$\eta_{n\lambda^n} = 0.30$$ is the labor supply elasticity, $$\rho = 1.34$$ is the demand elasticity across occupational outputs, and $$\lambda^k_o k_o / (\lambda^y_o y_o)$$ is the capital expenditure share. Exposure is positive (CETC raises labor demand) when $$\sigma_o < \rho$$ and negative when $$\sigma_o > \rho$$.

**General equilibrium quantification (Sections III-IV).** The model is parameterized to the US 1984-2015 period using two steps. First, $$\sigma_o$$ and the capital user costs $$\lambda^k_{ot}$$ come from Sections I-II. Second, the scale parameters $$T_{oht}$$ of the Frechet distribution are inferred from observed occupational choices and wages using the equilibrium conditions (equations (21) and (22)). The demand elasticity $$\rho = 1.34$$ is estimated from the regression (equation (14), p. 1664):

$$
\ln\!\frac{\lambda^y_{ot} y_{ot}}{\lambda^y_{o_0 t} y_{o_0 t}} = \beta_1 + \beta_{2o} t + \beta_3 \ln\!\frac{\lambda^y_{ot}}{\lambda^y_{o_0 t}} + \varepsilon_{ot}, \tag{14}
$$

instrumented by a Bartik-style shift in the average cost of capital by occupation. Counterfactuals are run by removing exogenous forces (CETC, demand, demographics, comparative advantage, group composition) one at a time in all orderings, then averaging the marginal contributions (Shapley decomposition approach).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| BEA Fixed-Asset Tables (24 equipment/software categories) | Quality-adjusted capital stocks by equipment category; investment series; depreciation rates | no page yet |
| O\*NET Tools and Technology module (2010s) | Occupational tool use for 2015 assignment of capital to occupations | no page yet |
| Dictionary of Occupational Titles (DOT, 1977) | NLP-extracted tool use for 1984 capital assignment; interpolated with O\*NET to build time series | no page yet |
| March Current Population Survey (CPS, Flood et al. 2019 / IPUMS) | Annual labor market statistics: employment shares, wages, full-time-equivalent workers, by occupation and demographic group, 1984-2015 | no page yet |
| October CPS computer supplement (1984, 2003) | External validation of computer tool assignment against workers' self-reported computer use at work | no page yet |

Sample: 324 3-digit census occupations (9 one-digit groups), 1984-2015, annual frequency. Capital stocks initialized 1984; base year for normalization is 1985.

## When to read the full paper

Read the [original](https://doi.org/10.1257/aer.20211478) when: constructing occupational-level capital exposure measures (Section I describes the data construction and NLP assignment in full detail); estimating occupation-specific factor substitution elasticities (the IV strategy and weak-instrument diagnostics in Appendix Tables B.II-B.III are essential for replication); building a multi-occupation Roy-model GE framework (Appendix A derives all equilibrium conditions); or studying the differential role of specific equipment categories (Table 3 decomposes CETC by computers, communication, and software). The replication data and code are at [zenodo.7591599](https://doi.org/10.5281/zenodo.7591599) and the occupational capital dataset is available at www.capitalbyoccupation.weebly.com.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(6), June 2023. Freely accessible via pubs.aeaweb.org after the AEA's 12-month delayed open-access period. No CC licence; redistribution is extract-only. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**.

> Caunedo, Julieta, David Jaume, and Elisa Keller. "Occupational Exposure to Capital-Embodied Technical Change." *American Economic Review* 113, no. 6 (June 2023): 1642-1685. DOI: 10.1257/aer.20211478.


==============================================================================
# Macroeconomics of the Greek Depression: Chodorow-Reich, Karabarbounis & Kekre (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/chodorow-reich-et-al-macroeconomics-greek-depression-2023/
# Distilled: An estimated structural dynamic general equilibrium model decomposes Greece's 1998-2017 boom-bust cycle. Tax policy accounts for the largest fraction of the production bust (-18 of -34 model log-point decline), while uninsurable idiosyncratic income risk drives the bust in consumption and wages. Spending-based fiscal consolidation would have reduced the output bust by roughly 7 log points. American Economic Review 2023, paywalled. Eight core results with source locators, the model equations, and the Bayesian estimation approach. LLM-distilled, not human-verified.
# Tags: paper-summary, macro, fiscal-policy, business-cycles, open-economy-macro
==============================================================================

**What this is.** The paper's core results, the structural model it builds, and
the Bayesian estimation approach with key equations: enough to know what Greece's
boom-bust cycle was driven by and how the model identifies those forces, without
reading all 47 pages. To replicate or extend it, read the full source at the
[original](https://doi.org/10.1257/aer.20210864).

## TL;DR

The paper develops and estimates a dynamic general equilibrium model of a
small open economy in a currency union to decompose the sources of Greece's
boom (1998-2007) and subsequent depression (2007-2017). It quantitatively
confirms the central finding of Gourinchas, Philippon and Vayanos (2016) that
fiscal consolidation drove about half of the bust in output, while substantially
extending it with endogenous TFP via variable utilization, banking sector
frictions, idiosyncratic income risk, and a more detailed tax structure. On the
production side, external demand and government non-traded consumption account
for essentially the entire boom; tax policy, amplified by a working capital
constraint on firms and variable factor utilization, accounts for the largest
fraction of the bust. On the consumption side, realized and anticipated EU
transfers fuel the boom; the rise in uninsurable idiosyncratic income risk,
tracked by the long-term unemployment rate, accounts for the largest fraction
of the decline in consumption, prices, and wages.

Unlike the standard boom-bust narrative emphasizing downward nominal wage
rigidity in a currency peg (Schmitt-Grohé and Uribe 2016), nominal rigidities
play only a moderate role: wages and prices fell substantially during the Greek
crisis, which this model attributes primarily to the rise of idiosyncratic risk
as a negative demand and positive labor-supply shock. The model also extends
the joint European boom-bust analysis of Martin and Philippon (2017) by adding
endogenous TFP movements, capital accumulation, a banking sector, time-varying
idiosyncratic risk, and multi-rate tax measurement.

Counterfactual experiments show that a more spending-based fiscal consolidation
would have reduced the output bust by roughly 7 log points, and that avoiding
the debt-financed boom in household transfers would have created fiscal space to
lower distortionary capital taxes in the crisis. External and bank bailouts
mitigated the depression; without the Economic Adjustment Programme, borrowing
costs would have spiked roughly 30 percentage points in 2012.

## Core results

Magnitudes and contributions are as reported in the source tables and figures.
All log deviations are expressed as differences from the 1998 baseline after
detrending at 1.6 percent per year (quantities) and 1 percent per year (prices
and wages). Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | External demand and government non-traded spending account for essentially all of the production boom | Table 3, p. 2443 | External demand $$\bar{a}_T$$ +0.04 log pts, government consumption $$g_N^c$$ +0.02 log pts out of 0.09 total model log-output boom; data boom = 0.14 |
| R2 | Realized EU structural transfers and anticipated transfers drive the consumption boom | Table 3, p. 2443 | $$T^s$$ +0.02 log pts, $$T^l$$ +0.01 log pts, external +0.05 log pts to log consumption out of 0.08 model boom; data consumption boom = 0.15 |
| R3 | Tax policy is the dominant driver of the bust in production | Table 4, p. 2445 | Tax policy contributes -0.18 log pts out of -0.34 model (data -0.40) in log output 2007-2017; $$\kappa_\tau$$ -0.07, $$\tau_N^k$$ -0.05, $$\tau^\ell$$ -0.03 |
| R4 | Uninsurable idiosyncratic risk is the dominant driver of the bust in consumption and wages | Table 4, p. 2445; text p. 2444 | Idiosyncratic risk $$\pi^\theta$$ contributes -0.14 out of -0.28 model log-consumption bust; accounts for 10 pp of price decline and 18 pp of wage decline |
| R5 | Spending-based consolidation would have reduced the output bust by 7 log points | Figure 5, p. 2448 | Shifting all fiscal adjustment from taxes to spending cuts raises log output by +7 log pts by 2017 vs baseline; roughly half via TFP gains from lower taxes |
| R6 | Fiscal discipline in the boom and capital tax cuts in the bust raise output by 16 pp by 2017 | Figure 6, p. 2452 | Removing debt-financed transfers in boom and using freed resources to cut capital taxes: output +16 log pts, consumption +12 log pts by 2017; labor tax path adds only +2-4 log pts |
| R7 | Fiscal multipliers: most spending multipliers below 1 (nontraded investment $$g_N^x$$ = 1.24 exceeds 1); capital-tax multipliers are large | Table 6, p. 2450; text p. 2451 | $$g_N^c$$ output multiplier = 0.56; $$g_N^x$$ = 1.24 (nontraded investment); aggregate revenue-based tax multiplier = 1.34; capital tax cost-based multiplier $$\tau_H^k$$ = 4.46 |
| R8 | External bailout (EAP) prevented a 20 pp additional output shortfall; bank equity injections raised output 4 pp | Figures 7-8, pp. 2453-2454; text pp. 2414-2415 | Without EAP, borrowing cost rises ~30 pp in 2012; government bailout raised output ~20 pp and consumption 20-40 log pts in 2013 by preventing further spending cuts or tax hikes; bank bailout raised output ~4 pp by 2017 |

**Overall (paper's conclusion).** Greece's depression differs profoundly from
standard small-open-economy boom-bust narratives. The production bust was not
driven by nominal rigidities preventing wage adjustment, since wages and prices
fell substantially. Instead, tax increases, amplified through variable
utilization and a working capital constraint, drove most of the output decline.
The consumption decline, unusual in its persistence, reflected rising
idiosyncratic income risk. The composition of fiscal adjustment and the timing
of transfers in the boom period were as consequential as the aggregate size of
the consolidation.

## Theory / model

The model is a small open economy operating in a currency union, populated by
heterogeneous households, traded and nontraded goods firms, a banking sector,
and a government. Trend productivity grows at rate $$(1-\alpha)\mu$$, and all
variables are expressed in detrended stationary form (p. 2417).

**Households.** Workers $$\iota \in [0,1]$$ belong to two types: a fraction
$$\zeta$$ belongs to the rule-of-thumb household $$r$$ (more impatient, borrows at
capacity, does not hold firm shares) and a fraction $$1-\zeta$$ belongs to the
optimizing household $$o$$. Workers in the optimizing household face idiosyncratic
income risk. Worker $$\iota$$ in household $$h = \{r,o\}$$ values consumption and
labor via recursive preferences (equation (1), p. 2417):

$$
V_{it}^h = \left\{(c_{it}^h)^{1-\frac{1}{\rho}} \left[1 + \left(\frac{1}{\rho}-1\right) \frac{\chi(\ell_{it}^h)^{1+\frac{1}{\varepsilon}}}{1+\frac{1}{\varepsilon}}\right] + \beta^h e^{(1-\frac{1}{\rho})\mu}\left[E_{it}(V_{it+1}^h)^{1-\sigma}\right]^{\frac{1-\frac{1}{\rho}}{1-\sigma}}\right\}^{\frac{1}{1-\frac{1}{\rho}}} \tag{1}
$$

where $$\sigma > 0$$ governs risk aversion, $$\rho > 0$$ the intertemporal elasticity of
substitution (estimated: $$\hat{\rho} = 0.97$$), and $$\varepsilon > 0$$ the Frisch
elasticity (estimated: $$\hat{\varepsilon} = 1.16$$). Combining Epstein and Zin (1989)
preferences with constant Frisch elasticity separates risk aversion from
intertemporal substitution, which matters for the role of idiosyncratic
risk in the bust.

Idiosyncratic income shocks for the optimizing household follow a random walk
in logs (equation (4), p. 2418):

$$
\log \theta_{it+1}^o = \log \theta_{it}^o + \nu_{it+1}^\theta \tag{4}
$$

where innovations wash out at the household level, $$\int \exp(\nu_{it}^\theta) d\iota = 1$$.
A permanent income loss $$-\varphi^\theta$$ occurs with probability $$\pi_t^\theta$$,
measured by the long-term unemployment rate (rising from ~5% before the crisis
to ~20% during it, Figure 3 panel I). This uninsurable risk is central to the
model's transmission of the bust into consumption, prices, and wages.

**Firms.** Intermediate goods firms produce traded goods $$y_H$$ and nontraded
goods $$y_N$$ using Cobb-Douglas technology with variable utilization (equation
(9), p. 2420):

$$
y_{H,t} = z_{H,t} u_{H,t} (e^{-\mu} k_{H,t})^\alpha (\ell_{H,t})^{1-\alpha}, \quad y_{N,t} = z_{N,t} u_{N,t} (e^{-\mu} k_{N,t})^\alpha (\ell_{N,t})^{1-\alpha} \tag{9}
$$

where $$z_{H,t}$$, $$z_{N,t}$$ are exogenous productivity in each sector,
$$u_{H,t}$$, $$u_{N,t}$$ are endogenous utilization rates chosen by firms, and $$k$$
is capital (variable utilization raises depreciation, calibrated using firm
surveys). The endogenous utilization mechanism is central: without it ($$\xi_H = \xi_N = \infty$$),
the model would generate a bust in output and TFP more than 10 log points smaller
(Table 5, p. 2447).

Firms face a working capital constraint that links production decisions to the
endogenous borrowing cost $$i_t$$ (equation (12), p. 2421):

$$
B_{t+1}^f + \kappa_y(P_{H,t} y_{H,t} + P_{N,t} y_{N,t}) = \kappa_x(1+\tau_t^x) P_{x,t} x_t + \kappa_\ell W_t \ell_t + \kappa_{\tau,t} T_t^f + (1+i_t) e^{-\mu} B_t^f \tag{12}
$$

where $$\kappa_x$$, $$\kappa_\ell$$, $$\kappa_{\tau,t}$$ are the fractions of investment,
labor, and tax payments requiring working capital financing. The fraction
$$\kappa_{\tau,t}$$ rises from 50 to 100 percent during the crisis as firms are
required to prepay income taxes before revenues realize. This constraint amplifies
the production bust: without it, both the production boom and bust would have been
smaller.

**Banking sector.** Banks follow Gertler and Kiyotaki (2011) and Bocola (2016).
Incumbent banker net worth evolves as (equation (17), p. 2424):

$$
N_{t+1}^c = (1+\bar{i}_{t+1}) e^{-\mu} N_t + (i_{t+1} - \bar{i}_{t+1})(B_{t+1}^f + \zeta B_{t+1}^r) \tag{17}
$$

where $$\bar{i}$$ is the cost of funds from the rest of the world and $$i$$ is the
domestic lending rate. An incentive compatibility constraint (equation (18),
p. 2424) limits the lending spread via the threat of diversion:

$$
\kappa_b(B_{t+1}^f + \zeta B_{t+1}^r) \leq J_t^b \tag{18}
$$

where $$J_t^b$$ is bankers' continuation value proportional to net worth $$N_t$$.
Losses on sovereign debt (captured in $$T_{Gd,t}^b$$) erode bank net worth during
the crisis, raise the lending spread $$i_t - \bar{i}_t$$, and reduce firms' factor
demand through the working capital constraint.

**Driving forces.** The model organizes exogenous shocks into six categories
(p. 2425): (i) traded and nontraded productivity $$z_H$$, $$z_N$$; (ii) external demand
$$\bar{a}_T$$ and import prices; (iii) financial conditions (sovereign borrowing
limit, bank net worth shocks $$T_W^b$$, $$T_{Gd}^b$$, $$T_{Ge}^b$$); (iv) government
spending ($$g_T^c$$, $$g_N^c$$, $$g_T^x$$, $$g_N^x$$, transfers $$T^r$$); (v) tax policy
($$\tau^c$$, $$\tau^x$$, $$\tau^\ell$$, $$\tau_H^k$$, $$\tau_N^k$$, prepayment fraction
$$\kappa_{\tau}$$); and (vi) disaster risk (idiosyncratic $$\pi^\theta$$ and aggregate
$$\pi^a$$). All processes follow a VAR(1) (equation (21), p. 2426):

$$
\mathbf{z}_{t+1} = \bar{\mathbf{z}} + \mathbb{R}\, \mathbf{z}_t + \Sigma\, \nu_{t+1}, \quad \nu_{t+1} \sim \mathcal{N}(0, \mathbf{I}) \tag{21}
$$

## Method

The model is solved by a first-order perturbation around its steady state. It
is estimated by Bayesian techniques following the `bayesian-dsge-estimation`
approach (related to Smets and Wouters 2007). The key methodological
discipline is that the time series of exogenous processes $$\mathbf{z}$$ are fed
directly as observables without adding any measurement error; only the outcome
variables receive measurement errors. This restricts the shocks to account for
the data without slack from measurement noise, testing the model's fit more
stringently.

The estimation uses 16 observable outcome variables (equation (23), p. 2437):

$$
\mathbf{y} = \left(\log \ell_H,\; \log \ell_N,\; \log \text{TFP}_H,\; \log \text{TFP}_N,\; \log u_H,\; \log u_N,\; s,\; \log c,\right.
$$

$$
\left.\log(P_N c_N),\; \log x_T,\; \log x_N,\; \log P_H,\; \log P_N,\; \log W,\; \Pi^f/(P_y y),\; \log N\right) \tag{23}
$$

covering sectoral labor and TFP, utilization, capital share, aggregate and
sectoral consumption, investment, prices, wages, firm profits, and bank net worth.
The model achieves correlations with data above 0.9 for most variables (online
Appendix Table C.10, p. 2437 footnote).

Parameters are divided into three groups: (a) parameters set without solving the
model (Table 1, p. 2436: $$\sigma=3$$, trade elasticity $$\eta=1.65$$ estimated from
a regression of relative expenditure on relative prices, $$\varphi^a=0.24$$ from
Barro and Liao (2021)); (b) parameters calibrated from steady-state targets
(Table 2 Panel A: discount factors, capital share $$\alpha=0.44$$, banking
parameters); (c) parameters estimated by Bayesian MCMC from the time series
(Table 2 Panel B: $$\hat{\rho}=0.97$$, $$\hat{\phi}=3.17$$, $$\hat{\varepsilon}=1.16$$,
$$\hat{\zeta}=0.34$$, utilization elasticities $$\hat{\xi}_H=3.12$$, $$\hat{\xi}_N=3.75$$,
price and wage adjustment costs $$\hat{\psi}_{H,p}=79.3$$, $$\hat{\psi}_w=78.4$$).

Source decompositions (Tables 3-4) are computed by shutting off the time
evolution of each group of driving forces in turn, holding them constant at
their steady-state values. Positive entries in a row indicate the group
contributed to an increase in a variable; negative entries indicate it
contributed to a decrease. By construction, contributions sum to the model
total up to rounding.

## Empirical specifications

The paper's three main empirical exercises are source decompositions, structural
element comparisons, and policy counterfactuals.

**Source decompositions.** The model is run once for the full sample with all
shocks; then each group $$g$$ of driving forces is held at its steady-state mean
while all others are fed in. Changes in endogenous variables across these runs
identify the contribution of each group. Tables 3 and 4 (pp. 2443, 2445)
report changes in log output, log labor, log capital, log TFP, log consumption,
log traded and nontraded prices, log wage, and net-exports-to-GDP for the boom
(1998-2007) and bust (2007-2017) periods respectively.

**Structural element analysis.** Table 5 (p. 2447) re-runs the model with
alternative parameter values (e.g. $$\xi_H = \xi_N = \infty$$, $$\psi_p = \psi_w = 0$$,
$$\varphi^\theta = 0$$, no working capital) to identify which model features
account for the boom-bust dynamics. Variable utilization and idiosyncratic risk
are identified as the two structural elements that account for most of the boom-bust dynamics.

**Fiscal multipliers.** Fiscal multipliers are defined as the present-discounted
ratio of the output response to the present-discounted change in the fiscal
instrument, at a 7-year horizon (equation (25), p. 2448), discounted at the
steady-state private interest rate $$\bar{i} = 0.04$$:

$$
M_f^y(h) = \frac{\sum_{t=1}^h (1+\bar{i})^{1-t} \Delta y_t}{\sum_{t=1}^h (1+\bar{i})^{1-t} \Delta f_t} \tag{25}
$$

where the impulse is a 1-percentage-point change in the fiscal instrument $$f$$
initiated from its autoregressive process. Revenue-cost multipliers divide
$$M_f^y(h)$$ by the revenue counterpart $$M_f^r(h)$$ (defined symmetrically). Table 6
(p. 2450) reports output effects, revenue costs, and output-per-dollar-of-revenue
for ten instruments: four spending categories, transfers, and five tax rates.
The government nontraded investment multiplier $$g_N^x = 1.24$$ is the largest
individual spending multiplier (investment > consumption, nontraded > traded, per
the paper p. 2449); the nontraded consumption multiplier $$g_N^c = 0.56$$ equals
the aggregate spending-weighted multiplier because $$g_N^c$$ is the largest
spending category by expenditure share.
Capital tax multipliers are the largest tax multipliers, consistent with the
economy operating near the peak of the Laffer curve for capital income taxation
(capital income tax cut approximately revenue-neutral at the margin, p. 2450).

**Policy counterfactuals.** Figures 5-8 (pp. 2448, 2452, 2453, 2454) evaluate
three alternative scenarios: (i) shifting fiscal adjustment entirely from taxes
to spending cuts, holding tax rates at 2009 values while expanding government
spending innovations to balance the budget; (ii) eliminating debt-financed
transfers $$T^r$$ in the boom and using the freed fiscal space to reduce
distortionary taxes in the bust; (iii) removing the external bailout (EAP)
resources and instead forcing Greece to balance the budget via additional
spending cuts or tax hikes. All counterfactuals condition on the estimated
sequence of shocks and compare the model-generated paths to the baseline path
under observed fiscal policies.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Eurostat European System of Accounts (ESA) | Output, prices, consumption, investment, labor, TFP for Greece 1998-2017 (baseline observables for estimation) | No page yet |
| EU Joint Harmonised Commission Surveys (JCS) | Firm-level capacity utilization (manufacturing sector) and services survey (services sector) for $$u_H$$ and $$u_N$$ | No page yet |
| Bank of Greece Flow of Funds | Firm dividends $$\Pi^f$$ and bank net worth $$N$$; financial accounts | No page yet |
| Maastricht Treaty / OECD Economic Outlook | Government debt misreporting (anticipated transfers $$T^l$$ series from stated vs. revised deficits) | No page yet |
| Barro-Liao (2021) options-based disaster probability | Far-out-of-the-money put option prices on the Greek stock market for aggregate disaster probability $$\pi^a$$ | No page yet |

Sample: annual, 1998-2017 (20 years, Greece). Quantities detrended at 1.6%
per year, TFP at 0.7%, prices and wages at 1% (euro inflation average).

## When to read the full paper

Use the [original](https://doi.org/10.1257/aer.20210864) if you are:
studying the structural mechanisms that generate large depressions in currency
unions; replicating the source decompositions in Tables 3-4 (the replication
package is at the ICPSR repository linked in `replicationCode`); extending the
model to other periphery euro-area economies; or evaluating the design of fiscal
consolidation programs. The online appendix contains alternative specifications,
robustness checks, and a full set of parameter estimates and model validation.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(9), September 2023.
Published by the American Economic Association, paywalled. This distillation
was extracted by an LLM on 2026-06-25 and is **not human-verified or
independently reproduced**. Redistribution is extract-only; no verbatim PDF
is hosted here.

> Chodorow-Reich, Gabriel, Loukas Karabarbounis, and Rohan Kekre.
> "The Macroeconomics of the Greek Depression."
> *American Economic Review* 113, no. 9 (September 2023): 2411-2457.
> DOI: 10.1257/aer.20210864.


==============================================================================
# Nonlinear Pricing with Underutilization: Corrao, Flynn & Sastry (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/corrao-et-al-nonlinear-pricing-underutilization-theory-2023/
# Distilled: establishes that multi-part tariffs (price schedules with tiers of zero marginal price) are the optimal contract when buyers can freely underutilize purchases and usage generates revenue for the seller via advertising, data, or network effects. American Economic Review 113(3), 2023, paywalled. Six core theoretical results with proposition locators, the seller's problem, and the virtual surplus characterization. LLM-distilled.
# Tags: paper-summary, nonlinear-pricing, multi-part-tariffs, digital-markets, mechanism-design, screening, peer-reviewed
==============================================================================

**What this is.** The paper's core theoretical results and the formal model with its defining equations: enough to know what was proved and how, without reading all 25 pages. To replicate or extend the proofs, read the full source at the [original](https://doi.org/10.1257/aer.20220199).

## TL;DR

The paper studies a seller who cannot monitor or enforce how much a buyer actually consumes of what they purchase (free disposal / noncontractibility of usage), while actual usage generates revenue for the seller via advertising clicks, data collection, or network effects. The classical nonlinear pricing literature of Mussa and Rosen (1978) and Wilson (1993) predicts smooth, continuously increasing price schedules. This paper's main result is that once buyers can freely underutilize, the optimal price schedule is a multi-part tariff: it features at least one tier where the marginal price is zero. The mechanism is that sellers would benefit from paying buyers to use the product more (a negative marginal price), but noncontractibility makes this unenforceable, so zero marginal pricing is the constrained optimum. The model rationalizes free products (search engines, social media), free trials, unlimited subscriptions, and introductory offers, matching the observed pricing of major digital platforms. The welfare analysis shows that perfect contractibility would benefit both consumers and producers, but the same technological barrier that prevents extracting full usage value also prevents compensating users for it.

## Core results

Locators point into the source PDF.

| # | Result | Locator | Statement |
|---|---|---|---|
| R1 | Optimal consumption is the min of producer-optimal and consumer-optimal (bliss point) levels; the optimal price schedule is uniquely given by an integral over marginal willingness to pay | Prop. 1, pp. 844-845 | Characterization of optimal contract under free disposal |
| R2 | H(x) > 0 is sufficient for the price schedule to be flat at x (a multi-part tariff tier); H(x) < 0 is sufficient for a strictly positive marginal price | Prop. 2, p. 848 | Multi-part tariffs arise when marginal usage revenue dominates marginal information rents |
| R3 | Four pricing schemes rationalized by the sign of H: regular (H < 0 everywhere), fixed/free (H >= 0 everywhere), premium-tier (H changes sign from negative), introductory-offer (H changes sign from positive) | Cor. 1, p. 850 | Corollary applies when H crosses zero at most once |
| R4 | Sufficient conditions for unlimited subscriptions (marginal usage revenue positive at the highest-type bliss point) and for free trials (total marginal usage revenue at the lowest-type bliss point exceeds information rent) | Cor. 2, pp. 852-853 | Free trials and unlimited subscriptions co-occur when usage revenue is high at both ends of the type distribution |
| R5 | Under any fixed price schedule, free disposal weakly improves consumer welfare; but under the seller-reoptimized schedule, perfect contractibility strictly improves both consumer and producer welfare for all types | Prop. 3, p. 856 | Noncontractibility reduces both consumer and producer welfare relative to the contractibility benchmark |
| R6 | When usage becomes more profitable, both consumer and producer welfare increase, but by less than under perfect contractibility; free disposal dampens welfare gains from improved advertising or data-collection technology | Prop. 4, p. 857 | Free disposal reduces the sensitivity of welfare to changes in usage-based revenue |

**Overall (paper's conclusion).** The mechanism generating multi-part tariffs is the collision between two constraints: sellers would like to charge negative marginal prices to encourage valuable usage, but noncontractibility prevents this, making zero marginal pricing the constrained optimum. As a normative corollary, users of digital platforms would be better off if usage were perfectly contractible, but the same technological barrier that prevents full extraction also prevents compensation for usage.

## Theory / model

There is a single good consumed in amounts $$x \in X = [0, \bar{x}]$$. A unit measure of buyers have privately known type $$\theta \in \Theta = [0,1]$$ drawn from distribution $$F \in \Delta(\Theta)$$ with density $$f$$ bounded away from zero (p. 840). Buyer utility is quasilinear: $$u(x,\theta) - t$$, where $$u$$ satisfies strict single crossing ($$u_{x\theta} > 0$$) and strict quasiconcavity in $$x$$ for all $$\theta$$. The outside option is normalized to zero: $$u(0,\theta) = 0$$ for all $$\theta$$.

**Underutilization (p. 840):** A buyer who purchases $$y$$ can consume any $$x \in [0,y]$$. This models noncontractibility in digital markets: a newspaper can check if an article was loaded but not whether it was read; Google can verify a search was submitted but not that a human performed it. The AllAdvantage.com case study (p. 840) illustrates this directly: the platform paid users to view banner ads but was defrauded by automated click bots.

**Usage-derived revenue (pp. 841-842):** The seller receives both transfer payments and usage-derived revenue captured by a continuously differentiable function $$\pi: X \times \Theta \to \mathbb{R}$$ with $$\pi(0,\theta) = 0$$. The function $$\pi$$ encompasses advertising revenue, data collection value, network effects, and future addiction revenue. The seller values total revenue $$\pi(\phi(\theta),\theta) + T(\xi(\theta))$$ from both usage and transfers.

**Seller's problem (p. 842):** The seller designs a total-revenue-maximizing price schedule $$T: X \to \bar{\mathbb{R}}$$ anticipating that each type $$\theta$$ will choose purchase quantity $$\xi(\theta)$$ and consumption $$\phi(\theta) \in [0, \xi(\theta)]$$ optimally:

$$\sup_{\phi,\xi,T} \int_{\Theta} \left[\pi(\phi(\theta),\theta) + T(\xi(\theta))\right] dF(\theta) \tag{1}$$

subject to three constraints:
- **(O) Obedience:** each buyer chooses optimal consumption given their purchase: $$\phi(\theta) \in \arg\max_{x \in [0,\xi(\theta)]} u(x,\theta)$$
- **(IC) Incentive compatibility:** each buyer chooses optimal purchase given the price schedule and their ability to underutilize: $$\xi(\theta) \in \arg\max_{y \in X} \left\{\max_{x \in [0,y]} u(x,\theta) - T(y)\right\}$$
- **(IR) Individual rationality:** $$u(\phi(\theta),\theta) - T(\xi(\theta)) \geq 0$$ for all $$\theta \in \Theta$$

**Key objects (p. 843):** The consumer-optimal (bliss point) consumption is:

$$\phi^A(\theta) = \arg\max_{x \in X} u(x,\theta), \tag{2}$$

which is unique and increasing by strict quasiconcavity and strict single crossing. The virtual surplus net of information rents is:

$$J(x,\theta) = \pi(x,\theta) + u(x,\theta) - \frac{1-F(\theta)}{f(\theta)}\,u_{\theta}(x,\theta). \tag{3}$$

Under the assumption that $$J$$ satisfies strict single crossing in $$(x,\theta)$$ and strict quasiconcavity in $$x$$, the producer-optimal consumption maximizing virtual surplus is:

$$\phi^P(\theta) = \arg\max_{x \in X} J(x,\theta). \tag{4}$$

**Proposition 1 (Optimal Pricing, pp. 844-845):** In any optimal contract, consumption is the bliss-point-capped virtual surplus maximizer:

$$\phi^* = \min\{\phi^P, \phi^A\}. \tag{5}$$

The optimal price schedule on $$X^* = [\phi^*(0), \phi^*(1)]$$ is uniquely determined by the standard envelope formula:

$$T^*(x) = u(\phi^*(0),0) + \int_{\phi^*(0)}^{x} u_x\!\left(z,\,\phi^{*-1}(z)\right) dz. \tag{6}$$

Intuition: forcing consumption beyond the bliss point violates (O) because buyers would dispose, so $$\phi \leq \phi^A$$ is necessary. Combined with monotonicity required by (IC), capping at $$\phi^A$$ is both necessary and sufficient for obedience and incentive compatibility. The price formula (6) follows from local (IC) binding.

**Proposition 2 (Multi-part Tariffs, p. 848):** The constrained marginal revenue function $$H: X^* \to \mathbb{R}$$ maps each outcome level to the net marginal gain from additional usage for the type whose bliss point is $$x$$:

$$H(x) = J_x\!\left(x,\,(\phi^A)^{-1}(x)\right). \tag{13}$$

The sufficient condition for $$H(x) > 0$$ is that marginal revenue from usage strictly dominates marginal information rents at the relevant type $$\theta = (\phi^A)^{-1}(x)$$:

$$\underbrace{f(\theta)\,\pi_x(x,\theta)}_{\text{marginal revenue from usage}} > \underbrace{(1-F(\theta))\,u_{x\theta}(x,\theta)}_{\text{marginal information rent}}, \tag{14}$$

where the left side is the per-type marginal profit from usage and the right side is the information rent that must be paid to all higher types to induce truthful purchase revelation. If $$H(x) > 0$$, then $$T^*$$ is flat at $$x$$ (zero marginal price, a multi-part tariff tier). Conversely, if $$T^*$$ is flat at $$x$$, then $$H(x) \geq 0$$. The logic: when $$H(x) > 0$$ the seller would prefer a negative marginal price to incentivize usage, but free disposal makes this unenforceable (buyers would underutilize to capture a negative price without delivering usage value), so zero is the binding constrained optimum.

The closest predecessor, Grubb (2009), demonstrates optimality of three-part tariffs in a model with overconfident consumers. This paper shows that overconfidence maps to a specific external revenue function $$\pi$$, and the framework with free disposal generalizes his result to a broader class of revenue functions and pricing structures.

**Welfare (pp. 855-857):** Consumer welfare under free disposal for type $$\theta$$ is:

$$V(\theta;T) = \sup_{y \in X,\, x \in [0,y]} \left\{u(x,\theta) - T(y)\right\}, \tag{19}$$

and producer welfare is total revenue from type $$\theta$$:

$$\Pi(\theta;T) = \pi(\phi(\theta;T),\theta) + T(\xi(\theta;T)). \tag{20}$$

Let $$V_N$$ and $$\Pi_N$$ denote the corresponding quantities under perfect contractibility of usage (no free disposal). Proposition 3 (p. 856) establishes: for any fixed $$T$$, $$V(\theta;T) \geq V_N(\theta;T)$$ for all $$\theta$$, but under the reoptimized price schedules $$V^*(\theta) \leq V^*_N(\theta)$$ and $$\Pi^*(\theta) \leq \Pi^*_N(\theta)$$ for all $$\theta$$. Proposition 4 (p. 857) shows that when usage becomes more profitable ($$\tilde{\pi}_x \geq \pi_x$$ pointwise) and demand weakens ($$\bar{F}$$ hazard-rate dominates $$F$$), welfare increases for both parties but the gain is bounded above by the gain under perfect contractibility.

## Method

The analysis uses the virtual surplus characterization standard in mechanism design and nonlinear pricing, building on the Mussa and Rosen (1978) framework. The key analytical steps are:

1. **Relaxed problem:** Impose only local (IC) and (O) constraints and derive a pointwise maximization in $$J$$ at each $$\theta$$, giving $$\phi^P$$ as the solution were free disposal absent. The key departure from standard screening is that the obedience constraint (O) is now active: it forces $$\phi \leq \phi^A$$.

2. **Binding obedience:** For types where $$\phi^P(\theta) > \phi^A(\theta)$$, the obedience constraint binds. Monotonicity of $$\phi^*$$ (required by IC) and quasiconcavity of $$J$$ together ensure that capping at $$\phi^A$$ is globally optimal: $$\phi^* = \min\{\phi^P, \phi^A\}$$. The price formula (6) follows from integrating the binding local (IC).

3. **Flatness characterization:** Differentiating (6) gives $$T^{*\prime}(x) = u_x(x,\phi^{*-1}(x))$$ on regions where $$\phi^* < \phi^A$$ (standard positive marginal pricing). On regions where $$\phi^* = \phi^A$$ (obedience binds), the seller is constrained to offer zero marginal prices. The constrained marginal revenue $$H$$ in (13) captures the seller's net gain from this constraint, yielding Proposition 2.

4. **Four pricing schemes (Corollary 1, p. 850):** When $$H$$ crosses zero at most once on $$X^*$$, four cases arise: $$H < 0$$ everywhere (regular pricing); $$H \geq 0$$ everywhere (fixed/free pricing); $$H$$ crosses from negative to positive at $$\hat{x}$$ (premium-tier: positive marginal prices for $$x < \hat{x}$$, zero thereafter); $$H$$ crosses from positive to negative at $$\hat{x}$$ (introductory-offer: zero marginal prices for $$x \leq \hat{x}$$, positive thereafter).

5. **Unlimited subscriptions and trials (Corollary 2, p. 852):** An unlimited subscription ($$T^*$$ flat at $$\phi^*(1)$$, the top of the consumption range) requires $$\pi_x(\phi^A(1),1) > 0$$: marginal usage revenue at the highest-type bliss point is positive, so information rents vanish at the top and usage incentives dominate. A free trial ($$T^*$$ flat at $$\phi^*(0)$$) requires total marginal usage revenue at the lowest-type bliss point to exceed marginal information rents paid to higher types. These conditions are mutually compatible, generating two-tier pricing with both features (Example 2, p. 853).

6. **Bunching extension (online Appendix B.1):** When $$J$$ fails strict single crossing in $$(x,\theta)$$, Nöldeke and Samuelson (2007)'s assignment approach applies. The conclusion of Proposition 2 extends: $$T^*$$ is flat whenever the obedience constraint binds.

7. **Perfect competition (online Appendix B.2):** Under a zero-profit constraint for the monopolist, the equilibrium price schedule maximizes total surplus instead of virtual surplus. Total surplus is maximized at a higher consumption level than virtual surplus (no information rent deduction), making $$\phi^* = \phi^A$$ bind more often. Multi-part tariffs are therefore more prevalent under perfect competition than under monopoly.

## Empirical specifications

This is a pure theory paper with no empirical estimation. The paper includes three closed-form illustrative examples calibrated to digital goods settings:

**Example 1 (Digital platform with advertisements, p. 845):** Quadratic utility $$u(x,\theta) = \theta x - x^2/2$$, uniform types on $$[0,1]$$, and linear-quadratic advertising revenue $$\pi(x,\theta) = \alpha x - (\beta/2)x^2$$ where $$\alpha = pk - c$$ (revenue per unit time net of production cost) and $$\beta = 2ph$$ (user fatigue parameter). The consumer-optimal and producer-optimal consumption functions are:

$$\phi^A(\theta) = \theta, \qquad \phi^P(\theta) = \max\!\left\{0,\,\min\!\left\{1,\,\frac{\alpha + 2\theta - 1}{\beta + 1}\right\}\right\}. \tag{10}$$

Restricting to $$\alpha \leq 1$$ and $$\beta < 1$$, the constrained marginal revenue $$H(x) = (\alpha - \beta x) - (1-x)$$ crosses zero once, generating a premium-tier tariff with a threshold at $$x = (1-\alpha)/(1-\beta)$$ and price schedule:

$$T^*(x) = \begin{cases} \frac{1-\alpha}{2}x - \frac{1-\beta}{4}x^2, & x < \frac{1-\alpha}{1-\beta} \\ \frac{(1-\alpha)^2}{4(1-\beta)}, & x \geq \frac{1-\alpha}{1-\beta}. \end{cases} \tag{12}$$

Figure 3 (p. 853) illustrates all four cases of Corollary 1 as $$(\alpha, \beta)$$ vary.

**Example 2 (Online newspaper subscriptions, p. 853):** Same demand, but exponential advertising revenue $$\pi(x,\theta) = \alpha(1 - e^{-\lambda x})$$ (ads noticed according to a Poisson process with hazard rate $$\lambda$$, one click per consumer). The constrained marginal revenue $$H(x) = \lambda\alpha e^{-\lambda x} - (1-x)$$ can cross zero twice, generating two-tier pricing with both a free trial and an unlimited subscription for $$\lambda = 2.5$$, $$\alpha = 0.5$$ (Figure 4, p. 855). This matches the Wall Street Journal's pricing structure.

**Example 3 (Arbitrary-part tariffs, p. 854):** Revenue $$\pi(x,\theta) = x(1-\theta) - (k/2\pi\omega)[\cos(2\pi\omega x) - 1]$$ is constructed so that $$H(x) = k\sin(2\pi\omega x)$$ crosses zero $$\omega$$ times, generating $$\omega + 2$$ part tariffs. Figure 5 (p. 856) plots three-, four-, and five-part tariffs for $$\omega \in \{1,2,3\}$$.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| None | Pure theory paper; results established analytically | n/a |

The illustrative examples (Examples 1-3) use closed-form parametric functions, not empirical data sources. All results are theoretical propositions.

## When to read the full paper

Read the [original](https://doi.org/10.1257/aer.20220199) if you are: deriving the optimal contract for a specific digital product (online Appendix A has complete proofs); analyzing welfare effects of data-privacy regulation that reduces advertising revenue (Propositions 3-4 give the formal comparison); extending the model to partial contractibility or more general competition (the conclusion maps open directions); checking the bunching case where virtual surplus fails single crossing (online Appendix B.1); or studying competitive equilibrium pricing (online Appendix B.2 shows multi-part tariffs are more prevalent under perfect competition than monopoly).

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(3), March 2023. Published under AEA copyright; paywalled. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. Extract-only redistribution.

> Corrao, Roberto, Joel P. Flynn, and Karthik A. Sastry. "Nonlinear Pricing with Underutilization: A Theory of Multi-Part Tariffs." *American Economic Review* 113, no. 3 (March 2023): 836-860. DOI: [10.1257/aer.20220199](https://doi.org/10.1257/aer.20220199).


==============================================================================
# Relinquishing Riches: Covert & Sweeney (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/covert-sweeney-relinquishing-riches-auctions-versus-2023/
# Distilled: Auctioned oil and gas leases in Texas generate 53 log points more in up-front bonus payments and 39 log points more output than informally negotiated leases, measured using a natural experiment from early-twentieth-century Texas land allocation decisions. American Economic Review 2023, paywalled. Six core results with source locators, datasets used, the identification strategy, and the estimating equations.
# Tags: paper-summary, auction-theory, market-design, natural-resources, oil-and-gas
==============================================================================

**What this is.** The paper's core results, identification strategy, and estimating equations: enough to know what it found and how, without reading all 36 pages. To replicate or extend, read the full source at [https://doi.org/10.1257/aer.20191594](https://doi.org/10.1257/aer.20191594).

## TL;DR

Covert and Sweeney compare outcomes on oil and gas leases on Texas Permanent School Fund (PSF) land, where early-twentieth-century legislative decisions quasi-randomly assigned parcels to two allocation mechanisms: informal bilateral negotiations (RAL parcels) and centralized first-price auctions (state auction parcels). Using data from over 1,500 leases signed during the 2004-2016 shale boom, they find that auctioned leases pay 53 log points more in up-front bonus payments and produce 39 log points more output than comparable negotiated leases signed in the same location and time. The combined revenue gain amounts to roughly $341,000 more per average lease. Auctions do not generate more leases overall; the gains arise entirely from better firm-parcel matching, evidenced by within-firm auction premiums and large gaps in bidder values within auctions. The findings imply large potential gains from replacing informal allocation in the roughly $3 trillion private US mineral rights market.

## Core results

Magnitudes and significance are as reported; `\*\*`/`\*\*\*` = 5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Auctioned leases pay **53 log points more** in bonus payments per acre than comparable negotiated leases (main specification) | Table 4 col 2, p. 643 | 0.53 (SE 0.06)\*\*\*; range 0.44-0.59 across all nine specifications; $185,000 more per average RAL lease |
| R2 | Auctioned leases produce **39-45 log points more** oil and gas output per acre (pseudo-Poisson) | Table 6 Panel B col 2, p. 648 | Auction-output = 0.45 (SE 0.17)\*\*\*; consistent across 10-mile grid, 20-mile grid, and DML specifications |
| R3 | Total seller revenue (bonus + royalties) is **$341,000 higher** per average auctioned lease | Table 6 col 2, p. 648 | $1.15k/acre more (SE $0.39k)\*\*\*; Panel B: 0.43 log points (SE 0.12)\*\*\*; average negotiated lease generates $5,780/acre |
| R4 | Within-firm: **auction premium persists** after conditioning on firm identity, ruling out firm-composition as the driver | Table 9 col 2 and 4, p. 654 | Auction-bonus with firm FE = 0.58 (SE 0.08)\*\*\*; within-firm output gap = 1.583 barrels/acre (SE 0.681)\*; point estimates larger than the pooled baseline |
| R5 | **No differential leasing rate**: auction and negotiation parcels are equally likely to be under lease in 50 of 52 sample quarters | Figure 4, p. 650 | Point estimates tightly centered on zero; cannot reject equal hazard rates; both mechanisms generate similar reserve prices (Table 12 cols 3-4) |
| R6 | Large **allocative efficiency gains**: auction winner's inferred value is 50-107 log points above competing bidders' values | Table 11, p. 656 | 1-to-2 allocative gain = 50 log points; 1-to-N gain = 107 log points (3+ bidders); consistent with output difference in Panel B of Table 6 |

**Overall (paper's conclusion).** The allocation mechanism for mineral leases, which is determined by pre-fracking land privatization dates rather than by observed quality, generates large and robust differences in both payments and output. The gains arise from better horizontal matching of parcels to the firms that can use them most productively, not from a differential probability of transacting. The implied annual gains from formalizing the private mineral leasing market are economically large relative to the cost of operating electronic auction platforms.

## Theory / model

The paper has no formal structural model. The tested hypothesis is that a centralized, first-price sealed-bid auction (administered by the Texas General Land Office) generates higher payments and more output than the informal bilateral negotiation process used on RAL parcels, when both mechanisms face the same distribution of bidder values and the same underlying resource quality.

The mechanism-design benchmark is the classic Bulow and Klemperer (1996) result that an English auction with one extra bidder dominates optimal bilateral bargaining, and the Roberts and Sweeting (2013) result that a sequential mechanism can outperform a simultaneous auction when entry is selective. Prior empirical work by Hendricks and Porter (1988) on US government Gulf of Mexico mineral lease auctions showed that centralized auctions capture most of the surplus in symmetric information environments; this paper extends the comparison to informal private-land negotiations. Larsen (2021) provides analogous reduced-form evidence in wholesale used-auto markets, finding large gaps between actual bargaining outcomes and second-best mechanisms. Salz (2022) models intermediary-organized auctions in waste collection; the paper uses that framework to interpret negotiations as an auction with fewer participants. Kong (2020) studies auction design in neighboring New Mexico mineral lease auctions and provides complementary evidence on bidder uncertainty. The paper tests these predictions directly rather than estimating a structural model of the negotiation process, because the informal process leaves no observable record at the transaction level that would anchor a structural estimate.

**Identification logic (pp. 637-640).** Parcels inside the PSF were privatized at different times: RAL parcels were sold to private surface owners before 1931 (granting them rights to negotiate mineral leases on behalf of the state); all remaining PSF parcels transact via GLO auction. The privatization dates are determined before the fracking boom, and, within narrow geographic areas, the RAL/auction status is argued to be uncorrelated with shale rock quality, verified via balance tests on shale thickness (Table 3, p. 640). The identifying assumption is thus:

Within a 10-mile geographic grid cell, and conditional on the year-quarter of lease signing, the assignment of a parcel to the RAL (negotiation) or auction mechanism is as good as random.

## Method

The paper applies a natural-experiment design with location-by-time fixed effects and several additional robustness layers, including double/debiased machine learning (DML) following Chernozhukov et al. (2018).

**Core estimator.** The primary object is the average treatment effect of auction assignment on lease outcomes, estimated by ordinary least squares. The technique builds on `panel-regression` and uses the location-by-time fixed-effect structure to absorb unobserved geological and market variation. For outcomes measured in levels (dollar amounts, output quantities), the linear estimator is:

$$Y_i = \tau \, \text{Auction}_i + X_i \beta + \delta_{L(i),\, T(i)} + \varepsilon_i \tag{1}$$

where $$Y_i$$ is the lease outcome, $$\text{Auction}_i$$ is an indicator equal to one for state-auction leases, $$X_i$$ includes lease size and (in extended specifications) additional surface and geological controls, and $$\delta_{L(i),\, T(i)}$$ is a fixed effect for the location bin $$L(i)$$ (10- or 20-mile square grid cell containing the lease centroid) crossed with the time bin $$T(i)$$ (year-quarter of lease signing). Standard errors are clustered at the grid level. The parameter of interest $$\tau$$ is interpreted as the average causal effect of auction assignment on the outcome, under the identifying assumption above.

**Pseudo-Poisson estimator.** For heavily right-skewed outcomes with a mass of zeros (most leases are never drilled), a pseudo-Poisson quasi-maximum likelihood estimator is used:

$$\log E[Y \mid \text{Auction}_i,\, X_i,\, L_i,\, T_i] = \tau \, \text{Auction}_i + X_i \beta + \delta_{L,\, T}$$

This projects the log of the expected outcome on the same controls, accommodating proportional effects and the zeros without requiring a log transformation (p. 648).

**DML nonparametric controls.** As an alternative to grid fixed effects, the Robinson (1988) partially linear model estimated via DML (Chernozhukov et al. 2018) is used to control nonparametrically for location and time. The cross-fitted empirical analog of the orthogonality condition is:

$$E\!\left[\bigl(Y - \gamma(L,T,X) - \tau(D - \delta(L,T,X))\bigr)(D - \delta(L,T,X))\right] = 0$$

where $$D = \text{Auction}$$, $$\gamma(L,T,X) = E[Y \mid L,T,X]$$, and $$\delta(L,T,X) = E[D \mid L,T,X]$$ are estimated by random forest. The DML estimates (columns labelled "DML" in Tables 4, 6, and 7) are stable across all specifications (p. 641, footnote 22).

## Empirical specifications

**Bonus regressions (R1).** The dependent variable is $$\log(\text{bonus payment per acre})$$. All models include a spline in lease size. Specifications in Table 4 (p. 643) vary the spatial grid (10 vs. 20 miles), the time control (quarter-of-sample Q vs. grid-by-year-quarter GY,Q vs. grid-by-year-and-quarter-of-sample GYQ), and whether additional surface and geological controls ("Extra") or private-surface-only sample restrictions ("Private only") are included. The 9 specifications all yield $$\tau \in [0.44, 0.59]$$ with SEs of 0.05-0.10, all significant at 1%.

**Output and revenue regressions (R2, R3).** Dependent variables are discounted lease revenue (in $k/acre), discounted output (hundreds of BOE/acre), and discounted seller revenue (in $k/acre). Table 6 Panel A uses the linear model (1); Panel B uses the pseudo-Poisson. Fixed effect models drop grids and time periods with no variation in outcome. The sample is restricted to leases whose primary term ends before March 2019 to reduce right-censoring of realized production (p. 647). Main specification output result: Auction coefficient = 0.45 log points (SE 0.17, Panel B col 2); seller revenue = 0.43 log points (SE 0.12, Panel B col 2).

**Royalty rate and primary term regressions (R5, not in findings).** Table 5 (p. 645) estimates equation (1) with the royalty rate (in percentage points) and primary term (in years) as dependent variables. Auctions generate approximately 0.82-0.95 pp higher royalty rates and 0.85-1.00 year longer primary terms, significant at 1%, across all specifications. These are intermediate outcomes that also favor the mineral owner.

**Firm fixed-effect regressions (R4).** Table 9 (p. 654) adds indicators for each E&P firm identity to equation (1). The within-firm auction premium for bonus is 0.58 (SE 0.08), larger than the pooled 0.53, confirming that the auction gains are not driven by a different composition of firms.

**Parcel-level regressions.** Table 7 (p. 652) re-estimates equation (1) at the parcel level, including parcels that never sign a lease (all outcomes discounted to 2004). This addresses the selection concern that only successful transactions are observed at the lease level. Parcel results confirm the lease-level findings, though noisier due to the additional zeros.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Texas General Land Office (GLO) oil and gas lease records, 2004-2016 | Bonus payments, primary terms, royalty rates for 4,460 RAL leases + 694 auction leases; monthly royalty production data through March 2019 | No page yet |
| GLO auction bid notices and bid data | Auction reserve prices and all bids above reserve; used to infer bidder values and allocative efficiency (Tables 10-12) | No page yet |
| P2 Energy Solutions Texas PSF parcel map (2018) | GIS shapefile linking historical PSF parcel records to geographic boundaries; used to match leases to parcels and compute parcel-level outcomes (Section V) | No page yet |
| US Energy Information Administration (EIA) price and shale data | Henry Hub gas prices, WTI oil prices (for output-revenue conversion); shale formation boundaries and isopachs defining shale thickness | [No page yet](/wiki/datasets/) |
| US Census Bureau TIGER shapefiles (2017) | Texas county boundaries for spatial intersection | No page yet |
| US Geological Survey National Hydrography + Land Cover (2019, 2021) | Distance to water and land cover measures used as surface quality controls in robustness checks (Table 3) | No page yet |

Sample: leases signed 2004-2016 on PSF land overlying shale formations, with size 10-1,000 acres and single-ownership. The main bonus sample contains 1,515 leases; output sample is restricted to leases whose primary term ends by March 2019 (1,259 leases).

## When to read the full paper

Read the [original](https://doi.org/10.1257/aer.20191594) if you are: (a) designing or evaluating a formal mechanism for natural resource or real-estate allocation in a market currently served by informal bilateral negotiation; (b) studying the auction theory literature on auctions vs. negotiations and want direct empirical evidence; (c) extending the identification strategy to other settings where assignment to formal/informal mechanisms is determined by historical institutional decisions; or (d) replicating, with the replication data at [Harvard Dataverse](https://doi.org/10.7910/DVN/9WJ3JK) and [ICPSR](https://doi.org/10.3886/E181143V1). The locators above point to the exact tables and figures.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(3), March 2023. Paywalled; no open-access licence found in Crossref metadata. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. Extract-only; the verbatim article is available via AEA or institutional access.

> Covert, Thomas R., and Richard L. Sweeney. "Relinquishing Riches: Auctions versus Informal Negotiations in Texas Oil and Gas Leasing." *American Economic Review* 113, no. 3 (March 2023): 628-663. DOI: 10.1257/aer.20191594.


==============================================================================
# Old Boys' Club: Cullen & Perez-Truglia (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/cullen-perez-truglia-old-boys-club-schmoozing-2023/
# Distilled: Face-to-face social interactions with managers give same-gendered employees a promotion advantage at a large anonymous commercial bank in Southeast Asia, with quasi-random manager rotations providing causal identification; the male-to-male advantage accounts for about 40 percent of the gender pay gap in promotions at this firm. American Economic Review 2023, paywalled. Eight core results with source locators, datasets used, the event-study design, and the empirical specifications with equations. LLM-distilled, not human-verified.
# Tags: paper-summary, labor-economics, gender-gap, promotions, social-interactions
==============================================================================

**What this is.** The paper's core results, the event-study identification design, and the empirical specifications with exact estimating equations: enough to understand what was found and how, without reading all 38 pages. To replicate or extend, read the full source at [doi:10.1257/aer.20210863](https://doi.org/10.1257/aer.20210863).

## TL;DR

Cullen and Perez-Truglia exploit quasi-random manager rotation at a large anonymous commercial bank in Southeast Asia to show that employees who share a social trait with their manager (smoking habit or gender) get promoted faster. The causal effect operates through face-to-face social interactions: after gaining a same-type manager, treated employees share significantly more work breaks with that manager, while otherwise similar employees gaining a different-type manager do not change their break-sharing rate. The male-to-male advantage in promotions accumulates to 0.54 pay grades after 2.5 years and is concentrated among employees who work in physical proximity to their manager. Back-of-the-envelope calculations attribute roughly 40 percent of the unconditional gender gap in promotions to this mechanism, a magnitude comparable to estimates of the motherhood penalty from Kleven, Landais, and Søgaard (2019). No effects are found on effort, sales performance, or retention, ruling out productivity differences as the mechanism. The paper is the first to provide causal evidence on the old boys' club hypothesis, extending the correlational evidence of Kunze and Miller (2017) and contributing to the gender pay gap literature of Goldin (2014) and Bertrand, Goldin, and Katz (2010).

## Core results

Magnitudes and significance are as reported. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Smoker-to-smoker **promotion advantage**: smoking employees promoted faster under a smoking manager | Figure 1, panel B, p. 1719 | Double-difference = 0.63 pay grades at 10 quarters (p = 0.035), ~15 percent salary increase |
| R2 | Smoking employees **share significantly more breaks** with a smoking manager | Figure 2, panel A, p. 1721 | +24 pp (38% to 62% of breaks, p = 0.002); no effect for nonsmoking employees |
| R3 | Male employees **promoted faster under a male manager** (double-differences) | Figure 7, panel A, p. 1729 | Male-to-male double-difference = 0.65 pay grades at 10 quarters (p < 0.001) |
| R4 | Male-to-male advantage (dual-double-differences, gain and lose) | Figure 7, panel C, p. 1729 | 0.54 pay grades at 10 quarters (p < 0.001) |
| R5 | Male employees **share significantly more breaks** with a male manager | Figure 2, panel C, p. 1721 | +14.5 pp (46.7% to 61.2%, p = 0.017); no robust effect for female employees |
| R6 | Male-to-male advantage **concentrated in high-proximity** positions | Figure 8, p. 1732 | 0.76 pay grades high-proximity (p < 0.001) vs 0.21 low-proximity (p = 0.178); difference p = 0.013 |
| R7 | Male-to-male advantage **accounts for 40 percent** of the gender gap | Section IV, p. 1734 | 0.54 x 0.66 = 0.36 pay grade reduction; gap falls from 0.90 to 0.54 pay grades |
| R8 | **No effect** on effort (days worked, hours), sales performance, or firm exit | Figure 3, p. 1723 | All coefficients near zero and insignificant (e.g., attrition at 10 quarters = -0.010, p = 0.887) |

**Overall (paper's conclusion).** Manager-employee social interactions generate a durable and economically large promotion advantage for co-typed employee-manager pairs (same smoking status or same gender). The effect builds gradually over two years as more employees cycle through promotion opportunities, is concentrated in positions requiring physical proximity to the manager, and is not accompanied by any gain in measured productivity. At this firm, the male-to-male advantage through social interactions can account for about 40 percent of the observed gender gap in pay grades, comparable in magnitude to the motherhood penalty.

## Theory / model

The paper has no formal economic model. It tests two related hypotheses about face-to-face social interactions and promotions.

**Hypothesis 1 (smoker-to-smoker advantage).** Employees who smoke and gain a smoking manager have more shared smoking breaks, leading the manager to favor them in promotion decisions or to learn more about their effort and potential through increased contact. This advantage should be absent for nonsmoking employees gaining a smoking manager, and larger among employees who work in physical proximity to their manager.

**Hypothesis 2 (male-to-male advantage).** Male employees broadly have more opportunities for social interaction with male managers than female employees do, not limited to smoking breaks. Switching from a female manager to a male manager therefore raises male employees' promotion prospects but not female employees'. The effect should again be larger in high-proximity positions.

The identification logic rests on the quasi-random rotation of managers as part of the firm's standard practice of rotating personnel across teams to give them broad exposure (pp. 1704-1705, pp. 1712-1713). The paper tests this assumption via parallel pre-trends and via a falsification exercise using the reverse transition direction (losing vs. gaining a male manager), which should and does produce mirror-image effects. An affinity-channel falsification using shared demographic traits (same province, same college, or close in age, covering 16, 8, and 43 percent of pairs respectively) finds negligible effects on pay grade at 10 quarters (0.05 pay grades, p = 0.478) compared to the smoker-to-smoker estimate of 0.63, ruling out group identity as the main channel. Evidence from Bandiera, Barankay, and Rasul (2009) on manager social connections in a different workplace setting is cited for context.

## Method

The estimator is a two-way fixed-effects event-study exploiting manager rotation events, introduced in Section IIA (p. 1716). Let $$y_{i,t}$$ be the outcome for employee $$i$$ at month $$t$$. Let $$S_i \in \{0,1\}$$ indicate whether the employee smokes, and let $$J_S = \{N2S, N2N, S2S, S2N\}$$ denote the four types of manager transitions (N2S = nonsmoking-to-smoking manager, etc.). For event-time leads and lags $$\mathcal{E} = \{-30, \ldots, -4, 0, +1, \ldots, +30\}$$ (monthly, aggregated to quarterly for presentation), the baseline specification is (equation 1, p. 1717):

$$
y_{i,t} = \sum_{j \in J_S} \sum_{e \in \mathcal{E}} \beta_{j,e}^S \cdot S_i \cdot D_{i,t+e}^j + \sum_{j \in J_S} \sum_{e \in \mathcal{E}} \beta_{j,e}^N \cdot (1-S_i) \cdot D_{i,t+e}^j + \gamma_i + \eta_{i,t} + \delta_t^S + \delta_t^N + \epsilon_{i,t} \tag{1}
$$

where $$D_{i,t+e}^j = 1$$ if employee $$i$$ experiences a type-$$j$$ manager transition at time $$t+e$$, $$\gamma_i$$ are employee fixed effects, $$\eta_{i,t}$$ are manager fixed effects, and $$\delta_t^S$$, $$\delta_t^N$$ are separate month effects for smokers and nonsmokers. Standard errors cluster two-way by manager and employee. The omitted category is the quarter prior to the transition event.

The key estimands are:

- *Single-difference* for smokers: $$\beta_{N2S,e}^S - \beta_{N2N,e}^S$$, the gain from acquiring a smoking manager vs. acquiring any nonsmoking manager.
- *Double-difference*: $$\left(\beta_{N2S,e}^S - \beta_{N2N,e}^S\right) - \left(\beta_{N2S,e}^N - \beta_{N2N,e}^N\right)$$, the differential effect of a smoking manager on smoking vs. nonsmoking employees (R1).

The male-to-male specification is identical to equation (1) but replaces the smoking indicator $$S_i$$ with a female indicator $$F_i$$ and the transition set $$J_S$$ with $$J_G = \{F2M, F2F, M2F, M2M\}$$ (p. 1727). A symmetric *dual-double-difference* for the male-to-male analysis (R4) averages the double-difference from gaining a male manager (panel A of Figure 7) with the negative of the double-difference from losing a male manager (panel B of Figure 7), using a disjoint set of transition events as a sharp robustness check.

For social interactions, the outcome $$\text{Share}_{i,m}$$ is the fraction of work breaks employee $$i$$ took with manager $$m$$. Since this is measured as a cross-section of employee-manager pairs rather than a monthly panel, the specification collapses to (equation 2, p. 1720):

$$
\text{Share}_{i,m} = \sum_{j \in J_S} \beta_{j,\text{post}}^S S_i D_{i,m}^j + \sum_{j \in J_S} \beta_{j,\text{post}}^N (1-S_i) D_{i,m}^j + \sum_{j \in J_S} \beta_{j,\text{pre}}^S S_i D_{i,m+1}^j + \sum_{j \in J_S} \beta_{j,\text{pre}}^N (1-S_i) D_{i,m+1}^j + \mathbf{X}_{i,m} \boldsymbol{\gamma} + \epsilon_{i,m} \tag{2}
$$

where $$D_{i,m}^j = 1$$ if employee $$i$$ experienced a type-$$j$$ transition from manager $$m-1$$ to $$m$$, and $$D_{i,m+1}^j = 1$$ for the upcoming transition from $$m$$ to $$m+1$$ (used as a placebo pre-trend test). Controls $$\mathbf{X}_{i,m}$$ include unit size, manager pay grade, and position-title dummies.

## Empirical specifications

**Smoker-to-smoker advantage (Section II, R1-R2).** The sample is male employees and male managers with assignable smoking status: 2,907 unique employees, 997 unique managers, 1,798 manager transition events, 94,728 employee-month observations. Outcome for R1: monthly pay grade (range 41-66, January 2015 to December 2018). Outcome for R2: share of breaks with the manager (manager relationship survey, cross-section of employee-manager pairs). The headline estimate is the double-difference at 10 quarters posttransition.

**Male-to-male advantage (Section III, R3-R7).** The full panel covers 14,638 unique employees, 1,269 unique managers, 8,670 transition events, and 380,959 employee-month observations (65 percent female). The specification is identical to equation (1) with gender replacing smoking. The primary headline estimate (R4) uses the dual-double-differences from panel C of Figure 7, averaging gains and losses to control for mean-reversion and maximize precision.

**Physical proximity heterogeneity (R6).** The dual-double-differences model is re-estimated separately for high-proximity and low-proximity subsamples. High-proximity classification uses card-swipe floor-sharing data (headquarters employees, 45 percent of the sample) and survey-reported daily proximity (sales and distribution employees); roughly half of employees fall in each group. The 0.76 vs. 0.21 pay grade contrast (p = 0.013 for the difference) confirms the face-to-face interaction channel.

**Effort, performance, and retention (R8, Figure 3).** The same event-study specification as equation (1) is run with four alternative dependent variables: log days worked (HR absence records), log daily hours worked (card-swipe data, headquarters only), monthly sales revenue index (sales-role employees, normalized to mean 100), and a firm-exit dummy. All post-transition coefficients are close to zero and precisely estimated, with pre-trends also flat.

**Affinity-channel falsification (Section IIE, Figure 4).** The event-study is re-run using transitions in which an employee gains or loses a manager with a shared demographic trait. The posttransition pay grade coefficient at 10 quarters is 0.05 (p = 0.478), well below the 0.63 smoker-to-smoker estimate, ruling out pure demographic affinity as the driver.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Anonymous commercial bank HR records (pay grades, manager assignments, organizational chart, HR absence records) | Primary outcome (monthly pay grade), manager assignment construction, effort and retention outcomes; 2015-2018 | No page yet (proprietary-confidential) |
| Anonymous commercial bank annual health exam (smoking status, 2017) | Smoking status for 59% of employees (59% classified from exam, 41% from supplementary surveys) | No page yet (proprietary-confidential) |
| Two supplementary smoking surveys (February 2018 and December 2017) | Imputed smoking status for employees not in the health exam | No page yet (proprietary-confidential) |
| Manager relationship survey (share of breaks with manager, December 2017) | Social interactions measure; 3,345 of 4,847 invited employees responded (69%) | No page yet (proprietary-confidential) |
| Card-swipe security data (floor occupancy by employee) | Physical proximity classification for headquarters employees (45% of sample) | No page yet (proprietary-confidential) |

Sample: January 2015 to December 2018 (48 months). All data originate from a single anonymous commercial bank in Southeast Asia; no external public sources used in the main analysis. The firm identity is withheld by agreement; the paper refers to it throughout as "the firm."

## When to read the full paper

Use the [original](https://doi.org/10.1257/aer.20210863) if you are: tracing the mechanism in detail (the affinity-channel tests in Section IIE and proximity heterogeneity in Sections IIF and IIIE provide the richest identification evidence); applying the same manager-rotation design to a new organizational dataset (Sections IIB and IIIB lay out the parallel-trends and reverse-transitions validation); assessing how the male-to-male advantage interacts with occupational proximity or cultural norms across different settings (Section V); or running the replication code (data: [doi:10.3886/E182243V1](https://doi.org/10.3886/E182243V1), noting the firm identity remains anonymous).

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(7), July 2023. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. The article is paywalled under AEA standard terms; no CC license was found in Crossref metadata. Replication data are publicly archived at [doi:10.3886/E182243V1](https://doi.org/10.3886/E182243V1).

> Cullen, Zoë, and Ricardo Perez-Truglia. "The Old Boys' Club: Schmoozing and the Gender Gap." *American Economic Review* 113, no. 7 (July 2023): 1703-1740. DOI: 10.1257/aer.20210863. Extracted here under fair use for educational and research purposes; no verbatim reproduction of extended passages.


==============================================================================
# Subjective Performance Evaluation and Influence Activities: de Janvry et al. (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/de-janvry-et-al-subjective-performance-evaluation-influence-2023/
# A randomized field experiment among 3,785 Chinese civil servants shows that revealing the evaluator's identity induces evaluator-specific influence activities, creating a 0.311-point asymmetry in supervisor assessments (0.24 SD) that disappears under a masked scheme. Masking the evaluator's identity improves colleague assessments, supervisor assessments, and objective performance pay. American Economic Review vol. 113(3), 2023, paywalled. 8 core results with source locators, datasets used, the model, and the method. LLM-distilled.
# Tags: paper-summary, bureaucracy, incentives, public-sector, field-experiment
==============================================================================

**What this is.** This page is a distilled skeleton of the paper. Read the original at [https://doi.org/10.1257/aer.20211207](https://doi.org/10.1257/aer.20211207) to replicate or extend.

## TL;DR

De Janvry, He, Sadoulet, Wang, and Zhang (2023) run a randomized field experiment with 3,785 college graduate civil servants ("CGCSs") in two Chinese provinces. The experiment randomizes whether each civil servant learns the identity of her performance evaluator at the start of the evaluation cycle (the "revealed" scheme, mimicking the status quo) or only learns that one of her two supervisors will be randomly selected as evaluator at the end of the year (the "masked" scheme). Under the revealed scheme the evaluating supervisor gives 0.311 higher assessment score points than the nonevaluating supervisor (0.24 SD; DV SD = 1.31): consistent with the CGCS engaging in evaluator-specific influence activities to improve her evaluation outcome, which determines her promotion to a permanent civil service position. Switching to the masked scheme eliminates this assessment asymmetry and improves multiple performance indicators: colleague assessments rise by 0.22 points on a 7-point scale, nonevaluator assessments rise by 0.22 points, and performance-linked monthly wages increase by roughly 2.3 percent. The results show that a low-cost modification of the evaluation scheme can improve bureaucratic work performance.

## Core results

| # | Result | Locator | Magnitude |
|---|--------|---------|-----------|
| R1 | Evaluator gives higher assessment than nonevaluator under revealed scheme | Table 2, col 1, p.781 | 0.311 (SE 0.082); 0.24 SD (paper text p.780; DV SD=1.31) |
| R2 | Evaluator-nonevaluator asymmetry disappears under masked scheme | Table 2, col 2, p.781 | -0.097 (SE 0.121), not significant |
| R3 | Masked scheme increases colleague assessment score | Table 3, Panel A, col 1, p.782 | +0.217 on 1-7 scale (SE 0.035) |
| R4 | Masked scheme increases probability rated top 10% by colleagues | Table 3, Panel A, col 2, p.782 | +7.7 pp (SE 1.3) |
| R5 | Masked scheme increases nonevaluator supervisor assessment | Table 3, Panel B, col 3, p.782 | +0.215 on 1-7 scale (SE 0.059) |
| R6 | Masked scheme increases performance-linked monthly wage | Table 3, Panel C, col 1, p.782 | +48.81 yuan (~2.3%) (SE 22.41) |
| R7 | One-point increase in evaluator score raises promotion probability | Table 4, col 1, p.787 | +7.3 pp (SE 1.1) |
| R8 | Hometown tie with evaluator raises evaluator assessment under revealed scheme only | Table 7, Panel A, col 2, p.791 | +0.189 (SE 0.067); null in masked scheme (-0.067, SE 0.088) |

**Overall.** Evidence from Tables 2-7 consistently supports the existence of evaluator-specific influence activities under the revealed scheme and shows that the masked scheme eliminates such activities while improving actual job performance. The hometown favoritism result (R8) isolates "bottom-up" influence activities from "top-down" evaluator preferences: hometown favoritism appears only under the revealed scheme, where the CGCS knows who the evaluator is and can direct influence efforts accordingly.

## Theory / model

The paper has no full structural model; Section II (pp. 777-779) presents a conceptual framework to rationalize the experimental design and derive testable propositions.

A CGCS allocates effort across three types of activity: $$X$$ (common productive tasks valued by both supervisors), $$x_j$$ (supervisor-$$j$$-specific productive influence activities, i.e., tasks assigned or observed mainly by supervisor $$j$$), and $$u_j$$ (nonproductive influence activities directed at supervisor $$j$$, e.g., personal favors). Following Milgrom and Roberts (1988), $$x_j$$ are "productive influence activities" and $$u_j$$ are "nonproductive influence activities."

The organization's performance measure uses only productive activities (p.778):

$$P = X + x_1 + x_2 \tag{1}$$

Supervisor $$j$$'s subjective assessment score is (p.778):

$$Y_j = \alpha X + x_j + u_j, \quad j = 1, 2 \tag{2}$$

where $$\alpha > 0$$ is the relative weight the supervisor places on common productive activities over supervisor-specific influence activities.

Each CGCS maximizes utility subject to a total time constraint of $$T$$ (p.778):

$$\max_{X,\, x,\, u} V = \alpha X + \sum_{j \in \{1,2\}} s_j (x_j + u_j) - G(X) - g\!\left(\sum_j x_j\right) - h\!\left(\sum_j u_j\right) \tag{3}$$

subject to $$X + \sum_j x_j + \sum_j u_j = T$$, $$X, x_j, u_j \in [0, T]$$,

where $$s_j$$ is the probability supervisor $$j$$'s assessment determines the CGCS's reward ($$\sum_j s_j = 1$$), and $$G$$, $$g$$, $$h$$ are strictly convex cost functions. Under the revealed scheme $$s_1 = 1$$, $$s_2 = 0$$; under the masked scheme $$s_1 = s_2 = 1/2$$.

Two propositions follow from solving the CGCS's maximization problem (pp.778-779):

**Proposition 1.** Under the revealed scheme, the CGCS engages in evaluator-specific influence activities ($$x_j > 0$$, $$u_j > 0$$), and the evaluating supervisor gives a higher assessment ($$Y_j$$) than the nonevaluating supervisor.

**Proposition 2.** Compared to the revealed scheme, the masked scheme increases common productive effort ($$X$$) and improves overall work performance ($$P$$). The masked scheme raises the nonevaluator's assessment unambiguously, but its effect on the evaluator's assessment is ambiguous (the evaluator benefits from more $$X$$ but loses evaluator-specific influence).

## Method

The primary identification strategy is random assignment of CGCSs to the revealed vs. masked evaluation scheme. In collaboration with two Chinese provincial governments in 2017, the authors randomized all 3,785 CGCSs employed in that year across 788 townships (Section I.B-C, pp.773-775). Two-thirds were assigned to the revealed scheme and one-third to the masked scheme. Randomization was conducted at the work-unit level; since 83.9 percent of units had only one CGCS, this is statistically nearly equivalent to individual-level randomization.

Each CGCS reports to a party leader and an administrative leader under China's dual-leadership governance structure (Shirk 1993). One of the two supervisors was randomly selected as evaluator. In the revealed scheme, the CGCS was notified of the evaluator's identity at the start of the evaluation year. In the masked scheme, the CGCS was told only that one supervisor would be randomly selected at year-end. Neither supervisor was informed of the selection. Official government notifications with formal stamps were sent to all CGCSs to establish credibility.

The benchmark performance measure is the average colleague assessment, collected through anonymous surveys of coworkers who have no incentive to inflate or deflate CGCS evaluations (they are not in the CGCS's evaluation chain and do not compete with her for promotion). Performance is further benchmarked against both supervisor assessments and administrative salary records verified by the provincial governments.

The method builds on the principal-agent framework of Baker, Gibbons, and Murphy (1994), operationalized as an RCT in the spirit of Finan, Olken, and Pande (2015) on public employee incentives. Lazear and Oyer (2012) survey the theoretical literature motivating the empirical test. Prendergast and Topel (1996) develop the theoretical foundations of favoritism in organizations under subjective evaluation. Wu (2017) provides a related natural experiment varying authority allocation in Chinese media, complementing this paper's approach of directly cross-randomizing the employee's knowledge of the evaluator's identity.

## Empirical specifications

**Specification 1 (Proposition 1 test).** Using the revealed-scheme subsample, the paper estimates (p.780, eq.1):

$$\text{Sup1\_Edge}_{icst} = \alpha \times \text{Sup1\_Eval}_i + \gamma_c + \lambda_s + \phi_t + \varepsilon_{icst} \tag{4}$$

where $$\text{Sup1\_Edge}_{icst}$$ is Supervisor 1's assessment minus Supervisor 2's assessment for CGCS $$i$$ in county $$c$$, CGCS type $$s$$, cohort $$t$$. $$\text{Sup1\_Eval}_i$$ is a dummy for whether Supervisor 1 is the randomly selected evaluator. $$\gamma_c$$, $$\lambda_s$$, $$\phi_t$$ are county, CGCS-type, and cohort fixed effects. Standard errors are clustered at the work-unit level. Because the evaluator is chosen randomly, $$\alpha$$ causally identifies the additional positiveness of the evaluating supervisor's assessment due to influence activities. The same regression is estimated on the masked-scheme sample (Table 2, cols 1 and 2, p.781) to check that the asymmetry disappears when evaluator identity is withheld.

**Specification 2 (Proposition 2 test).** Using the full sample, the paper estimates (p.782, eq.2):

$$Y_{icst} = \alpha \times \text{Mask}_i + \gamma_c + \lambda_s + \phi_t + \varepsilon_{icst} \tag{5}$$

where $$Y_{icst}$$ is a performance measure (colleague assessment, supervisor assessment, performance pay) and $$\text{Mask}_i$$ is a dummy for being assigned to the masked scheme. Same fixed effects and clustering. Random scheme assignment means $$\alpha$$ identifies the causal effect of masking on performance (Table 3, p.782).

**Balance.** Table 1 (p.776) shows no statistically significant differences in CGCS characteristics (age, gender, college type, major, party membership, CEE score, risk aversion, local birth) across the two schemes; the joint F-statistic is 0.90 (p = 0.54).

**Robustness.** The paper controls for LASSO-selected covariates (online Appendix Tables A7, A13), applies Lee (2009) bounds for non-random attrition (online Appendix Tables A9, A15), and uses an interaction approach on the full sample (online Appendix Table A10). Results are stable across these checks.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---------|--------------|-----------|
| Author-collected CGCS baseline and endline surveys (Sep 2017, Jun 2018) | Primary performance measures: colleague assessments, supervisor assessments, self-assessments, job-task allocation, influence activity proxies | no page yet |
| Chinese provincial government administrative records (2017-2018) | Promotion outcomes (permanent civil service placement) and salary data verified against administrative records | no page yet |

Sample: 3,785 CGCSs ("College Graduate Civil Servants" hired through China's "3+1 Supports" program) in two provinces (Province A coastal, Province B inland), cohorts admitted 2016 and 2017. Endline: 2,854 CGCSs after 24.5 percent attrition, primarily from reassignment between townships (14.9 percent) and voluntary exits to graduate school or civil service exams (7.4 percent). Position types: township government clerks (poverty alleviation and agricultural support), primary school teachers, and township clinic nurses. Randomization at the work-unit level across 788 townships.

## When to read the full paper

Read Section II for formal proofs of the two propositions and model extensions in online Appendices C-E. Read Section III.A (Table 2, p.781) for the evaluator-asymmetry test. Read Section III.B (Table 3, p.782) for the performance-improvement results and Section III.C (Table 4, p.787) for the promotion-weight evidence confirming the stakes are real. Read Section IV for the mechanism analysis: Table 5 (p.789) for productive influence activities (task reallocation toward evaluator-assigned tasks), Table 6 (p.789) for indirect proxies of nonproductive influence activities, and Table 7 (p.791) for hometown favoritism as a test of bottom-up vs. top-down favoritism. Read Section IV.C-D (pp.791-796) for the full battery of robustness checks ruling out evaluator behavioral change and information-quality alternative explanations.

Useful for: researchers studying subjective performance evaluation, influence activities in bureaucracies, and personnel economics of the public sector; practitioners designing evaluation systems in organizations with multiple supervisors or dual-leadership structures.

## Attribution and rights

This paper is published in the *American Economic Review* 113(3), 2023 under AEA standard copyright. No CC license was found in Crossref metadata (checked 2026-06-25). Extract-only.

> de Janvry, Alain, Guojun He, Elisabeth Sadoulet, Shaoda Wang, and Qiong Zhang. "Subjective Performance Evaluation, Influence Activities, and Bureaucratic Work Behavior: Evidence from China." *American Economic Review* 113, no. 3 (March 2023): 766-799. https://doi.org/10.1257/aer.20211207

Replication data: de Janvry et al. (2023). *Replication Data for: Subjective Performance Evaluation, Influence Activities, and Bureaucratic Work Behavior: Evidence from China.* AEA/ICPSR. https://doi.org/10.3886/E182787V1

LLM-distilled by paper-distiller (claude-sonnet-4-6), 2026-06-25. Not human-verified. Not reproduced.


==============================================================================
# Evidence and Lessons on Health Impacts of Public Health Funding: Dillender (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/dillender-evidence-lessons-health-impacts-2023/
# Distilled: Exploiting staggered variation in Ryan White CARE Act Title I eligibility, this paper finds that federal HIV/AIDS funding to US cities reduced HIV/AIDS death rates by 15-17 percent, saved approximately 57,000 lives through 2018 at a cost of $334,000 per death avoided, and reduced HIV prevalence by 36-40 percent. American Economic Review 2023, open (AEA). Six core results with source locators, datasets used, identification strategy, and the estimating equations. LLM-distilled; not human-verified.
# Tags: paper-summary, public-health, hiv-aids, federal-funding, place-based-policy
==============================================================================

**What this is.** The paper's core results, the epidemiological model of HIV dynamics, and the research designs (difference-in-differences and regression discontinuity) with their estimating equations: enough to understand what the paper found and how it identified the causal effect of federal HIV/AIDS funding. To replicate or extend the results, read the full source at the [original](https://doi.org/10.1257/aer.20220089).

## TL;DR

This paper estimates the health impact of Ryan White CARE Act Title I funds, which are federal grants directed to US cities to help low-income HIV-positive people access treatment and support services. Identification exploits two quasi-experimental sources of variation. First, the original 1990 Ryan White legislation granted cities Title I status after they reported at least 2,000 cumulative AIDS cases by March 31 of a given year, while a 1996 rule change (combined with a grandfather clause) froze the set of eligible cities just as effective antiretroviral treatment emerged, generating staggered treatment timing across cities with similar baseline HIV/AIDS trajectories. Second, the sharp discontinuity in Title I funding at the 2,000-case threshold supports a regression discontinuity design to estimate effects on HIV prevalence.

Comparing 25 cities that qualified for Title I status under the original rules with the 25 cities that had the most AIDS cases but fell just below the threshold, the paper finds that Title I status reduced annual HIV/AIDS death rates by about 15-17 percent on average. Annual AIDS case rates fell by roughly 20-25 percent. A regression discontinuity at the 2,000-case threshold finds 36-40 percent fewer people living with HIV in Title I cities by 2008, indicating the funding reduced HIV spread as well as deaths. The implied cost per HIV/AIDS death avoided is $334,000, and the program's benefit-cost ratio is approximately 30 at a $10 million value of statistical life. Total lives saved are estimated at approximately 57,000 through 2018.

## Core results

Magnitudes and significance are as reported. Locators point to the source PDF (American Economic Review 113(7): 1825-1887).

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Title I status reduced HIV/AIDS death rates (1988-2006) | Table 2, col 1, p. 1847 | DiD coefficient: -0.185 log points (SE 0.069, p=0.010); 50 cities, 950 city-year obs |
| R2 | Effect on HIV/AIDS death rates through 2018 | Table 2, col 2, p. 1847 | DiD coefficient: -0.163 log points (SE 0.075, p=0.036); 50 cities, 1,550 obs |
| R3 | Title I reduced annual rates of new AIDS cases | Table 8, col 1, p. 1870 | DiD coefficient: -0.227 log points (SE 0.063, p=0.001); ~25% reduction |
| R4 | Title I reduced people living with HIV (RDD) | Table 9, Panel A, col 1, p. 1875 | RDD: -0.510 log points (SE 0.152, p=0.002); -40.0% in 2008 |
| R5 | Title I reduced new HIV diagnoses (RDD) | Table 9, Panel B, col 1, p. 1875 | RDD: -0.628 log points (SE 0.219, p=0.006); -46.7% in 2008 |
| R6 | Cost per HIV/AIDS death avoided; total lives saved | Table 6, p. 1860 | $334,000 per death avoided; 9,421 lives in sample; ~57,000 total through 2018; BCR = 30 |

**Overall (paper's conclusion).** Federal HIV/AIDS funding allocated to cities through Ryan White Title I had large health impacts: reducing HIV/AIDS deaths, new AIDS cases, and HIV prevalence. The funding disparities that emerged from the 1996 Ryan White reauthorization rules, which effectively froze Title I eligibility just as effective treatment arrived, are responsible for a large share of the divergent HIV/AIDS trajectories across US cities. The cost per life saved is low relative to other health programs, which the paper attributes to Ryan White targeting a vulnerable population with a deadly infectious disease for which effective treatment exists.

## Theory / model

The paper has no formal economic theory model. The theoretical content consists of the identification framework and an epidemiological susceptible-infected-removed (SIR) model used to analyze HIV transmission dynamics under Title I funding.

**HIV dynamics (SIR model, pp. 1871-1872).** Let S, I, and R denote the susceptible, infected, and removed (deceased) populations in a city. Absent Title I, HIV-positive individuals transmit at rate trans and die from HIV/AIDS at rate death. Title I changes transmission by factor p and death rates by factor q (q < 0 since Title I reduces deaths). The annual transitions are:

$$
S_{t+1} = S_t - (1+p)\times \text{trans}\times I_t\times S_t \tag{SIR-S}
$$

$$
I_{t+1} = I_t + (1+p)\times \text{trans}\times I_t\times S_t - (1+q)\times \text{death}\times I_t \tag{SIR-I}
$$

$$
R_{t+1} = (1+q)\times \text{death}\times I_t \tag{SIR-R}
$$

The number of people living with HIV after t periods of Title I status is (equation 3, p. 1871):

$$
I_t = I_0\times\prod_{j=1}^{t}\left[1 + (1+p)\times\text{trans}\times S\!\left(t;\,I_0,S_0,\text{trans},p,\text{death},q\right) - (1+q)\times\text{death}\right] \tag{3}
$$

Taking the log difference between Title I and non-Title I cities gives the main identifying expression (equation 4, p. 1872):

$$
\gamma_t = \log\!\left(I_t^{\text{Title1}}\right) - \log\!\left(I_t^{\text{NoTitle1}}\right)
= \sum_{j=1}^{t}\log\!\left\{\frac{1 + (1+p)\times\text{trans}\times S(\cdot;\,p,q) - (1+q)\times\text{death}}{1 + \text{trans}\times S(\cdot;\,p{=}0,q{=}0) - \text{death}}\right\} \tag{4}
$$

Because Title I reduces death rates (q < 0), and because the paper documents a reduction in people living with HIV in 2008 (Table 9, R4), equation (4) implies that p < 0 as well: Title I must have reduced HIV transmission rates, not only death rates. This rules out the scenario proposed by Lakdawalla, Sood, and Goldman (2006), in which providing treatment to HIV-positive people increases HIV spread through behavioral responses.

**Identification logic.** The identifying assumption is parallel trends: absent Title I, cities that qualified under the original rules would have trended similarly in HIV/AIDS outcomes to cities that fell just below the 2,000-case threshold. Event-study plots (Figure 3) confirm that treatment and control cities tracked each other in log HIV/AIDS death rates before treatment cities gained Title I status. Robustness to Callaway and Sant'Anna (2021) reweighting methods, matching on 1995 AIDS rates or population, and state-by-year fixed effects (Table 3) supports the identifying assumption.

## Method

The main estimating strategy is a staggered difference-in-differences using variation arising from three features of the Ryan White CARE Act. First, the original 1990 legislation granted Title I status to any city reporting at least 2,000 cumulative AIDS cases to the CDC by March 31 of a given year. Second, Title I status, once obtained, was not lost even if a city's AIDS burden fell below the threshold. Third, a 1996 reauthorization changed eligibility from a cumulative to a five-year rolling count, but included a grandfather clause allowing cities that had qualified by March 31, 1995 to retain Title I status regardless. Since effective antiretroviral treatment emerged in 1996, the combined rule change effectively froze Title I eligibility for the next decade, creating persistent large funding differences between cities just above and just below the 2,000-case threshold (treatment cities averaged $68.9 million in Title I funds from 1996 to 2006; control cities averaged $3.9 million).

This builds on `difference-in-differences` for the main estimates (equation 1), `panel-regression` with two-way fixed effects for the within-city-year estimator, `matching` for robustness checks that pair each treated city with control cities having similar baseline AIDS rates or AIDS trends (equation 2), and `regression-discontinuity-design` for the HIV stock and transmission analysis (equation 5).

For the regression discontinuity, the running variable is the log of cumulative AIDS cases by March 31, 1995. The 2,000-case threshold is credible because cities could not manipulate their AIDS case counts ex ante (a McCrary density test fails to reject smoothness at the cutoff; p-value 0.37), and the significance of crossing 2,000 cases by March 31, 1995 only became clear after the 1996 rule change and treatment emergence.

## Empirical specifications

**Main DiD specification (equation 1, p. 1839).**

$$
y_{jt} = \gamma_j + \delta_t + \mathbf{X}_{jt}\alpha_t + \text{Title1}_{jt}\,\beta + \varepsilon_{jt} \tag{1}
$$

where j indexes cities, t indexes years; $$y_{jt}$$ is the log of HIV/AIDS deaths per 100,000 people (or log AIDS cases, or other health outcomes); $$\gamma_j$$ are city fixed effects; $$\delta_t$$ are fiscal-year fixed effects; $$\mathbf{X}_{jt}$$ is a vector of demographic controls (shares male, younger than 18, older than 64, Black, Hispanic) with coefficients $$\alpha_t$$ allowed to vary by year; and $$\text{Title1}_{jt}$$ is an indicator equal to one for city j having qualified for Title I status under the original Ryan White rules by year t. Standard errors are clustered by city. The coefficient $$\beta$$ is the average causal effect of Title I status on the outcome.

The baseline sample is 50 cities (25 treatment, 25 control), yielding 950 observations for the 1988-2006 window and 1,550 for 1988-2018. All city-year observations for death rates come from restricted-use Vital Statistics data. Alternative samples expanding to all AIDS Public Information Dataset cities confirm results (Table 3, cols 9-10).

**Matching robustness (equation 2, p. 1856).** For each treated city, control cities are selected by nearest-neighbor matching on 1995 AIDS rates per 100,000, 1995 population, or 1990-to-1991 changes in HIV/AIDS death rates or AIDS cases, creating matched groups g:

$$
y_{gjt} = \gamma_j + \delta_{gt} + \mathbf{X}_{jt}\alpha_t + \text{Title1}_{jt}\,\beta + \varepsilon_{gjt} \tag{2}
$$

where $$\delta_{gt}$$ are group-by-year fixed effects; identification comes entirely from within-matched-group variation in Title I status. Results are robust across all four matching approaches (Table 5).

**HIV stock regression discontinuity (equation 5, p. 1874).** For the cross-sectional analysis of 2008 HIV outcomes:

$$
\log(\text{Num\_HIV}_{j,2008}) = \lambda + f(\text{AIDS\_Cases}_{j,1995}) + \text{Title1}_j\,\gamma + \eta_j \tag{5}
$$

where f is a linear polynomial in the log of AIDS cases ever reported by March 31, 1995, fit separately on each side of the 2,000-case cutoff; $$\text{Title1}_j$$ is an indicator for cities above the threshold; and robust standard errors are used. The baseline specification uses 46 cities (all main-sample cities with nonmissing 2008 HIV data); local linear regression with the Calonico, Cattaneo, and Titiunik (2014) optimal bandwidth uses 15 cities on each side (Table 9).

**Decomposition of Title I's effect on new HIV transmissions (equation 6, p. 1875).** The effect on 2008 new diagnoses decomposes as:

$$
\tau = \log\!\left(\text{Num\_Trans}_{2008}^{\text{Title1}}\right) - \log\!\left(\text{Num\_Trans}_{2008}^{\text{NoTitle1}}\right)
= \underbrace{\log(1+p)}_{1} + \underbrace{\hat{\gamma}}_{2} + \underbrace{\log\!\left[\frac{S(t{=}2008;\ldots;\,p,q)}{S(t{=}2008;\ldots;\,p{=}0,q{=}0)}\right]}_{3} \tag{6}
$$

Term 1 is the direct effect on HIV transmissibility; term 2 is the estimated reduction in HIV-positive people (from Table 9, Panel A); term 3 is the offsetting effect of a larger susceptible population (fewer past infections means more people at risk). Estimates from Table 9 indicate that 81 to 85 percent of the reduction in new diagnoses in 2008 is accounted for by term 2 alone (fewer people living with HIV), with term 3 partially offsetting.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Vital Statistics Multiple Cause of Death Files (restricted-use) | Annual HIV/AIDS death rates per 100,000 people for all US civilians, 1988-2018; primary outcome (R1, R2) | No page yet |
| AIDS Public Information Dataset (CDC) | Annual cumulative and annual AIDS cases by city, 1988-2002; used to determine Title I eligibility and as an outcome (R3) | No page yet |
| CDC HIV surveillance data (special request) | City-level HIV diagnoses and HIV prevalence in 2008 (46 cities); used for RDD analysis (R4, R5) | No page yet |
| SEER population data | Annual city populations and demographic denominators, 1988-2018; used to convert counts to rates | No page yet |
| Ryan White Title I funding (assembled) | Annual city-level Title I allocations 1991-2018, assembled from GAO reports, HRSA releases, and federal grant databases; used to construct funding treatment variable and cost-per-life estimate (R6) | No page yet |

Sample: 50 US cities, annual data 1988-2018 (1,550 city-year observations for main sample). Treatment cities received on average $68.9 million in Title I funds from 1996 to 2006; control cities received on average $3.9 million over the same period (Table 1, p. 1844).

## When to read the full paper

Read the [original](https://doi.org/10.1257/aer.20220089) if you are: estimating causal effects of place-based health funding programs; studying the determinants of HIV/AIDS disparities across US cities; evaluating the efficiency of federal public health spending relative to Medicaid or community health centers (Bailey and Goodman-Bacon (2015) comparison, pp. 1861-1862); or assessing whether "treatment as prevention" works in a real-world public health setting. The regression discontinuity design for HIV stock (Table 9 and Figure 12, pp. 1872-1875) provides especially clean quasi-experimental evidence on spillover effects of HIV treatment on HIV transmission, directly bearing on the debate opened by Miller, Johnson, and Wherry (2021) and others on optimal public health targeting.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(7), July 2023. This distillation was extracted by an LLM (claude-sonnet-4-6) on 2026-06-25 and is **not human-verified or independently reproduced**. Replication data are available at [openICPSR doi:10.3886/E184821V1](https://doi.org/10.3886/E184821V1). The paper is freely accessible at [pubs.aeaweb.org](https://doi.org/10.1257/aer.20220089) under AEA standard terms (no CC; redistribution restricted to extract-only).

> Dillender, Marcus. "Evidence and Lessons on the Health Impacts of Public Health Funding from the Fight against HIV/AIDS." *American Economic Review* 113, no. 7 (July 2023): 1825-1887. DOI: 10.1257/aer.20220089.


==============================================================================
# Persuasion through Slanted Language: Djourelova (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/djourelova-persuasion-slanted-language-evidence-2023/
# Distilled: Djourelova (2023) exploits the AP's April 2013 ban on the term "illegal immigrant" to show that slanted language diffuses from news wires to local media and causally lowers public support for restrictive immigration policies. American Economic Review 113(3), 2023, AEA open access. Eight core results with source locators, datasets used, identification design, and estimating equations.
# Tags: paper-summary, political-economy, media-economics, immigration
==============================================================================

**What this is.** The core results, identification design, and estimating equations from this paper on slanted language and media persuasion: enough to know what it found and how, without reading all 36 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1257/aer.20211537).

## TL;DR

The paper studies whether slanted language in news media causally affects readers' policy views, using the Associated Press (AP) ban on the term "illegal immigrant" in April 2013 as a natural experiment. The ban exploits supply-side variation: the AP distributes a common feed to member outlets, and outlets differ in how heavily they rely on AP copy (their AP intensity). The paper has three parts: (i) the ban caused near-complete disappearance of "illegal immigrant" from AP dispatches, while leaving other dimensions of immigration coverage unchanged; (ii) the ban diffused into local media, with 1-standard-deviation higher AP intensity associated with a 3.5 percentage point decline in use of the term; and (iii) readers of locally circulated newspapers in high-AP-intensity counties showed significantly lower support for restrictive immigration policies after the ban, with effects concentrated among moderates and those in areas with few immigrants. The results are specific to immigration and do not generalize to other policy issues, consistent with a persuasion rather than a social-signaling mechanism.

## Core results

Magnitudes and significance as reported; `\*\*`/`\*\*\*` = 5%/1%. Locators point to the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Ban diffused to local media**: 1-SD higher AP intensity associated with 3.5 p.p. decline in use of "illegal immigrant" by media outlets | Table 2, col 1, p. 814 | Coef. on PostBan x IHS(AP-intensity) = -1.613 (0.171)\*\*\*; effect for 1-SD = -3.53 p.p. (17% of mean 20.28); robust across 7 specifications including outlet-specific trends and DMA x year-month FEs |
| R2 | **Diffusion driven by AP-sourced content**, not original reporting; effect on original articles small and only marginally significant for print newspapers | Figure 5, Panel C, p. 815 | AP-sourced: large, sharp post-ban decline; original articles: near zero; visual evidence complemented by online Appendix Table B2 |
| R3 | **Left-leaning outlets show ~2x larger diffusion** than right-leaning ones; significant in 3 of 4 ideology quartiles | Figure 6, Panel A, pp. 817-818 | Most left-leaning (Q1): coefficient approximately twice as large as most right-leaning (Q4); declines significant for Q1-Q3; effect for Q4 near zero and insignificant |
| R4 | **Ban reduced support for border security** by 0.8 p.p. per 1-SD AP intensity increase (intention-to-treat) | Table 4, col 1, p. 824 | Reduced-form coef = -0.0051 (0.002)\*\*; effect for 1-SD = -0.0083; mean dep. var. = 0.55; robust to county-specific trends, DMA x year FEs, and state x year FEs (cols 2-5) |
| R5 | **LATE (2SLS)**: 10 p.p. increase in "illegal immigrant" usage in local media associated with 4.5 p.p. more support for border security | Table 4, 2SLS panel, p. 824 | 2SLS coef = 0.0057 (0.002)\*\*\*; first-stage F = 10.53; persuasion rate 1.9-4.4% for newspaper readers (equation 8, p. 830) |
| R6 | **Effects specific to immigration**: null effect on index of 9 non-immigration CCES policy questions (falsification test) | Table 6, col 8, p. 828 | Reduced-form coef = -0.0002 (0.001); 2SLS coef = 0.0002 (0.001); both statistically zero; covers abortion, gay marriage, healthcare, and economic policy questions |
| R7 | **Stronger effects for moderates** than liberals or conservatives; differences statistically marginal | Table 7, cols 4-6, p. 829 | Moderates: -0.0071 (0.003)\*\*; liberals: -0.0029 (0.003); conservatives: -0.0014 (0.002); p-value for equality of moderate vs extreme groups = 0.224 |
| R8 | **Stronger effects in counties with fewer immigrants** (low share of foreign-born) | Table 7, cols 7-8, p. 829 | Low-foreign-born (bottom quartile): -0.0087 (0.003)\*\*\*; high-foreign-born (top quartile): -0.0001 (0.004); p-value for equality = 0.054 |

**Overall (paper's conclusion).** The AP ban on "illegal immigrant" propagated into local media language and shifted public opinion on immigration policy. The pattern of heterogeneity (pronounced for moderates, larger in low-immigrant areas) is consistent with a persuasion mechanism in which readers with weaker priors on the issue are more susceptible to media framing of immigration. Effects do not transfer to other policies, ruling out a general leftward drift in political views.

## Theory / model

The paper has no structural economic model. The empirical strategy rests on two design features and a tested set of hypotheses.

**Hypotheses tested.** The AP functions as a centralized editor of news language: its guidelines affect a shared input in the editorial production of thousands of member outlets. The ban provides an exogenous variation in that input, orthogonal to outlet-specific demand, for the following reasons (pp. 801-802): (i) the AP is a cooperative that balances the interests of politically diverse member outlets, so the decision to ban reflects organizational consistency policy rather than ideological pressure; (ii) the ban did not coincide with any change in immigration volume, slant (excluding the banned term), or sentiment in AP dispatches; (iii) Reuters, a competing news wire that did not change its guidelines, serves as a placebo.

The tested hypotheses are:

1. **Diffusion**: Outlets with higher pre-ban reliance on AP copy (AP intensity) will show larger post-ban declines in use of "illegal immigrant" (Sections I-II).
2. **Persuasion**: Exposure to the reduced use of the term through locally circulated newspapers will lower support for restrictive immigration policies (Section III).
3. **Mechanism specificity**: Effects should be specific to immigration (not other policies), and stronger for politically persuadable readers (moderates, low-prior areas), consistent with a persuasion rather than social-signaling mechanism (Section III.E).

**Identification assumption.** Support for restrictive immigration policies in counties with high versus low AP intensity would have followed parallel trends in the absence of the ban. The identifying variation is the county-level AP intensity of locally circulated newspapers interacted with the post-ban indicator, a shift-share design where PostBan is the aggregate shock and IHS(APintensity) is local exposure (p. 823).

## Method

The analysis uses three estimators, all built on the DiD principle of comparing high-AP-intensity units to low-AP-intensity units before and after the ban.

**Diffusion (Section II).** The approach uses `difference-in-differences` with continuous treatment, estimating equation (2) (p. 813) at the outlet x month level:

$$
\frac{\text{Illimm}}{\text{Imm}}_{mt} = \alpha_m + \beta_t + \rho \cdot \text{IHS}(\text{APintensity}_m) \times \text{PostBan}_t + \varepsilon_{mt} \tag{2}
$$

where $$\text{Illimm}/\text{Imm}_{mt}$$ is the share of articles in outlet $$m$$ and month $$t$$ using "illegal immigrant" relative to "immigrant"; $$\text{IHS}(\text{APintensity}_m)$$ is the inverse hyperbolic sine of AP intensity (share of AP-sourced articles, measured in the 12 months before the ban); $$\text{PostBan}_t$$ is a dummy for April 2013 onward; $$\alpha_m$$ and $$\beta_t$$ are outlet and year-month fixed effects; standard errors are clustered at the outlet level. Observations are weighted by the number of articles using "immigrant." The parameter $$\rho$$ estimates how much a 1-unit increase in IHS-transformed AP intensity shifts the share after the ban.

**County-level AP intensity and outcome aggregation.** Newspapers' AP intensity is aggregated to the county level as the circulation-weighted average across newspapers serving county $$c$$ (equation 3, p. 820):

$$
\text{APintensity}_c = \frac{\sum_m \left( \text{circ}_{mc} \times \text{APintensity}_m \right)}{\sum_m \text{circ}_{mc}} \tag{3}
$$

The circulation-weighted share of "illegal immigrant" articles in county $$c$$ and year $$y$$ is analogously defined as equation (4) (p. 820).

**Reduced-form effect on views (Section III, equation 5).** The intention-to-treat effect on immigration policy views $$X_{icy}$$ for respondent $$i$$ in county $$c$$ and survey year $$y$$ is estimated as:

$$
X_{icy} = \alpha_c + \beta_y + \rho \cdot \text{IHS}(\text{APintensity}_c) \times \text{PostBan}_y + \varepsilon_{icy} \tag{5}
$$

where $$\alpha_c$$ and $$\beta_y$$ are county and survey-year fixed effects; standard errors are clustered by county. Respondent characteristics (age, gender, education, race, immigration status, household income) are included as controls (p. 822).

**Local average treatment effect (2SLS, equations 6-7).** To estimate the LATE for readers of newspapers that changed their language solely because of the AP input, the paper uses a shift-share IV strategy. The second stage is (p. 822):

$$
X_{icy} = \mu_c + \nu_y + \gamma \cdot \widehat{\text{Illimm}/\text{Imm}_{cy}} + \eta_{icy} \tag{6}
$$

instrumented by the first stage:

$$
\text{Illimm}/\text{Imm}_{cy} = \alpha_c + \rho_y + \gamma \cdot \text{IHS}(\text{APintensity}_c) \times \text{PostBan}_y + \varepsilon_{cy} \tag{7}
$$

The excluded instrument is the interaction of county AP intensity with PostBan. The first stage at the county x year level replicates the diffusion result: 1-SD higher AP intensity is associated with an 8% reduction in "illegal immigrant" use after the ban (p. 824).

**Persuasion rate.** Following Gentzkow and Shapiro (2010) and DellaVigna and Gentzkow (2010), the paper converts magnitudes into a persuasion rate: the share of exposed readers who changed their survey response because of the treatment (equation 8, p. 830):

$$
f = \frac{db}{de} \cdot \frac{1}{1 - b_0} \tag{8}
$$

where $$b$$ is support for restricting immigration, $$e$$ is exposure to "illegal immigrant," and $$b_0$$ is the share who would oppose restrictive policy absent the treatment. Applied to the full-sample ITT and average newspaper readership, this yields a persuasion rate of 1.9-4.4%. For comparison, DellaVigna and Kaplan (2007) estimate a 12% persuasion rate for access to Fox News, and Chiang and Knight (2011) estimate 2% for expected and 6.5% for surprising newspaper endorsements. The estimates here are in the lower range, consistent with the milder nature of a language-change treatment.

## Empirical specifications

All regressions use CCES data for 192,635 respondents across survey waves 2009-2017 (main sample) with additional 2007 and 2018-2020 waves for long-run trends.

**Diffusion regressions (R1-R3, Table 2 and Figure 5-6).** Equation (2), estimated at the outlet x month level. The full sample has 2,385 outlets and 139,523 outlet-month observations. The subsample of daily print newspapers has 853 outlets. Columns in Table 2 (p. 814) progressively add: state x year-month FEs (col 2), outlet-specific linear trends (col 3), DMA x year-month FEs (col 6), and linear trends combined with DMA FEs (col 7). The coefficient on PostBan x IHS(AP-intensity) ranges from -1.3 to -1.8 across all seven columns. The key channel test (Figure 5, Panel C) splits the dependent variable into AP-sourced versus original articles; the sharp post-ban decline appears only for AP-sourced content.

**Intention-to-treat regressions on immigration views (R4, Table 4).** Equation (5), estimated at the respondent level. Main outcome is support for increasing border security (mean = 0.55). Columns in Table 4 add county characteristics x year (col 2), state x year FEs (col 3), DMA x year FEs (col 4), and county-specific linear trends (col 5). The reduced-form coefficient is stable at -0.004 to -0.006, corresponding to 0.77-0.92 p.p. per 1-SD AP intensity increase. The 2SLS estimates in Table 4 (bottom panel) yield a coefficient of 0.0045-0.0071 on the share of "illegal immigrant" articles (LATE), with first-stage F-statistics ranging from 10.5 to 29.8.

**Other immigration policy outcomes (R6, Table 6).** Equation (5) replicated for all five immigration policy questions available in pre- and post-ban waves, and for two aggregate indexes. Four of five individual questions show significant effects (border security, employer sanctions, police questioning, prohibiting services); legalization support shows no significant effect. The falsification test replicates the specification with a nine-item non-immigration policy index (column 8): coefficient -0.0002 (0.001) in the reduced form, 0.0002 (0.001) in 2SLS, both indistinguishable from zero.

**Heterogeneity regressions (R7-R8, Table 7).** Equation (5) estimated within subsamples split by self-reported newspaper readership (did not read, read any, read print), respondent ideology quartile (liberal/moderate/conservative), and county foreign-born population quartile. All specifications include respondent controls, baseline county characteristics x year, and county and survey-year FEs.

**Reuters placebo.** Tables 3 and 5 replace AP intensity with Reuters intensity (share of Reuters-credited articles per 10,000). Reuters did not change its guidelines on "illegal immigrant." The PostBan x IHS(Reuters-intensity) coefficient is 0.292 (0.311) in the diffusion regression (Table 3, col 4) and -0.0019 (0.0038) in the views regression (Table 5, col 4), both statistically zero, supporting the identification assumption.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Factiva (Dow Jones) | AP dispatches, June 2009-June 2017: ~28,000 dispatches using "immigrant" (8,000 using "illegal immigrant"); date, headline, full text | No page yet |
| NewsLibrary (NewsBank) | Media content: ~2,566 print and online outlets, July 2009-July 2017; article-level keyword search for "immigrant" and "illegal immigrant" | [Historical newspaper archives](/wiki/commercial/newsbank-newspaper-archive/) (licensed) |
| ProQuest (Newsstream) | 125 major newspapers missing from NewsLibrary, 2009-2017; complements NewsLibrary for diffusion analysis | [Historical newspaper archives](/wiki/commercial/newsbank-newspaper-archive/) (licensed) |
| CCES (Cooperative Congressional Election Study) | Immigration policy views (5 questions) and other policy preferences across pre- and post-ban survey waves, 2006-2020; ~50,000-60,000 respondents per election-year wave | No page yet |
| Alliance for Audited Media (AAM) | Circulation by newspaper and ZIP code (Fall 2012 GEO/Circ report); used to aggregate AP intensity and outcomes to county level | No page yet |
| Gentzkow-Shapiro ideology index | Newspaper ideological lean for ~370 newspapers (based on congressional speech phrase scores, 2005 data); used for heterogeneity by outlet ideology | No page yet |
| American Community Survey | County-level demographics (share college educated, share foreign-born, log income per capita), 2012 five-year estimates | [ACS](/wiki/datasets/acs/) |
| Editor and Publisher yearbook | Universe of daily print newspapers; baseline circulation for low-circulation outlets not in AAM | No page yet |

Sample: AP dispatches June 2009 to June 2017. Media content July 2009 to July 2017. CCES survey waves 2006-2020 (election years 2008-2016 as main sample; 2018 and 2020 for long-run trends). County-level analysis covers ~2,300 counties matched to ~800 daily newspapers.

## When to read the full paper

Use the [original](https://doi.org/10.1257/aer.20211537) if you are: designing a research project on media language and political attitudes; replicating the diffusion or persuasion estimates (the Internet Appendix contains the plagiarism-detection algorithm, AP-intensity distribution, and additional robustness tables); studying the AP as a supply-side source of variation in media content; or comparing persuasion rates across media interventions (Tables 4-5 contain all the inputs for equation 8). The locators above point directly to the key tables and figures.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(3). This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. No Creative Commons licence is asserted; the PDF is freely readable at aeaweb.org under AEA copyright. Redistribution of this distillation is extract-only.

> Djourelova, Milena. "Persuasion through Slanted Language: Evidence from the Media Coverage of Immigration." *American Economic Review* 113, no. 3 (March 2023): 800-835. DOI: 10.1257/aer.20211537.


==============================================================================
# Optimal Policy under Dollar Pricing: Egorov & Mukhin (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/egorov-mukhin-optimal-policy-dollar-pricing-2023/
# Distilled: In a generalized sticky-price open economy model with dollar currency pricing, targeting domestic inflation is robustly optimal for non-US central banks, capital controls cannot improve welfare unilaterally, and US monetary policy deviates from domestic price stabilization to manipulate global demand. American Economic Review 113(7) 2023, paywalled. Eight core results with source locators, model equations (open-economy DGE with DCP), and the planner Lagrangian method.
# Tags: paper-summary, macro, monetary-policy, exchange-rates, international-finance
==============================================================================

**What this is.** The paper's key propositions, model equations, and welfare results, extracted from the source. To replicate or extend, read the original at [doi.org/10.1257/aer.20200636](https://doi.org/10.1257/aer.20200636).

## TL;DR

Egorov and Mukhin (2023) characterize optimal monetary, macroprudential, and trade policies in a general open-economy model where all export and import prices are invoiced in US dollars (dollar currency pricing, DCP). Their central finding is that targeting domestic producer price stability is robustly optimal for non-US central banks, even though this policy cannot implement the first-best allocation. The mechanism is that DCP makes it impossible for monetary policy to close both the local wedge (domestic inefficiency) and the external wedge (export sector inefficiency) simultaneously, but the external wedge is constrained efficient because individual exporters cannot affect aggregate world demand. Capital controls cannot improve welfare unilaterally because they cannot influence this external wedge either. US monetary policy, by contrast, deviates from domestic inflation targeting: the dominance of the dollar gives the Fed leverage over global demand for dollar-invoiced goods and over international asset prices, creating incentives to overstimulate exports and lower borrowing costs. International cooperation improves global welfare but requires the United States to sacrifice domestic objectives, making it not in the self-interest of the United States.

## Core results

Propositions R1-R5 are analytical; R6-R8 are from the numerical simulation (Section IV). Magnitudes from Table 1 are consumption-equivalent welfare losses in percent.

| # | Result | Locator | Magnitude as reported |
|---|---|---|---|
| R1 | **Proposition 1 (Non-US Monetary Policy):** Targeting domestic producer prices is robustly optimal for any non-US economy under DCP and is time consistent | §II.A, Prop. 1, p. 1795; eq. (11) | pi\_iit = 1 for all parameter values, asset-market structures, preference forms, and shock distributions |
| R2 | **Proposition 2 (Capital Controls):** Given the optimal monetary policy, private risk sharing is constrained efficient and no capital controls are used by the planner | §II.D, Prop. 2, p. 1806 | tau\^h\_it = 0 for all internationally traded assets; holds under general DCP even though monetary policy cannot implement the first-best |
| R3 | **Proposition 3 (Trade Policy):** A revenue-neutral mix of monetary policy, export tax, and exporter subsidy implements the first-best allocation | §II.E, Prop. 3, p. 1807 | Monetary policy targets domestic prices; export tax tau\^E\_it stabilizes destination prices in domestic currency (tau\^E\_it times E\_it = 1); subsidy tau\^R\_it stabilizes exporters' dollar prices |
| R4 | **Proposition 4 (US Policy Rule):** The optimal US monetary policy deviates from inflation targeting and involves three motives: external wedge, export price adjustment costs, and asset-price manipulation | §III.A, Prop. 4, pp. 1809-1810; eq. (12) | US rule (eq. 12) has three terms: the zero-markup external wedge, a cost-of-price-adjustment term, and a current-account weighted asset-price elasticity term; the policy is not time consistent |
| R5 | **Proposition 5 (Cooperative Policy):** Under cooperation, US monetary policy stabilizes the global external wedge; non-US monetary policy continues to target domestic prices; capital controls are generically nonzero | §III.B, Prop. 5, pp. 1812-1813; eq. (13) | US targets the global demand-weighted external wedge equal to zero; cooperative capital controls fight the aggregate demand externality that is absent in the noncooperative case |
| R6 | **Approximation quality of inflation targeting:** Under the optimal non-US policy, inflation volatility is 1-3 orders of magnitude smaller than output gap volatility, confirming inflation targeting is a near-exact approximation even when export prices are inefficient | Figure 2, p. 1816 | Ratio std(pi\_iit)/std(y\_iit) ranges from 0.001 to 0.04 across openness (gamma) and demand-elasticity (epsilon) values; the ratio converges to zero as gamma approaches 0 |
| R7 | **Welfare cost comparison (non-US):** The optimal non-US policy reduces total welfare loss to 3.05% vs. 5.18% for output gap targeting; local financial shocks account for most of the difference (2.92% vs. 3.40%) | Table 1, p. 1819 | Non-US optimal (col 1): 3.05% total; output gap targeting (col 2): 5.18% total; financial-shock component: 2.92% vs. 3.40% consumption equivalents |
| R8 | **DCP welfare asymmetry (US gains, non-US loses relative to PCP):** Under optimal policies, the United States gains 0.34 ppts and other economies lose 0.12 ppts in consumption equivalents relative to the PCP benchmark | Table 1 discussion, p. 1820 | US: 2.59% loss (DCP optimal) vs. 2.93% (PCP), net gain = 0.34 ppts; non-US: 3.05% (DCP optimal) vs. 2.93% (PCP), net loss = 0.12 ppts |

**Overall (paper's conclusion).** Domestic inflation targeting is robustly optimal for non-US central banks under DCP, independently of asset-market structure, preference parameters, and shock distribution, and the policy is time consistent. Macroprudential policies cannot help unilaterally but become useful under international cooperation. The United States benefits from dollar dominance in trade, but its optimal policy deviates from domestic stabilization to extract global rents, and international cooperation to improve global welfare is not in its self-interest.

## Theory / model

The baseline model augments the canonical sticky-price small open economy of Gali and Monacelli (2005) with a dominant-currency international price system: all export and import prices are sticky in dollars, while domestic prices are set in local currency (§I, p. 1788). Gopinath (2016) provides the empirical motivation for this assumption, showing it is a good first-order approximation to the actual international price system.

**World structure.** Time is discrete, infinite horizon. The world consists of a continuum of symmetric small open economies $$ i \in [0,1] $$, with the United States indexed by $$ i = 0 $$. Each country is populated by identical households that consume domestic goods $$ C_{iit} $$, foreign goods $$ C^*_{it} $$, and supply labor $$ L_{it} $$. Households maximize expected lifetime utility (p. 1788):

$$
E\sum_{t=0}^{\infty} \beta^t U(C_{iit}, C^*_{it}, L_{it}, \xi_{it}),
$$

where $$ \xi_{it} $$ captures intratemporal (labor supply) and intertemporal (discount) shocks. The import bundle aggregates products from all country-pairs and varieties (p. 1788):

$$
C^*_{it} = \left(\int C_{jit}^{\frac{\varepsilon-1}{\varepsilon}}\,dj\right)^{\frac{\varepsilon}{\varepsilon-1}}, \qquad C_{jit} = \left[\int C_{jit}(\omega)^{\frac{\varepsilon-1}{\varepsilon}}\,d\omega\right]^{\frac{\varepsilon}{\varepsilon-1}},
$$

where $$ \varepsilon > 1 $$ is the micro elasticity of substitution between varieties and differs from the macro elasticity between home and foreign goods.

**Household optimality conditions.** The static FOCs for labor supply and relative demand (eqs. 1-2, p. 1789) are:

$$
-\frac{U_{L_{it}}}{U_{C_{iit}}} = \frac{W_{it}}{P_{iit}}, \tag{1}
$$

$$
\frac{U_{C^*_{it}}}{U_{C_{iit}}} = \frac{\mathcal{E}_{it} P^*_t}{P_{iit}}, \tag{2}
$$

where $$ P_{iit} $$ is the domestic price index, $$ W_{it} $$ the nominal wage, $$ \mathcal{E}_{it} $$ the nominal exchange rate (local currency per dollar), and $$ P^*_t $$ the import price index in dollars (common across all countries under DCP). The Euler equation for nominal bonds (eq. 3, p. 1789) is:

$$
E_t \Theta_{it,t+1} R_{it} = 1, \quad \text{where} \quad \Theta_{it,t+\tau} \equiv \beta^\tau \frac{U_{C_{iit+\tau}}}{U_{C_{iit}}} \frac{P_{iit}}{P_{iit+\tau}}, \tag{3}
$$

and no-arbitrage conditions for internationally traded assets (eq. 4, p. 1789) generalize the standard Euler equation to all traded securities.

**Firms.** Monopolistic competitors in country $$ i $$ produce variety $$ \omega $$ with production function $$ Y_{it} = A_{it} N_{it} $$. Firms face price-adjustment costs $$ \Omega(\cdot) $$ satisfying $$ \Omega(\cdot) \geq 0 $$, $$ \Omega(1) = 0 $$. Domestic firms set prices in local currency to maximize profits net of adjustment costs (eq. 5, p. 1790). Exporters set a single dollar price for all foreign markets (eq. 6, p. 1790-1791), minimizing the costs of adjusting their terms of trade $$ S_{it} \equiv P^*_{it}/P^*_t $$ relative to the dollar import price index.

**Market clearing and budget constraint.** The resource constraint (eq. 7, p. 1790) requires that labor be allocated across domestic production, exports, and price adjustment:

$$
A_{it} L_{it} = C_{iit} + h(S_{it}) C^*_t + A_{it}\!\left[\Omega(\pi_{iit}) + \Omega^*\!\!\left(\frac{S_{it}}{S_{it-1}} \pi^*_t\right)\right], \tag{7}
$$

where $$ h(\cdot) $$ is the demand function for country $$ i $$ exports, $$ \pi_{iit} = P_{iit}/P_{iit-1} $$ is the domestic inflation index, and $$ \pi^*_t = P^*_t/P^*_{t-1} $$ is the global export-price inflation index. The budget constraint (eq. 8, p. 1790) relates net exports to changes in the net foreign asset position, including "valuation effects" from exchange rate movements.

**The two key wedges.** In a static setup, the first-best requires closing two distortions (p. 1796):

$$
\bar{\tau}_{iit} \equiv 1 + \frac{1}{A_{it}} \frac{U_{L_{it}}}{U_{C_{iit}}} = 0 \quad (\text{local wedge}),
$$

$$
\bar{\tau}^*_{it} \equiv 1 + \frac{\varepsilon}{\varepsilon-1} \frac{1}{A_{it} S_{it}} \frac{U_{L_{it}}}{U_{C^*_{it}}} = 0 \quad (\text{external wedge}).
$$

The local wedge measures inefficiency in the trade-off between domestic consumption and leisure; the external wedge measures the inefficiency in the trade-off between leisure and foreign consumption via exports. Under flexible prices and PCP (producer currency pricing), both wedges can be closed with a single monetary instrument. Under DCP they cannot, because domestic prices and export dollar prices respond to different price indices.

The central insight (pp. 1797-1800) is that DCP makes export prices **constrained efficient**: because each small open economy's exports constitute a zero measure of global demand $$ C^*_t $$, individual adjustments in $$ S_{it} $$ do not affect the aggregate import price index. Hence, the social and private benefits of changing export prices coincide, the external wedge is beyond monetary control, and the planner focuses exclusively on closing the local wedge by targeting domestic price stability. This argument holds generically, even when export prices respond endogenously to monetary policy (via the Calvo or Rotemberg mechanism), because the adjusting exporters' prices are still constrained efficient at the private margin. Corsetti, Dedola, and Leduc (2020) established inflation targeting optimality in knife-edge cases (fully sticky export prices, or log-linear preferences); this paper proves it holds generically for arbitrary preferences, technologies, asset markets, and shock distributions. A further implication (Proposition 2) is that capital controls cannot improve welfare unilaterally; this contrasts with the general lesson from Farhi and Werning (2016) that macroprudential interventions are useful whenever monetary policy falls short of first-best, and the paper clarifies why the DCP external-wedge channel nullifies that general argument.

**Equilibrium and game structure.** The paper defines equilibrium as a subgame-perfect Nash equilibrium in which the United States moves first as a Stackelberg leader (internalizing effects on other economies) and non-US economies best-respond taking all foreign variables as given (Definition, p. 1792). Lemma 2 (p. 1792) establishes that the equilibrium outcome is the same under simultaneous play and under non-US discretionary policy, making the results independent of timing assumptions.

## Method

The optimal policy is derived using the **primal approach**: the planner chooses allocations subject to implementability conditions (Lemma 1, p. 1792) rather than choosing prices or interest rates directly. This approach solves the planner's problem without relying on second-order approximations, allowing a full nonlinear stochastic characterization.

**Proof of Proposition 1 (pp. 1799-1800).** The proof considers a relaxed version of the non-US planner's problem that drops the price-setting constraints and retains only market clearing (eq. 7) and the budget constraint (eq. 8). The Lagrangian is:

$$
\mathcal{L} = E\!\sum_{t=0}^{\infty} \beta^t \left\{ U(C_{iit}, C^*_{it}, L_{it}, \xi_{it}) + \lambda_{it}\!\left[A_{it}L_{it} - C_{iit} - A_{it}\Omega(\pi_{iit})\right] + \mu_{it}(\psi_{it} - C^*_{it}) + [\ldots] \right\},
$$

where $$ \lambda_{it} $$ is the multiplier on market clearing for domestic goods and $$ \mu_{it} $$ on the balance-of-payments constraint. Taking FOCs with respect to $$ C_{iit} $$, $$ C^*_{it} $$, and $$ L_{it} $$ yields multiplier values $$ \lambda_{it} = U_{C_{iit}} $$, $$ \mu_{it} = U_{C^*_{it}} $$, and the efficient labor condition $$ -U_{L_{it}}/U_{C_{iit}} = A_{it} $$. The FOC with respect to $$ \pi_{iit} $$ (the domestic inflation rate) implies zero domestic price-adjustment cost, hence $$ \pi_{iit} = 1 $$ (zero inflation). The FOC with respect to $$ S_{it} $$ (the terms of trade) coincides with the private export price-setting condition (eq. 6), confirming the price-setting constraint is not binding. Thus all constraints of the original planner's problem are satisfied, and $$ \pi_{iit} = 1 $$ is globally optimal. The resulting allocation is generically inefficient because $$ \Omega^*(\cdot) > 0 $$ (non-zero export adjustment costs remain), but monetary policy can do no better.

**Proposition 4 (US Policy Rule, eq. 12, pp. 1809-1810).** The US planner maximizes welfare over all prices and quantities in the world economy, taking non-US best-response as given. Assuming flexible domestic prices in the US (to isolate the new DCP motives), the optimal US policy rule is:

$$
S_{it} h(S_{it}) C^*_t \,\tilde{\tau}^*_{it} - E_t \sum_{k=0}^{\infty} \Theta^*_{it,t+k}\!\left(\frac{W_{it+k}}{P^*_{t+k}} \Omega^{*\prime}_{it+k}\right)\!\left(\pi^*_{it+k} \frac{\partial \log \pi^*_{t+k}}{\partial \log C^*_t}\right)
$$

$$
- E_t \sum_{k=0}^{\infty} \Theta^*_{it,t+k}\!\left(\sum_h \omega^h_{it}\frac{\partial \log \mathcal{Q}^h_{t+k}}{\partial \log C^*_t}\right)\! CA_{it+k} = 0, \tag{12}
$$

where $$ \tilde{\tau}^*_{it} \equiv 1 + (1/A_{it} S_{it})(U_{L_{it}}/U_{C^*_{it}}) $$ is the zero-markup external wedge, $$ CA_{it} \equiv \sum_h \mathcal{Q}^h_t (B^h_{it+1} - B^h_{it}) $$ is the US current account, $$ \Theta^*_{it,t+k} \equiv \beta^k (U_{C^*_{it+k}}/U_{C^*_{it}}) $$ is the real stochastic discount factor, and $$ \omega^h_{it} $$ are asset-portfolio weights summing to one (p. 1809). The three terms capture: (i) the net benefit of stimulating exports by closing the external wedge; (ii) the cost of export price adjustment; and (iii) the effect of US monetary policy on international asset prices, which the US exploits to raise foreign asset returns when running a current-account deficit.

**Proposition 5 (Cooperative Policy, eq. 13, pp. 1812-1813).** The global planner's problem adds country-specific demand shifters $$ \varpi_{jit} $$ within import baskets. US monetary policy is used to target a weighted average of external wedges across all countries:

$$
\int v_{it} \frac{U_{C^*_{it}}}{\mathcal{P}^*_{it}} \tilde{\tau}^*_{it}\,di = 0, \tag{13}
$$

where $$ v_{it} $$ is an invariant measure reflecting what fraction of an additional dollar printed by the United States is spent on exports of country $$ i $$. This rule differs from the noncooperative US case: under cooperation, the United States targets the global external wedge rather than extracting rents from the rest of the world.

**Numerical solution (Section IV, p. 1814).** The calibrated model uses the CRRA-CES utility function:

$$
U(C_{iit}, C^*_{it}, L_{it}) = \frac{C_{it}^{1-\sigma}-1}{1-\sigma} - \frac{L_{it}^{1+\phi}}{1+\phi},
\quad C_{it} = \left[(1-\gamma)^{\frac{1}{\theta}} C_{iit}^{\frac{\theta-1}{\theta}} + \gamma^{\frac{1}{\theta}} C^{*\frac{\theta-1}{\theta}}_{it}\right]^{\frac{\theta}{\theta-1}},
$$

with Cobb-Douglas production $$ Y_{it} = A_{it} N^{1-\alpha}_{it} X^\alpha_{it} $$ where $$ X_{it} $$ is intermediate inputs. The model is log-linearized around the symmetric deterministic steady state and solved using standard perturbation methods; welfare losses are computed via second-order approximations of the value function.

## Empirical specifications

Section IV (pp. 1814-1819) calibrates and simulates the model. One period is one quarter.

**Parameter values (p. 1815):**
- Preferences: $$ \beta = 0.99 $$, $$ \sigma = 2 $$ (inverse EIS), $$ \phi = 2 $$ (inverse Frisch elasticity)
- Technology: $$ \alpha = 0.5 $$ (intermediate-goods share in production)
- Trade elasticities: $$ \theta = 1.5 $$ (macro), $$ \eta = 4 $$ (bilateral import), $$ \varepsilon = 11 $$ (variety, implying a 10% markup)
- Openness: $$ \gamma = 0.15 $$ (import-to-GDP ratio 0.3), $$ n = 0.2 $$ (US share in the world economy)
- Shock processes: AR(1) with persistence $$ \rho = 0.97 $$ for both productivity $$ a_{it} $$ and wealth shocks $$ \psi_{it} $$
- Price adjustment: quadratic Rotemberg costs calibrated to match the slope of the New Keynesian Phillips curve for a Calvo model with average price duration of three quarters
- Calibration targets: annualized std($$ \Delta e_{it} $$) = 10% (bilateral exchange rate), std($$ \Delta c_{it} $$) = 2% (US consumption), corr($$ \Delta c_{it}, \Delta c_{jt} $$) = 0.3

**Impulse responses.** Figure 1 (p. 1816) shows that responses under the optimal policy and under inflation targeting are nearly indistinguishable for non-US economies following a local financial shock $$ \psi_{it} $$. Figure 3 (p. 1817) shows the non-US policy response to a US financial shock under four invoicing structures (PCP; DCP in imports only; DCP in exports only; full DCP): whether non-US interest rates rise or fall depends on the relative strength of the import and export channels. Countries with a stronger import channel (e.g., emerging economies importing sticky-price manufactures) tighten monetary policy when the dollar appreciates; countries with a stronger export channel ease policy. Figure 4 (p. 1818) shows US impulse responses, confirming that the optimal US policy deviates from inflation targeting by raising interest rates sharply to generate a global recession and attract savings from the rest of the world following a US wealth shock.

**Welfare decomposition (Table 1, p. 1819).** Columns: (1) non-US optimal; (2) non-US output gap targeting; (3) US optimal; (4) US targeting domestic inflation; (5) PCP benchmark. Rows: local and foreign productivity shocks $$ a_{it} $$, local and foreign financial shocks $$ \psi_{it} $$, and global shocks. Welfare costs of productivity shocks are small across all policies and similar to the PCP benchmark. Financial shocks dominate and vary substantially: the optimal non-US response to local financial shocks costs 2.92% vs. 3.40% under output gap targeting; and the gains from the optimal policy over naive targeting equal 2.13 percentage points in total welfare (3.05% vs. 5.18%).

## Datasets used

This is a theoretical paper with a calibrated numerical illustration. No specific empirical dataset is used; the model is calibrated to standard macroeconomic moments drawn from the prior literature.

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Standard macro calibration moments (exchange rate and consumption volatility from existing empirical literature) | Calibrate shock volatilities and cross-country correlation targets (std = 10%, 2%; corr = 0.3) | No page yet |

## When to read the full paper

Use the [original](https://doi.org/10.1257/aer.20200636) if you are: (i) characterizing optimal monetary policy in an open economy with DCP and need the complete proofs (the online appendix contains formal derivations and extensions for sticky wages, market power, endogenous currency choice, and intermediate goods); (ii) evaluating the role of capital controls and trade policy under dollar pricing (Propositions 2-3 and Sections II.D-E); (iii) studying international spillovers and the gains from cooperation (Section III.B and Proposition 5 with the cooperative capital-controls formula eq. 14); or (iv) running the calibrated model (replication data and code at [doi.org/10.3886/E184741V1](https://doi.org/10.3886/E184741V1)).

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(7), July 2023. Published by the American Economic Association. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. The AER is paywalled; this page contains extracted text and equations only (extract-only redistribution).

> Egorov, Konstantin, and Dmitry Mukhin. "Optimal Policy under Dollar Pricing." *American Economic Review* 113, no. 7 (July 2023): 1783-1824. DOI: 10.1257/aer.20200636.


==============================================================================
# Confidence, Self-Selection, and Bias in the Aggregate: Enke, Graeber & Oprea (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/enke-et-al-confidence-self-selection-bias-2023/
# Distilled: Using 15 cognitive tasks and 2,153 participants in betting market, auction, and committee experiments, Enke, Graeber, and Oprea document that social institutions filter some biases strongly and others barely at all, with the cross-task variation explained almost entirely by the within-task confidence-performance correlation (r = 0.76 to 0.93). American Economic Review 2023, AEA copyright. Seven core results with source locators, the theoretical framework, the experimental design equations, and the datasets used.
# Tags: paper-summary, behavioral-economics, cognitive-biases, self-selection
==============================================================================

**What this is.** The paper's core results, the theoretical framework linking confidence to institutional filtering, and the three experimental institutions (parimutuel betting market, discriminatory auction, committee voting) with their defining equations: enough to know what it found and how, without reading all 34 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1257/aer.20220915).

## TL;DR

Enke, Graeber, and Oprea run a large preregistered online experiment on Prolific (2,153 subjects, June 2021) exposing participants to 15 canonical cognitive biases from behavioral economics and three simple social institutions (betting markets, auctions, committees) that allow voluntary self-selection. They find that institutions filter biases on average, but with large cross-task variation: exponential growth bias (EGB) is reduced by roughly 17 percentage points, while base-rate neglect and correlation neglect are barely affected, and the winner's curse is even amplified. Almost all of this cross-task heterogeneity (r = 0.76 to 0.93) is explained by a single sufficient statistic: the within-task Pearson correlation between subjects' stated confidence and their decision optimality. When better performers are also more confident, they self-select more intensively and the institution de-biases effectively. When confidence and performance are uncorrelated or negatively correlated, the institution cannot filter, regardless of average overconfidence levels.

## Core results

Magnitudes as reported; Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Positive self-selection in all three institutions on average across tasks: optimal decision makers bet, bid, and vote more intensively than suboptimal ones | Figure 2, p. 1952; p. 1953 | Betting: 64.8 avg. bet (optimal) vs 47.4 (suboptimal), 37% more; Auction: 56.4 vs 43.6, 29% more; Committee: 75 vs 57.9 votes, 29% more |
| R2 | Large cross-task variation in institutional filtering: EGB and iterated reasoning (IR) strongly improved, some tasks near-zero or negative | Figure 3, p. 1954 | EGB: ~17 pp improvement across institutions; IR: ~8 pp; RM, AC, EQ near-zero or negative (approx. -4 to 0 pp); pairwise correlations across institutions 0.85-0.91 |
| R3 | Confidence-performance correlation varies widely across tasks: from negative (RM = -0.13, TM significantly negative) to moderately positive (GF = 0.39); 6 of 15 tasks negative | Figure 4, p. 1956 | Pearson r ranges -0.13 (RM) to 0.39 (GF); N = 334 in Confidence treatment; no task exceeds r = 0.5 |
| R4 | Confidence-performance correlation strongly predicts institutional improvement across the 15 tasks | Figure 5, p. 1957 | r = 0.76 (between-subjects), r = 0.93 (within-subjects); robust to leave-two-out: between-subjects range 0.61-0.83 (mean 0.76) |
| R5 | Predictive power is consistent across all three institutions | p. 1958 | r\^auction = 0.69, r\^betting = 0.73, r\^committee = 0.77 (between); r\^auction,within = 0.90, r\^betting,within = 0.90, r\^committee,within = 0.91 |
| R6 | Confidence-performance correlation predicts institutional efficiency (fraction of theoretically possible improvement realized) even more strongly | p. 1959 | r = 0.87 (between-subjects), r = 0.94 (within-subjects) |
| R7 | Average overconfidence (d = c - p) shows a weak, statistically insignificant negative relationship with institutional improvement | p. 1961 | r = -0.34 (between-subjects), r = -0.32 (within-subjects); neither significantly different from 0 at conventional levels |

**Overall (paper's conclusion).** Average overconfidence, the traditional focus of most confidence research, is largely irrelevant for predicting whether markets and organizations de-bias economic aggregates. The relevant object is the confidence-performance correlation, which determines whether the biased individuals who self-select out of institutions are actually the ones making worse decisions. This implies a simple methodological blueprint: researchers studying cognitive biases can estimate the likely institutional impact by appending an unincentivized confidence question and reporting the resulting correlation with performance.

## Theory / model

The paper lays out a simple analytical framework (Section II, pp. 1947-1950) to derive testable predictions about institutional filtering. There is no fully specified equilibrium model; the framework is a linear approximation designed to connect observable confidence and performance to institutional outcomes.

**Setup.** Each of $$N$$ agents forms a judgment on a cognitive task. Agent $$i$$'s solution is optimal ($$X_i = 1$$) with probability $$p_i$$ and incorrect ($$X_i = 0$$) with probability $$1 - p_i$$. Pre-institutional aggregate performance equals the raw optimality rate

$$
\Theta^{\text{pre}} = \frac{1}{N}\sum_{i=1}^N X_i, \qquad \theta^{\text{pre}} \equiv E[\Theta^{\text{pre}}] = \frac{1}{N}\sum_{i=1}^N p_i.
$$

Each agent then makes an institutional decision $$k_i \in [0, 1]$$, representing bet intensity, bid size, or vote share, depending on the institution. Institutional filtering $$\mathbb{G} = \theta^{\text{post}} - \theta^{\text{pre}}$$ is positive when the institution makes aggregate outcomes appear as if participants were more rational than they actually are.

**Institutional performance metrics (pp. 1948).** For betting markets and committees, the performance metric is a weighted average of agents' optimality, weighted by their participation intensity:

$$
\theta^{\text{post}}_{\text{bet,com}} = \frac{\sum_i k_i p_i}{\sum_i k_i}. \tag{5}
$$

The expected institutional gain is

$$
\mathbb{G}_{\text{bet,com}} = \theta^{\text{post}} - \theta^{\text{pre}} = \frac{\sum_i p_i (k_i - \bar{k})}{N\bar{k}}. \tag{6}
$$

This expression is positive if and only if better-performing agents participate more intensively than the average. For auctions, only the top-5 bidders win and the metric is their average optimality rate, so

$$
\mathbb{G}_{\text{auc}} = \frac{1}{|W|}\sum_{j \in \Omega} p_j - \frac{1}{N}\sum_i p_i, \tag{7}
$$

where $$\Omega$$ is the set of winners (highest five bids).

**Confidence and performance (p. 1949).** Institutional self-selection is assumed to depend on agents' stated confidence $$c_i$$ about the ex ante optimality of their decision. The paper models the within-task confidence-performance relationship as approximately linear:

$$
c_i = \alpha + \beta \cdot p_i. \tag{8}
$$

The slope $$\beta$$ is the confidence-performance correlation (the key object). Average overconfidence is $$d = \bar{c} - \bar{p} = \alpha + (\beta - 1)\bar{p}$$, which combines both the intercept $$\alpha$$ and the slope $$\beta$$. Institutional self-selection is taken to be proportional to confidence:

$$
k_i = \omega \cdot c_i \in [0,1]. \tag{9}
$$

Here $$\omega > 0$$ captures the degree to which self-selection actually depends on confidence as opposed to other factors.

**Predictions.** Substituting equations (8) and (9) into (6) and (7) yields two preregistered predictions (pp. 1950):

- **Prediction 1**: If $$\beta > 0$$, then $$\mathbb{G} > 0$$ (institutions filter biases). Institutional improvement $$\mathbb{G}$$ increases in the confidence-performance correlation $$\beta$$.
- **Prediction 2**: The effect of average overconfidence $$d$$ on $$\mathbb{G}$$ is ambiguous. In auctions, there is no relationship (only the ordering of bids matters, not their level). In betting and committees with $$\beta > 0$$, the effect of $$d$$ is weakly negative.

Fehr and Tyran (2005) provide foundational evidence that individual irrationality sometimes survives in aggregate market outcomes, and sometimes does not; this framework clarifies that the confidence-performance correlation is the sufficient statistic for predicting which case applies. The paper complements List (2003), who shows that market experience reduces anomalies through learning; here the channel is purely self-selection, which operates even in the absence of feedback or repeated play.

## Method

The experiment (Section I, pp. 1938-1946) implements three maximally simple static variants of canonical economic institutions to isolate the self-selection mechanism. All three are implemented on Prolific with identical slider-based interfaces (0-100) so the self-selection decision is comparable across institutions. The three institutions and their performance metrics are:

**Betting market (parimutuel), p. 1942.** Ten subjects are grouped into a parimutuel betting market. Each subject $$i$$ bets $$b_i \in [0, 100]$$ ECUs on the proposition that her own part-1 response was optimal. The market price on the optimal-decision security is

$$
\theta^{\text{Betting}} = \frac{\sum_{i=1}^{10} x_i b_i}{\sum_{i=1}^{10} b_i} \in [0,1]. \tag{1}
$$

If subject $$i$$'s part-1 decision was optimal, her payoff is

$$
\pi_i^{\text{Betting}} = \frac{b_i}{\theta^{\text{Betting}}} + (100 - b_i). \tag{2}
$$

The market price in equation (1) is simply a reweighting of individual part-1 decisions $$x_i$$ by how much each subject bets; if no self-selection occurs (everyone bets equally), the price equals the raw optimality rate.

**Discriminatory auction (5 winners), p. 1943.** Subjects submit sealed bids $$b_i \in [0, 100]$$ ECUs. The five highest bidders win and receive a bonus of 100 ECUs if their part-1 decision was optimal. The institutional performance metric is the optimality rate among winners:

$$
\theta^{\text{Auction}} = \frac{\sum_{i \in \Omega} x_i}{5}, \tag{3}
$$

where $$\Omega$$ is the set of five highest bidders. Under standard assumptions this auction implements an efficient allocation to the highest-value bidders (Krishna 2009).

**Utilitarian committee voting, p. 1943.** Each subject receives 100 votes and submits $$v_i \in [0, 100]$$ votes in favor of her own part-1 answer. The fraction of votes on the optimal answer is

$$
\theta^{\text{Committee}} = \frac{\sum_{i=1}^{10} X_i v_i}{\sum_{i=1}^{10} v_i} \in [0,1]. \tag{4}
$$

All subjects earn $$100 \times \theta^{\text{Committee}}$$ regardless of their own vote.

**Confidence elicitation and treatments.** After each part-1 cognitive task, subjects in the Confidence treatment (N = 334) and the Within treatments (N = 314) are asked an unincentivized slider question: "How certain are you that your decision in Part 1 was optimal?" (0-100%). In between-subjects treatments (Betting, Auction, Committee), confidence is never elicited; the confidence-performance correlation is measured in a separate Confidence treatment. Table 2 (p. 1945) summarizes the full experimental design: Betting (N = 387), Auction (N = 323), Committee (N = 337), Confidence (N = 334) between-subjects; and Betting Within (N = 105), Auction Within (N = 105), Committee Within (N = 104) within-subjects.

The 15 cognitive tasks (Table 1, p. 1940) cover information processing and statistical reasoning (base rate neglect, correlation neglect, balls-and-urns belief updating, gambler's fallacy, sample size neglect, regression to mean), logic (Wason task, cognitive reflection test), strategic reasoning (backward induction, equilibrium reasoning), constrained optimization (knapsack), and financial reasoning (thinking at the margin, portfolio choice, exponential growth bias, acquiring a company).

## Empirical specifications

The main empirical analysis works at the task level (15 observations) rather than at the individual level. There is no single regression equation; the core result is a cross-task bivariate correlation.

**Step 1: Measure the confidence-performance correlation per task.** For each of the 15 tasks $$k$$ in the Confidence treatment (between-subjects) or the Within treatments (within-subjects), compute the Pearson correlation between the binary optimality indicator $$x_i$$ (part 1) and stated confidence $$c_i$$ (part 2):

$$
\hat{\beta}_k = \text{Corr}(x_i, c_i) \quad \text{in task } k.
$$

This is computed separately for each task across the 334 (between) or 314 (within) subjects who see the confidence elicitation.

**Step 2: Measure institutional improvement per task.** For each task $$k$$ and institution $$j \in \{\text{Betting, Auction, Committee}\}$$, simulate 10,000 random 10-subject cohorts by drawing with replacement from the pool of part-1 and part-2 decisions. Compute $$\theta^{\text{post}}_{k,j}$$ for each cohort using equations (1)-(4), and compare to the cohort's raw optimality rate $$\theta^{\text{pre}}_k$$. The institutional improvement is the mean of $$\theta^{\text{post}} - \theta^{\text{pre}}$$ over the 10,000 cohorts. Standard errors are computed conservatively as the standard deviation of cohort-level improvements divided by $$\sqrt{N/10}$$, where $$N$$ is the treatment sample size (e.g., $$387/10 = 38.7$$ cohorts in Betting; Figure 3 notes, p. 1954).

**Step 3: Cross-task regression.** The main result (Figure 5, p. 1957) is the Pearson correlation between $$\hat{\beta}_k$$ (step 1) and the average institutional improvement in task $$k$$ (step 2) across the 15 tasks. The between-subjects correlation pools improvements across Betting, Auction, and Committee. The within-subjects correlation uses the same subjects for both confidence and institutional decisions.

**Average overconfidence check.** As an ancillary test, the paper replaces $$\hat{\beta}_k$$ with task-level average overconfidence $$\hat{d}_k = \bar{c}_k - \bar{p}_k$$ and repeats step 3. This directly tests Prediction 2: the resulting correlation is weakly negative (r = -0.34 between, r = -0.32 within) but not statistically distinguishable from zero (p. 1961).

**Expert survey.** A separate sample of 38 behavioral economists (CESifo/VIBES panel, November 2021) predicted institutional improvements and confidence differences for 7 of the 15 tasks in the Auction treatment. Experts' median forecasts are compared to actual outcomes in Figure 6 (p. 1962). The analysis uses a paired comparison of forecast vs. actual for each of the 7 tasks; no regression is reported. Camerer and Lovallo (1999) document overconfidence in entry decisions; the expert results here parallel that finding in the prediction domain. Moore and Healy (2008) provide the taxonomy of overconfidence types that the paper uses to frame what experts miss. Kendall and Oprea (2018) study the market selection hypothesis in a related laboratory design.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Primary experimental data (Prolific online experiment, June 2021) | 15 cognitive tasks x 3 institutions x 2,153 subjects; ~70,000 individual decisions | No page yet |
| Expert survey (Social Science Prediction Platform, November 2021) | 38 behavioral economists predicting institutional filtering and confidence differences for 7 tasks | No page yet |

Sample: 1,381 subjects in between-subjects treatments (Betting, Auction, Committee, Confidence); 314 subjects in within-subjects treatments (Betting Within, Auction Within, Committee Within); June 2021 on Prolific. Replication data available at ICPSR (https://doi.org/10.3886/E185741V1).

## When to read the full paper

Use the [original](https://doi.org/10.1257/aer.20220915) if you are: studying which specific cognitive biases survive market aggregation (the paper gives results for all 15 tasks by institution); designing experiments to study self-selection through institutions (the exact institution implementations are in online appendices); investigating what determines the confidence-performance correlation itself and why it varies across tasks (Section IV.C); or replicating the expert-survey methodology (the SSPP survey instruments are reproduced in online appendices). Locators above point to the exact figures and pages.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(7), July 2023. Published under AEA copyright with 12-month delayed open access. No CC license detected. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. Redistribution is extract-only.

> Enke, Benjamin, Thomas Graeber, and Ryan Oprea. "Confidence, Self-Selection, and Bias in the Aggregate."
> *American Economic Review* 113, no. 7 (July 2023): 1933-1966.
> DOI: 10.1257/aer.20220915.
> Replication data: https://doi.org/10.3886/E185741V1 (AEA/ICPSR).
> This page is an extraction by the Institute for Automated Research: core results
> and equations summarized; **not reproduced**.


==============================================================================
# Optimal Contracting with Altruistic Agents: Gaynor, Mehta & Richards-Shubik (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/gaynor-et-al-optimal-contracting-altruistic-agents-2023/
# Distilled: A structural screening model estimated on 2008-2009 Medicare EPO claims shows that optimal nonlinear payment contracts for dialysis providers eliminate all medically excessive dosages and reduce spending by 12-48%, for aggregate gains of roughly $300 million per year. American Economic Review 2023, paywalled. Seven core results with source locators, the model, the method (demand profile approach for supply contracting), and the empirical specifications with equations.
# Tags: paper-summary, health-economics, optimal-contracting, mechanism-design, structural-estimation, peer-reviewed, unreplicated, data:cms-medicare
==============================================================================

**What this is.** The paper's core results, the structural model (provider utility and government objective), the method (demand profile approach for supply contracting), and the empirical specifications with equations: enough to know what it found and how, without reading all 42 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1257/aer.20210208).

## TL;DR

The paper estimates a structural screening model of health-care provider behavior using 2008-2009 Medicare claims for epoetin alfa (EPO), an expensive drug used to treat anemia in dialysis patients with end-stage renal disease. Dialysis providers are heterogeneous in their degree of altruism toward patients and their marginal costs of administering EPO, both unobservable to Medicare. Using natural variation in patient hematocrit levels and quarterly variation in national Medicare payment rates, the paper recovers the joint distribution of provider types. It then derives optimal nonlinear payment contracts using the demand profile approach of Goldman, Leland, and Sibley (1984) and Wilson (1993), which handles multidimensional heterogeneity tractably. The optimal contracts completely eliminate medically excessive dosages (present for 75-86% of providers under the observed linear contract), reduce Medicare spending by 12-48%, and improve the government's welfare objective by $87-$220 per patient per month. Aggregate gains are approximately $300 million per year. Like Clemens and Gottlieb (2014), who examine Medicare payment incentive effects broadly, this paper pushes further to derive and characterize the optimal contract for a specific treatment.

## Core results

Magnitudes as reported; `\*\*`/`\*\*\*` = 5%/1%.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Optimal nonlinear contract **eliminates medically excessive dosages** for all provider types | Table 5, p. 1560 | Share with medically excessive dosages: 82% / 75% / 86% under observed contract -> 0% / 0% / 0% under optimal nonlinear (low / medium / high hematocrit intervals); optimal linear contract reduces but does not eliminate this inefficiency (19% and 45% share remain in medium and high intervals) |
| R2 | Optimal nonlinear contract **reduces mean Medicare spending by 12-48%** | Table 5, p. 1560 | Mean monthly payments per patient: $744 -> $388 (-48%, low hematocrit); $541 -> $392 (-27%, medium); $437 -> $384 (-12%, high) |
| R3 | **Government welfare objective improves by $87-$220 per patient per month** under optimal nonlinear contract | Table 5, p. 1560 | Gains vs observed: $220 (low), $124 (medium), $87 (high) per patient per month; optimal linear achieves 70-85% of these gains but leaves medically excessive dosages in the medium and high intervals |
| R4 | **Aggregate gains from better contracting estimated at ~$300 million per year** | Section VI, p. 1563 | Rough approximation multiplying per-patient gains by patient-months; Medicare spent ~$2 billion per year on EPO during the study period |
| R5 | Providers **respond significantly to reimbursement rates**: dosage rises by 6,390 units per $1 payment rate increase | Table 2, p. 1552 | OLS reduced form: $$\beta_2$$ = 9.53 (SE 3.11), 6.39 (SE 2.12), 3.92 (SE 1.89) thousand units per $1/1,000u payment rate in low / medium / high hematocrit intervals; SEs clustered on dialysis center, 250 bootstrap replications |
| R6 | **Losses from asymmetric information** about provider types are $1,739-$3,752 per patient per month | Section VB, p. 1561 | Difference between full-information government objective and second-best achievable gains; approximately 8-43 times the gains achievable through optimal contracting ($220/$124/$87 in low/medium/high intervals), indicating costs of asymmetric information dwarf the gains from better contracting alone |
| R7 | Optimal nonlinear contract **reduces unjustified dosage variation by 26-52%** | Table 5, p. 1560 | Std dev of dosage: 9.7 -> 7.2 thousand units (-26%) for medium hematocrit; 5.2 -> 2.5 thousand units (-52%) for high; variation reduction reflects elimination of type-heterogeneity-driven overprovision |

**Overall (paper's conclusion).** The observed Medicare fee-for-service contract, which pays a constant marginal rate per EPO unit regardless of dosage, cannot be rationalized as optimal for any value of the government's health weight given the estimated structural parameters. Moving to an optimal nonlinear contract with declining marginal payments would eliminate medically excessive dosages, reduce both mean and variance of treatment amounts, and improve the government's welfare objective by hundreds of millions of dollars per year. The demand profile approach, applied here for the first time to supply contracting, is the tool used to handle the multidimensional provider heterogeneity (joint unobservability of altruism and marginal cost) that is central to this and many similar health-care payment settings.

## Theory / model

The framework is a static screening model where the government (principal) pays a dialysis provider (agent) to treat a patient with end-stage renal disease (Section II, p. 1537). The patient arrives with baseline hematocrit $$b$$ and observed characteristics $$x$$; the provider chooses EPO dosage $$a$$ (total units administered per month). All three are observable by the government because providers report them on Medicare insurance claims.

**Provider utility** (equation 1, p. 1539): the provider values patient health, weighted by altruism $$\alpha$$, less the cost $$za$$ of administering the drug, plus the government's payment $$P(a; b, x)$$:

$$
u(a;\,\alpha,z,b,x,P) \;\equiv\; \alpha\, h(a;b,x) \;-\; za \;+\; P(a;b,x). \tag{1}
$$

The provider utility specification follows the altruistic physician model of Ellis and McGuire (1986), extended here to allow heterogeneity in both altruism and costs. The health production function $$h(a; b, x)$$ is twice differentiable and strictly concave in $$a$$; it first increases then decreases in dosage. Dosages with $$h'(a; b, x) < 0$$ are "medically excessive." The provider's type $$(\alpha, z)$$ is unobserved by the government (asymmetric information): $$(\alpha, z)$$ has joint density $$f(\alpha, z)$$ on a compact support $$[\underline{\alpha}, \bar{\alpha}] \times [\underline{z}, \bar{z}]$$.

**Government objective** (equation 2, p. 1539): the government maximizes patient health (weighted by its own health preference $$\alpha_g$$) minus payments to the provider:

$$
u_g(a;\,b,x,P) \;\equiv\; \alpha_g\, h(a;b,x) \;-\; P(a;b,x). \tag{2}
$$

The government sets a potentially nonlinear payment policy $$\{P(a; b, x)\}$$ before provider types and patient health states are realized. Given $$(b, x)$$, the government maximizes the expectation of (2) over the distribution of types and their resulting treatment choices, subject to incentive compatibility (IC) and voluntary participation (VP) for each type $$(\alpha, z)$$:

$$
\max_{P \in \mathcal{P}} \int_{\alpha,z} \Bigl[\alpha_g\,h\bigl(a^*(\alpha,z;b,x,P);\,b,x\bigr) - P\bigl(a^*(\alpha,z;b,x,P);\,b,x\bigr)\Bigr]\,f(\alpha,z)\,d\alpha\,dz,
$$

subject to

$$
\text{IC:}\quad a^*(\alpha,z;b,x,P) = \arg\max_{a \geq 0}\,u(a;\alpha,z,b,x,P),\quad \forall\,\alpha,z,
$$

$$
\text{VP:}\quad u\!\bigl(a^*(\alpha,z;b,x,P);\,\alpha,z,b,x,P\bigr) \;\geq\; \underline{u},\quad \forall\,\alpha,z.
$$

**Full-information first best** (equation 3, p. 1540): under full information, the optimal treatment equates the government's marginal benefit to the agent's net marginal cost:

$$
\alpha_g\,h'(a^{*FI}(\alpha,z)) \;=\; z - \alpha\,h'(a^{*FI}(\alpha,z)). \tag{3}
$$

Altruism reduces the agent's effective marginal cost, so first-best treatment amounts are higher with altruism than without. The full-information allocation never produces medically excessive dosages (where $$h' < 0$$), because both $$\alpha$$ and $$\alpha_g$$ are positive.

## Method

The paper uses the demand profile approach of Goldman, Leland, and Sibley (1984) and Wilson (1993) to solve the optimal contracting problem with two-dimensional unobserved heterogeneity (Section IIC, p. 1541). Standard methods based on the revelation principle (Myerson (1981); Maskin and Riley (1984)) require a strict ordering of agent types so that the binding IC constraints reduce to adjacent-type comparisons; under multidimensional heterogeneity such a reduction is generally unavailable. The demand profile approach instead reformulates the government's problem in terms of setting the marginal payment for each treatment amount, and separates it into independent subproblems.

**Provider first-order condition** under any differentiable contract $$P$$ (equation 4, p. 1541): the provider equates the net marginal cost to the marginal payment:

$$
\underbrace{z - \alpha\,h'(a^*)}_{\text{nc}(a^*;\alpha,z)} \;=\; \underbrace{\dfrac{\partial P(a^*)}{\partial a}}_{p(a^*)}. \tag{4}
$$

The net marginal cost $$\text{nc}(a;\alpha,z) = z - \alpha h'(a)$$ is upward sloping in $$a$$ (since $$h'' < 0$$). If the marginal payment curve is downward sloping, each net marginal cost curve intersects it at most once from below, which is the key regularity condition for the demand profile approach.

**Demand profile** (equation 6, p. 1542): $$S(p, a)$$ is the probability (over the type distribution) that the provider supplies at least amount $$a$$ when the marginal payment at $$a$$ equals $$p$$:

$$
S(p, a) \;\equiv\; \Pr\!\bigl\{p(a) \;\geq\; z - \alpha\,h'(a)\bigr\}. \tag{6}
$$

**Decomposed government problem** (equations 5 and 7, pp. 1540-1544): because of the regularity condition and quasilinearity of provider preferences, the government's objective separates into independent maximizations, one for each treatment amount $$a \in A$$:

$$
\max_{p(a)\in\mathbb{R}}\; S(p(a),\,a)\,\bigl[\alpha_g\,h'(a) - p(a)\bigr]. \tag{7}
$$

**Optimal contract first-order condition** (equation 8, p. 1544):

$$
\frac{\partial S(p^*(a),a)}{\partial p(a)}\,\bigl[\alpha_g\,h'(a) - p^*(a)\bigr] \;=\; S(p^*(a),\,a). \tag{8}
$$

This equates the marginal benefit of raising the marginal payment (the change in the probability of provision times the government's marginal health valuation) to the marginal cost (the probability that $$a$$ is already being provided). The optimal total payment $$P^*$$ is recovered by integrating $$p^*(a)$$ over dosage. The optimal marginal payment declines toward and past the health-maximizing dosage level, ensuring that no medically excessive dosages arise in the second-best allocation (a standard no-distortion-at-the-top result holds at the highest treatment amount; all others are distorted downward).

## Empirical specifications

Estimation proceeds in three steps (Section IV, p. 1547).

**Step 1: Reduced-form OLS.** The health function is quadratic (equation 9, p. 1547):

$$
h(a;b,x) \;=\; H \;-\; \tfrac{1}{2}(\delta\,a + b - \tau'x)^2, \tag{9}
$$

where $$\delta$$ converts EPO units into hematocrit points and $$\tau'x$$ is a patient-characteristics index. Under a linear contract with constant marginal payment $$p_1$$, the provider's first-order condition (4) yields (equation 10, p. 1547):

$$
a^*(\alpha,z;b,x,P^L) \;=\; \frac{\tau'x - b}{\delta} + \frac{p_1 - z}{\alpha\,\delta^2}. \tag{10}
$$

Decomposing the marginal cost as $$z_{ik} = \mu_z + \zeta_{ik}$$ and adding an idiosyncratic shock $$\eta_{ijt}$$, the estimating equation within hematocrit interval $$k$$ is (equation 11, p. 1549):

$$
a_{ijt} \;=\; \underbrace{\!\left[\frac{-1}{\delta_k}\right]\!}_{\beta_1^k}\! b_{jt} \;+\; \underbrace{\!\left[\frac{1}{\alpha_{ik}\delta_k^2}\right]\!}_{\beta_2^k}\![p_{1t} - \mu_z] \;+\; \underbrace{\!\frac{\tau_k'}{\delta_k}\!}_{\beta_3^k}\! x_{jt} \;+\; \underbrace{\!\left[\frac{-\zeta_{ik}}{\alpha_{ik}\delta_k^2}\right]\!}_{\nu_i^k} \;+\; \eta_{ijt}. \tag{11}
$$

This is estimated by OLS separately within each of three hematocrit intervals ($$b \in (30, 33]$$, $$(33, 36]$$, $$(36, 39]$$). The regression includes age, sex, CCI indicators, and month and year dummies. Standard errors are clustered on dialysis center (250 bootstrap replications). Identification rests on: (i) natural month-to-month variation in patient hematocrit $$b_{jt}$$ (not manipulated by providers), which identifies $$\beta_1^k$$; and (ii) quarterly variation in the national Medicare payment rate $$p_{1t}$$, set by an administrative formula (106% of average sales price lagged six months) that no individual facility influences, which identifies $$\beta_2^k$$. The mean marginal cost $$\mu_z = \$8.58$$ per 1,000 units is set externally from facility cost reports (acquisition cost $7.53 + administration cost $1.05).

**Step 2: Structural parameter recovery.** Structural parameters $$\delta_k$$, $$\tau_k$$, and the joint distribution $$F_k(\alpha, z)$$ are recovered analytically from the reduced-form moments within each hematocrit interval. The joint distribution of $$(\ln\alpha, z)$$ is bivariate normal with four unknown parameters per interval. Using Stein's lemma and properties of the log-normal distribution, these are identified from the first and second moments of the random coefficient $$\beta_2^k$$ and random effect $$\nu^k$$ in equation (11), estimated via a semiparametric auxiliary regression of the residuals (Section IVB, p. 1549-1550; Online Appendix F for full details).

**Step 3: Optimal contract construction.** The government's health weight $$\alpha_g = 52.6$$ is calibrated from a statistical life year value and EPO dose-response estimates from clinical trials (Online Appendix G.2). Type distributions are truncated at the 0.5th and 99.5th percentiles to ensure compact support. The demand profile $$S(p, a)$$ is computed from the estimated distributions; optimal marginal payments $$p^*(a)$$ are solved numerically from equation (8) for each treatment amount and hematocrit interval and integrated to obtain $$P^*$$. The paper verifies that the regularity condition (no provider type has a net marginal cost curve with multiple intersections with the optimal marginal payment curve) holds in the estimated model (Online Appendix I).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Medicare outpatient claims (CMS 2008-2009b, 20% sample) | Primary estimation data: monthly EPO dosages, baseline hematocrit, patient demographics (age, sex), Charlson Comorbidity Index, payment rates; 919,745 claims after exclusions | No page yet |
| Renal Dialysis Facilities Cost Report Data (CMS 2008-2009a) | Annual per-facility EPO acquisition costs used to set mean marginal cost $$\mu_z$$; publicly available from CMS | No page yet |
| Medicare Part B ASP Drug Pricing Files (CMS 2008, 2009) | National quarterly payment limits for EPO (the source of payment rate variation); public administrative data | No page yet |
| Medicare Beneficiary Summary File (MedPAR; CMS 2007-2009a) | Patient age and sex, linked to claims | No page yet |

Sample: January 2008 to December 2009 (monthly, US). Final sample restricts to patients with hematocrit in 30-39 percent range: 919,745 claims, 74,260 unique patients, 5,148 dialysis providers.

## When to read the full paper

Read the [original](https://doi.org/10.1257/aer.20210208) if you are:
designing optimal nonlinear payment contracts for provider-administered drugs or any supply-contracting problem with multidimensional unobserved agent heterogeneity;
applying the demand profile approach beyond its original monopoly-pricing context;
estimating structural models of provider behavior with altruism, including the identification argument and semiparametric moment-recovery details (Section IVB and Online Appendices E, F);
or studying the welfare costs of asymmetric information in health-care reimbursement. The locators above point to the exact tables and figures.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(6), June 2023. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. The paper is paywalled; only text extracts are reproduced here under extract-only terms.

> Gaynor, Martin, Nirav Mehta, and Seth Richards-Shubik. "Optimal Contracting with Altruistic Agents: Medicare Payments for Dialysis Drugs." *American Economic Review* 113, no. 6 (June 2023): 1530-1571. DOI: 10.1257/aer.20210208.


==============================================================================
# Optimal Insurance: Gershkov, Moldovanu, Strack & Zhang (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/gershkov-et-al-optimal-insurance-dual-utility-2023/
# Distilled: Characterizes profit-maximizing insurance menus under adverse selection with dual-utility (Yaari 1987) agents and random losses: optimal contracts are layer contracts where the retention slope is 0 or 1 almost everywhere, deductibles arise when private information concerns loss probability, and coverage limits when it concerns loss magnitude. American Economic Review 2023, paywalled. Seven core theoretical results with source locators, the model, and the solution method.
# Tags: paper-summary, insurance, adverse-selection, mechanism-design, contract-theory, peer-reviewed, unreplicated
==============================================================================

**What this is.** The paper's core theoretical results, the dual-utility insurance model, and the mechanism design method: enough to understand what was found and how, without reading all 34 pages. To replicate or extend the results, read the full source at the [original](https://doi.org/10.1257/aer.20221247).

## TL;DR

Gershkov, Moldovanu, Strack, and Zhang study a monopoly insurance market under adverse selection where agents have dual utility (Yaari 1987), meaning they weight probabilities via a distortion function rather than taking expectations, and face random losses whose distribution is correlated with their privately known risk type. This generalizes Rothschild and Stiglitz (1976) and Stiglitz (1977) in two directions: from expected utility to dual utility, and from a single fixed loss to a random distribution of losses correlated with the agent's type.

The central result (Theorem 1) characterizes the optimal retention function as a "layer contract": for each loss level, the agent either retains the entire marginal loss or the insurer covers it entirely, so the derivative of the retention function is in {0, 1} almost everywhere. Whether the optimal menu uses deductibles or coverage limits depends on the direction of private information. When the agent's type governs loss probability, the virtual value single-crosses from below, and deductibles are optimal. When the type governs loss magnitude, coverage limits are optimal, even though they are the worst possible contract for any given expected cost (Theorem 2). The welfare gain from reduced information rents dominates the efficiency loss from offering the worst contract form.

In contrast to Chade and Schlee (2012) under expected utility, where full insurance is never optimal, full insurance to some types can be optimal here because dual utility agents exhibit first-order risk aversion.

## Core results

All results are theoretical; locators point to theorems, propositions, and examples in the source PDF.

| # | Result | Locator | Key statement |
|---|---|---|---|
| R1 | **Layer contract structure**: the optimal retention function satisfies $$\partial R / \partial l \in \{0,1\}$$ almost everywhere; each marginal dollar of loss is either fully retained by the agent or fully covered by the insurer | Theorem 1, pp. 2595-2596 | The profit-maximization objective is linear in $$\partial R / \partial l$$; extreme points of the feasible set of Lipschitz-1 retention functions satisfy $$\partial R / \partial l \in \{0,1\}$$ a.e. (Bauer's maximum principle) |
| R2 | **SOSD ranking of contract forms**: for any strongly risk-averse agent (averse to mean-preserving spreads), the deductible contract second-order stochastically dominates any doubly monotone contract with the same expected cost, which dominates the coverage limit contract | Theorem 2, p. 2600 | $$R_C(\cdot,\theta) \leq_{\text{SOSD}} R_I(\cdot,\theta) \leq_{\text{SOSD}} R_D(\cdot,\theta)$$ for any contract $$I$$ with the same expected cost; deductibles are welfare-maximizing at fixed cost |
| R3 | **Deductible menu is optimal** when the virtual value $$J(l,\theta)$$ crosses zero from below in $$l$$: the insurer offers a menu of deductible-premium pairs $$(D(\theta), t(\theta))$$ | Theorem 3(i), p. 2600; Example 2, pp. 2600-2601 | Holds when private information concerns loss probability (equation (1)); the optimal deductible $$D^*(\theta)$$ is nonincreasing in the agent's degree of loss aversion |
| R4 | **Coverage limit menu is optimal** when $$J(l,\theta)$$ crosses zero from above in $$l$$: the insurer offers a menu of cap-premium pairs $$(C(\theta), t(\theta))$$ | Theorem 3(ii), p. 2600; Example 3, p. 2602 | Holds when private information concerns loss magnitude; in Example 3, $$C^*(\theta) = \theta^2 f(\theta) / [2(1 - F(\theta))]$$ is increasing in $$\theta$$ |
| R5 | **Higher risk aversion raises insurer profit**: if $$g_2(p) < g_1(p)$$ for all $$p \in (0,1)$$ (agent 2 is more risk averse than agent 1), the insurer's profit under $$g_2$$ is strictly higher | Proposition 2, p. 2599 | More risk-averse agents value coverage more; for any fixed retention function the premium extractable from a more risk-averse agent is higher, and adjusting to the optimal retention amplifies the gain |
| R6 | **Finite losses**: with $$n$$ possible loss levels $$l_1 < \dots < l_n$$, the optimal deductible menu uses at most $$n+1$$ contracts; each offered deductible equals one of the $$n$$ loss values or 0 | Proposition 3, p. 2603 | The optimal mechanism is a basic deductible-premium contract plus a finite ladder of add-on fees that progressively reduce the deductible; full insurance is one possible top rung |
| R7 | **Single fixed loss (Stiglitz 1977 case)**: the optimal menu offers either full insurance (deductible = 0) or no insurance (deductible = $$\bar l$$); partial insurance is never optimal | Corollary 1, p. 2604 | Follows from Proposition 3 with $$n = 1$$; extends the DeFeo and Hindriks (2014) result to dual utility; contrasts with Chade and Schlee (2012) under EU where full insurance is never optimal |

**Overall (paper's conclusion).** Layer contracts (deductibles and coverage limits) emerge endogenously from first principles under dual utility with adverse selection. The form of the optimal menu depends on which component of the agent's private information drives the loss distribution, explaining the structural difference observed in practice between property/casualty insurance (deductibles) and medical malpractice insurance (coverage limits).

## Theory / model

**Setup.** An agent faces a random loss $$L$$ distributed on $$[0, \bar L]$$. The agent's private type $$\theta \in \Theta = [\underline\theta, \bar\theta]$$ parameterizes the conditional loss distribution $$H_\theta : \mathbb{R}_+ \to [0,1]$$ via $$H_\theta(l) = \Pr(L \leq l \mid \theta)$$. Higher types face stochastically larger losses in first-order stochastic dominance; $$H_\theta$$ is decreasing in $$\theta$$. Types are distributed according to $$F$$ with density $$f$$ (p. 2587).

Two canonical cases illustrate the model. When the type is the accident probability (Rothschild and Stiglitz 1976; Stiglitz 1977 classical setting):

$$
H_\theta(l) = (1 - \theta) + \theta Q(l), \tag{1}
$$

where $$Q$$ is a fixed conditional loss distribution given an accident (p. 2588, equation 1). When the type scales the loss magnitude:

$$
H_\theta(l) = Q(l / \theta),
$$

so all types face the same accident probability but higher types incur proportionally larger losses (p. 2588).

**Dual (Yaari) utility.** Agents have Yaari (1987) dual utility determined by a probability distortion function $$g : [0,1] \to [0,1]$$, increasing, absolutely continuous, with $$g(p) \leq p$$ (weak risk aversion). For a random total loss $$x$$ distributed according to $$H$$, the certainty equivalent is (p. 2589):

$$
CE(x) = -\int_0^\infty \left[1 - g(H(s))\right] ds. \tag{2}
$$

Dual utility modifies the expectation operator by reweighting each loss level $$s$$ by $$g'(H(s))$$; the condition $$g(p) \leq p$$ implies the agent overweights the probability of large losses, generating first-order risk aversion (the risk premium is proportional to the standard deviation of the loss, not its variance). A key property for tractability is additivity in nonrandom transfers: $$CE(x + t) = CE(x) + t$$, which makes the mechanism design problem separable in premia (p. 2589).

Koszegi and Rabin (2006) loss-averse preferences with linear utility over outcomes correspond to the distortion $$g(p) = (2-\lambda)p + (\lambda-1)p^2$$ for $$\lambda \in (1,2]$$, a special case of dual utility (p. 2590).

**Insurance contracts.** A direct mechanism offers a menu of retention functions $$R(\cdot, \theta)$$ and premia $$t(\theta)$$, where $$R(l, \theta) \in [0, l]$$ is the share of loss $$l$$ retained by type $$\theta$$. Two ex post moral hazard conditions (Assumption 1, p. 2591) restrict the retention slope: $$(i)$$ $$\partial R(l,\theta)/\partial l \geq 0$$ (the agent cannot inflate a reported loss to reduce retention), and $$(ii)$$ $$\partial(l - R(l,\theta))/\partial l = 1 - \partial R/\partial l \geq 0$$ (the agent cannot hide part of a loss to claim higher indemnity). Together these require $$\partial R(l,\theta)/\partial l \in [0,1]$$ almost everywhere.

Under the additivity of dual utility, the certainty equivalent of contract $$(R(\cdot,\theta), t(\theta))$$ to type $$\theta$$ is (p. 2591):

$$
U(\theta) = -t(\theta) - \int_0^{\bar L} \left[1 - g(H_\theta(l))\right] \frac{\partial R(l,\theta)}{\partial l}\, dl. \tag{3}
$$

## Method

**Incentive compatibility envelope.** Dual utility's linearity in $$\partial R / \partial l$$ permits a simple envelope characterization (Proposition 1, p. 2593). Any incentive-compatible mechanism satisfies:

$$
U(\theta) = U(\underline\theta) + \int_{\underline\theta}^{\theta} \left[\int_0^{\bar L} \frac{\partial R(l,s)}{\partial l}\, g'(H_s(l))\, \frac{\partial H_s(l)}{\partial s}\, dl \right] ds. \tag{4}
$$

Integration by parts yields the insurer's expected profit as a functional of $$R$$ alone:

$$
\pi(R) = \int_{\underline\theta}^{\bar\theta} \left[-E[L(\theta)] - \int_0^{\bar L} \frac{\partial R(l,\theta)}{\partial l}\, J(l,\theta)\, dl \right] f(\theta)\, d\theta - U(\underline\theta), \tag{5}
$$

where the **virtual value** (analogous to the Myerson virtual value in mechanism design) is (p. 2593):

$$
J(l,\theta) = H_\theta(l) - g(H_\theta(l)) + \frac{1-F(\theta)}{f(\theta)}\, g'(H_\theta(l))\, \frac{\partial H_\theta(l)}{\partial \theta}. \tag{6}
$$

The first two terms $$H_\theta(l) - g(H_\theta(l)) \geq 0$$ measure the efficiency gain from covering the marginal loss at level $$l$$ for type $$\theta$$ (the agent's valuation exceeds the insurer's cost because of risk aversion). The third term is the information rent cost: since $$\partial H_\theta / \partial\theta < 0$$, the product $$g'(H_\theta(l))(\partial H_\theta/\partial\theta)$$ is nonpositive, so higher types require larger information rents that reduce the effective profit from insuring them.

**Pointwise linear optimization.** For each $$\theta$$, the integrand in (5) is linear in $$\partial R(l,\theta)/\partial l$$. The feasible set of retention functions satisfying Assumption 1 is convex. By Bauer's maximum principle, the maximum is attained at an extreme point of this set, and extreme points of the unit ball of Lipschitz-1 functions on $$[0, \bar L]$$ have derivative in $$\{0,1\}$$ almost everywhere (p. 2608). The pointwise optimum therefore sets:

$$
\frac{\partial R^*(l,\theta)}{\partial l} = \mathbf{1}\{J(l,\theta) \leq 0\}. \tag{7}
$$

When $$J(l,\theta)$$ is nondecreasing in $$\theta$$ for all $$l$$ (the regularity condition of Theorem 1), the resulting $$R^*$$ is submodular: $$\partial R(l,\theta') / \partial l \leq \partial R(l,\theta) / \partial l$$ for $$\theta' > \theta$$, meaning higher-risk types receive more coverage at each loss level. Proposition 1(ii) shows submodularity of $$R$$ is sufficient for incentive compatibility; thus the pointwise solution is also the global optimum of the original screening problem.

**Deductible and coverage limit forms.** A deductible $$D(\theta)$$ and a coverage limit $$C(\theta)$$ correspond to (p. 2599):

$$
R_D(l,\theta) = \begin{cases} l, & l < D(\theta) \\ D(\theta), & l \geq D(\theta) \end{cases}, \qquad R_C(l,\theta) = \begin{cases} 0, & l \leq C(\theta) \\ l - C(\theta), & l > C(\theta) \end{cases}. \tag{8}
$$

Whether $$J(l,\theta)$$ crosses zero from below (deductibles) or from above (coverage limits) as a function of $$l$$ is governed by the direction of private information. In the loss-probability case (equation (1)), $$J$$ typically crosses from below, yielding deductibles optimal. In the loss-magnitude case, $$J$$ crosses from above, yielding coverage limits optimal. Gershkov et al. (2022) develop related tools for nonexpected utility in auction settings.

Liang, Zou, and Jiang (2022) study a two-type variant with distortion risk measures; the full continuum is handled here by the pointwise separability of the profit functional.

## Empirical specifications

This paper presents no empirical analysis; all results are theorems, propositions, and corollaries derived from the model. Examples 1 through 4 illustrate the general results with specific parametric choices for the distortion function $$g$$ and the type-conditional loss distribution $$H_\theta$$, but use no data. The paper cites empirical patterns from prior literature (household deductible choices in Sydnor 2010; Barseghyan et al. 2013; medical malpractice coverage limits in Silver et al. 2015) for motivation only.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| None | Pure theoretical analysis; no datasets collected or analyzed | n/a |

This is a theory paper. Empirical patterns cited in the introduction and related literature (household insurance data, malpractice claims) come from prior published work, not new data collection.

## When to read the full paper

Use the [original](https://doi.org/10.1257/aer.20221247) if you are: extending the mechanism design approach to other nonexpected utility frameworks beyond dual utility; studying competitive (multi-insurer) adverse selection with probability-distorting agents; analyzing the welfare implications of mandating deductibles vs coverage limits in regulated insurance markets; examining when coinsurance (linear contracts, which are never optimal under this framework) is appropriate; or working through the formal proofs and examples in the appendix and online appendices.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(10), October 2023. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. The paper is under AEA copyright; no CC license was found. Extract-only: do not reproduce the full text.

> Gershkov, Alex, Benny Moldovanu, Philipp Strack, and Mengxi Zhang.
> "Optimal Insurance: Dual Utility, Random Losses, and Adverse Selection."
> *American Economic Review* 113, no. 10 (October 2023): 2581–2614.
> DOI: 10.1257/aer.20221247. © 2023 American Economic Association.


==============================================================================
# Voice of Monetary Policy: Gorodnichenko, Pham & Talavera (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/gorodnichenko-et-al-voice-monetary-policy-2023/
# Distilled: A deep learning model detects emotions in Fed chair voices during FOMC press conference Q&A sessions; a more positive voice tone raises S&P 500 returns by roughly 100 basis points over five days, reduces VIX, lowers inflation expectations, and appreciates the dollar against the euro, after controlling for policy actions and text sentiment. American Economic Review 113(2) 2023, paywalled. Seven core results with source locators, the emotion-detection model, VoiceTone construction, and the local-projections specification. LLM-distilled, not human-verified, not reproduced.
# Tags: paper-summary, monetary-policy, central-banking, text-as-data
==============================================================================

**What this is.** The paper's core results, the VoiceTone construction and the deep-learning emotion model, and the local-projections specification: enough to know what it found and how, without reading all 37 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1257/aer.20220129) or the [preprint](https://doi.org/10.2139/ssrn.3809564).

## TL;DR

Gorodnichenko, Pham, and Talavera (2023) build a deep learning model that classifies emotions from audio features of Fed chair answers during FOMC press conference Q&A sessions (April 2011 to June 2019) into positive, negative, or neutral. They aggregate these into a VoiceTone measure, then run local projections following Jorda (2005) of daily returns on 14 financial outcomes against VoiceTone and TextSentiment, controlling for Swanson (2021) policy shocks (FFR, FG, AP) and the Wu and Xia (2016) shadow rate. A one-unit increase in VoiceTone raises S&P 500 returns by roughly 100 basis points over five days, reduces VIX, lowers inflation expectations, and appreciates the dollar against the euro. The bond market shows no significant response. Text sentiment (BERT-based) does not have a statistically significant effect on stock prices in this 36-meeting sample. The results are robust to Fed chair fixed effects, alternative text sentiment measures (RoBERTa, FinBERT, search-and-count, human classification per Cieslak, Morse, and Vissing-Jorgensen (2019) style controls), and a high-frequency intraday analysis showing a +1 bp immediate impact on impact. A concurrent paper by Curti and Kazinnik (2021) confirms nonverbal effects through facial expressions.

## Core results

Estimates from nonparametric accelerated bootstrap with 90% bias-corrected confidence intervals. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Positive voice tone raises S&P 500 (SPY) returns**: effect builds over days, significant at 10% | Figure 3 Panel B, p. 562 | +~100 bps at h=5 days for +1 unit VoiceTone; +~75 bps per +1 SD (abstract, p. 550) |
| R2 | **Positive voice tone reduces stock market volatility (VIX)**: economically significant negative effect, builds over 15 days | Figure 4 Panel B, p. 563-564 | Reduction roughly equal to increase from a 1-SD FG shock (paper's comparison, p. 563) |
| R3 | **Bond market (GOVT ETF) does not respond significantly to voice tone**: statistically insignificant at all horizons | Figure 7 Panel B, p. 567 | Point estimates fluctuate near zero; 90% CI contains zero throughout h=0-15 days |
| R4 | **Positive voice tone raises GOVT-TIP spread** (signals lower inflation expectations): near zero on impact, builds and peaks at h~10 days | Figure 10, p. 570 | Gradual increase in spread, statistically significant at longer horizons; GLD (gold) price shows similar pattern (Figure 11, p. 571) |
| R5 | **Positive voice tone leads to dollar appreciation vs. euro** (EUR/USD coefficient positive): builds over 5-10 days | Figure 13, p. 573 | ~5-7 bps (from figure scale), marginally significant; JPY/USD shows no significant response |
| R6 | **Text sentiment (BERT) does not significantly move SPY**: positive point estimates but inside CI throughout | Figure 3 Panel C, p. 562 | Positive point estimates across all horizons, statistically insignificant in this sample |
| R7 | **Intraday (flow spec): positive voice tone raises SPY by +1 bp on impact**, answer by answer | Figure 15 top panels, p. 576-577 | +1 bp at h=0 minutes, statistically significant; cumulative spec shows gradual build during Q&A session |

**Overall (paper's conclusion).** Voice tone conveys economically significant information that moves financial markets across multiple asset classes, beyond the content of spoken words and beyond the Fed's explicit policy actions. The effect builds gradually over days, consistent with information frictions and slow incorporation of nonverbal signals into prices via continued media coverage (Figure 16, p. 578-580).

## Theory / model

The paper has no formal structural model. The empirical strategy tests whether voice tone carries market-moving information orthogonal to text and policy actions, under two interpretations that the data cannot distinguish.

**Interpretation 1 (forward-guidance signaling, pp. 576-577).** A positive voice tone signals that the Fed is unlikely to tighten policy in the near future. Under the zero lower bound, this attenuates rate risk, reduces perceived future volatility (VIX falls), lowers discount uncertainty (equity prices rise), and, since monetary tightening is a lower-probability event, reduces the expected future path of inflation (GOVT-TIP spread rises). Responses to voice tone are qualitatively similar to those for the FG shock from Swanson (2021), consistent with this interpretation.

**Interpretation 2 (Fed information effect, Romer and Romer 2000, p. 577).** The Fed chair's positive tone signals superior private information about a stronger economic outlook. A more upbeat assessment reduces uncertainty, may lower the rate-risk component of yields, and raises equity valuations. The paper tests this by checking whether positive-tone press conferences are followed by more dovish Fed tweets on days h=1 to 15; Figure 16 (p. 578-579) confirms this pattern, consistent with information revelation.

**Identification.** The design controls for observable channels: VoiceTone enters after partialling out TextSentiment (BERT-based, built from statements, remarks, and Q&A responses), three identified policy shocks from Swanson (2021) (FFR, FG, AP), the Wu and Xia (2016) shadow rate, and an indicator for meetings without press conferences. Voice tone and text sentiment are only weakly correlated ($$\rho = 0.37$$ (Pearson; Spearman = 0.30) for statement text sentiment vs. voice tone; $$\rho = 0.29$$ for Q&A text sentiment vs. voice tone; pp. 557-558), providing independent variation. Hansen and McMahon (2016) document that FOMC textual content carries significant information; controlling for it isolates the vocal dimension. The design is selection-on-observables: no instrument for voice tone is used, and the identifying assumption is that residual VoiceTone variation is uncorrelated with omitted drivers of financial outcomes after the controls listed above.

## Method

**Step 1: Emotion detection and VoiceTone construction (pp. 552-555).** Each FOMC press conference audio is manually split into individual answers (692 segments across 36 press conferences). For each audio segment, three audio feature sets are extracted via Librosa:

$$
\text{Features}_s = \bigl[\underbrace{128 \text{ mel spectrogram coeff.}}_{\text{loudness at frequency}} ,\; \underbrace{12 \text{ chroma coeff.}}_{\text{melodic/harmonic energy}} ,\; \underbrace{40 \text{ MFCCs}}_{\text{spectral envelope}}\bigr] \quad (180 \text{ features total})
$$

A fully connected neural network with architecture: Input(180) -> Linear(200) -> Linear(200) -> Linear(200) -> Softmax(5) [with dropout rate 0.3 after each hidden layer] is trained on RAVDESS and TESS labeled emotion databases (80% train / 20% test). Accuracy on the test set is 84% overall (equation 1, p. 553):

$$
\text{Accuracy}(y,\hat{y}) = \frac{1}{n} \sum_{i=1}^{n} \mathbf{1}\{\hat{y}_i = y_i\} \tag{1}
$$

with per-class accuracy: angry 87%, sad 84%, neutral 74%, surprised 87%, happy 80%. Each segment is then collapsed to positive (happy or pleasantly surprised), negative (sad or angry), or neutral. Press-conference-level VoiceTone is (equation 2, p. 554):

$$
\text{VoiceTone} = \frac{\text{Positive answers} - \text{Negative answers}}{\text{Positive answers} + \text{Negative answers}}, \qquad \text{VoiceTone} \in [-1, +1] \tag{2}
$$

The measure is constructed for three Fed chairs: Bernanke (12 press conferences), Yellen (16), and Powell (8). Mean VoiceTone is +0.64 (Bernanke), -0.13 (Yellen), -0.30 (Powell); within-chair variation is substantial (Table 1, p. 558).

**Step 2: TextSentiment via BERT (pp. 556-557).** BERT base model (12 layers, 768 hidden states, 110M parameters) generates word embeddings for each paragraph of FOMC policy texts. A neural network trained on hand-scored statements from 1997-2010 (scored -10 to +10 on dovish-hawkish axis by research assistants; 81% accuracy) classifies each paragraph as dovish (score $$\geq +0.5$$), hawkish (score $$\leq -0.5$$), or neutral. TextSentiment aggregates over statements, opening remarks, and Q&A responses (equation 3, p. 557):

$$
\text{TextSentiment} = \frac{\text{Dovish text} - \text{Hawkish text}}{\text{Dovish text} + \text{Hawkish text}}, \qquad \text{TextSentiment} \in [-1, +1] \tag{3}
$$

A positive value signals a dovish/expansionary stance. TextSentiment and VoiceTone are weakly correlated ($$\rho \approx 0.29$$ for Q&A text vs. voice tone; $$\rho \approx 0.37$$ (Pearson; Spearman = 0.30) for statement text vs. voice tone), confirming partial independence (pp. 557-558; Figure 1, p. 559).

This method builds on `text-classification` for the BERT-based text scoring and `speech-emotion-recognition` (proposed) for the audio emotion classification.

## Empirical specifications

**Main specification (local projections, Jorda 2005).** For each horizon $$h = 0, 1, \ldots, 15$$ days, equation (4) from p. 560 is estimated separately by OLS:

$$
\text{Outcome}_{t,t+h} = b_0^{(h)} + b_1^{(h)} \text{VoiceTone}_t + b_2^{(h)} \text{TextSentiment}_t + b_3^{(h)} \text{FFRShock}_t + b_4^{(h)} \text{FGShock}_t + b_5^{(h)} \text{APShock}_t + b_6^{(h)} \text{ShadowRate}_t + b_7^{(h)} \mathbf{1}\{\text{NoPressConference}_t\} + \text{error}_t^{(h)} \tag{4}
$$

where $$t$$ indexes FOMC meetings; $$\text{Outcome}_{t,t+h}$$ is the log-return from open on FOMC day $$t$$ to close on day $$t+h$$ (e.g. $$\log(\text{SPY}_{t+h}^{\text{close}}) - \log(\text{SPY}_t^{\text{open}})$$); FFRShock, FGShock, APShock are Swanson (2021) policy shocks (normalized to unit variance over the pre-ZLB period); ShadowRate is from Wu and Xia (2016); and $$\mathbf{1}\{\text{NoPressConference}\} = 1$$ when no press conference was held. The coefficient path $$\{b_1^{(h)}\}_{h=0}^{15}$$ traces the impulse response to a unit increase in voice tone. Standard errors are from nonparametric accelerated bootstrap (90% bias-corrected confidence intervals; p. 561). The sample has 68 FOMC meetings, 36 with press conferences.

Results (R1)-(R6) in the Core results table are from specification (4). Outcomes studied: SPY (R1, R6), VIX (R2), VIXY, VIXM (Figure 5, p. 564-565), GOVT (R3), GOVT-TIP spread (R4), EUR/USD (R5), and others (LQD, GLD, IVR, JPY).

**Robustness variants (pp. 570-575; Figure 14, p. 574).** Adding the Citigroup Economic Surprise Index (Panel A), controlling for volume of corporate earnings announcements (Panel B), adding pre-FOMC media sentiment (Panel C), adding Fed chair fixed effects (Panel D), allowing sign asymmetry in voice tone (Panel E), and replacing BERT with RoBERTa (Panel F), FinBERT (Panel G), search-and-count (Panel H), or human classification (Panel I) leave the voice tone coefficient on SPY stable and statistically significant.

**High-frequency (intraday) specifications (pp. 575-577).** Audio segments are timed to the second; SPY prices are matched to each answer. "Flow" specification (eq. 6', p. 575), answer by answer:

$$
\text{Outcome}_{t \to t+h,m,s} = b_0^{(h)} + b_1^{(h)} \text{VoiceTone}_{t,m,s} + b_2^{(h)} \text{TextSentiment}_{t,m,s} + \lambda_m + \gamma_s + \text{error}^{(h)} \tag{6'}
$$

"Cumulative" specification (eq. 6'', p. 575), cumulating from press conference start $$t_0$$:

$$
\text{Outcome}_{t_0 \to t+h,m,s} = b_0^{(h)} + b_1^{(h)} \text{VoiceTone}_{t_0 \to t,m,s} + b_2^{(h)} \text{TextSentiment}_{t_0 \to t,m,s} + \lambda_m + \gamma_s + \text{error}^{(h)} \tag{6''}
$$

where $$\lambda_m$$ is a meeting fixed effect and $$\gamma_s$$ is a question-order fixed effect; $$h$$ is in minutes; bootstrap clusters at the meeting level. Result R7 uses specification (6').

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| FOMC press conference audio (Fed YouTube, 36 conferences, April 2011-June 2019) | Construct VoiceTone: 692 Q&A audio segments, split per answer; author-introduced dataset | No page yet |
| FOMC statements, opening remarks, Q&A transcripts (Fed website) | Construct TextSentiment; training corpus from 1997-2010 statements | No page yet |
| RAVDESS (Ryerson Audio-Visual Database of Emotional Speech and Song) | Train emotion-detection neural network (80% of combined training set) | No page yet |
| TESS (Toronto Emotional Speech Set) | Train emotion-detection neural network (joint with RAVDESS) | No page yet |
| ETF prices: SPY, VIX, VIXY, VIXM, GOVT, TIP, LQD, LQDH, IVR, GLD (Thomson Reuters, Yahoo Finance, Tiingo) | Outcome variables: daily log returns from open on FOMC day to close on day t+h | No page yet |
| Swanson (2021) policy shocks (FFR, FG, AP) | Controls for monetary policy actions in specification (4) | [No page yet](/wiki/datasets/) |
| Wu and Xia (2016) shadow policy rate | Control for ZLB period and monetary policy cycle in specification (4) | No page yet |
| Nexis Uni news database (23,275 FOMC-related articles, Jan 2011-Jul 2019) | Robustness: MediaSentiment control for media framing before each FOMC meeting | No page yet |
| Federal Reserve Twitter (Board of Governors + regional Fed accounts) | Robustness: post-FOMC tweet sentiment tracks positive voice tone (Figure 16, p. 578-579) | No page yet |

Sample: 68 FOMC meetings, 36 press conferences, April 2011 to June 2019. VoiceTone mean = 0.09, SD = 0.75 across all meetings (Table 1, p. 558).

## When to read the full paper

Use the [original](https://doi.org/10.1257/aer.20220129) (or [preprint](https://doi.org/10.2139/ssrn.3809564)) if you are: studying nonverbal central bank communication as a channel for market expectations; building or extending an emotion-detection model for audio from policy events; evaluating whether text-only NLP models fully capture FOMC communication; checking exact robustness specifications and the Internet Appendix with additional results by ETF maturity and currency; or replicating via the [ICPSR replication data](https://doi.org/10.3886/E178302V1). The locators above point to the exact figures and tables.

## Attribution and rights

Source: peer-reviewed, *The American Economic Review* 113(2), February 2023. Paywalled; preprint available at SSRN. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. Rights are held by the American Economic Association; no CC licence was found in the Crossref record. Reproduce and adapt only per AEA terms.

> Gorodnichenko, Yuriy, Tho Pham, and Oleksandr Talavera. "The Voice of Monetary Policy." *The American Economic Review* 113, no. 2 (February 2023): 548-584. DOI: 10.1257/aer.20220129.


==============================================================================
# Micro Anatomy of Macro Consumption Adjustments: Guntin, Ottonello & Perez (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/guntin-et-al-micro-anatomy-macro-consumption-2023/
# Distilled: Documents that consumption-income elasticities are near unity across all income groups during macro consumption crises (Euro crisis and emerging-market sudden stops), including among top-income and asset-rich households; a calibrated heterogeneous-agent model shows the permanent-income view explains the micro patterns while credit-tightening theories predict a cross-sectional pattern inconsistent with the data. American Economic Review 2023, AEA copyright (free-to-read after 12-month embargo). Seven core results with source locators, datasets used, the model, and the empirical specifications.
# Tags: paper-summary, macro, household-finance, sudden-stops, business-cycles, panel-regression
==============================================================================

**What this is.** The paper's core results, the model (heterogeneous-agent small open economy with borrowing constraints), and the empirical specifications, with exact source locators. To replicate or extend, read the full source at the [original](https://doi.org/10.1257/aer.20201931).

## TL;DR

This paper documents the cross-sectional patterns of consumption adjustment during five episodes of large aggregate consumption decline: the Euro crisis in Italy and Spain, and three emerging-market sudden stops (Mexico 1994, Mexico 2008, Peru 2008). The central finding is that consumption-income elasticities are near unity across all income groups, including top-income and asset-rich households, contradicting the expectation from credit-tightening theories that wealthy households should smooth consumption. A calibrated heterogeneous-agent open-economy model shows the permanent-income view of crises, in the tradition of Aguiar and Gopinath (2007), can account for the micro-level patterns. Credit-tightening theories, as in Mendoza (2005) and Eggertsson and Krugman (2012), predict a decreasing elasticity pattern across the income distribution that is at odds with the data. The divergence between the two views has direct implications for fiscal transfer policy effectiveness.

## Core results

Magnitudes from source tables; locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Average consumption-income elasticity is near 1.0 across all five episodes; large consumption adjustments observed throughout the income distribution | Table 1, Panel A, p. 2209 | Average across episodes: 0.92; by episode: Italy 1.13, Spain 0.97, Mexico 1994 0.78, Mexico 2008 0.73, Peru 0.99 |
| R2 | Top-income household elasticities are similar to or larger than the economy average; income-rich households do not smooth consumption during these crises | Table 1, Panel A, p. 2209 | Top-decile mean: 0.93; by episode: Italy 0.95, Spain 0.90, Mexico 1994 0.79, Mexico 2008 0.88, Peru 1.15 |
| R3 | Households holding liquid assets show consumption-income elasticities near the average, ruling out a hand-to-mouth interpretation for the top-income result | Table 1, Panel B, p. 2209 | Liquid-asset holders: average elasticity 0.86, top-income elasticity 1.01; defined as holding liquid assets exceeding two weeks of income per Kaplan, Violante, and Weidner (2014) |
| R4 | High consumption-income elasticities appear across all observable household characteristics: age group, education level, geography, employment status, and economic sector | Table 2, p. 2212 | All subgroups show elasticities broadly near or above 1; no systematic pattern concentrating the result in a specific demographic group |
| R5 | The permanent-income (PI) model calibrated for Italy reproduces a flat elasticity pattern close to 1 for all income deciles, matching the data; the result is robust to multiple extensions | Figure 5, p. 2220; Table 3 (calibration), p. 2218 | PI model predicts elasticities close to 1 across all deciles; pattern robust to heterogeneous income loadings, negative asset revaluations, and uncertainty shocks (Panels A-D) |
| R6 | The credit-tightening (CT) model predicts a decreasing elasticity pattern across the income distribution (rich smooth, poor adjust more), at odds with the observed flat or increasing pattern | Figure 7, Panel B, p. 2226 | CT model predicts rich-household elasticities near 0 and poor-household elasticities well above 1; data show the opposite |
| R7 | Fiscal transfer stimulus is less effective under the PI crisis experiment than under the CT crisis; the MPC from a one-time transfer is positive but decreasing in income in all scenarios | Figure 8, p. 2228 | MPC from transfer is highest under the CT crisis (borrowing-constrained households have high MPC), lowest under the PI crisis; PI-crisis MPC close to steady-state transitory-shock MPC |

**Overall (paper's conclusion).** The consumption-income elasticities observed during these crises are large and broadly uniform across the income distribution, including for households with liquid assets that should be able to smooth under borrowing-constraint theories. The permanent-income view of crises can account for these patterns analytically and quantitatively. Credit-tightening theories face a challenge explaining why income-rich households adjust consumption as much as the average. The difference has policy bite: fiscal transfers are less effective in stimulating consumption when the crisis reflects a permanent income decline than when it stems from a borrowing-constraint tightening.

## Theory / model

The model is a heterogeneous-agent small open economy with a continuum of households (pp. 2215-2216). Each household has preferences over an infinite consumption stream (equation 1, p. 2215):

$$
E_0 \sum_{t=0}^{\infty} \beta^t u(c_{it}), \tag{1}
$$

where $$u(\cdot)$$ is increasing and concave, $$c_{it}$$ is household $$i$$'s consumption in period $$t$$, and $$\beta \in (0,1)$$ is the discount factor. Each period the household receives an endowment $$y_{it} = h(\mu_{it}, Y_t)$$, where $$\mu_{it}$$ is idiosyncratic and $$Y_t$$ is aggregate income with $$\int h(\mu_{it}, Y_t)\, di = Y_t$$; the baseline sets $$y_{it} = \mu_{it} Y_t$$. Asset markets are incomplete; households save and borrow only in a riskless bond. The budget constraint and borrowing constraint are (equations 2-3, p. 2215):

$$
c_{it} = y_{it} - a_{i,t+1} + (1+r)\, a_{it}, \tag{2}
$$

$$
a_{i,t+1} \geq -\kappa, \quad \kappa > 0, \tag{3}
$$

where $$a_{it}$$ are bond holdings and $$r$$ is the international interest rate.

**Analytical characterization (Proposition 1, pp. 2216-2217).** Under quadratic utility $$u(c) = ac - bc^2$$ and proportional endowment structure, iterating the Euler equation yields optimal consumption (equation 4, p. 2216):

$$
c_{it} = r a_{it} + \frac{r}{1+r} E_t\!\left[\sum_{s=0}^{\infty} \frac{y_{it+s}}{(1+r)^s}\right] - \frac{r}{1+r} E_t\!\left[\sum_{s=0}^{\infty} \frac{\lambda_{it+s}}{(1+r)^s}\right], \tag{4}
$$

where $$\lambda_{it}$$ is the Lagrange multiplier on the borrowing constraint. For a permanent aggregate income shock ($$Y_{t+h} = Y_t < Y_{ss}$$ for all $$h \geq 0$$) and small interest rates ($$r \to 0$$), Proposition 1 states that the consumption-income elasticity is $$\varepsilon_{cy} = 1$$ for both constrained and permanently unconstrained households: the proportional aggregate shock reduces permanent income of all households proportionally, generating a flat cross-sectional elasticity near 1.

**Credit-tightening extension (Proposition 2, pp. 2224-2225).** The CT crisis uses a borrowing constraint that depends on aggregate income (equation 6, p. 2224):

$$
a_{i,t+1} \geq -\kappa\, f(Y_t), \tag{6}
$$

where $$f(Y_t) \geq 0$$ is non-decreasing; calibrated as $$f(Y_t) = Y_t^{\nu}$$ with $$\nu = 2.7$$. Under a mean-reverting transitory income shock plus constraint tightening, Proposition 2 shows that permanently unconstrained households have $$\varepsilon_{cy} < 1$$ (close to 0 for a highly transitory shock) while constrained households have $$\varepsilon_{cy} = g(\varepsilon_{fY}) > 1$$. Since income-rich households are more likely to be permanently unconstrained, the CT view generates a decreasing elasticity pattern across the income distribution: rich households smooth, poor households adjust.

**Emerging-market extension with nonhomotheticities (p. 2222).** To account for the increasing elasticity pattern in emerging markets, where many households are close to subsistence consumption, the model adopts Stone-Geary preferences:

$$
u(c_{it}) = \frac{(c_{it} - \underline{c})^{1-\gamma}}{1 - \gamma},
$$

where $$\underline{c}$$ is the subsistence consumption level. Low-income households near $$\underline{c}$$ have a strong desire to smooth and therefore a lower consumption-income elasticity, generating the increasing pattern with income observed in Mexico and Peru.

## Method

The quantitative model uses CRRA utility $$u(c) = c^{1-\gamma}/(1-\gamma)$$ with $$\gamma = 2$$ and an AR(1) idiosyncratic income process in logs (p. 2217):

$$
\ln\mu_{it} = \rho_\mu \ln\mu_{i,t-1} + \sigma_\mu\, \varepsilon_{it}, \quad \varepsilon_{it} \sim N\!\left(-\frac{\sigma_\mu}{2(1+\rho_\mu)},\, 1\right).
$$

The model is solved via `value-function-iteration` on a discrete state space. Steady-state calibration targets two moments from Italian SHIW data: the liquid wealth-to-income ratio (0.87) and the hand-to-mouth share (0.23) (Table 4, p. 2219), yielding $$\beta = 0.90$$, $$r = 0.02$$, $$\rho_\mu = 0.88$$, $$\sigma_\mu = 0.26$$, and $$\kappa = 0.23$$ (Table 3, p. 2218). The model is assessed against untargeted moments including income and wealth distribution statistics (Table 4, p. 2219).

The heterogeneous-loading extension (equation 5, p. 2219) replaces $$y_{it} = \mu_{it} Y_t$$ with:

$$
y_{it} = \mu_{it} Y_t^{\Gamma(\mu_{it})},
$$

where $$\Gamma(\mu_{it})$$ is estimated nonparametrically from the income dynamics of each decile in Italian crisis data.

The empirical measurement methodology follows Blundell, Pistaferri, and Preston (2008): income and consumption are residualized by projecting on household observables (family size, number of children, head's sex, age, education, and geographic dummies) and time trends before computing group-level averages.

## Empirical specifications

**Consumption-income elasticity (baseline measurement, p. 2208).** For income group $$j$$, the consumption-income elasticity is:

$$
\hat{\varepsilon}^j_{cy} = \frac{\Delta_h \log\bar{c}_{j,\tau+h}}{\Delta_h \log\bar{y}_{j,\tau+h}},
$$

where $$\bar{c}_{j,t} \equiv \frac{1}{n_{j,t}} \sum_{i \in \mathcal{I}_{j,t}} c_{i,t}$$ and $$\bar{y}_{j,t} \equiv \frac{1}{n_{j,t}} \sum_{i \in \mathcal{I}_{j,t}} y_{i,t}$$ are group-level averages, $$\tau$$ is the output peak, and $$h$$ is the peak-to-trough interval. Income $$Y$$ is monetary after-tax nonfinancial income; consumption $$C$$ is expenditure on nondurable goods and services; both are deflated by CPI and residualized from household observable characteristics. Confidence intervals use 2,000 bootstrap replications. Synthetic income-group cohorts allow application to countries with only cross-sectional data; results hold for fixed households where panel data exist (Italy, Peru). Episode windows: Italy 2006-2014; Spain 2008-2013; Mexico 1994-1996; Mexico 2006-2010; Peru 2007-2010 (footnote 5, p. 2208).

**Business-cycle comparison (p. 2214; Figure 4).** For Italy across biennial periods, and analogously for the US using CEX data (1980-2010), the specification is:

$$
\Delta \ln c_{q,t} = \alpha_q + \beta_q\, \Delta \ln y_{q,t} + \varepsilon_{q,t},
$$

where $$c_{q,t}$$ and $$y_{q,t}$$ are average residualized consumption and income in quintile $$q$$ at year $$t$$. Estimates of $$\beta_q$$ are close to 1 for all quintiles in Italy, and range from 0.2 to 0.6 for the United States, consistent with the aggregate evidence that Italy exhibits less consumption smoothing than the US.

**Crisis experiments.** The model replicates the same elasticity statistic computed from the data. Under the PI experiment: aggregate income follows $$\log Y_t = \log Y_{t-1} + \rho_g^t \varepsilon_Y$$ with $$\varepsilon_Y = -0.15$$ and $$\rho_g = 0.24$$, calibrated to match the aggregate elasticity from Section I. Under the CT experiment: income is transitory (persistence $$\rho_Y = 0.9$$) and the borrowing constraint tightens via $$f(Y_t) = Y_t^{\nu}$$ with $$\nu = 2.7$$; the sensitivity of the constraint to aggregate income is identified by the aggregate consumption-income elasticity (Figure 7, p. 2226).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Survey on Household Income and Wealth (SHIW), Banca d'Italia | Italy: household income, consumption, wealth, demographics; main calibration and crisis episode 2006-2014 | no page yet |
| Encuesta de Presupuestos Familiares (EPF), INE Spain | Spain: household income and nondurable consumption cross-section; crisis episode 2008-2013 | no page yet |
| Encuesta Financiera de las Familias (EFF), Banco de Espana | Spain: supplement for household asset holdings and debt data | no page yet |
| Encuesta Nacional de Ingresos y Gastos de los Hogares (ENIGH), INEGI Mexico | Mexico: household income and consumption; two episodes (Mexico 1994-1996 and 2006-2010) | no page yet |
| Encuesta Nacional de Hogares (ENAHO), INEI Peru | Peru: household income and consumption; episode 2007-2010 | no page yet |
| FRED and OECD | Aggregate output and consumption series for macro context and episode identification (Figure 2 sources) | [FRED](/wiki/datasets/fred/) |
| Consumer Expenditure Survey (CEX), BLS USA | US comparison of consumption-income elasticities, 1980-2010 | no page yet |

Total sample: 90,199 household-observations across five episodes (Italy 7,067; Spain 21,802; Mexico 1994 13,122; Mexico 2008 27,038; Peru 21,170; Table 1, p. 2209). Data are annual or biennial depending on the survey.

## When to read the full paper

Read the [original](https://doi.org/10.1257/aer.20201931) if you are: building or calibrating heterogeneous-agent open-economy models (online Appendix D has full calibration details including aggregate risk, closed-economy variants, and the interest-rate shock extensions); comparing micro distributional evidence across crisis types; assessing credit-tightening models against consumption survey data from Europe and Latin America; or designing fiscal transfer policies for macro crises and want the formal policy-experiment details (Section IIIb and Appendix D4).

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(8), August 2023. Copyright 2023 American Economic Association; article freely readable at [doi.org/10.1257/aer.20201931](https://doi.org/10.1257/aer.20201931) after the AEA 12-month embargo; no Creative Commons licence assigned.

This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**.

> Guntin, Rafael, Pablo Ottonello, and Diego J. Perez. "The Micro Anatomy of Macro Consumption Adjustments." *American Economic Review* 113, no. 8 (August 2023): 2201-2231. DOI: 10.1257/aer.20201931.


==============================================================================
# Not Too Late: Guryan, Ludwig et al. (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/guryan-et-al-too-late-improving-academic-2023/
# Distilled: Two large-scale RCTs (n=5,343) of high-dosage tutoring with paraprofessional tutors in Chicago public high schools find math test score gains of 0.18 SD (Study 1) and 0.40 SD (Study 2), persisting at 0.23 SD one to two years later. American Economic Review 2023, paywalled. Nine core results with source locators, datasets used, the Lazear-based classroom model, and ITT/TOT regression specifications.
# Tags: paper-summary, education, human-capital, inequality, tutoring, adolescents
==============================================================================

**What this is.** A distilled skeleton of Guryan, Ludwig et al. (2023). Read the
[original article](https://doi.org/10.1257/aer.20210434) to replicate or extend; this
page records the headline results with PDF locators, the model equations, and the
datasets used, as extracted by an LLM and not yet human-verified.

## TL;DR

Two separate randomized controlled trials (RCTs) of high-dosage tutoring for
disadvantaged high school students in Chicago test whether paraprofessional tutors
working at a 2:1 student-to-tutor ratio for 50 minutes per school day can raise math
achievement. Study 1 (n = 2,633, 2013-2014) and Study 2 (n = 2,710, 2014-2015) both
find large positive treatment effects on math test scores (0.18 SD and 0.40 SD,
respectively) and math course grades, with no detectable effect on arrests or disciplinary
outcomes. Pooling the two studies, the treatment-on-the-treated (TOT) effect on math
test scores is 0.28 SD. Effects persist: one to two years after tutoring, math test
scores remain 0.23 SD higher in eleventh grade. The benefit-cost ratio (2.4-8.0 depending
on study) is in the range of well-known early childhood programs such as the Abecedarian
Project and the Perry Preschool Program. The results are consistent with Lazear (2001)'s
model of classroom production and the hypothesis that personalization of instruction
drives the gains, rather than a generic mentoring channel.

## Core results

| # | Result | Locator | Magnitude as reported |
|---|--------|---------|----------------------|
| R1 | TOT effect on math test score, Study 1 Year 1 | Table 3, p. 749 | TOT = 0.179 SD (0.066)\*\*\*, ITT = 0.091 (0.035)\*\*\*; CCM = -0.111 |
| R2 | TOT effect on math GPA, Study 1 Year 1 | Table 3, p. 749 | TOT = 0.571 GPA pts (0.079)\*\*\*; CCM = 1.617 (~C- to ~C+) |
| R3 | TOT decline in math course failure rate, Study 1 Year 1 | Table 3, p. 749 | TOT = -0.086 (0.026)\*\*\*; 48% decline from CCM 0.178 |
| R4 | TOT effect on math test score, Study 2 Year 1 | Table 4, p. 750 | TOT = 0.398 SD (0.105)\*\*\*, ITT = 0.135 (0.036)\*\*\*; CCM = -0.172 |
| R5 | TOT effect on math test score, pooled Year 1 | Table 5, p. 751 | TOT = 0.282 SD (0.059)\*\*\*; CCM = -0.143 |
| R6 | TOT effect on math GPA, pooled Year 1 | Table 5, p. 751 | TOT = 0.516 GPA pts (0.069)\*\*\*; CCM = 1.675 |
| R7 | TOT decline in math course failure rate, pooled Year 1 | Table 5, p. 751 | TOT = -0.086 (0.022)\*\*\*; 47% decline from CCM 0.184 |
| R8 | Persistent TOT on 11th grade math test score | Table 7, p. 753 | TOT = 0.232 SD (0.065)\*\*\*; CCM = -0.147 |
| R9 | TOT effect on on-time high school graduation | Table 7, p. 753 | TOT = 1.3pp (3.2pp); CCM = 78.3%; FDR q = 0.677 (null) |

**Overall (paper's conclusion).** High-dosage tutoring with paraprofessional tutors
raises math test scores by 0.18-0.40 SD within one academic year for disadvantaged
high school students - effect sizes comparable to what Fryer (2014) finds for tutoring
as a component of no-excuses charter schools. Effects persist at 0.23 SD in eleventh
grade math test scores and 0.25 GPA points (Table 7). No statistically significant
effects emerge on disciplinary outcomes, arrests, or graduation (the graduation
point estimate is positive but imprecise at 1.3pp). Benefit-cost ratios of 2.4-8.0
are comparable to the Abecedarian Project (1.9-2.2) and Perry Preschool (3.9-6.8).
The evidence is consistent with personalization of instruction as the primary mechanism,
supported by Banerjee et al. (2007) and by heterogeneity analysis showing larger
gains in classrooms with more heterogeneous math achievement levels (Figure 3, p. 759).
No detectable treatment effects on grit, conscientiousness, or locus of control rule
out a generic mentoring mechanism (p. 758).

## Theory / model

The paper adapts the Lazear (2001) educational production model to compare
whole-class instruction (with a credentialed teacher) against small-group tutoring
(with a lower-paid paraprofessional tutor). The model helps predict when tutoring
produces the largest gains relative to classroom instruction.

The school has $$S$$ students and a budget $$M$$ to spend on teachers. Teacher
quality depends on the wage according to $$V(w)$$, with $$V'(w) > 0$$. Hiring
$$M/w$$ teachers yields an average class size of

$$n = \frac{wS}{M}. \tag{4.1}$$

Each student's skill level is drawn independently from $$N(\mu, \sigma^2)$$. Following
Lazear (2001), students only learn when there are no classroom disruptions, which occurs
with probability $$p^n$$ where $$p$$ depends on classroom achievement heterogeneity:

$$p(\sigma^2) = \frac{e^{-\sigma^2}}{1 + e^{-\sigma^2}}. \tag{4.2}$$

The school chooses the teacher wage $$w$$ to solve (p. 756)

$$\max_w S V(w) \, p(\sigma^2)^{wS/M}. \tag{4.3}$$

The key comparative static is (p. 756)

$$\frac{\partial w^*}{\partial \sigma^2} = \frac{\left\{\tfrac{S}{M}\left[1 - p(\sigma^2)\right] V(w^*)^2\right\}}{V(w^*) V''(w^*) - V'(w^*)^2}. \tag{4.4}$$

This is negative whenever teacher quality is not too convex in wages - meaning the
optimal wage (and hence class size) decreases as classroom achievement heterogeneity
$$\sigma^2$$ rises. Equivalently, shifting budget toward smaller classes (or tutoring)
becomes more valuable as students become more heterogeneous. The model thus predicts
tutoring gains should be larger in classrooms with higher dispersion in student
achievement levels (confirmed in Figure 3, p. 759), rather than in classrooms with
higher prevalence of behavioral disruptions (Figure 2, p. 758).

The paper also develops a "mentoring" alternative hypothesis - that tutors build
adult relationships that improve noncognitive skills - and uses survey evidence to
rule it out (no detectable effect on grit, conscientiousness, locus of control, or
number of caring adults, p. 758).

## Method

**Estimation.** The paper estimates both the intention-to-treat (ITT) and the
treatment-on-the-treated (TOT) effect. The ITT comes from a simple OLS regression of
the outcome on the randomization indicator (equation (1) in the paper, p. 745):

$$Y_i = \pi_0 + \pi_1 Z_i + X_i \pi_2 + B_i + \varepsilon_i \tag{1}$$

where $$Y_i$$ is a post-randomization outcome for student $$i$$, $$Z_i$$ is an
indicator for assignment to the tutoring offer, $$X_i$$ is a vector of baseline
controls (sociodemographics, prior test scores, GPA, days absent, disciplinary
incidents, arrest history), and $$B_i$$ is a full set of randomization block fixed
effects. Standard errors are heteroskedasticity-robust (clustered by individual in
Study 2 due to the 65-student duplicate overlap).

**TOT via 2SLS.** Because take-up rates are 37-40%, the ITT understates the
per-participant effect. The paper uses random assignment $$Z_i$$ as an instrument
for actual participation $$D_i$$ (having attended at least one tutoring session), as
in Angrist, Imbens, and Rubin (1996) and H. Bloom (1984). The first-stage equation
is (p. 747):

$$D_i = \gamma_0 + \gamma_1 Z_i + X_i \gamma_2 + B_i + \mu_i \tag{2}$$

and the structural equation of interest is:

$$Y_i = \beta_0 + \beta_1 \hat{D}_i + X_i \beta_2 + B_i + \vartheta_i \tag{3}$$

where $$\hat{D}_i$$ is the fitted value from (2). The TOT coefficient $$\beta_1$$
identifies the local average treatment effect (LATE) for compliers.

**Multiple testing.** Outcomes are grouped into four families: (i) mathematics
achievement, (ii) nonmath academic achievement, (iii) school behavior, and
(iv) arrests. The paper reports false discovery rate (FDR) $$q$$-values using
Benjamini and Hochberg (1995). Nonparametric permutation tests (100,000 randomizations)
are also reported to guard against finite-sample inference issues (Young 2019).

## Empirical specifications

**Study 1 sample (2013-2014).** Of 2,633 randomized ninth and tenth grade male
students in 12 Chicago Public Schools, 2,103 enrolled in a study school. The
sample is almost entirely low-income Black and Hispanic students. Study 1 used a
2x2 factorial design that also independently randomized students to a metacognitive
behavioral intervention (Becoming a Man, BAM), which had been separately evaluated in
Heller et al. (2017). Take-up rate for the tutoring offer was 40.2%. First-stage
impact on participation: $$\gamma_1$$ in equation (2) is approximately 0.40
(Table 6, p. 753).

**Study 2 sample (2014-2015).** 2,710 ninth and tenth graders (male and female)
in 15 schools; 36.9% take-up. Study 2 replicated Study 1 in response to preliminary
Study 1 results and public-sector support from the city of Chicago.

**Primary outcomes.** CPS standardized math test scores (EXPLORE for ninth grade,
PLAN for tenth grade, both by ACT Inc.) are the primary outcome, expressed as
CPS-wide $$z$$-scores. Math course GPA and math course failure rate (share of math
courses with a failing grade) are secondary course-record outcomes from CPS
administrative data. In Study 1, a supplemental math test administered by the
Institute for Social Research (ISR) at the University of Michigan corroborates the
CPS test score results (ISR TOT = 0.199 SD, Table 3).

**Specification.** The ITT specification (equation 1) controls for school-level
randomization block fixed effects (Study 1) or school-grade-gender block fixed effects
(Study 2), plus: sociodemographic controls, baseline test scores, prior year GPA,
days absent, suspension days, disciplinary incidents, and arrest history. Missing
covariates are imputed to zero with a missingness indicator included.

**Follow-up outcomes.** Table 7 (p. 753) reports 11th grade outcomes pooling both
studies. The estimating equations are the same as above, but outcomes are measured
one to two years after the intervention year, in the academic year when students
would be in 11th grade. On-time graduation is estimated pooling all students in
both samples (N = 3,594 for on-time graduation, N = 3,614 for ever-graduated).

**Heterogeneity.** Subgroup ITT effects by baseline achievement quartile (Figure 1,
p. 756) show positive math GPA effects across all four quartiles, but math test score
gains only for the top three quartiles - consistent with floor effects in the test for
the lowest-achieving students. Classroom-level heterogeneity interactions (Figures 2
and 3, pp. 758-759) show larger tutoring gains in more heterogeneous classrooms
(by math achievement dispersion), consistent with the personalization channel of
the Lazear (2001) model.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---------|--------------|-----------|
| Chicago Public Schools (CPS) Student Administrative Records | Primary outcomes (test scores, GPA, course failures, attendance, disciplinary actions); enrollment and school records; baseline covariates for Study 1 and Study 2 | no page yet |
| CPS standardized tests (EXPLORE, PLAN by ACT Inc.) | Primary math outcome, expressed as CPS-wide z-scores | no page yet |
| Chicago Police Department (CPD) Arrest Records | Secondary outcome family (violent, property, drug, other arrests) | no page yet |
| ISR Survey Data (Institute for Social Research, Univ. of Michigan) | Math achievement test administered by research team; survey measures of noncognitive skills, adult relationships, risky behavior | no page yet |
| Saga Education internal records | Tutoring attendance and dosage; tutor characteristics; Saga internal math assessments | no page yet |

**Sample scope.** Study 1: 2,633 ninth and tenth grade male students in 12 CPS
high schools, 2013-2014 academic year. Study 2: 2,710 ninth and tenth grade
students (male and female) in 15 CPS high schools, 2014-2015. Pooled N = 5,343.
Average baseline math score in both samples was 8-15 percentile points below the
CPS-wide average (Table 1, p. 746).

## When to read the full paper

Read the original if you are: (a) designing or scaling a tutoring program for
secondary students and need the benefit-cost analysis (Section IV, pp. 759-761);
(b) studying the mechanisms of educational production (Lazear (2001) model,
Section III, pp. 755-759); (c) assessing the credibility of ITT/TOT estimates
in large-scale school RCTs (Section II.D, pp. 745-748 for the analysis plan and
multiple-testing corrections); or (d) building on the evidence for high-dosage
tutoring reviewed in Nickow, Oreopoulos, and Quan (2020). Table 3 and Table 4
are the primary results tables by study; Table 5 (pooled) and Table 7 (persistent
effects) are the synthesis tables of most interest for policy.

## Attribution and rights

Guryan, Jonathan, Jens Ludwig, Monica P. Bhatt, Philip J. Cook, Jonathan M. V.
Davis, Kenneth Dodge, George Farkas, Roland G. Fryer Jr., Susan Mayer, Harold
Pollack, Laurence Steinberg, and Greg Stoddard. 2023. "Not Too Late: Improving
Academic Outcomes among Adolescents." *American Economic Review* 113(3): 738-765.
https://doi.org/10.1257/aer.20210434

Replication data: https://doi.org/10.3886/E182903V1

This page contains LLM-distilled extracts only (extract-only; paywalled article).
Not human-verified. Not reproduced. Read the original for any replication or
downstream use.


==============================================================================
# Too Much Benchmarking in Asset Management: Kashyap, Kovrijnykh, Li & Pavlova (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/kashyap-et-al-there-too-much-benchmarking-2023/
# Distilled: A tractable general equilibrium model shows that incentive contracts for fund managers create a pecuniary externality through equilibrium asset prices: benchmarking inflates the risky asset price, crowds trades, and reduces contract effectiveness for other investors, so the socially optimal contract has less skin in the game and less benchmarking than the privately optimal one. American Economic Review 2023, AEA copyright. Six core results with source locators, the model equations, and the method.
# Tags: paper-summary, asset-pricing, asset-management, benchmarking, general-equilibrium, mechanism-design, theory, peer-reviewed, unreplicated
==============================================================================

**What this is.** The paper's core results, the model it builds on (a two-period CARA general equilibrium with delegated asset management), and the method (analytical optimal contracting): enough to know what it found and how, without reading all 30 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1257/aer.20210476).

## TL;DR

The paper proposes a tractable two-period general equilibrium model of delegated asset management in which benchmarking arises endogenously. When fund managers incur a private, noncontractible cost to manage portfolios, optimal incentive contracts reward them for absolute performance and for performance relative to a benchmark. In general equilibrium, these contracts create a pecuniary externality: benchmarking raises the collective demand for the risky asset, inflating its price and reducing its expected return, which in turn reduces the value of benchmarking for all other fund investors. Because individual fund investors take the stock price as given, they do not internalize this crowding effect and over-incentivize their managers. A constrained social planner, who internalizes the externality, chooses less skin in the game and less benchmarking, and delivers lower asset management costs and a lower (more correctly priced) risky asset.

## Core results

All results are theoretical propositions; magnitudes are qualitative inequalities. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Privately optimal contracts always include benchmarking | Proposition 1(ii), p. 1126 | b\* > 0 whenever x̄ + λ\_D(Δ − ψ)/(γσ²) > 0 |
| R2 | Privately optimal skin in the game lies strictly between perfect risk sharing and full internalization | Proposition 1(i), p. 1126 | a\* ∈ (1/2, 1) |
| R3 | Social planner uses less skin in the game than private equilibrium | Proposition 2(i), p. 1131 | a\*\* < a\* |
| R4 | Social planner uses less benchmarking than private equilibrium | Proposition 2(ii), p. 1131 | b\*\* < b\* (and b\*\*/a\*\* < b\*/a\*) under the same condition as R1 |
| R5 | Private equilibrium inflates the risky asset price above the social optimum | Proposition 3(i), p. 1132 | p\*\* < p\* |
| R6 | Private equilibrium generates excessive risky asset holdings and asset management costs | Proposition 3(ii), p. 1132 | x^{M\*\*} < x^{M\*} and ψx^{M\*\*} < ψx^{M\*} |

**Overall (paper's conclusion).** When all fund investors use incentive contracts, they collectively increase demand for the risky asset, raise its price, and lower the expected return, making the marginal benefit of benchmarking lower for everyone else. Individual investors fail to account for this. A social planner, recognizing the crowding, opts for less incentive provision and less benchmarking. The planner also delivers lower asset management costs and a lower (better priced) risky asset.

## Theory / model

The model is a two-period ($$t = 0, 1$$) general equilibrium with one risky asset (stock) paying dividend $$\tilde{D} \sim N(\mu, \sigma^2)$$ at $$t = 1$$, in net supply $$\bar{x} > 0$$, and one risk-free bond at zero interest in infinite supply. The stock price $$p$$ clears the market at $$t = 0$$.

**Agents.** Three types, population normalized so $$\lambda_D + 2\lambda_M = 1$$:

- Direct investors (fraction $$\lambda_D$$): manage own portfolios.
- Fund investors (fraction $$\lambda_M$$): delegate to managers; can only buy the bond themselves.
- Fund managers (mass $$\lambda_M$$): each works for one fund investor; restricted to investing personal wealth in the bond.

All agents have CARA utility $$U(W) = -e^{-\gamma W}$$ ($$\gamma > 0$$).

**Manager's return.** Managers can access return-augmenting strategies (securities lending, market making, liquidity provision) unavailable to direct investors. The fund's per-share return (equation 1, p. 1118):

$$
r_x = x(\Delta + \tilde{D} - p) + \varepsilon, \tag{1}
$$

where $$x$$ is the manager's position in the risky asset, $$\Delta \geq 0$$ is the expected abnormal return, and $$\varepsilon \sim N(0, \sigma_\varepsilon^2)$$ is idiosyncratic noise from the return-augmenting activities. The manager incurs a private, noncontractible portfolio management cost $$x\psi$$ ($$\psi > 0$$) per share.

**Compensation contract.** Fund investors design linear contracts (equation 2, p. 1119):

$$
w = \hat{a}r_x + b(r_x - r_b) + c = ar_x - br_b + c, \tag{2}
$$

where $$r_b = \tilde{D} - p$$ is the benchmark return (one share of the risky asset), $$a = \hat{a} + b$$ is "skin in the game" (sensitivity to absolute performance), $$b \geq 0$$ is the benchmark sensitivity (relative performance fee), and $$c$$ is a fixed component. Benchmarking shields the manager from dividend variance while still incentivizing risky-asset investment, because performance relative to the benchmark is insensitive to the aggregate dividend shock.

**Equilibrium conditions.** An equilibrium with privately optimal contracts is a contract $$(a^*, b^*, c^*)$$, portfolio choices $$(x^D, x^{M*})$$, and price $$p^*$$ such that: (i) direct investors and managers optimize given $$p^*$$; (ii) fund investors optimize contracts given $$p^*$$ and the manager's incentive constraint (her first-order condition); and (iii) the stock market clears: $$\lambda_D x^D + \lambda_M x^{M*} = \bar{x}$$ (Definition 1, p. 1135). The equilibrium with socially optimal contracts replaces (ii) with a social planner who internalizes the price externality (Definition 2, p. 1135).

## Method

The paper solves both equilibria analytically using first-order conditions. CARA utility with normally distributed returns reduces every agent's problem to an equivalent mean-variance program, yielding closed-form portfolio demands and equilibrium prices. This builds on `principal-agent` and `mechanism-design` primitives and on the `cara-mean-variance-optimization` technique (proposed vocab).

**Portfolio demands and market-clearing price (Lemma 1, p. 1121).** For a given contract $$(a, b, c)$$:

$$
x^D = \frac{\mu - p}{\gamma\sigma^2} \tag{3}
$$

$$
x^M = \frac{\Delta - \psi/a + \mu - p}{a\gamma\sigma^2} + \frac{b}{a} \tag{4}
$$

$$
p = \mu - \gamma\sigma^2\Lambda\!\left(\bar{x} - \lambda_M\frac{b}{a}\right) + \Lambda\frac{\lambda_M}{a}\!\left(\Delta - \frac{\psi}{a}\right) \tag{5}
$$

where $$\Lambda \equiv (\lambda_M/a + \lambda_D)^{-1}$$ is the inverse of the market's effective risk aversion. From (4)-(5), an increase in $$b/a$$ raises manager demand, which raises the equilibrium price and lowers the expected return. This is the price-externality channel: benchmarking inflates the stock price.

**Private equilibrium (Lemma 2, p. 1125).** The fund investor maximizes her expected utility subject to the manager's participation constraint $$U^M \geq u_0$$ and the manager's incentive constraint (her FOC, equation 8, p. 1123):

$$
y = \frac{\Delta - \psi/a + \mu - p}{\gamma\sigma^2}
$$

where $$y = ax - b$$ is the manager's effective risky-asset exposure. The fund investor's FOC with respect to $$b/a$$ (equation 9) equates the marginal benefit of inducing more risky investment against the variance cost. The FOC with respect to $$a$$ (equation 11) trades incentive provision against risk sharing. After substituting the equilibrium price (5), $$a^*$$ solves:

$$
(1 - a^*)\frac{\psi^2}{\gamma\sigma^2 a^{*3}} - (2a^* - 1)\gamma\sigma_\varepsilon^2 = 0. \tag{12}
$$

The first term is the marginal benefit of raising $$a$$ (it reduces the manager's effective cost $$\psi/a$$, incentivizing more risky-asset investment); the second is the marginal cost (higher $$a$$ exposes the manager to more idiosyncratic risk $$\sigma_\varepsilon^2$$). The solution satisfies $$a^* \in (1/2, 1)$$ (Proposition 1(i), p. 1126), because at $$a = 1/2$$ the first term dominates and at $$a = 1$$ the second dominates. The equilibrium benchmark parameter $$b^*$$ follows from equation (13) and the price from:

$$
p^* = \mu - \gamma\sigma^2\bar{x} + \lambda_M\!\left(2\Delta - \psi - \frac{\psi}{a^*}\right) \tag{14}
$$

Benchmarking is optimal (b\* > 0) whenever $$\bar{x} + \lambda_D(\Delta - \psi)/(\gamma\sigma^2) > 0$$ (Proposition 1(ii)), which follows from Holmstrom (1979)'s sufficient-statistic logic: the benchmark return is a signal correlated with the manager's performance, so including it in the contract is always (weakly) beneficial for the principal.

**Social planner (Lemma 4, p. 1130).** The planner maximizes a weighted sum of fund investors' and direct investors' utilities, treating $$p$$ as a function of the contract parameters. The planner's FOC for $$b/a$$ adds a "contracting pecuniary externality" term to equation (9), yielding (equations 18-19, p. 1129):

$$
\Delta - \frac{\lambda_M/a + \lambda_D}{\lambda_M + \lambda_D}\,\psi + \mu - p - \gamma\sigma^2 z = 0. \tag{19}
$$

Compared to the private FOC (equation 9), the cost of incentive provision is now $$[(\lambda_M/a + \lambda_D)/(\lambda_M + \lambda_D)]\psi$$, which exceeds $$\psi$$ whenever $$a < 1$$. The planner perceives benchmarking as more expensive because she accounts for the price inflation it creates. Substituting the equilibrium price, $$a^{**}$$ solves:

$$
(1 - a^{**})\frac{\psi^2}{\gamma\sigma^2 a^{**3}} \cdot \frac{\lambda_D}{\lambda_M + \lambda_D} - (2a^{**} - 1)\gamma\sigma_\varepsilon^2 = 0. \tag{23}
$$

Comparing (12) and (23): the coefficient $$\lambda_D/(\lambda_M + \lambda_D) < 1$$ in (23) makes the first term smaller for the planner, delivering $$a^{**} < a^*$$ (Proposition 2(i)). A parallel argument shows $$b^{**} < b^*$$ (Proposition 2(ii)). Proposition 3 follows by substituting $$(a^{**}, b^{**})$$ into the equilibrium price (equation 25) and holdings (equation 26) and comparing with the private equilibrium.

The paper's mechanism unifies three strands of prior work. Holmstrom (1979)'s sufficient-statistic result explains why benchmarking enters the contract; Holmstrom and Milgrom (1991)'s tractable CARA contracting framework enables the closed-form analysis; and Brennan (1993)'s two-period finding that benchmarking lowers expected returns is replicated and embedded in a welfare framework. The crowding externality is the contracting analogue of the collateral externality in Davila and Korinek (2018). Lorenzoni (2008) also finds a decentralized equilibrium between constrained and unconstrained optima, but with the price ordering reversed; here $$p^{**} < p^* < p^{FB}$$ (Remark 2, p. 1133). Basak and Pavlova (2013) show the same price-inflating effect of benchmarking in dynamic models; this paper adds the optimal-contracting and welfare dimensions.

## Empirical specifications

This is a pure theory paper. There are no empirical specifications, datasets, or estimation procedures. All propositions (Propositions 1-3 and Lemmas 1-4) are derived analytically; all results are qualitative inequalities among equilibrium quantities under privately and socially optimal contracts. The paper does not calibrate to data or estimate model parameters. Online Appendix D provides a tax-implementation analysis of the social optimum. Online Appendix E analyzes extensions, including an effort-based private cost and an endogenous abnormal return from securities lending.

## Datasets used

This paper uses no empirical datasets. All results follow from theoretical propositions about a stylized two-period economy.

## When to read the full paper

Read the [original](https://doi.org/10.1257/aer.20210476) if you are: (i) designing or evaluating incentive contracts for fund managers and need the welfare benchmark (Propositions 2-3 give the comparison); (ii) studying the asset-pricing implications of institutional delegation (Lemma 2, equation 14 gives the equilibrium price formula); (iii) extending the model to passive funds, multiple risky assets (Remark 5, p. 1134), or ESG benchmarks; or (iv) analyzing tax implementations of the social optimum (Online Appendix D). The locators in the Core results table point to the exact propositions.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(4), April 2023. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. The article is freely accessible on the AEA website after the 12-month embargo (now elapsed); redistribution and reproduction rights are reserved by the American Economic Association. Extract-only.

> Kashyap, Anil K., Natalia Kovrijnykh, Jian Li, and Anna Pavlova.
> "Is There Too Much Benchmarking in Asset Management?"
> *American Economic Review* 113, no. 4 (April 2023): 1112-1141.
> DOI: 10.1257/aer.20210476. Copyright 2023 American Economic Association.


==============================================================================
# Banning Gendered Job Ads: Kuhn & Shen (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/kuhn-shen-what-happens-employers-can-2023/
# Distilled: When XMRC.com (a Chinese job board) removed explicit gender requests from all job ads overnight in March 2019, women's share of callbacks to previously male-requesting jobs rose by 61 percent and men's share of callbacks to previously female-requesting jobs rose by 146 percent. The ban generated a large increase in gender-mismatched applications that employers treated relatively well, suggesting gender requests often reflected weak preferences or outdated stereotypes. American Economic Review 2023, paywalled. Nine core results with source locators, datasets used, and the regression-discontinuity estimating equations. LLM-distilled, not human-verified.
# Tags: paper-summary, labor-economics, gender-discrimination, hiring, job-ads
==============================================================================

**What this is.** This is a distilled skeleton of the paper. Read the original
([doi:10.1257/aer.20211127](https://doi.org/10.1257/aer.20211127)) to replicate
or extend the analysis.

## TL;DR

On March 1, 2019, XMRC.com (a private Chinese job board serving the Xiamen
metropolitan area) removed its standardized "preferred gender" field from all
job ads overnight, without advance notice to employers or workers. Using
internal job-board records spanning a full year around the ban, Kuhn and Shen
(2023) estimate the causal effects of this sudden policy change on workers'
application behavior and the gender composition of successful (callback)
applicant pools. The ban raised women's share of callbacks to jobs that had
previously requested men by 2.94 percentage points (61 percent), and raised
men's share of callbacks to jobs that had previously requested women by 9.91
percentage points (146 percent). The main mechanism was a large surge in
gender-mismatched applications, which employers treated relatively well:
workers who applied to jobs of the "wrong" gender received callbacks at
65 to 87 percent of the rate of gender-matched applicants, both before and
after the ban. The ban did not increase matching frictions: aggregate
application match quality rose slightly and callback rates per application
were unaffected. The effect was asymmetric: men entered formerly female jobs
at a far greater rate than women entered formerly male jobs, which the paper
links to gender differences in ambiguity aversion and the greater
industry-specificity of the most male-dominated job titles. Integration
concentrated in relatively low-wage positions; the most strongly gendered
titles (drivers, electricians, nurses, receptionists) were essentially
untouched.

## Core results

| # | Result | Locator | Magnitude as reported |
|---|---|---|---|
| R1 | Women's share of callbacks to M (male-requesting) jobs rose after ban | Table 2 col 4, p. 1027 | +2.94 pp (+61%), from base 4.79% |
| R2 | Men's share of callbacks to F (female-requesting) jobs rose after ban | Table 2 col 4, p. 1027 | +9.91 pp (+146%), from base 6.78% |
| R3 | Women's share of applications to M jobs rose after ban | Table 1 col 4, p. 1026 | +4.01 pp (+72%), from base 5.59% |
| R4 | Men's share of applications to F jobs rose after ban | Table 1 col 4, p. 1026 | +13.40 pp (+133%), from base 10.06% |
| R5 | Total daily applications to F jobs increased | Table 4 panel C col 2, p. 1038 | +0.0228/day (+12.2%); all new applications gender-mismatched |
| R6 | Total daily applications to M jobs increased | Table 4 panel D col 2, p. 1038 | +0.0165/day (+8.5%); all new applications gender-mismatched |
| R7 | Women reduced their applications to F jobs after ban removed female invitation | Table 4 panel C col 3, p. 1038 | -0.0160 applications/day (sig.); men did not reduce M-job applications |
| R8 | Mean application match quality rose slightly | Table 5 panel A col 2, p. 1040 | +0.0169 SD (significant, 5% level) |
| R9 | Callback rate per application submitted: null result | Table 6 panel A col 2, p. 1041 | -0.0028 (SE 0.0026), not significant |

**Overall (paper's conclusion).** The ban integrated gender-previously-segregated
applicant pools and callback pools without measurable harm to application quality
or workers' success rates per application. The integrating effects were
broad-based across ads, firms, and job titles, but were concentrated in
lower-wage positions and did not penetrate the most gender-stereotyped job titles
(drivers, electricians, nurses, receptionists). The ban's effects were considerably
stronger for men than for women, which the paper attributes to gender differences
in ambiguity aversion and the greater specificity of skills required for highly
male-dominated occupations.

## Theory / model

The paper has no formal theoretical model. Building on Kuhn and Shen (2013),
the paper lays out the following testable hypotheses and identifies two channels
through which a gendered-ad ban could affect labor markets.

**Hypotheses.** (i) Banning explicit gender requests will change the gender
composition of applicant pools to formerly gendered jobs. (ii) The direction
and magnitude of the change depend on how employers used gender requests before
the ban: if requests reproduced the incumbent workforce gender mix (reinforcing
segregation), the ban integrates; if requests were affirmative-action
("lean against the wind"), the ban impedes integration. (iii) The ban may
generate additional frictions (harder to find good matches) or may be benign
if workers self-select appropriately despite the absence of a gender signal.

**Identification.** The policy shock is a sharp RD in time. On the night of
February 28, 2019, XMRC removed the preconfigured "desired gender" dropdown
field from all job ads, effective March 1, 2019, without advance notice.
The removal was sudden (employers had not pre-emptively reduced their use of
gender requests between 2016 and the ban date), unexpected, and limited to
the standardized field (text-embedded gender preferences in job descriptions
were not removed). This allows a regression-discontinuity-in-time design:
the outcome is compared in the same job ads before and after March 1, 2019.
The parallel DiD robustness check uses equivalent weeks in 2018 as controls
for 2019 (Delgado Helleseter, Kuhn, and Shen (2020) provides the baseline
for the time-series context).

**Evidence on the mechanism.** Figure 3 (p. 1029) shows that, overwhelmingly,
employers on XMRC used gender requests to reinforce the incumbent gender mix
of the job title (not for affirmative-action purposes), confirming that the
ban had an integrating effect in principle. The asymmetry between men and
women (R4 vs R3, R2 vs R1) is consistent with two channels: (a) the
female-dominated jobs on XMRC are less industry-specific ("administration and
reception" titles appear across all sectors), so men could quickly qualify;
(b) women are more deterred than men by ambiguous signals - removing an
explicit female invitation reduced women's own-gender applications (R7),
while removing an explicit male invitation did not reduce men's own-gender
applications, consistent with Card, Colella, and Lalive (2021) and gender
differences in ambiguity aversion documented in other settings.

## Method

The paper applies two regression models, both adapted to the regression-
discontinuity-in-time structure.

**Equation (1): gender-share outcomes (weekly ad-week cells).** The outcome
$$Y_{jt}$$ is the female share of applications or callbacks to job $$j$$ in week $$t$$.
The estimating equation is (p. 1025):

$$Y_{jt} = \beta^1 \text{Post}_t \cdot F_j + \beta^2 \text{Post}_t \cdot M_j + \beta^3 \text{Post}_t + \beta^4 F_j + \beta^5 M_j + \beta^6 \mathbf{X}_j + \varepsilon_{jt} \tag{1}$$

where $$F_j$$ and $$M_j$$ indicate that job $$j$$ had a female (male) gender request
when first posted; $$\text{Post}_t$$ equals one for weeks on or after March 1, 2019;
and $$\mathbf{X}_j$$ is a vector of controls including an intercept. Nongendered
(N) jobs in the preban period are the reference category. The treatment effects
are $$\beta^1$$ (ban's effect on female share in F jobs relative to N jobs) and
$$\beta^2$$ (ban's effect in M jobs relative to N jobs). A quartic in calendar
weeks and a quartic in job age (weeks since posting) control for secular trends
and duration dependence within recruiting spells. The most saturated specification
adds job-ad fixed effects (column 4 of Tables 1 and 2), absorbing all
time-invariant job characteristics. Observations are weighted by total
applications received; standard errors cluster by firm ID throughout.

**Equation (2): application arrival rates (daily ad-day cells, 30-day window).**
For outcomes with non-smooth seasonal trends around the Spring Festival, the
paper fits local linear regressions to daily data within a 30-day window on
either side of the ban (p. 1037):

$$Y_{jt} = \alpha + \beta \text{Post}_t + \delta^1 t + \delta^2 t \cdot \text{Post}_t + \theta \mathbf{X}_{jt} + \varepsilon_{jt} \tag{2}$$

where $$t$$ indexes days relative to the ban date (March 1, 2019 = day 0),
$$Y_{jt}$$ is the number of applications received by job $$j$$ on day $$t$$,
$$\beta$$ is the size of the discontinuity on the first treatment day, and
$$\delta^1, \delta^2$$ allow different linear time trends on either side of the ban.
Controls $$\mathbf{X}_{jt}$$ include day-of-week fixed effects, the number of
vacancies specified in the ad, and dummies for the first three days of an
ad's life (to account for application-arrival spikes at posting). Job-ad
fixed effects are added in column 2 of Table 4, giving the tightest
within-ad estimate of the ban's daily application effect. Regressions for
match quality (Table 5) use the same specification as equation (2) with the
normalized match quality as $$Y_{jt}$$. The callback-per-application regressions
(Table 6) replace job-ad fixed effects with applicant fixed effects and define
the outcome at the application level (application $$i$$ ever receives a callback).

## Empirical specifications

**Tables 1 and 2: gender composition of applicant and callback pools.**
Equation (1) is estimated on ad-week cells with at least one application
(N = 1,428,768 for applications; 214,585 for callbacks). The preferred
specification is column 4 (job-ad fixed effects), which implies:

- R3: ban raised female applicant share in M jobs by $$\hat{\beta}^2 + \hat{\beta}^3 = 0.0348 + 0.0053 = 4.01$$ pp, from a preban base of 5.59% (Table 1, p. 1026).
- R4: ban raised male applicant share in F jobs by $$|{\hat{\beta}^1 + \hat{\beta}^3}| = |{-0.1393} + 0.0053| = 13.40$$ pp, from 10.06% (Table 1, p. 1026).
- R1: ban raised female callback share in M jobs by $$2.46 + 0.48 = 2.94$$ pp (61%), from 4.79% (Table 2 col 4, p. 1027).
- R2: ban raised male callback share in F jobs by $$10.39 - 0.48 = 9.91$$ pp (146%), from 6.78% (Table 2 col 4, p. 1027).

All four estimates are robust across five specifications (columns 1-5 of Tables
1 and 2) including quartic time trends, calendar-week fixed effects, and
job-ad fixed effects.

**Tables 4-6: application flows, match quality, and callback yield.**
Equation (2) identifies the ban's effect on application arrival rates using
the 30-day daily window (N = 3,514,552 application-day cells for Table 4
panel A). The ban raised total applications by 3.2 percent in aggregate,
concentrated in F and M jobs (R5, R6). All of the new applications to F
and M jobs were gender-mismatched (columns 3-4, Table 4): women exclusively
drove the increase to M jobs, and men drove the increase to F jobs, while
women reduced applications to their own F jobs (R7, coefficient -0.0160/day,
significant), consistent with greater female ambiguity aversion when an
explicit gender invitation is removed. Match quality rose slightly (R8,
Table 5), suggesting the new gender-mismatched applications were not low
quality. Callback rates per application were statistically indistinguishable
from zero in all sixteen specification-by-subsample cells in Table 6 (R9),
confirming that the ban did not increase matching frictions for workers.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| XMRC.com internal job-board records | Primary: 3,133,603 applications by 204,407 workers to 117,390 ads by 15,902 firms over September 2018 to August 2019, including job gender labels, application timestamps, callback indicators, match scores, and worker resume data | no page yet |

Sample scope: private-sector vacancies in the Xiamen metropolitan area (Fujian,
China), skewed toward skilled workers (mean education 13.3 years, mean
requested age 29.8). The mean posted wage was 5,793 RMB/month for M jobs and
4,433 RMB/month for F jobs, reflecting a raw gender wage gap of 23.5 percent
in posted wages.

## When to read the full paper

- Read if you study gender discrimination in labor markets, effects of equal-opportunity legislation, or job-ad content on application behavior.
- Read Section III (Tables 1 and 2, Figure 1) for the core integration result with gender-share regressions.
- Read Section IV (Table 3, Figures 3 and 4) to understand which job titles and workplaces integrated and which did not.
- Read Section V (Tables 4-6, Figure 5) for the mechanism evidence: application flows, match quality, and callback yield.
- Read the Discussion (Section VII, pp. 1045-1046) for the policy limits: asymmetry, low-wage concentration, and the untouched most-gendered titles.

## Attribution and rights

Kuhn, Peter, and Kailing Shen. 2023. "What Happens When Employers Can No Longer
Discriminate in Job Ads?" *American Economic Review* 113(4): 1013-1048.
https://doi.org/10.1257/aer.20211127

Replication data: Kuhn and Shen (2023), openICPSR, https://doi.org/10.3886/E183021V1.

Paywalled; all rights reserved by the American Economic Association. Extract-only.
This page is an LLM-distilled summary (paper-distiller, claude-sonnet-4-6,
2026-06-25); it is not human-verified and the results have not been reproduced.


==============================================================================
# Law and Norms: Lane, Nosenzo & Sonderegger (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/lane-et-al-law-norms-empirical-evidence-2023/
# Distilled: Using incentivized vignette experiments and a legal-threshold identification strategy, Lane, Nosenzo, and Sonderegger show laws causally shape social norms, producing sharp discontinuities in perceived social appropriateness at legal thresholds across UK, US, and Chinese samples (n=7,000). American Economic Review 2023, paywalled. Eight core results with source locators, the social-image model, and the estimating regressions.
# Tags: paper-summary, law-and-economics, social-norms, behavioral-economics, experimental-economics, panel-regression, peer-reviewed, unreplicated
==============================================================================

**What this is.** This is an LLM-distilled skeleton of Lane, Nosenzo, and Sonderegger (2023); read the [original](https://doi.org/10.1257/aer.20210970) to replicate or extend.

## TL;DR

Lane, Nosenzo, and Sonderegger develop a legal-threshold identification strategy to test whether laws causally influence social norms. They exploit a special class of laws that regulate behavior via legal thresholds (age of consent, legal drinking age, maximum undeclared cash import, BAC drink-driving limit, motorway speed limit), measuring the social norm function S(o, 1) on both sides of each threshold using incentivized vignette experiments with 1,248 UK subjects across three samples. Under the identifying assumption that norms vary continuously near the threshold absent the law, any sharp discontinuity at the legal threshold is causally attributable to the law. Laws produce large, statistically significant drops in perceived social appropriateness for three of five behaviors (age of consent, alcohol to youth, cash at customs) and smaller effects for drink driving and speeding. Robustness experiments with 5,771 additional UK, US, and Chinese subjects show: (i) placebo thresholds produce no comparable discontinuities, ruling out focal-point and information-transmission alternatives; (ii) perceived prosocial traits (trustworthiness, honesty, altruism) shift discontinuously at legal thresholds in the same direction, consistent with a social-image signaling mechanism; and (iii) a "bad law" (CANO nuisance ordinance) reverses the direction of discontinuity in honesty, ruling out the meta-norm (rule-following) explanation.

## Core results

| # | Result | Locator | Magnitude as reported |
|---|--------|---------|----------------------|
| R1 | Age of consent: strong norm discontinuity at legal threshold | Table 2 col A1/B1/C1/D1, Fig 1, p. 1273 | beta2 = -0.778\*\*\* (students, SE 0.184); -0.890\*\*\* (gen pop 2019, SE 0.127); -0.803\*\*\* / -0.813\*\*\* (gen pop 2021 first/second order) |
| R2 | Alcohol to youth: largest norm discontinuity across all samples | Table 2 col A2/B2/C2/D2, Fig 1, p. 1273 | beta2 = -1.035\*\*\* (students, SE 0.138); -0.920\*\*\* (gen pop 2019, SE 0.118); up to -1.137\*\*\* (gen pop 2021 second order) |
| R3 | Cash at customs: strong norm discontinuity | Table 2 col A3/B3/C3/D3, Fig 1, p. 1273 | beta2 = -0.866\*\*\* (students, SE 0.132); -0.948\*\*\* (gen pop 2019, SE 0.124); -0.821\*\*\* / -0.971\*\*\* (gen pop 2021) |
| R4 | Drink driving: weaker norm discontinuity | Table 2 col A4/B4, p. 1273-1274 | beta2 = -0.326 (students, SE 0.178, p=0.068); -0.522\*\*\* (gen pop 2019, SE 0.143); roughly half of R1/R2/R3 magnitudes |
| R5 | Speeding: weakest effect, insignificant for student sample | Table 2 col A5/B5, p. 1273-1274 | beta2 = -0.103 (students, SE 0.107, n.s.); -0.461\*\*\* (gen pop 2019, SE 0.127) |
| R6 | Placebo thresholds: no systematic discontinuities | Fig 2, Appendix F, p. 1278-1280 | Krupka-Weber: placebo discontinuity = 0.22 vs legal threshold = 0.88 (difference p=0.000); no significant placebo for age-of-consent and alcohol-to-youth vignettes under either method |
| R7 | Perceived trustworthiness and honesty drop sharply at legal threshold (Exp 2) | Fig 3, Appendix G, p. 1282-1283 | Trustworthiness drops 0.45-0.63 units; honesty drops 0.51-0.81 units; altruism drops 0.26-0.42 units across high-effect vignettes (all p <= 0.009) |
| R8 | "Bad" law (CANO): upward discontinuity in honesty rules out rule-following mechanism (Exp 3) | Fig 4, Appendix I, p. 1286 | Honesty +0.36 (p=0.002); trustworthiness +0.11; altruism +0.17 (both n.s.); rule-compliance shows near-zero discontinuity (-0.04) |

**Overall.** Laws exert causal expressive power on social norms, creating sharp discontinuities in perceived social appropriateness at legal thresholds. The effect is large for behaviors with high perceived intentionality and high law-enforcement detectability (age of consent, alcohol to youth, cash at customs), and weaker for behaviors that are harder to detect or likely to be unintentional (drink driving, speeding). The pattern of results across all five robustness experiments is consistent with a social-image signaling mechanism (Benabou and Tirole 2011) in which laws make illegality a credible signal of low prosociality.

## Theory / model

The model (Section I, pp. 1260-1264) adapts the social-image framework of Benabou and Tirole (2006, 2011). An individual with privately known type theta (measuring prosociality) faces an opportunity o drawn from a distribution with density g(.) and full support [o_min, o_max] with o_min > 0, where o is the magnitude of the negative externality imposed on others. The individual can seize or leave the opportunity. A law sets a threshold o-bar above which seizing the opportunity is illegal; if caught (with probability p), the individual incurs a material penalty K.

**Utility.** Utility depends on the material payoff, the psychological cost of imposing externalities, and social esteem from observers' inferences about type (eq. 1, p. 1260):

$$u_a(o;\theta) = (t - \theta o - pK I_{o>\bar{o}})a + S(o,a), \tag{1}$$

where t is the material gain, $$I_{o>\bar{o}}$$ is an indicator for illegal action, a = 1 if the opportunity is seized (a = 0 otherwise), and S(o, a) is the social esteem accruing to the individual. Social esteem equals observers' expectation of the individual's type upon observing opportunity o and action a (pp. 1260-1261):

$$S(o,1) \equiv E(\theta \mid o, a=1), \qquad S(o,0) \equiv E(\theta \mid o, a=0).$$

**Absence of law (Proposition 1, p. 1262).** When no law sets a threshold (p = 0), equilibrium esteem is determined by the highest type theta-hat who seizes opportunity o, defined by the indifference condition (eq. 3, p. 1261):

$$t - \hat{\theta}_o \, o - \Delta(\hat{\theta}_o) = 0, \tag{3}$$

where $$\Delta(\hat{\theta}_o) \equiv \mathcal{M}^+(\hat{\theta}_o) - \mathcal{M}^-(\hat{\theta}_o)$$ is the esteem gap, with $$\mathcal{M}^-(\theta_o) \equiv E(\theta \mid \theta < \theta_o)$$ and $$\mathcal{M}^+(\theta_o) \equiv E(\theta \mid \theta > \theta_o)$$. Since theta-hat is continuously decreasing in o, S(o, 1) is also continuously decreasing in o: without law, norms do not create sharp distinctions between arbitrarily close behaviors.

**Presence of law (Proposition 2, p. 1263).** When a law sets a threshold o-bar, legal sanctions create a payoff discontinuity. The threshold type for illegal actions solves (eq. 5, p. 1263):

$$t - pK - \bar{\theta}_o \, o - \Delta(\bar{\theta}_o) = 0. \tag{5}$$

Since theta-hat (highest type taking legal action just below o-bar) always lies strictly above theta-bar (highest type taking illegal action just above o-bar), and $$\mathcal{M}^-(\cdot)$$ is increasing, the esteem function exhibits a downward discontinuity at o-bar (eq. 4, p. 1263):

$$S(o,1) = \begin{cases} \mathcal{M}^-(\hat{\theta}_o), & \text{if } o \le \bar{o}; \\ \mathcal{M}^-(\bar{\theta}_o), & \text{if } o > \bar{o}, \end{cases} \tag{4}$$

with jump:

$$\lim_{\epsilon\to 0}\bigl[S(\bar{o}-\epsilon,1) - S(\bar{o}+\epsilon,1)\bigr] = \mathcal{M}^-(\hat{\theta}_{\bar{o}}) - \mathcal{M}^-(\bar{\theta}_{\bar{o}}) > 0.$$

Proposition 3 (p. 1264) extends the result to distant observers (who see the criminal record but not the action directly): the discontinuity holds even when K = 0, since visibility of illegal behavior to a wider audience is itself a mechanism.

## Method

The empirical approach (Section II, pp. 1265-1270) exploits laws with threshold rules to identify the causal effect of law on norms, using a design analogous to regression discontinuity. The identifying assumption is that the social norm function S(.) is continuous in the vicinity of the legal threshold absent the law's expressive effect. Under this assumption, any discrete jump in measured norms at the legal threshold is causally attributable to the law (p. 1265).

**Norm elicitation.** Two incentivized methods from Krupka and Weber (2013) and Bursztyn, Gonzalez, and Yanagizawa-Drott (2020) are used:

- **Krupka-Weber method**: subjects rate vignette behavior on a four-point scale (very/somewhat socially appropriate or inappropriate) and earn a bonus when their response matches the modal response in their sample. This coordination game incentivizes truthful second-order belief reporting.
- **Opinion matching method**: a first group states personal appropriateness opinions (unincentivized); a second group guesses the first group's modal response for a financial bonus. This elicits second-order beliefs without legality as a coordination device.

**Experimental design.** Subjects are randomly assigned between-subjects to one of 4 or 8 versions of each vignette, varying the behavior's distance from the legal threshold (e.g., age of the younger person in the age-of-consent vignette being 1-4 months above or below the threshold). The random assignment of subjects to versions means the identification design is randomized, not observational; the RDD-analogy is in the analysis, not in whether subjects select their treatment. Prior work such as Tankard and Paluck (2017) and Casoria, Galeotti, and Villeval (2020) exploited natural changes in existing laws to study law's effect on norms; those designs face the challenge that simultaneous events may confound the measured effect. The present design instead uses legal thresholds, requiring only the local continuity assumption rather than assuming all other norm-relevant factors are unchanged by a legislative change.

**Samples.** The main experiment (Section II.C, pp. 1269-1270, Table 1, p. 1270) ran between September 2017 and March 2021 with 1,248 UK participants across three samples: 197 students (Krupka-Weber, 2017), 375 representative general-population subjects (Krupka-Weber, 2019), and 676 representative general-population subjects (opinion matching, 2021). Robustness experiments added 5,771 subjects (Table 3, p. 1277): 1,554 UK general-population (Exp 1 placebo), 2,767 UK general-population (Exp 2 prosocial traits), 1,202 US general-population (Exp 3 bad law), and 248 Chinese students (Exp 4 weak rule of law).

## Empirical specifications

**Main regression (eq. 7, p. 1273).** For each vignette and sample, OLS is estimated with heteroskedasticity-robust standard errors:

$$s(o_i) = \alpha + \beta_1(T - o_i) + \beta_2 \, \text{Illegal}_i + \beta_3(T - o_i) \times \text{Illegal}_i + \epsilon_i, \tag{7}$$

where $$s(o_i)$$ is subject i's social appropriateness evaluation (coded +1 to -1 for four-point scale), $$(T - o_i)$$ is the signed distance from the legal threshold T (positive for legal, negative for illegal actions), $$\text{Illegal}_i$$ is a dummy equal to one for the vignette version describing illegal behavior, and $$\beta_2$$ is the key coefficient of interest: the estimated norm discontinuity at the legal threshold (the causal effect of law on normative appropriateness). Regressions for the two general-population samples include demographic controls (age, gender, income); no controls were collected for the student sample.

**Identification check.** Chow tests confirm no statistically significant differences in $$\beta_2$$ across the three high-effect vignettes (age of consent, alcohol to youth, cash at customs) within each sample (all p >= 0.136), but statistically significant differences between the first group (high effect) and the second group (drink driving, speeding), supporting the model's prediction that intentionality and law-enforcement detectability moderate the effect size (p. 1273-1276, Appendix E).

**Placebo regression (Exp 1).** An extended version of equation (7) adds dummy variables for a placebo threshold placed 5-6 units from the legal threshold and its interaction with threshold distance. The placebo produces no significant discontinuities in four of five vignettes (Krupka-Weber method); the legal-threshold discontinuity is significantly larger than the placebo in all five cases where a comparison is possible (p. 1278-1280, Appendix F).

**Prosocial-traits regression (Exp 2).** The same regression specification is applied to subjects' ratings of the likelihood that the vignette person engages in trustworthy, honest, and altruistic behavior, using the opinion-matching method with 1,984 incentivized and 783 unincentivized UK general-population subjects. Each subject sees one version of one target vignette, randomly assigned. Results (Appendix G, p. 1282-1283): statistically significant discontinuities in trustworthiness and honesty across all high-effect vignettes (all p <= 0.009), confirming the social-image mechanism.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---------|---------------|-----------|
| Original vignette experiment, UK samples (1,248 subjects, 2017-2021) | Main experiment measuring social norm functions; Krupka-Weber and opinion-matching methods | no page yet |
| Robustness experiment 1 (1,554 UK subjects, 2021) | Placebo-threshold test via Prolific | no page yet |
| Robustness experiment 2 (2,767 UK subjects, 2021) | Prosocial-traits elicitation via Prolific | no page yet |
| Robustness experiment 3 (1,202 US subjects, 2022) | "Bad law" CANO vignette via Prolific | no page yet |
| Robustness experiment 4 (248 Chinese students, 2017) | Weak-rule-of-law generalizability test | no page yet |

All data are author-collected online vignette experiments; replication data are deposited at https://doi.org/10.3886/E182995V1.

## When to read the full paper

Read this paper to understand the expressive function of law from a causal empirical standpoint. The theoretical model (Section I) is compact and shows formally how a social-image mechanism (Benabou and Tirole 2011) predicts exactly the pattern observed. The norm-elicitation methods (Section II.B) using Krupka and Weber (2013) and Bursztyn, Gonzalez, and Yanagizawa-Drott (2020) are carefully contrasted. The online appendices (A-K) contain full proofs, vignette wordings, robustness analyses, regression tables, and an alternative conformity-based model. Figure 1 (p. 1272) gives a visual overview of all five norm functions across all three UK samples; Table 2 (pp. 1273-1275) gives the full regression results.

## Attribution and rights

Lane, T., Nosenzo, D., and Sonderegger, S. (2023). "Law and Norms: Empirical Evidence." *American Economic Review* 113(5): 1255-1293. https://doi.org/10.1257/aer.20210970

This page is an LLM-distilled extract (claude-sonnet-4-6, 2026-06-25). Not human-verified; not reproduced. The source article is paywalled; no CC licence is recorded in Crossref metadata. An open author manuscript is available at https://eprints.ncl.ac.uk/file_store/production/293182/ACD73151-8B74-4985-A1E4-B8D65A151BE8.pdf. Replication data are available at https://doi.org/10.3886/E182995V1.


==============================================================================
# Constrained-Efficient Capital Reallocation: Lanteri & Rampini (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/lanteri-rampini-constrained-efficient-capital-reallocation-2023/
# Distilled: In a heterogeneous-firm general equilibrium model with collateral constraints, the competitive equilibrium price of used capital is inefficiently high because distributive pecuniary externalities dominate collateral externalities by a factor of roughly 2.3 quantitatively, providing a new rationale for new-investment subsidies. American Economic Review 2023, paywalled. Six core results with source locators, the full theoretical model with equations, and calibrated quantitative welfare analysis.
# Tags: paper-summary, capital-reallocation, financial-frictions, pecuniary-externalities
==============================================================================

**What this is.** The paper's core propositions, the full equilibrium model with collateral constraints, and the quantitative calibration: enough to understand what was proved and why, without reading all 42 pages. To replicate or extend the results, read the original at [https://doi.org/10.1257/aer.20210902](https://doi.org/10.1257/aer.20210902) and the replication archive at [https://doi.org/10.3886/E180421V1](https://doi.org/10.3886/E180421V1).

## TL;DR

The paper characterizes constrained efficiency in an equilibrium model of investment and capital reallocation in which heterogeneous firms face collateral constraints. In competitive equilibrium, the resale price of used (old) capital is inefficiently high. Two pecuniary externalities pull in opposite directions: a collateral externality (a higher resale price relaxes borrowing capacity) and a distributive externality (buyers of old capital are more financially constrained than sellers, so a lower price redistributes resources toward higher-marginal-value firms). The main analytical result is that the distributive externality strictly dominates the collateral externality in stationary equilibrium, so that a lower price of old capital raises welfare. In a quantitative model calibrated to US firm dynamics, financial frictions cause an aggregate output loss of about 10 percent and a consumption loss of about 7 percent relative to first best. The constrained-efficient allocation, implemented via an average subsidy of 8.6 percent on new investment combined with an average tax of 103.7 percent on old capital purchases (both rebated lump-sum), recovers roughly 70 percent of these losses. The paper builds on the heterogeneous-firm capital reallocation model of Rampini (2019) and the externality-decomposition framework of Dávila and Korinek (2018).

## Core results

Magnitudes and locators are as reported in the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Distributive externality exceeds collateral externality** in stationary competitive equilibrium: marginal decrease in old-capital price raises welfare | Proposition 2, p. 368 | Sign proved analytically; inequality (24) $$\int k^O \phi_d\, d\pi > \theta \int k^N \lambda\, d\pi$$ holds for $$q > q^{FB}$$ and at $$q = q^{FB}$$ |
| R2 | **Distributive externality is about 2.3 times the collateral externality** quantitatively | p. 387, Figure 4 discussion | Distributive externality $$\approx 2.3\times$$ collateral externality in stationary equilibrium of the calibrated model |
| R3 | **Financial frictions cause ~10% output loss and ~7% consumption loss** relative to first best | Table 2, p. 389 | CE output = 89.9% of first best; CE consumption = 93.3% of first best |
| R4 | **Constrained-efficient allocation recovers ~70% of welfare losses**: +8% output and +5% consumption over CE | Table 2, p. 389 | Constrained-efficient output = 97.3% of first best; consumption = 98.3% of first best |
| R5 | **Old-capital price is inefficiently high in CE**; planner drives it to the scrap value floor | Table 2, p. 389; p. 386 | CE price $$q = 0.553 > q^{FB} = 0.547$$; constrained-efficient price = 0.100 (scrap value floor from calibration) |
| R6 | **Optimal policy: subsidy on new investment + tax on old capital** (both rebated lump-sum) | Table 2, p. 389; Section IIE, p. 371 | Average $$\tau^N = -8.6\%$$ (subsidy); average $$\tau^O = 103.7\%$$ (tax on old capital purchases) |

**Overall (paper's conclusion).** In the class of infinite-horizon heterogeneous-firm models with collateral constraints, the distributive pecuniary externality from the price of used capital dominates the collateral externality. This holds analytically in the stylized model (Propositions 2, 3, 4, 5) and quantitatively in the full model. New investment is welfare-improving beyond the individual incentive because it expands the future supply of old capital, benefiting the most financially constrained firms that are net buyers of old capital.

## Theory / model

### Environment (Section II.A, pp. 359-361)

Time is discrete and infinite ($$t = 0,1,2,\dots$$). A representative household with linear utility maximizes

$$\sum_{t=0}^{\infty} \beta^t C_t \tag{1}$$

where $$\beta \in (0,1)$$ is the discount factor and $$C_t$$ is aggregate consumption. Overlapping generations of firms are born each period with a continuum of measure one. Each firm lives two dates: it invests when young and produces when old. New capital $$k^N$$ has two productive periods (depreciates to become old capital); old capital $$k^O$$ has one residual productive period. Total capital $$k = k^N + k^O$$; the production function satisfies $$f(0) = 0$$, $$f_k > 0$$, $$f_{kk} < 0$$.

The aggregate resource constraint equates total output to consumption plus new-capital investment (eq. 2, p. 360):

$$\int f\!\left(k^N_{t-1}(w) + k^O_{t-1}(w)\right) d\pi(w) = C_t + \int k^N_t(w)\, d\pi(w) \tag{2}$$

and the market-clearing condition for old capital (eq. 3, p. 360):

$$\int k^N_{t-1}(w)\, d\pi(w) = \int k^O_t(w)\, d\pi(w) \tag{3}$$

### First-best allocation (Section II.B, p. 360)

In the frictionless benchmark, the representative household allocates capital optimally. Optimal conditions for new and old capital (eqs. 4-5, p. 360):

$$1 = \beta\!\left[f_k(k_t^{FB}) + q_{t+1}^{FB}\right] \tag{4}$$

$$q_t^{FB} = \beta f_k(k_t^{FB}) \tag{5}$$

where $$q_t^{FB}$$ is the shadow value of old capital (its first-best price). In stationary equilibrium, $$q^{FB} = 1/(1+\beta)$$ and all firms produce at the same efficient scale $$k^{FB}$$.

### Competitive equilibrium with collateral constraints (Section II.C, pp. 361-364)

Firms can borrow at the household's discount rate $$R = \beta^{-1}$$, but debt repayments cannot exceed a fraction $$\theta \in [0,1)$$ of the future resale value of new capital. Old capital has no future resale value, so it cannot be pledged. Firms can also issue equity at convex cost $$\phi(-d)$$ ($$\phi_d \geq 0$$, $$\phi_{dd} \geq 0$$).

Each firm maximizes the present value of dividends net of equity issuance costs (eq. 6, p. 361):

$$\max_{\{d_{0t},\, d_{1,t+1},\, b_t,\, k_t^N,\, k_t^O\}} d_{0t} - \phi(-d_{0t}) + \beta d_{1,t+1} \tag{6}$$

subject to the budget constraint at birth (eq. 7, p. 361):

$$w_{0t} + b_t = d_{0t} + k_t^N + q_t k_t^O \tag{7}$$

the budget constraint when old (eq. 8, p. 361):

$$f(k_t^N + k_t^O) + q_{t+1} k_t^N = d_{1,t+1} + \beta^{-1} b_t \tag{8}$$

and the collateral constraint (eq. 9, p. 361):

$$\theta q_{t+1} k_t^N \geq \beta^{-1} b_t \tag{9}$$

where $$\theta q_{t+1} k_t^N$$ is the maximum amount the firm can borrow against the future resale value of its new capital investment. This collateral structure, in which debt cannot exceed a fraction of the asset's future resale value, follows the framework of Kiyotaki and Moore (1997); its microfoundation from limited enforcement without exclusion is derived in Rampini and Viswanathan (2010). Old capital has no future resale value, so it carries a higher down payment; this induces the most financially constrained firms to prefer old capital.

After combining the first-order conditions (eqs. 10-12), the investment Euler equations can be expressed as user cost conditions (eqs. 15-16, p. 362):

$$u_N(w) \equiv 1 - \beta q + \phi_d(1 - \beta\theta q) = 1 - \beta q + \phi_d \varphi_N \geq \beta f_k(k) \tag{15}$$

$$u_O(w) \equiv q(1 + \phi_d) = q + \phi_d \varphi_O \geq \beta f_k(k) \tag{16}$$

where $$\varphi_N \equiv 1 - \beta\theta q$$ is the down payment per unit of new capital (price minus maximum borrowing against it) and $$\varphi_O \equiv q$$ is the full price of old capital (which cannot be pledged). More financially constrained firms (higher $$\phi_d$$) face a smaller user-cost difference between old and new capital, making old capital relatively cheaper for them.

**Proposition 1** (Stationary Competitive Equilibrium, p. 363): The equilibrium has (i) new capital with a higher down payment than old ($$\varphi_N > \varphi_O$$) but weakly lower user cost for an unconstrained firm; (ii) the price of old capital weakly exceeds first best ($$q \geq q^{FB}$$); and (iii) if $$q > q^{FB}$$, a threshold structure in which the most constrained firms invest only in old capital, intermediate firms invest in both, and unconstrained firms invest only in new capital.

### Constrained efficiency: sign of inefficiency (Section II.D, pp. 365-369)

The planner chooses investment allocations and old-capital prices to maximize aggregate dividends net of equity issuance costs (eq. 18, p. 365), subject to each firm's budget and collateral constraints and the old-capital market-clearing condition (3). The planner internalizes both pecuniary externalities through the market-clearing condition for old capital, which enters with multiplier $$\beta^t \eta_t$$.

The planner's first-order conditions for new and old capital (eqs. 19-20, p. 366) differ from the competitive equilibrium conditions (10-11) by the terms $$\beta^t \eta_{t+1}$$ and $$\eta_t$$:

$$1 + \phi_{d,t} = \beta\!\left[f_k(k_t) + q_{t+1}\right] + \beta\theta\lambda_t q_{t+1} + \underline{\nu}^N_t + \beta\eta_{t+1} \tag{19}$$

$$q_t(1 + \phi_{d,t}) + \eta_t = \beta f_k(k_t) + \underline{\nu}^O_t \tag{20}$$

The multiplier $$\eta_t$$ on the market-clearing condition (3) measures the shadow value of increasing the supply of old capital. The first-order condition for the old-capital price $$q_t$$ (eq. 21, p. 366), simplified using market clearing, becomes (eq. 23, p. 366):

$$\underbrace{\int k^O_t(w)\, \phi_{d,t}(w)\, d\pi(w)}_{\text{aggregate distributive externality}} = \underbrace{\theta \int k^N_{t-1}(w)\, \lambda_{t-1}(w)\, d\pi(w)}_{\text{aggregate collateral externality}} \tag{23}$$

The left side is the aggregate distributive externality: buyers of old capital (who have high $$\phi_d$$, i.e., high marginal value of net worth) value the expenditure saving from a lower price. The right side is the aggregate collateral externality: firms that purchased new capital last period and face binding collateral constraints value the higher borrowing capacity from a higher price.

**Proposition 2** (Sign of Constrained Inefficiency, p. 368): In stationary competitive equilibrium, the aggregate distributive externality exceeds the aggregate collateral externality:

$$\int k^O(w)\, \phi_d(w)\, d\pi(w) > \theta \int k^N(w)\, \lambda(w)\, d\pi(w) \tag{24}$$

A marginal decrease in the price of old capital induces a positive welfare gain. The proof uses three properties of stationary equilibrium: (a) buyers of old capital are more financially constrained than sellers ($$\phi_d$$ is decreasing in net worth $$w$$); (b) market clearing (eq. 3) implies aggregate purchases of old capital exceed aggregate new-capital purchases with binding constraints; and (c) $$\theta < 1$$, so the collateral externality is scaled down by $$\theta$$.

The same sign result holds under risk-averse entrepreneurs (Proposition 3, p. 374), heterogeneous productivity (Proposition 4, p. 375), and long-lived firms and capital with geometric depreciation (Proposition 5, p. 376).

### Ramsey implementation (Section II.E, pp. 369-371)

In the stylized model, the constrained-efficient old-capital price $$q^*$$ satisfies the optimality condition $$\int k^O \phi_d\, d\pi = \theta \int k^N \lambda\, d\pi$$. Setting $$\phi_d = 0$$ for all firms (all firms unconstrained) and solving, the constrained-efficient price is $$q^* = w_{\min}/k^{FB}$$. The tax rates that implement this as a competitive equilibrium with taxes rebated lump-sum (p. 371):

$$\tau^N = -\beta\eta = -\beta(q^{FB} - q^*), \qquad \tau^O = \frac{\eta}{q^*} = \frac{q^{FB}}{q^*} - 1$$

As $$\eta = \beta f'(k^{FB}) - q^* > 0$$ (old capital is scarce from the planner's perspective), $$\tau^N < 0$$ (a subsidy on new investment) and $$\tau^O > 0$$ (a tax on old capital purchases). Both rates are proportional and rebated lump-sum; the subsidy on new capital increases the future supply of old capital, reducing the old-capital price and benefiting the most constrained firms.

## Method

The analytical results use the Lagrangian formulation of the planner's problem (Appendix, p. 392). The Lagrangian assigns multipliers $$\beta^t \mu_{0t}$$ and $$\beta^{t+1} \mu_{1,t+1}$$ to the young and old budget constraints, $$\beta^{t+1}\lambda_t$$ to the collateral constraint, $$\beta^t \underline{\nu}^N_t$$ and $$\beta^t \underline{\nu}^O_t$$ to nonnegativity constraints on capital, and $$\beta^t \eta_t$$ to the market-clearing condition for old capital. The planner's FOCs are then compared with those of the competitive equilibrium to isolate the externality terms.

The key proof technique for Proposition 2 (pp. 367-368) is to bound the two integrals in (24) using properties of the stationary equilibrium from Proposition 1. In particular, using that $$\phi_d$$ is weakly decreasing in $$w$$ and that the optimality condition for debt implies $$\lambda(w) = \phi_d(w)$$, the collateral externality on the right of (24) can be rewritten as $$\theta \int k^N \phi_d\, d\pi$$. The distributive externality on the left is $$\int k^O \phi_d\, d\pi$$. The market-clearing condition (3) plus the equilibrium sorting (more constrained firms hold more old capital than new, unconditionally) then delivers the strict inequality.

The quantitative model (Section IV, pp. 379-383) introduces persistent idiosyncratic productivity shocks $$s_{it}$$ following an AR(1) in logs (discretized with two states), stochastic firm death probability $$\rho$$, and geometric depreciation for both new and old capital. The production function is a CES bundle of new and old capital with elasticity of substitution $$\epsilon = 5$$ and new-capital share $$\sigma^N = 0.5$$ (Section IV.A, p. 380):

$$k_{t-1}(s^a) = g\!\left(k^N_{t-1}(s^a),\, k^O_{t-1}(s^a)\right) = \left[\left(\sigma^N\right)^{1/\epsilon}\!\left(k^N\right)^{(\epsilon-1)/\epsilon} + \left(1-\sigma^N\right)^{1/\epsilon}\!\left(k^O\right)^{(\epsilon-1)/\epsilon}\right]^{\epsilon/(\epsilon-1)}$$

The collateral constraint in the quantitative model (eq. 44, p. 382) allows both new and old capital to serve as collateral:

$$\theta\!\left\{\!\left[1 - \delta^N(1-q_{t+1})\right] k_t^N(s^a) + q_{t+1}(1-\delta^O) k_t^O(s^a)\right\} \geq \beta^{-1} b_t(s^a) \tag{44}$$

The stationary constrained-efficient allocation is solved numerically (online Appendix C.1). The planner's optimal price satisfies the condition in eq. 51 (p. 384), which generalizes eq. 23 to the quantitative environment with productivity heterogeneity and long-lived capital, and the equilibrium price is bounded below by the scrap value $$\underline{q}$$.

## Empirical specifications

**Calibration (Section V.A, pp. 384-385, Table 1).** The quantitative model is calibrated at annual frequency ($$\beta = 0.96$$). Key parameter choices:

| Parameter | Value | Source / target |
|---|---|---|
| Capital curvature $$\alpha$$ | 0.6 | Capital share in firm-dynamics literature |
| CES elasticity $$\epsilon$$ | 5 | Lanteri (2018) |
| Depreciation $$\delta^N = \delta^O$$ | 0.2 | Average capital age new = 4 yr, old = 9 yr |
| Scrap value $$\underline{q}$$ | 0.1 | Interior solution for planner |
| Productivity persistence $$\chi_s$$ | 0.7 | Khan and Thomas (2013), Lanteri (2018) |
| Productivity std. dev. $$\sigma_s$$ | 0.12 | Firm-level investment rate std. dev. = 0.32 |
| Collateralizability $$\theta$$ | 0.5 | Li, Whited, and Wu (2016) |
| Equity cost $$\phi_0 = 0.1$$, $$\phi_1 = 5$$ | 0.1 / 5 | Hennessy and Whited (2007); premium on internal funds $$\approx 5\%$$ |
| Death probability $$\rho$$ | 0.1 | Decker et al. (2014) firm entry/exit rate |
| Initial net worth $$w_0$$ | 5 | $$\approx 9\%$$ of unconstrained-optimal capital for high-productivity firms |

Under the calibration, the standard deviation of firm-level investment rates in competitive equilibrium equals 0.32, close to Cooper and Haltiwanger (2006). The model matches the empirical relationship between firm age and capital age reported by Ma, Murfin, and Pratt (2022): age-0 firms buy capital that is on average 7.5 years old; age-10 firms buy capital averaging 6.4 years old.

**Quantitative results (Section V.B, pp. 386-388, Table 2).** The stationary competitive equilibrium price of old capital equals 0.553 versus the first-best price of 0.547. The planner drives the price to the scrap value floor 0.100, well below first best, because imperfect substitutability between new and old capital prevents the first-best scale from being achieved at a low price. Output, investment, and consumption under the three allocations (as fractions of first-best values):

| Variable | First best (level) | Competitive equilibrium | Constrained efficient |
|---|---|---|---|
| Output | (9.910) | 0.899 | 0.973 |
| Investment | (4.497) | 0.857 | 0.962 |
| Consumption | (5.413) | 0.933 | 0.983 |
| Price $$q$$ | (0.547) | 1.010 | 0.183 |
| Avg. tax $$\tau^N$$ | 0 | 0 | -8.6% |
| Avg. tax $$\tau^O$$ | 0 | 0 | 103.7% |

The distributive externality is approximately 2.3 times the collateral externality in stationary competitive equilibrium (Figure 4, p. 388), consistent with Proposition 2. Sensitivity analysis (Section VI.B) shows the sign result is robust across $$\theta \in \{0, 0.5, 0.75\}$$, $$\epsilon \in \{1, 5, 10\}$$, and different scrap values.

The paper also studies restricted policy instruments: a subsidy on new capital alone (without taxes on old capital) reduces the old-capital price by about 4 percent per 1 percent subsidy, and a balanced-budget policy with $$\tau^N = -0.03$$ and $$\tau^O = 0.073$$ (without lump-sum transfers) drives the price to $$q = 0.412$$ and raises aggregate welfare. Eisfeldt and Rampini (2006) and Eisfeldt and Rampini (2007) provide the underlying empirical facts about capital reallocation that motivate the model setup.

## Datasets used

This paper develops a calibrated theoretical model; it does not directly use external datasets. The calibration is based on empirical moments and parameter estimates from the published literature.

| Source | Role in paper | Wiki page |
|---|---|---|
| Published empirical moments (investment rates, firm entry/exit, capital age, equity cost estimates) | Calibration targets for Table 1 parameters | No page yet (multiple published sources) |

## When to read the full paper

Use the [original](https://doi.org/10.1257/aer.20210902) if you are: (i) studying the proofs of Propositions 3-5 (extensions to risk-averse entrepreneurs, heterogeneous productivity, and long-lived capital) or the conditions under which the opposite sign of inefficiency can arise (Section III.F); (ii) analyzing the transition dynamics of investment subsidies; (iii) examining sensitivity with respect to collateralizability, substitutability, or the scrap value floor (Table C1 in the online appendix); or (iv) extending the framework to settings with aggregate fluctuations, which the paper flags as an open direction. The replication archive at [https://doi.org/10.3886/E180421V1](https://doi.org/10.3886/E180421V1) contains the code for the quantitative model.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(2), February 2023. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. The paper is paywalled; no open-access licence was found in Crossref metadata. This page reproduces only excerpts (equations, numbers, and structural summaries) for educational and research reference purposes under extract-only terms.

> Lanteri, Andrea, and Adriano A. Rampini. "Constrained-Efficient Capital Reallocation."
> *American Economic Review* 113, no. 2 (February 2023): 354-395.
> DOI: 10.1257/aer.20210902.


==============================================================================
# Profits, Scale Economies, and Trade Gains: Lashkaripour & Lugovskyy (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/lashkaripour-lugovskyy-profits-scale-economies-gains-2023/
# Distilled: Second-best trade taxes are a poor substitute for Pigouvian industrial subsidies at correcting scale-economy misallocation, raising average real GDP by only 1.19 percent versus 3.05 percent under the first-best in a calibrated multi-country Krugman model. Unilateral corrective industrial policies trigger immiserizing growth (average -2.78 percent), while coordinated policies via a deep agreement deliver +3.42 percent gains. American Economic Review 113(10), 2023, paywalled. Five core results with source locators, datasets used, the model (generalized Krugman 1980 with nested CES preferences), and the estimation method (shift-share exchange rate IV on Colombian firm-level import data).
# Tags: paper-summary, international-trade, trade-policy, industrial-policy
==============================================================================

**What this is.** The paper's core results, the model it builds on (multi-industry multicountry Krugman 1980 with nested CES preferences), and the estimation method (shift-share exchange rate IV on Colombian firm-level import data) with the defining equations: enough to know what it found and how, without reading all 50 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1257/aer.20210419).

## TL;DR

The paper asks how effective trade and industrial policies are at correcting misallocation from scale economies and profit-generating markups. It characterizes optimal policies analytically in a multi-industry multicountry Krugman (1980) model and estimates the key structural parameters (scale elasticity μ_k and trade elasticity σ_k) from Colombian firm-level import data using a shift-share exchange rate instrument. The answer is sharp: second-best trade taxes are a poor substitute for Pigouvian industrial subsidies, raising average real GDP by only 1.19 percent versus 3.05 percent for the first-best (free entry). Worse, unilateral corrective industrial policies cause immiserizing growth in most countries (average -2.78 percent, free entry), because the estimated cross-industry covariance cov(σ_k, μ_k) ≈ -0.65 creates an irreconcilable tension between terms-of-trade (ToT) improvement and misallocation correction. Coordinated policies via a deep agreement deliver average real GDP gains of +3.42 percent (free entry), dominating any unilateral alternative even before partner retaliation.

## Core results

Welfare gains are simple averages across 43 WIOD countries. Results under both restricted and free entry are reported; restricted entry approximates short-run and free entry approximates long-run policy consequences.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Second-best trade policy (import taxes + export subsidies, no domestic subsidies) raises average real GDP but far below first-best | Table 4, p. 2796; §VI.B, p. 2799-2800 | avg 1.19% (free entry), 0.60% (restricted entry); vs 3.05% and 1.59% first-best; trade taxes replicate ~39% of first-best gains (1.19/3.05) |
| R2 | Third-best (import tariffs alone) is weaker still | Table 4, p. 2796 | avg 0.63% (free entry), 0.47% (restricted entry); ~21% of first-best gains (free entry) |
| R3 | Unilateral corrective industrial policies cause immiserizing growth | Table 5, p. 2801; §VI.C, p. 2800-2803 | avg -2.78% (free entry), -0.32% (restricted entry); most countries experience welfare losses |
| R4 | Coordinated industrial policies via a deep agreement reverse the immiserizing outcome | Table 5, p. 2801; Figure 2, p. 2802 | avg +3.42% (free entry), +1.67% (restricted entry); exceed any unilateral trade or industrial policy alternative |
| R5 | Micro-estimated cross-industry covariance of trade and scale elasticities is strongly negative | Table 3, p. 2792; §V.C, p. 2791-2792 | cov(σ_k, μ_k) ≈ -0.65; median σ_k - 1 = 3.9; median μ_k ≈ 0.20 |

**Overall (paper's conclusion).** The negative covariance (R5) is the empirical basis for why trade policy is ineffective (R1-R2) and industrial policy backfires unilaterally (R3). Industries with the most national product differentiation (high σ, high ToT leverage) are not the ones with the highest love-for-variety externalities (high μ, high misallocation). Correcting misallocation in high-μ industries expands exports in high-σ industries, worsening ToT. A deep agreement that coordinates Pigouvian subsidies globally avoids this race to the bottom and unlocks large gains (R4).

## Theory / model

The baseline is a multi-industry, multicountry Krugman (1980) model with semiparametric preferences. Countries are indexed by i, j, n ∈ C and industries by g, k ∈ K. Each country has L_i workers who supply labor inelastically; labor is the sole factor of production.

**Cross-industry preferences.** The representative consumer in country i maximizes a nonparametric utility U_i over industry-level bundles subject to a budget constraint (eq. 1, p. 2764):

$$
V_i(Y_i, \tilde{\mathbf{P}}_i) = \max_{\mathbf{Q}_i} U_i(\mathbf{Q}_i), \quad \text{s.t.} \quad \sum_{k \in \mathbb{K}} \tilde{P}_{i,k} Q_{i,k} = Y_i. \tag{1}
$$

**Within-industry CES aggregator (Assumption A1, p. 2765).** Within each industry k, the utility aggregator is nested CES:

$$
Q_{i,k} = \left(\sum_{j \in \mathbb{C}} Q_{ji,k}^{\frac{\sigma_k - 1}{\sigma_k}}\right)^{\!\frac{\sigma_k}{\sigma_k - 1}}, \qquad Q_{ji,k} = \left[\int_{\omega \in \Omega_{j,k}} \varphi_{ji,k}(\omega)^{\frac{1}{\gamma_k}} q_{ji,k}(\omega)^{\frac{\gamma_k - 1}{\gamma_k}} d\omega\right]^{\!\frac{\gamma_k}{\gamma_k - 1}},
$$

with γ_k ≥ σ_k > 1 and φ_{ji,k}(ω) > 0 a constant variety-specific taste shifter. The outer tier σ_k governs national-level market power (the trade elasticity); the inner tier γ_k governs firm-level market power and love for variety.

**Producer prices and the scale elasticity.** Under free entry, the producer price index for composite good ij,k aggregating over firm-level varieties from origin i is (eq. 4, p. 2766):

$$
P_{ij,k} = \frac{\gamma_k}{\gamma_k - 1}\, \tau_{ij,k}\, \bar{a}_{i,k}\, w_i\, M_{i,k}^{-\frac{1}{\gamma_k - 1}},
$$

where τ_{ij,k} is an iceberg trade cost, ā_{i,k} is average unit labor cost, and M_{i,k} is the mass of varieties. The *scale elasticity* (equal to the markup under restricted entry) is (p. 2767):

$$
\mu_k \equiv \frac{1}{\gamma_k - 1} \quad \sim \text{scale elasticity} \sim \text{markup}.
$$

Under free entry, μ_k measures the elasticity by which variety-adjusted TFP grows with industry employment (love for variety). Under restricted entry, 1 + μ_k = γ_k/(γ_k - 1) is the constant firm-level markup.

**Tax instruments.** The government in country i can levy an import tax t_{ji,k} on goods from origin j, an export subsidy x_{ij,k} on goods sold to market j, and an industrial subsidy s_{i,k} on domestic output. These create a wedge between consumer price P̃ and producer price P (eq. 7, p. 2768):

$$
\tilde{P}_{ji,k} = \frac{1 + t_{ji,k}}{(1 + x_{ji,k})(1 + s_{j,k})}\, P_{ji,k}, \qquad \forall j,i \in \mathbb{C},\; k \in \mathbb{K}. \tag{7}
$$

**Globally efficient policy.** The globally efficient allocation uses Pigouvian domestic subsidies that restore marginal-cost pricing and sets all trade taxes to zero (eq. 9, p. 2772):

$$
t^*_{ji,k} = x^*_{ji,k} = 0, \quad \forall\, ji,k; \qquad 1 + s^*_{i,k} = 1 + \mu_k, \quad \forall\, i,k. \tag{9}
$$

Governments deviate from this benchmark only to exploit terms-of-trade (ToT) gains vis-a-vis the rest of the world, as shown by Bagwell and Staiger (2001) in a closely related setting. The isomorphism established by Kucheryavyy, Lyn, and Rodriguez-Clare (2023a) between the Krugman model and the Eaton and Kortum model allows the policy theorems to extend to alternative trade frameworks; the paper also derives analogous results for the Melitz-Pareto model.

## Method

The paper proposes a **dual approach** (`optimal-trade-dual`) for characterizing unilaterally optimal trade and industrial policies in multi-country, multi-industry GE models. It builds on `instrumental-variables` (shift-share exchange rate IV) and `panel-regression` (first-difference estimation on a firm-level panel).

**Unilaterally optimal policy via sufficient statistics (Theorem 1, p. 2775).** The government's problem of maximizing welfare W_i(t_i, x_i, s_i; w) is reformulated as choosing price vectors P̃_i directly (Step 1). First-order conditions are derived using the envelope theorem (Step 2). Tax neutrality (Lemma 1) and homogeneity properties of Marshallian demand are invoked to eliminate redundant wage and income terms (Step 3). The result is a unique sufficient statistics formula:

$$
1 + s^*_{i,k} = (1 + \mu_k)(1 + \bar{s}_i), \tag{T1a}
$$

$$
1 + t^*_{ji,k} = (1 + \omega_{ji,k})(1 + \bar{t}_i), \tag{T1b}
$$

$$
\mathbf{1} + \mathbf{x}^*_{ij} = -\mathbf{E}_{ij}^{-1}\, \mathbf{E}_{ij}^{(-ij)}\,(\mathbf{1} + \mathbf{t}^*_i), \tag{T1c}
$$

where ω_{ji,k} is the good ji,k's conditional inverse export supply elasticity (eq. 10, p. 2773), E_{ij} ~ E_{ij}^{(ji)} and E_{ij}^{(-ij)} are matrices of Marshallian demand elasticities as defined under (D1), and s̄_i, t̄_i ∈ R_+ are uniform tax shifters that account for the multiplicity of optimal equilibria (Lemma 1). Intuitively: (T1a) is a Pigouvian domestic subsidy restoring marginal-cost pricing; (T1b) exploits the home country's collective import market power via ω; and (T1c) exploits export market power. First-best tariffs and export subsidies are misallocation-blind but depend on the overall magnitude of scale economies through ω.

**Second-best and third-best policies (Theorems 2-3, pp. 2778-2780).** When domestic subsidies are unavailable (second-best), both import taxes and export subsidies gain a misallocation-correcting component proportional to (1 + μ_k)/(1 + μ̄_i), so they favor high-μ_k industries to mimic Pigouvian subsidies. When export subsidies are also unavailable (third-best), a uniform import tariff shifter 1 + t̄*_i compensates for the missing export instruments via Lerner symmetry. These formulas extend to economies with IO linkages (Theorem 4, p. 2785) where first-best import tariffs remain IO-blind but export subsidies require upward adjustment to account for reimportation.

The Bartelme et al. (2019) concurrent analysis of first-best policy for small open economies is a special case (setting ω ≈ 0, λ ≈ 1), as is the Lashkaripour (2021) Nash tariff formula for monopolistic competition with restricted entry (a special case of Theorem 3).

**Tension between objectives (Conjectures 1-2, pp. 2781-2782).** A central feature of the sufficient statistics is the inverse export supply elasticity ω_{ji,k}: under free entry, ω_{ji,k} → 0 when μ_k is uniform (no misallocation), confirming that scale economies are the only source of ToT-misallocation tension under free entry. When cov(σ_k, μ_k) < 0, industries best positioned for ToT manipulation (high σ) are not those most in need of correction (high μ), making it impossible for trade policy to simultaneously achieve both objectives.

## Empirical specifications

**Estimation sample.** Demand parameters (σ_k, γ_k) are estimated from the universe of firm-level import transactions at the Colombian Customs Office, 2007-2013, accessed via DATAMYNE. The data include detailed information on HS10 product category, country of origin, importing and exporting firm IDs, quantity (f.o.b.), and c.i.f. value. A unique feature is the identification of individual foreign firms, enabling firm-product combinations as the variety unit rather than the standard country-product.

**Estimating equation (eq. 16, p. 2788).** Log-linearizing the nested CES demand function (A1) and taking first differences to eliminate time-invariant quality φ̄_{j,k}(ω):

$$
\Delta \ln \bar{x}_{j,kt}(\omega) = (1 - \sigma_k)\,\Delta \ln \bar{p}_{j,kt}(\omega) + \left(1 - \frac{\sigma_k - 1}{\gamma_k - 1}\right)\Delta \ln \lambda_{j,kt}(\omega) + \Delta D_{kt} + \Delta \ln \varphi_{\omega,jkt}, \tag{16}
$$

where x̄_{j,kt}(ω) is annual sales of variety ω (firm-product-country), p̄_{j,kt}(ω) is the variety's annual average price (quantity-weighted monthly average), λ_{j,kt}(ω) is the within-national expenditure share, D_{kt} = ln(P^{σ_k - 1}_{kt} Q_{kt}) is a product-year fixed effect, and Δ ln φ_{ω,jkt} is a variety-level demand shock. The coefficient on Δ ln p̄ identifies σ_k; the coefficient on Δ ln λ additionally pins down (σ_k - 1)/(γ_k - 1), from which μ_k = 1/(γ_k - 1) is recovered.

**Shift-share instrument (pp. 2789-2790).** Both Δln p̄ and Δln λ are endogenous. The price instrument uses lagged monthly export-share weights applied to current monthly exchange rate changes:

$$
z_{j,kt}(\omega) = \sum_{m \in \mathcal{M}} s_{j,kt-1}(\omega,m)\, \Delta \ln \mathcal{E}_{jt}(m), \tag{IV}
$$

where s_{j,kt-1}(ω,m) is the share of month m in variety ωjkt's annual export sales to Colombia in year t-1, and Δln E_{jt}(m) is the year-over-year change in origin j's exchange rate with the Colombian peso in month m. This shift-share design generates firm-level cost shocks via the monthly composition of each firm's prior export activity, circumventing the country-level tariff approach of Ossa (2014) (which cannot discriminate across firms within a country-product). The within-national market share instrument follows Khandelwal (2010): annual changes in the total number of origin j firms and in the total number of HS10 categories served by firm ω.

Standard errors are clustered two-way by product-year and origin-product, following Adao, Kolesár, and Morales (2019), to account for cross-cluster correlation in shift-share designs. The pooled Kleibergen-Paap Wald rk F-statistic is 259, well above Stock-Yogo critical values.

**Welfare quantification (eqs. 17-21, pp. 2797-2799).** Policy welfare gains are computed via the hat-algebra technique, avoiding numerical optimization. Counterfactual changes in wages ŵ_i, expenditure shares λ̂_{ji,k}, industry sales shares ρ̂_{i,k}, and optimal taxes are jointly solved from a system of 2N + NK + [2(N-1)+1]K equations and unknowns derived from the optimal policy formulas (Theorems 1-3) plus labor-market clearing (eq. 20) and balanced budget conditions (eq. 21). Macro data on production and bilateral trade come from the 2014 WIOD, aggregated to 15 traded industries across 43 countries. The Cobb-Douglas cross-industry utility assumption (U_i = Π_k Q^{e_{i,k}}_{i,k}) is imposed for the quantitative analysis.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Colombian Customs Office import transactions, 2007-2013 (via DATAMYNE) | Firm-level structural estimation of demand parameters (σ_k, γ_k) and hence scale and trade elasticities; 225,000+ importing firms from 251 countries, HS10 product level | no page yet |
| 2014 World Input-Output Database (WIOD) | Macro-level calibration for policy quantification: bilateral trade shares, industry production and expenditure shares, applied tariff rates; 43 countries, 56 industries | no page yet |
| Bank of Canada monthly exchange rates, 2007-2013 | Construction of the shift-share exchange rate instrument for IV identification | no page yet |

Sample for estimation: firm-product-year level, 2007-2013, Colombia as the importer; one-time exporters dropped; observations trimmed at the 1st and 99th percentiles of price changes per HS10-year. Sample for quantification: 2014 cross-section, 43 countries (all 27 EU members plus 16 other major economies).

## When to read the full paper

Use the [original](https://doi.org/10.1257/aer.20210419) if you are:
verifying the sufficient statistics derivations (Theorems 1-3 with formal proofs in online appendices B, I, J);
implementing the optimization-free welfare quantification procedure (Proposition 1, eqs. 17-21) for other trade models;
studying the IO linkage extension (Theorem 4, §IV.B) or the political economy extension (§IV.C) where politically adjusted subsidies replace Pigouvian ones;
reading country-specific welfare decompositions (Table 4, Figure 2) for all 43 countries;
checking robustness of the elasticity estimates to alternative IV lags, two-way FE, firm-size trimming, or Melitz-Pareto reparametrization (online appendices P, Q, Y).

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(10), October 2023. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. The paper is paywalled; no CC licence found in Crossref metadata. Extract only.

> Lashkaripour, Ahmad, and Volodymyr Lugovskyy. "Profits, Scale Economies, and the Gains from Trade and Industrial Policy." *American Economic Review* 113, no. 10 (October 2023): 2759-2808. DOI: 10.1257/aer.20210419.


==============================================================================
# Trade with Correlation: Lind & Ramondo (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/lind-ramondo-trade-correlation-2023/
# Distilled: A Ricardian trade model where productivity across countries follows a max-stable multivariate Frechet distribution with a general correlation function, spanning the full class of GEV import demand systems. A latent factor model (LFM) estimated on four-digit SITC trade and tariff data finds 7 technology factors and wide heterogeneity in correlation: countries with more dissimilar technology gain up to 90% more from trade; LFM gains dispersion is an order of magnitude larger than sectoral gravity (SD 2.6 vs 0.07). American Economic Review 2023, paywalled. Seven core results with source locators, the CNCES/GEV model equations, the LFM estimator, and datasets used.
# Tags: paper-summary, international-trade, trade-policy, factor-models, structural, peer-reviewed, unreplicated, data:comtrade, data:wiod
==============================================================================

**What this is.** This is the LLM-distilled skeleton of Lind and Ramondo (2023). Read the [original paper](https://doi.org/10.1257/aer.20190781) to replicate or extend; this page records the model equations, estimator, and quantitative results with PDF locators.

## TL;DR

Lind and Ramondo develop a Ricardian model of trade where the joint distribution of productivity across countries is a max-stable multivariate Frechet distribution with a general correlation function $$G^d$$. This spans the full class of generalized extreme value (GEV) import demand systems and nests Eaton and Kortum (2002) as the independence special case. A cross-nested CES (CNCES) correlation function that can approximate any correlation function enables tractable counterfactuals and a flexible estimation procedure.

For estimation they propose a latent factor model (LFM) that compresses four-digit SITC bilateral trade flow and tariff data for 31 countries and 787 sectors into 7 latent technology classes via non-negative matrix factorization with a pseudo-Poisson criterion. The LFM finds wide heterogeneity in correlation: Factor 1 (apparel and textiles) has $$\rho_1 = 0.927$$; Factor 7 (energy and minerals) has $$\rho_7 = 0.0$$. Countries with relatively dissimilar technology (low correlation) gain much more from trade: Canada gains about 90% more than Germany despite similar self-trade shares. Controlling for self-trade, LFM gains dispersion is an order of magnitude larger than the sectoral gravity model (standard deviation 2.6 vs 0.07).

## Core results

| # | Result | Locator | Magnitude as reported |
|---|---|---|---|
| R1 | Optimal number of latent factors: LR test selects K = 7 | Table 1, p. 335 | p-value for K = 7 vs K = 8 equals 1.0; K = 8 adds no significant fit |
| R2 | LFM with 7 factors explains bilateral trade flow variation | Table 1, p. 335 | R\^2 = 0.937 overall; 0.334 within origin-destination |
| R3 | Factor elasticities and correlation coefficients are highly heterogeneous | Table 2, p. 336 | sigma_k in [0.375 (F7), 5.175 (F1)]; rho_k in [0.0, 0.927]; theta = 0.375 |
| R4 | Expenditure-weighted avg. elasticities differ sharply between LFM and SGM | Figure 2, p. 339 | LFM: 1.5 (India) to ~3 (Turkey); SGM: near-uniform 2.7-3.2 across countries |
| R5 | Canada gains ~90% more from trade than Germany despite equal self-trade | Figure 5, p. 342 | LFM: Canada ~90% higher gains; SGM: near-identical gains for the two countries |
| R6 | LFM gains from trade are an order of magnitude more dispersed than SGM | p. 343 | SD of log gains (controlling self-trade): 2.6 LFM vs 0.07 SGM |
| R7 | US welfare cost of China tariffs is roughly 2x larger in LFM than SGM | Figure 6, p. 345 | Total log real wage at 50pp tariff: ~-0.017 (LFM) vs ~-0.008 (SGM) |

**Overall.** The model shows that correlation in productivity matters quantitatively for gains-from-trade calculations and for counterfactual tariff analysis. Standard models assuming independence (EK/ACR/SGM) understate the heterogeneity in gains across countries and mischaracterize the structure of import demand. The LFM estimate of 7 technology factors that are broadly shared across sectors implies nonzero cross-sector substitution elasticities absent from gravity models.

## Theory / model

The model is a global economy of N countries trading a continuum of goods $$v \in [0,1]$$. Consumers have CES preferences with elasticity $$\eta > 1$$. Each good is produced with one-factor (labor) constant-returns technology

$$Y_{od}(v) = Z_{od}(v)\, L_{od}(v), \tag{tech}$$

where $$Z_{od}(v)$$ is productivity for origin o delivering to destination d, absorbing both efficiency and delivery costs. The key departure from Eaton and Kortum (2002): the joint distribution of productivity across origins is **max-stable multivariate Frechet** with a general correlation function $$G^d$$. The joint CDF is (eq. 1, p. 321):

$$\Pr\!\left[Z_{1d}(v) \leq z_1, \ldots, Z_{Nd}(v) \leq z_N\right] = \exp\!\left[-G^d\!\left(T_{1d} z_1^{-\theta}, \ldots, T_{Nd} z_N^{-\theta}\right)\right], \tag{1}$$

where $$T_{od} > 0$$ is the scale parameter (absolute advantage) and $$\theta > 0$$ controls productivity dispersion. The function $$G^d: \mathbb{R}_+^N \to \mathbb{R}_+$$ is the **correlation function** (a max-stable copula generator). When $$G^d = \sum_o x_o$$ (additive, the independence case), the model reduces exactly to EK with CES import shares. Nonlinear $$G^d$$ introduces correlation and departures from IIA.

**CNCES correlation function.** The cross-nested CES (CNCES) form (eq. 6, p. 323) is the foundation for estimation:

$$G^d(x_1, \ldots, x_N) = \sum_{k=1}^{K}\left[\sum_{o=1}^{N}\!\left(\omega_{kod}\, x_o\right)^{\!\frac{1}{1-\rho_k}}\right]^{\!1-\rho_k}, \tag{6}$$

where $$\rho_k \in [0,1)$$ is the within-nest correlation and $$\omega_{kod} > 0$$ are nest weights. Proposition 1 (p. 323) shows any correlation function can be uniformly approximated by a CNCES on compact sets, so the CNCES is without loss of generality.

**Expenditure shares and prices.** Under max-stability, Proposition 2 (p. 325) gives the closed-form expenditure share of destination d on goods from origin o and the price index (eqs. 8-9, pp. 325-326):

$$\pi_{od} \equiv \frac{X_{od}}{X_d} = \frac{P_{od}^{-\theta}\, G_o^d(P_{1d}^{-\theta}, \ldots, P_{Nd}^{-\theta})}{G^d(P_{1d}^{-\theta}, \ldots, P_{Nd}^{-\theta})}, \quad P_d = G^d\!\left(P_{1d}^{-\theta}, \ldots, P_{Nd}^{-\theta}\right)^{-1/\theta}, \tag{8,9}$$

where $$P_{od} \equiv \gamma T_{od}^{-1/\theta} W_o$$ and $$G_o^d \equiv \partial G^d/\partial x_o$$. The cross-price elasticity $$\varepsilon_{oo'd} = -\theta\, P_{o'd}^{-\theta} G_{oo'}^d / G_o^d \geq 0$$ is nonnegative (gross substitutes), and is zero when $$G^d$$ is additive (the CES/IIA case).

**Gains from trade.** The real wage of country d relative to autarky is (eq. 16, p. 328):

$$\frac{W_d/P_d}{W_d^A/P_d^A} = \left(\tilde{\pi}_{dd}\right)^{-1/\theta}, \tag{16}$$

where $$\tilde{\pi}_{dd} \equiv \pi_{dd}/G_d^d(P_{1d}^{-\theta}, \ldots, P_{Nd}^{-\theta})$$ is the **correlation-adjusted self-trade share**. Under independence $$\tilde{\pi}_{dd} = \pi_{dd}$$ and (16) collapses to the Arkolakis, Costinot, and Rodriguez-Clare (2012) formula. With correlation, two countries sharing the same self-trade share can have different gains depending on how similar their technology is to trading partners.

The CNCES closed-form gains from trade (eq. 17, p. 328) are:

$$\frac{W_d/P_d}{W_d^A/P_d^A} = \pi_{dd}^{-1/\theta}\left[\sum_{k=1}^{K}\!\left(\pi_{kdd}^W\right)^{\!1-\rho_k}\!\pi_{kd}^B\right]^{-1/\theta}, \tag{17}$$

where $$\pi_{kdd}^W$$ is the within-factor self-trade share and $$\pi_{kd}^B$$ is the between-factor share. Higher $$\rho_k$$ (more correlation in factor k) reduces gains from trade for given within-factor expenditure; the ACR formula is the special case $$\rho_k = 0$$ for all k.

## Method

The LFM estimation procedure builds on Adao, Costinot, and Donaldson (2017) by compressing disaggregate sectoral trade data into K latent technology classes. In the multisector version (Section III, p. 330), goods are assigned to S observable sectors, but each sector can use multiple latent factors, relaxing the assumption that technology classes equal observed sectors.

Under the separability condition on factor-level scale parameters (eq. 21, p. 331),

$$T_{ksod}^* = (B_{sk} A_{kod})^{\theta}, \tag{21}$$

sectoral expenditure shares decompose into a sum over latent factors (eq. 22, p. 332):

$$\pi_{sod} = \sum_{k=1}^{K}\!\left(\frac{t_{sod}}{t_{kod}^*}\right)^{\!\!-\sigma_k} \lambda_{sk}\, \pi_{kod}^*, \tag{22}$$

where $$\sigma_k \equiv \theta/(1-\rho_k)$$ is the within-factor elasticity of substitution, $$\lambda_{sk} \equiv B_{sk}^{\sigma_k}/\sum_{s'} B_{s'k}^{\sigma_k}$$ are sector-factor weights (time-invariant), and $$t_{kod}^* \equiv (\sum_s t_{sod}^{-\sigma_k}\lambda_{sk})^{-1/\sigma_k}$$ is a factor-level tariff index. The cross-price elasticity between any two sector-origin pairs so and s'o' (eq. 20, p. 331) is:

$$\varepsilon_{sos'o'd} = \theta\sum_{k=1}^{K}\frac{\rho_k}{1-\rho_k}\,\pi_{ksod}^W\,\pi_{ks'o'd}^W\,\pi_{kd}^B \geq 0. \tag{20}$$

This is zero when all $$\rho_k = 0$$ or sectors share no latent factors (the sectoral gravity model, SGM). Nonzero values arise when two sector-origin pairs rely on factors with high within-factor correlation and similar within-factor expenditure shares.

The LFM is estimated by minimizing the pseudo-Poisson deviance via non-negative matrix factorization (Lee and Seung 1999, 2000; Fu et al. 2019). For a given K, the joint estimation problem (eq. 29, p. 334) is:

$$\hat{\Sigma},\,\hat{\Lambda},\,\hat{\Phi}^* = \arg\min_{\Sigma \geq 0,\,\Lambda \geq 0,\,\Phi^* \geq 0}\;\sum_{s,o,d,t}\ell\!\left(\pi_{sodt},\;\sum_k t_{sodt}^{-\sigma_k}\lambda_{sk}\phi_{kodt}^*\right), \tag{29}$$

where $$\ell(x, \hat{x}) = 2[x\ln(x/\hat{x}) - (x - \hat{x})]$$ is the Poisson deviance. Non-negativity of $$\Lambda$$ and $$\Phi^*$$ ensures uniqueness of the factorization (up to permutation and scale) under general conditions (Fu et al. 2019). The number of factors K is chosen via likelihood ratio tests comparing specifications; K = 7 is selected because K = 8 yields p-value = 1.0 (Table 1, p. 335).

The shape parameter $$\theta$$ is estimated as $$\theta = \min_{k}\hat{\sigma}_k = 0.375$$, the conservative upper bound consistent with all $$\rho_k \geq 0$$ (p. 335). Factor correlation coefficients are then $$\rho_k = 1 - \theta/\sigma_k$$.

## Empirical specifications

The baseline estimation uses four-digit SITC bilateral trade flow and tariff data from Comtrade combined with WIOD aggregate sectoral expenditure data, covering 31 countries and S = 787 sectors over 1999-2007 (5,528,764 sector-origin-destination-year observations; p. 333 and online Appendix O.9). Factor weights $$\lambda_{sk}$$ and within-factor elasticities $$\sigma_k$$ are assumed time-invariant across the sample period; factor-level expenditures $$\phi_{kodt}^*$$ can vary over time.

The **sectoral gravity model** (SGM) restricts each latent factor to one sector ($$B_{sk} = 0$$ for $$s \neq k$$, so $$\lambda_{sk} = \mathbf{1}\{k = s\}$$), yielding the sector-level gravity specification (eq. 26, p. 332) used as a benchmark. SGM implies $$\varepsilon_{sos'o'd} = 0$$ for $$s \neq s'$$ (no cross-sector substitution) and $$\varepsilon_{soo'd} = (\sigma_s - \theta)\pi_{sod}^W$$ for within-sector pairs. The CES model further restricts all $$\rho_k = 0$$, recovering the ACR sufficient-statistic result.

Counterfactuals use hat-algebra applied to the CNCES gains-from-trade formula (17). For the US protectionism exercise, the total effect on US real wages of a tariff increase on China by $$\Delta t$$ is decomposed (eq. 32, p. 344) into:

$$\frac{d\ln(W_d/P_d)}{d\ln t_{o'd}} = \underbrace{(1 - \pi_{dd})\frac{d\ln(W_d/W_{o'})}{d\ln t_{o'd}}}_{\text{domestic wage effect}} + \underbrace{\sum_{o \neq d,\,o \neq o'}\pi_{od}\frac{d\ln(W_o/W_{o'})}{d\ln t_{o'd}}}_{\text{third-party effect}} + \underbrace{\pi_{o'd}}_{\text{direct tariff effect}}. \tag{32}$$

The US welfare cost of a 50pp China tariff is roughly 2x larger under LFM than SGM (Figure 6, p. 345), because LFM implies US consumers substitute less toward domestic goods and more toward third-party suppliers when China is taxed (smaller domestic wage effect, larger third-party effect; the direct effect is larger in LFM as it is proportional to expenditure shares that shrink more slowly in LFM).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| UN Comtrade (4-digit SITC bilateral trade flows) | Sectoral expenditure shares $$\pi_{sodt}$$ for LFM estimation; 787 sectors, 31 countries, 1999-2007 | no page yet |
| UN Comtrade / UNCTAD-TRAINS (tariff schedules) | Tariff rates $$t_{sodt}$$ used to identify within-factor elasticities $$\sigma_k$$ from within-sector variation | no page yet |
| World Input-Output Database (WIOD) | Aggregate sectoral expenditure data to scale factor-level shares (online Appendix O.9) | no page yet |

Sample: 31 countries, 787 four-digit SITC sectors, annual 1999-2007, 5,528,764 bilateral-sector-year observations (p. 333). Rank condition (eq. 24, p. 332) requires $$K \leq S \times N^2/(S + N^2) < S$$; with S = 787 and N = 31 up to 432 factors could be fit.

## When to read the full paper

Read Lind and Ramondo (2023) if you need: (a) the proofs for Propositions 1-2 and the gains-from-trade derivation (Appendices A-C, pp. 346-351), including the connection to max-stable processes and GEV discrete choice; (b) the full NMF algorithm with missing-data extensions and identification conditions (online Appendix O.10); (c) country-by-country gains-from-trade estimates and factor-level export patterns (Figure 5, Table 2, online Appendix O.11); (d) reduced-form evidence on departures from IIA within and across sectors (online Appendix O.6); (e) robustness to the alternative two-step $$\theta$$ estimation using between-factor gravity variation (online Appendix O.8); or (f) the three-country analytical example showing how correlation affects gains (pp. 329-330).

## Attribution and rights

Nelson Lind and Natalia Ramondo, "Trade with Correlation," *American Economic Review* 113, no. 2 (February 2023): 317-353. DOI: [10.1257/aer.20190781](https://doi.org/10.1257/aer.20190781). Replication data deposited at ICPSR: [https://doi.org/10.3886/E173601V1](https://doi.org/10.3886/E173601V1).

This page is an LLM-distilled extract prepared by claude-sonnet-4-6 on 2026-06-25. Not human-verified; not reproduced. Rights held by the American Economic Association; extract-only under fair use.


==============================================================================
# Optimal Procurement with Quality Concerns: Lopomo, Persico & Villa (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/lopomo-et-al-optimal-procurement-quality-concerns-2023/
# Distilled: This paper derives the optimal procurement mechanism when low-cost suppliers are also low-quality (adverse selection), finding that a lowball lottery auction (LoLA) with a floor price and a reserve price maximizes any weighted average of buyer surplus and social surplus subject to incentive compatibility. Applied to Italian government procurement data, the buyer-optimal LoLA yields up to 15 percent higher buyer surplus than a first-price auction. American Economic Review 2023, paywalled. Seven core results with source locators, the mechanism design model, and LoLA with its defining equations. LLM-distilled.
# Tags: paper-summary, auction-theory, mechanism-design, procurement, adverse-selection
==============================================================================

**What this is.** A distilled skeleton of the paper for rapid orientation. Read the [original (doi:10.1257/aer.20211437)](https://doi.org/10.1257/aer.20211437) to replicate or extend the results.

## TL;DR

When quality is noncontractible and low-cost suppliers tend to be low-quality (a "lemons" problem in procurement), standard first-price or second-price auctions perform poorly. Lopomo, Persico, and Villa characterize the optimal mechanism: a lowball lottery auction (LoLA) with a floor price $$p_L$$ and a reserve price $$p_H$$. Bidders with costs below $$p_L$$ pool at that price and one is selected randomly; bidders with costs in $$[p_L, p_H]$$ compete as in a standard second-price auction; bids above $$p_H$$ are excluded. Under a mild regularity condition, the LoLA maximizes any weighted average of buyer surplus and social surplus subject to incentive compatibility and individual rationality. The optimal floor price is independent of the number of suppliers and rises with the severity of the lemons problem. A counterfactual calibration using Italian government procurement data from Decarolis (2018) finds buyer surplus gains of up to 15 percent over first-price auctions at high levels of quality concern.

## Core results

| \# | Result | Locator | Magnitude as reported |
|---|---|---|---|
| R1 | LoLA with optimal $$p_L^*$$ and $$p_H^*$$ solves the weighted welfare maximization problem; sincere bidding is an equilibrium in weakly dominant strategies | Theorem 1, §III, p.1514 | Analytical: LoLA implements the constrained-optimal mechanism for any $$\beta \in [0,1]$$ under Assumption 1 |
| R2 | Optimal floor and reserve prices are independent of the number of suppliers $$N$$; floor price is nondecreasing in the severity of the lemons problem $$\xi$$ for any $$\beta$$; social planner prefers a higher floor price than the buyer | Proposition 1, §III, p.1515-1516 | Comparative static; independent of $$N$$ by conditions (11)-(12) |
| R3 | Increasing $$N$$ raises the weighted welfare generated by the optimal LoLA (unlike standard auctions under adverse selection, where welfare can decrease in $$N$$) | Proposition 2, §III, p.1516-1517 | Analytical; contrast: standard FPA expected surplus $$E[w(c^{(1)})]$$ falls as $$N$$ grows under adverse selection |
| R4 | Sincere-bidding equilibrium is unique almost surely when $$p_H < c_H$$ and there are at least three bidders | Proposition 3, §III, p.1517 | Almost-sure uniqueness; follows from Blume and Heidhues (2004) Vickrey-auction uniqueness result |
| R5 | Italian calibration: buyer surplus up to 15 percent higher in buyer-optimal LoLA than in first-price auction when $$\xi = 1$$; gain is approximately 2.5 percent even at $$\xi \approx 0.5$$ | Figure 7, §VD, p.1527-1528 | 15% buyer surplus gain at $$\xi=1$$; 2.5% gain at $$\xi \approx 0.5$$ |
| R6 | Italian calibration: social surplus improvement up to approximately 20 percent over first-price auction at $$\xi = 1$$; supplier profit improvement exceeds 100 percent | Figure 7, §VD, p.1527-1528 | ~20% social surplus gain; ~100% supplier profit gain at $$\xi=1$$ |
| R7 | Illustrative example (§I): buyer-optimal LoLA ($$p_L^* = 3/4$$) achieves buyer surplus more than 10 percent above the second-price auction ($$p_L = 0$$) and the random assignment mechanism ($$p_L = 1$$) | Figure 2, §I, p.1511 | $$V(3/4) \approx 0.37$$; $$V(0) = V(1) \approx 0.33$$; gain >10% |

**Overall.** The LoLA is a practical mechanism (a reverse second-price auction with a price floor) that is simultaneously optimal for the buyer and for the social planner, differing only in the level of the optimal floor price. The theoretical characterization generalizes both Myerson (1981) (standard auctions optimal when no lemons problem) and Manelli and Vincent (1995) (random assignment optimal under extreme lemons problem) as limiting cases.

## Theory / model

The model (§II, p.1512) has one buyer with known type $$\xi$$ and $$N > 1$$ symmetric suppliers. Supplier $$i$$ has privately known cost $$c_i$$ drawn i.i.d. from density $$f$$ on $$[c_L, c_H]$$. Costs are private and quality is noncontractible. The buyer's value from procuring from a supplier with cost $$c$$ is $$v(c, \xi)$$, which is assumed to be increasing in $$c$$ (the lemons problem: higher-cost suppliers provide higher expected quality). The parameter $$\xi$$ encodes the severity of quality concerns, with $$v_{c\xi}(c, \xi) \geq 0$$. The supplier's profit when selected at payment $$m$$ is $$m - c$$; the buyer's surplus is $$v(c, \xi) - m$$.

The **virtual valuation function** (equation (4), p.1512) is:

$$w(c; \xi, \beta) \equiv v(c; \xi) - c - \beta \frac{F(c)}{f(c)} \tag{4}$$

The ratio $$F(c)/f(c)$$ is the information rent earned by a type-$$c$$ supplier. The parameter $$\beta \in [0,1]$$ encodes the designer's weight on buyer surplus relative to social surplus: $$\beta = 1$$ gives buyer-surplus maximization (Myerson (1981) in reverse) and $$\beta = 0$$ gives social surplus maximization.

**Assumption 1 (Regularity):** $$w(c; \xi, \beta)$$ is quasiconcave in $$c$$. This allows $$w$$ to first decrease then increase in $$c$$ (i.e., a lemons problem can be present) while remaining single-peaked. It is satisfied when $$v$$ is concave and $$F/f$$ is convex (which holds for power, Pareto, and exponential distributions of costs).

A **direct mechanism** specifies, for each supplier $$i$$ and any reported type profile $$c$$, the probability $$q_i(c_i, c_{-i})$$ that supplier $$i$$ is selected and the expected payment $$m_i(c_i, c_{-i})$$ it receives (equation (5), p.1513). By the revelation principle, the optimal mechanism is a truth-telling equilibrium of a direct mechanism.

## Method

The **weighted welfare maximization problem** (equations (6)-(10), §III, p.1513-1514) is:

$$\max_{q,m} \int_{[c_L,c_H]^N} \left\{ \sum_{i=1}^N \left[(v(c_i, \xi) - (1-\beta) \cdot c_i) \cdot q_i(c_i, c_{-i}) - \beta \cdot m_i(c_i, c_{-i})\right] \right\} \prod_{j=1}^N f(c_j)\, dc_j \tag{6}$$

subject to feasibility $$\sum_i q_i \leq 1$$, non-negativity $$q_i \geq 0$$, interim IC (equation (9)), and interim IR (equation (10)).

The paper shows that this problem is solved by a **Lowball Lottery Auction (LoLA)**: a reverse second-price sealed-bid auction with floor price $$p_L$$ and reserve price $$p_H \geq p_L$$, in which bids below $$p_L$$ and above $$p_H$$ are not allowed, and ties at $$p_L$$ are broken uniformly at random (Definition, p.1514).

**Theorem 1** (Optimality of LoLA, p.1514): Under Assumption 1, the LoLA implements the solution to the optimization problem (6)-(10) when the reserve price and floor price are set to:

$$p_H^* = \sup\{c \in [c_L, c_H] \text{ such that } w(c; \xi, \beta) > 0\} \tag{11}$$

$$p_L^* = \max\{p \in [c_L, c_H] \text{ such that } w(p; \xi, \beta) \geq E[w(c; \xi, \beta) \mid c \leq p]\} \tag{12}$$

The reserve price $$p_H^*$$ is the type at which the virtual valuation turns negative (identical to Myerson's reserve price). The floor price condition (12) equates the virtual valuation at $$p_L^*$$ to the average virtual valuation conditional on costs being at or below $$p_L^*$$: this reflects the optimal way to offer the same interim allocation to all types in $$[c_L, p_L^*]$$ simultaneously.

Equilibrium bidding is sincere: suppliers with cost $$c \in [p_L^*, p_H^*]$$ bid their cost $$c$$; suppliers with $$c < p_L^*$$ bid $$p_L^*$$; suppliers with $$c > p_H^*$$ do not bid. The proof builds on `mechanism-design` duality methods and explicitly solves for the shadow prices of the monotonicity constraints (via Lemma 4 in the online appendix), because standard approaches that sidestep monotonicity constraints do not apply under the lemons problem.

**Proposition 4** (FPLoLA equivalence, §IVD, p.1520): The sincere equilibrium of any LoLA can also be implemented by a first-price LoLA (FPLoLA) with the same reserve price and a suitably chosen minimum bid $$b_L \geq p_L$$, where $$b_L$$ is up to 24 percent higher than $$p_L$$ in the Italian calibration.

## Empirical specifications

The Italian calibration (§V, p.1523-1528) illustrates the gains from using the LoLA relative to the first-price auction (the format Italian government procurement actually uses). The buyer payoff function is calibrated using structural estimates from Decarolis (2018) and Decarolis (2019).

**Buyer payoff function (equation (18), p.1523):**

$$v(c, \xi) = \text{const} - K E[D(c, \xi) + O(c, \xi)] \tag{18}$$

where $$D(c, \xi)$$ is the delivery delay ratio, $$O(c, \xi)$$ is the cost overrun ratio, and both are unobserved random variables that depend on the winning supplier's cost $$c$$ and the quality-concern parameter $$\xi$$. After calibration using the empirical marginal distributions $$g_D$$ and $$g_O$$ (Figure 5, p.1524), the calibrated payoff simplifies to:

$$\hat{v}(c, \xi) = \text{const}(\xi) - \xi K[\delta(c) + \omega(c)] \tag{19}$$

where $$\delta(c) = G_D^{-1}([1 - \hat{F}(c)]^N)$$ and $$\omega(c) = G_O^{-1}([1 - \hat{F}(c)]^N)$$.

**Counterfactual computation.** The calibrated virtual valuation $$\hat{w}(c; \xi, \beta) \equiv \hat{v}(c; \xi) - c - \beta \hat{F}(c)/\hat{f}(c)$$ (equation (20), p.1525) is positive for all $$c$$ and $$\beta$$ at the estimated parameters, implying it is optimal to set no reserve price in the LoLA. Optimal floor prices $$p_L^*$$ are then computed from condition (12) for each $$(\xi, \beta)$$ pair (Figure 6, p.1526). For each value of $$\xi \in [0,1]$$, the paper computes expected buyer surplus, supplier profit, and social surplus under the buyer-optimal LoLA and under the first-price auction (Figure 7, p.1527).

The virtual valuation satisfies Assumption 1 (quasiconcavity) for all four displayed values of $$\xi \in \{0, 0.33, 0.67, 1\}$$, confirming that LoLA is optimal in the calibrated setting (Figure 6, p.1526).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Italian government procurement auctions (Decarolis 2019, "Dati Aste") | Estimated cost density $$\hat{f}$$, delay distribution $$g_D$$, and overrun distribution $$g_O$$ used to calibrate the buyer payoff function | no page yet |

Sample: Italian public procurement auctions. Cost units are $$10^5$$ euros. Distributions estimated structurally by Decarolis (2018) and provided to the authors by email (May 10, 2019). Replication data including calibrated distributions are publicly available at [ICPSR E182801V1](https://doi.org/10.3886/E182801V1).

## When to read the full paper

Read Lopomo, Persico, and Villa (2023) to:

- **Understand the full proof** of Theorem 1, particularly the dual solution approach and the shadow prices of monotonicity constraints (online Appendix A).
- **Extend the mechanism** to asymmetric bidders, descending-clock formats, or first-price implementations (Section IV, pp.1517-1522).
- **Use the software** applications (available on GitHub, footnote 7, p.1507) that compute buyer-optimal procurement mechanisms given any cost distribution and value function $$v(c, \xi)$$, including non-LoLA cases.
- **Calibrate the framework** to other procurement settings using the semiparametric calibration method of Section VB, which constructs $$v(c, \xi)$$ from empirical quality distributions conditional on cost.

The calibration results (Figure 7, p.1527) are the entry point for policy analysis; the asymmetric-bidder numerical results (Section IVE, p.1521-1522) are the entry point for applied mechanism designers.

## Attribution and rights

This page is an LLM-distilled extract; it is not human-verified and the results have not been reproduced. Cite the original:

> Lopomo, Giuseppe, Nicola Persico, and Alessandro T. Villa. 2023. "Optimal Procurement with Quality Concerns." *American Economic Review* 113(6): 1505–1529. https://doi.org/10.1257/aer.20211437

Replication data: Lopomo, Persico, and Villa (2023), [ICPSR E182801V1](https://doi.org/10.3886/E182801V1), American Economic Association / ICPSR.

Access: paywalled (AEA subscription). Extract-only; no PDF hosted here.


==============================================================================
# Worth Your Weight: Macchi (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/macchi-worth-weight-experimental-evidence-2023/
# Distilled: Two field experiments in Kampala, Uganda show that obesity functions as a wealth signal in low-income countries, raising credit access by an amount equivalent to a 60 percent increase in self-reported income, driven by statistical discrimination that weakens when financial information is provided. AER 2023, paywalled. Seven core results with source locators, the experimental designs, and the regression specifications.
# Tags: paper-summary, development-economics, discrimination, credit-markets, information-economics, health-economics, experimental, panel-regression, developing-countries, peer-reviewed, unreplicated
==============================================================================

**What this is.** The paper's core results, the hypotheses it tests, and the two experimental designs with their regression specifications: enough to know what it found and how, without reading all 36 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1257/aer.20211879).

## TL;DR

Macchi (2023) provides field-experimental evidence that obesity functions as a wealth signal in Uganda and confers tangible market benefits through statistical discrimination. Two complementary field experiments are set in Kampala. The first, a beliefs experiment with 511 residents rating 34 weight-manipulated portrait pairs, shows that obese portraits are rated 0.70 standard deviations wealthier than their nonobese counterparts, while obesity has no effect on perceived beauty, health, longevity, self-control, or trustworthiness. The second, a credit experiment with 238 professional loan officers evaluating 6,645 hypothetical borrower profiles, shows that obese borrowers receive significantly better credit ratings and are three percentage points more likely to be referred, equivalent in magnitude to a 60 percent raise in self-reported monthly income. The premium falls by 50 to 70 percent when borrowers provide self-reported financial information, consistent with loan officers using body size as a proxy for wealth under asymmetric information (statistical discrimination, following Akerlof (1976)). A third exercise shows that people overestimate both the obesity credit premium (by a factor of two to four) and the income difference between obese and normal-weight people (by two to three times), suggesting beliefs about obesity benefits are inaccurate and market distortions follow.

## Core results

Magnitudes and significance are as reported. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Obesity raises perceived wealth by 0.70 SD; the signal holds when other wealth signals (car, slum) are present | Table 2 Panel A, p. 2298 | Obese coeff = 0.699 SD (SE=0.077, p=0.000); Obese × MultiSignals = -0.190 (SE=0.104, p>0.05) |
| R2 | Obesity has no effect on perceived beauty, health, life expectancy, self-control, ability, or trustworthiness | Table 2 Panel A, p. 2298 | Coefficients range from -0.072 (life expectancy) to 0.113 (beauty), all p > 0.05; trustworthiness -0.358 (SE=0.691, n=679); wealth is the sole affected trait |
| R3 | Loan officers rate obese borrowers as significantly more creditworthy, financially able, and likely to be approved | Table 3 col. 1-3, p. 2306 | Approval: +0.199 SD (p=0.00); Financial ability: +0.180 SD (p=0.00); Creditworthiness: +0.151 SD (p=0.00) |
| R4 | Loan officers are more likely to request a real meeting with obese borrowers (a real-stakes choice) | Table 3 col. 4, p. 2306 | +0.066 SD (p=0.04); approx 3 percentage points vs. a 70.5% base referral rate |
| R5 | Providing self-reported financial information reduces the obesity premium by 50-70% (mechanism: statistical discrimination) | Table 3, p. 2306 | Obese × FinancialInformation = -0.129 (SE=0.038) for approval; premium drops approx 70% for approval likelihood (p=0.041) |
| R6 | Obesity premium concentrated among lower-quality borrowers (high DTI); insignificant for high-quality (low-DTI) borrowers | Table 4, p. 2308 | Obese × Low DTI ratio = -0.152 (SE=0.045) for approval; p-value: Obese + Obese×Low DTI = 0 is 0.149 |
| R7 | Laypeople overestimate the obesity credit premium by 2-4× and overestimate the income advantage of obese individuals by 2-3× | Figure 5, p. 2311; Figure 6, p. 2313 | Perceived credit premium: 2× actual (approval), 4× actual (referral); perceived income diff: approx US$230/month vs. actual approx US$110/month |

**Overall (paper's conclusion).** Obesity is a status symbol in Kampala that provides economically large benefits in credit markets because loan officers use it as a proxy for wealth under asymmetric information. The premium is consistent with statistical discrimination: it falls by 50 to 70 percent when financial information is provided and is concentrated among lower-quality borrowers where uncertainty about creditworthiness is greatest. People broadly overestimate both the obesity wealth signal and the credit market benefits, pointing to market distortions from inaccurate beliefs.

## Theory / model

The paper has no formal structural model. It tests the statistical discrimination hypothesis from Akerlof (1976): when credible information about creditworthiness is costly or unavailable, decision-makers rationally use observable signals that correlate with the underlying trait. In Uganda, body mass and wealth are positively correlated across all wealth quintiles (Figure 1, p. 2292), so Bayesian agents who observe obesity should update toward wealth. The credit market is a natural setting: loan applications in Kampala are dealt in person, and loan officers in poor countries face both moral hazard and adverse selection, as described by Karlan and Zinman (2009) (p. 2299).

The empirical test has two ingredients. First, the beliefs experiment isolates whether obesity specifically signals wealth (and not beauty, health, or trust), which is a necessary condition for a statistical discrimination interpretation. Second, the credit experiment cross-randomizes the degree of information asymmetry: under statistical discrimination, providing verified financial information should reduce the obesity premium, whereas taste-based discrimination (preference for obese borrowers independent of their wealth signal) should be unaffected.

Beliefs accuracy is assessed by comparing the obesity premium estimated among loan officers with the same premium as guessed by the general population. Overestimation is consistent with stereotyping models (Bordalo et al. (2016)), where a visible trait associated with a group becomes overweighted as a signal. Pluralistic ignorance (people think others rely on obesity more than themselves) is also consistent with the second-order beliefs results (Table 2 Panel B, p. 2298). The broader motivation for studying status-signal benefits draws on Bursztyn et al. (2017), who provide field-experimental evidence that demand for visible status goods generates tangible social and economic returns.

## Method

Two OLS panel regressions with fixed effects. The paper adapts the incentivized resume rating (IRR) design of Kessler, Low, and Sullivan (2019) and extends the correspondence study approach of Bertrand and Mullainathan (2004) from labor markets to credit markets in a developing country. The key methodological innovation over standard correspondence studies is the within-person portrait variation: rather than comparing different people at different body masses, each portrait pair is the same person morphed to a thinner (normal-weight) and a fatter (obese) version, eliminating confounding from other visible characteristics.

**Beliefs experiment.** Respondents rate portrait pairs along seven outcomes on a 1-4 scale. The estimating equation (p. 2296) is:

$$
Y^k_{ij} = \beta_0 + \beta_1 \text{Obese}_{ij} + \beta_2 \text{MultiSignals}_{j} + \beta_3 \text{Obese}_{ij} \times \text{MultiSignals}_{j} + \alpha_i + \gamma_j + u_{ij} \tag{1}
$$

where $$Y^k_{ij}$$ is respondent $$j$$'s rating of portrait $$i$$ on outcome $$k$$ (standardized), $$\text{Obese}_{ij}$$ is a dummy for the fatter (obese) version, $$\text{MultiSignals}_j$$ indicates the respondent was shown a second wealth signal (car ownership or slum residence), $$\alpha_i$$ are portrait-pair fixed effects, and $$\gamma_j$$ are respondent fixed effects. Standard errors are clustered at the respondent level. The coefficient of interest is $$\beta_1$$, capturing the causal effect of obesity on ratings, controlling for portrait-specific characteristics and respondent rating tendencies.

**Credit experiment.** Loan officers evaluate 30 borrower profiles each (238 officers, 6,645 evaluations). The credit regression (p. 2303) is:

$$
Y^k_{ij} = \beta_0 + \beta_1 \text{Obese}_{ij} + \beta_2 \text{FinancialInformation}_{ij} + \beta_3 \text{Obese}_{ij} \times \text{FinancialInformation}_{ij} + \delta_i + \gamma_j + u_{ij} \tag{2}
$$

where $$Y^k_{ij}$$ is loan officer $$j$$'s rating of profile $$i$$ on outcome $$k$$ (standardized), $$\text{Obese}_{ij}$$ is a dummy for the borrower's portrait being the obese version, $$\text{FinancialInformation}_{ij}$$ indicates whether the profile displays self-reported financial information (occupation, collateral, monthly revenue and profits), $$\delta_i$$ are borrower profile fixed effects, and $$\gamma_j$$ are loan officer fixed effects. Standard errors are clustered at the loan officer level. $$\beta_1$$ captures the obesity premium absent any financial information; $$\beta_3$$ captures the reduction in that premium when information asymmetry is reduced by financial disclosure.

## Empirical specifications

**Beliefs experiment design.** A 2×3 factorial structure crosses obesity status (obese vs. nonobese portrait) with the number of wealth signals available (no signal, one signal, two signals). Portrait pairs are 30 Kampala residents plus 4 White-race (computer-generated) portraits, each morphed to a thinner (normal-weight, BMI 18-26) and a fatter (obese, BMI 30-46) version using photo-morphing software, for a total of 34 portrait pairs (Figure A1, p. 2316). Each respondent rates 4 portraits drawn randomly from the set, always seeing one version of the portrait (obese or nonobese) within each pair. Stratified sample: 511 Kampala residents from the Greater Kampala Metropolitan Area, balanced by age, gender, and socioeconomic status (Table 1, p. 2294). Outcome variables are standardized; standard errors are clustered at the respondent level.

**Credit experiment design.** A 2×3 design cross-randomizes obesity (obese vs. nonobese portrait of the borrower) with financial information (no financial info / self-reported info with low DTI ratio / self-reported info with high DTI ratio). The 30 hypothetical borrower profiles are built from 187 real prospective borrowers and information from loan officer focus groups (p. 2302). Each profile is associated with a weight-manipulated portrait (same portrait set as the beliefs experiment), a name, passport number, nationality, date of birth, and loan information; profiles displaying self-reported financial information also show monthly revenue, profits, occupation, and collateral. Loan officers evaluate 30 profiles during working hours, with real incentives (they are referred to real borrowers matching their choices at study end, following Kessler, Low, and Sullivan (2019)). The first 10 profiles per loan officer show no financial information; the last 20 show self-reported financial information. Standard errors are clustered at the loan officer level. Robustness: randomization inference with 5,000 replications (online Appendix Figure G5).

**Identification.** The within-portrait-pair random assignment (same person, different body mass) is the core identification strategy: observable characteristics (face, age, gender) are held constant across conditions. In the credit experiment the 2×3 design additionally provides within-profile variation in financial information availability, so the interaction term $$\beta_3$$ in equation (2) identifies the change in the obesity premium as asymmetric information is reduced. The credit experiment uses borrower profile and loan officer fixed effects, so identification comes from within-profile variation in the obesity treatment.

**Heterogeneity analysis.** The mechanism test allows for heterogeneity in the DTI ratio (Table 4, p. 2308): most of the obesity premium is driven by lower-quality borrowers (high DTI), consistent with statistical discrimination theory (Bertrand and Mullainathan (2004)). A complementary test shows that the obesity premium for profiles rated as having more reliable financial information is smaller (Table 3 col. 5, p. 2306), further supporting the information asymmetry channel over taste-based explanations.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Author-collected weight-manipulated portrait pairs | 30 Kampala resident + 4 White-race (computer-generated) portraits, morphed to normal-weight and obese versions; 34 portrait pairs total; stimulus in both experiments | No page yet |
| Author-collected beliefs experiment survey (511 Kampala residents) | Beliefs experiment: rating perceived wealth and other traits from portrait pairs; 1,699 evaluations | No page yet |
| Author-collected credit experiment (238 loan officers, 6,645 profiles) | Credit experiment: loan officer evaluations of 30 hypothetical borrower profiles cross-randomized by body mass and financial information | No page yet |
| Uganda National Panel Survey (UNPS) 2019-2020 | Auxiliary: correlates BMI with credit access in nationally representative Ugandan data (Appendix Table A3, p. 2320) | No page yet |

Primary data: original field-experimental data collected November 2019 in Kampala, Uganda, in partnership with IPA Uganda. Deposited at AEA/ICPSR (Macchi 2023, https://doi.org/10.3886/E181481V1). The experiments were preregistered on the AEA registry (Macchi 2019a and 2019b).

## When to read the full paper

Use the [original](https://doi.org/10.1257/aer.20211879) if you are: studying field-experimental tests of statistical discrimination in credit markets (the full design and robustness checks, including the rural Malawi replication and the randomization inference); designing portrait-based or correspondence experiments for non-labor-market settings (the IRR adaptation by Kessler, Low, and Sullivan (2019) creates real stakes without deception); analyzing obesity or malnutrition policy in developing countries (Section IV discusses implications for optimal sin taxes and anti-malnutrition programs); or interested in beliefs accuracy and overestimation of status signals (Section III and Figure 5). The Internet Appendix contains all experimental instruments and additional robustness tables. The locators in the Core results table above point to the exact tables and figures.

## Attribution and rights

Source: peer-reviewed, *American Economic Review* 113(9), September 2023. Published by the American Economic Association. No open-access license detected (Crossref metadata, AEA publisher site, and OpenAlex all confirm paywalled). This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. Extract-only: the PDF is not hosted here.

> Macchi, Elisa. "Worth Your Weight: Experimental Evidence on the Benefits of Obesity in Low-Income Countries." *American Economic Review* 113, no. 9 (September 2023): 2287-2322. DOI: 10.1257/aer.20211879. © 2023 American Economic Association.


==============================================================================
# Value of Working Conditions: Maestas et al. (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/maestas-et-al-value-working-conditions-united-2023/
# Distilled: Using a new nationally representative stated-preference survey (AWCS, 2015-16, N = 1,738 US workers), this paper estimates willingness to pay for nine nonwage job amenities; a switch from the worst to the best amenity bundle equals 55 percent of the wage. Accounting for amenity incidence and preference heterogeneity attenuates the gender wage gap by 24 percent, widens the race compensation gap by 27 percent, and increases the 90-10 wage inequality measure. American Economic Review 2023, AEA copyright. Ten core results with source locators, datasets used, the indirect utility model, and the stated-preference logit estimation method with equations.
# Tags: paper-summary, labor-economics, wages, wage-inequality, working-conditions
==============================================================================

**What this is.** This is a distilled skeleton of Maestas et al. (2023). Read the
original at [https://doi.org/10.1257/aer.20190846](https://doi.org/10.1257/aer.20190846) to replicate or extend.

## TL;DR

Maestas et al. (2023) field the American Working Conditions Survey (AWCS), a new
nationally representative survey covering 1,738 employed Americans, and use
embedded stated-preference experiments to estimate how much workers are willing
to pay for nine nonwage job amenities. Across all amenities, switching from the
worst to the best job is equivalent to a 55 percent wage increase, confirming
that nonwage job attributes are a central component of total compensation. Workers
differ widely in their valuations by gender, race, education, and age: older
workers and women place especially high value on physical job demands and paid time
off. Incorporating both the incidence and the valuation of amenities into standard
wage differentials attenuates the gender gap (24 percent) but widens the race and
education gaps, and raises overall wage inequality. The paper builds on the
compensating differentials framework of Rosen (1986) and extends the experimental
approach of Mas and Pallais (2017) to a nationally representative sample and a
broader set of amenities. Related prior work includes Wiswall and Zafar (2018) on
stated-preference evidence for job attributes among students, and Pierce (2001) on
compensation inequality once fringe benefits are included.

## Core results

| # | Result | Locator | Magnitude as reported |
|---|---|---|---|
| R1 | Best vs worst amenity bundle: total WTP | Table 2 col 5, p. 2025 | 55.0% wage equivalent |
| R2 | Paid time off: WTP for 10 days vs none | Table 2 col 5, p. 2025 | 16.4% wage equivalent |
| R3 | Paid time off: WTP for 20 days vs none | Table 2 col 5, pp. 2025-2027 | 23.0% wage equivalent |
| R4 | Physical demands: WTP for moderate vs heavy activity | Table 2 col 5, pp. 2025-2027 | 14.5% wage equivalent |
| R5 | Schedule flexibility: WTP for setting own schedule | Table 2 col 5, pp. 2025-2027 | 8.9% wage equivalent |
| R6 | Work arrangement: WTP for working alone (vs team-evaluated team) | Table 2 col 5, pp. 2025-2027 | 8.6% wage equivalent |
| R7 | Gender log compensation gap with preference heterogeneity | Table 8 Panel A col 3, p. 2040 | -0.142 log pts (vs -0.192 unadjusted); 24% reduction |
| R8 | Race log compensation gap with preference heterogeneity | Table 8 Panel A col 3, pp. 2040-2041 | -0.274 log pts (vs -0.208 unadjusted); 27% widening |
| R9 | Education log compensation gap (HS or less vs college) | Table 8 Panel A col 3, p. 2041 | -0.667 log pts (vs -0.559 unadjusted); 19% widening |
| R10 | Overall wage inequality: 90-10 log wage gap | Table 8 Panel C col 3, p. 2042 | 1.769 (vs 1.664 unadjusted); +10.5 log pts |

**Overall (paper's conclusion).** Working conditions vary widely across
demographic groups and throughout the wage distribution. Workers have measurable
willingness to pay for most job amenities studied. Accounting for both the
incidence of amenities and heterogeneity in valuations changes standard measures
of the wage structure: the gender gap narrows, the race and education gaps widen,
and overall wage inequality increases. Contrary to the conclusion of Krueger and
Summers (1988), accounting for the value of working conditions widens rather than
narrows interindustry wage differentials.

## Theory / model

The paper builds on the competitive compensating differentials framework of
Rosen (1986). In a long-run competitive equilibrium, workers sort into jobs that
equate, at the margin, their willingness to pay for an amenity with the market
wage-amenity trade-off required by firms. Observed wages therefore understate total
compensation for workers in jobs with desirable nonwage attributes.

There is no formal general equilibrium model estimated in the paper, but the
theoretical logic motivates the empirical setup: if workers trade wages for amenities,
then the true compensation differential between two workers must add back the market
value of the amenities each holds. The indirect utility function of individual $$i$$
over job alternative $$j$$ in choice pair $$t$$ is specified as (p. 2021, equation in
estimation section):

$$V_{ijt} = \alpha + A'_{ijt}\,\beta_i + \delta_i \ln w_{ijt} + \varepsilon_{ijt} \tag{1}$$

where $$A_{ijt}$$ is the vector of nonwage job characteristics (length $$R$$), $$w_{ijt}$$
is the offered wage, and $$\beta_i$$ and $$\delta_i$$ allow heterogeneous marginal
utilities across individuals. The error $$\varepsilon_{ijt}$$ is i.i.d. Extreme Value
Type I, yielding a logit choice probability.

The identification logic is that the stated-preference experiments vary $$A_{ijt}$$
and $$w_{ijt}$$ independently and randomly across choice pairs for each respondent, so
the wage-amenity trade-off is observed directly without the selection confounds that
plague hedonic regressions estimated from observational job choices.

## Method

**Estimating equation.** Under the logit assumption, the probability that individual
$$i$$ prefers job $$j$$ over job $$k$$ in choice pair $$t$$ is (p. 2021):

$$\Pr(V_{ijt} > V_{ikt}) = \frac{\exp\!\bigl[(A'_{ijt} - A'_{ikt})\,\beta_i + \delta_i\,(\ln w_{ijt} - \ln w_{ikt})\bigr]}{1 + \exp\!\bigl[(A'_{ijt} - A'_{ikt})\,\beta_i + \delta_i\,(\ln w_{ijt} - \ln w_{ikt})\bigr]} \tag{2}$$

The authors estimate two versions: (i) a **standard logit** where $$\beta_i = \beta$$
and $$\delta_i = \delta$$ for all $$i$$ (so WTP does not vary across individuals except
through their wage level), and (ii) a **mixed logit** where $$\beta_i \sim N(\beta,\Sigma_\beta)$$
to allow unobserved preference heterogeneity. Results are similar across both
specifications; the standard logit is used for subgroup and robustness analyses.

**Willingness-to-pay derivation.** Individual $$i$$ is indifferent between not having
attribute $$r$$ at wage $$w_i$$ and having it at wage $$w_i - WTP_i^r$$. Setting the
two utility levels equal (p. 2022, eq. 1):

$$\delta_i \ln w_i = \beta_i^r + \delta_i \ln\!\bigl[w_i - WTP_i^r\bigr] \tag{3}$$

Solving for willingness to pay (p. 2022, eq. 2):

$$WTP_i^r = w_i\Bigl[1 - e^{(-\beta_i^r/\delta_i)}\Bigr] \tag{4}$$

This is reported as $$100\times[1 - e^{(-\beta^r/\delta)}]$$ percent of the wage for
the standard logit.

**Best-to-worst WTP.** The total value of the best amenity bundle relative to the
worst is (p. 2022, eq. 3):

$$WTP_i^{\text{FULL}} = w_i\Bigl[1 - e^{(-\sum_r \beta_i^r / \delta_i)}\Bigr] \tag{5}$$

where the sum is over the most-preferred value of each attribute. This yields 55
percent of the wage for the full sample (Table 2 col 5, p. 2025).

**Log total compensation** for the wage-structure analysis is (p. 2039):

$$\ln\!\Bigl(w_i + w_i\Bigl[1 - e^{(-\sum_r A_{ir}\,\beta_i^r / \delta_i)}\Bigr]\Bigr) \tag{6}$$

where $$A_{ir}$$ is an indicator for whether respondent $$i$$'s current job has
attribute $$r$$.

Standard errors use the delta method for the WTP estimates and are clustered by
respondent throughout.

## Empirical specifications

**Stated-preference experiment design.** The AWCS administered 10 stated-preference
experiments per respondent (December 2015-February 2016). In each experiment,
respondents chose between two hypothetical jobs (Job A and Job B). Job attributes were
drawn from the respondent's own current job as a baseline, with two nonwage attributes
randomly selected to vary between jobs. The offered wage $$w_{ijt}$$ was
$$\theta \cdot w_i$$ where $$\theta \sim N(1, 0.1^2)$$, truncated to $$[0.75, 1.25]$$,
ensuring wage variation of at most 50 percent of the current wage. Respondents chose
from four options: Strongly Prefer A, Prefer A, Prefer B, Strongly Prefer B, which
are aggregated into a binary indicator for Job A preference (p. 2021). This design
means identification of $$\beta / \delta$$ (the WTP ratio) comes from within-respondent
random variation in both wages and attributes across the 10 choice pairs.

**Subgroup WTP regressions.** To document heterogeneity in valuations, the authors
estimate the standard logit model separately for subgroups defined by gender (Table 4),
race (Table 5), education (Table 6), and age (Table 7), computing $$WTP^r$$
for each amenity within each subgroup.

**Wage and compensation differential regressions (Section V, Table 8).** For each
measure of compensation, the paper estimates separate regressions of log compensation
on indicator variables for demographic group and industry (aggregated to 11 NAICS
supersectors), with no constant (demeaned within supersector for industry analysis):

$$\ln(\text{compensation}_i) = \sum_k \phi_k D_{ik} + \text{controls} + e_i \tag{7}$$

where $$D_{ik}$$ are indicators for demographic group $$k$$ (female, non-White, education
group, age group) and $$\phi_k$$ is the log compensation differential relative to the
omitted group. Three versions are estimated: (i) log wage only, (ii) log total
compensation holding valuations at full-sample estimates from Table 2 col 5, and
(iii) log total compensation allowing valuations to differ by gender, race, education,
and age. The 90th, 50th, and 10th percentile differences are computed from these
regressions. Standard errors and confidence intervals use a block (by respondent)
bootstrap with 500 iterations (p. 2039).

**Sorting validation.** To test internal validity of the stated-preference estimates,
the authors use the 2018 AWCS follow-up wave ($$N = 977$$ matched respondents) to
test whether individuals who hold a given amenity in 2015 value it more in the
experiments than those without it. Those with a desired attribute in 2015 value it
4.1 percentage points more than those who transition away from it ($$p < 0.01$$),
consistent with preference-driven sorting (Table 3, p. 2032).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| American Working Conditions Survey (AWCS), waves 2015 and 2016 | Primary data source: incidence of 9 job attributes by demographic group (wave 1, July-October 2015); stated-preference experiments for WTP estimation (wave 2, December 2015-February 2016); N = 1,738 workers | no page yet |
| American Working Conditions Survey (AWCS), follow-up wave 2018 | Longitudinal follow-up for sorting validation and preference-transition analysis; N = 977 matched respondents | no page yet |
| RAND American Life Panel (ALP) | Nationally representative probability-based panel that served as the sampling frame for all AWCS waves | no page yet |
| Current Population Survey (CPS) | Used to generate survey weights matching AWCS to US working population demographics | no page yet |

Sample: 1,738 employed workers ages 25-71, from the RAND ALP, weighted to match the
US working population via CPS. The AWCS data are available publicly at
[https://www.rand.org/pubs/tools/TL269.html](https://www.rand.org/pubs/tools/TL269.html).
Replication data are archived at ICPSR (Maestas et al. 2023, DOI 10.3886/E184378V1).

## When to read the full paper

Read the source if you are:

- Estimating compensating wage differentials or the value of specific job amenities
  (Tables 2-7 provide WTP estimates by amenity and demographic group with standard errors).
- Adjusting wage gaps (gender, race, education, interindustry) for nonwage job
  attributes; Table 8 and Section V detail the methodology and results.
- Designing stated-preference experiments for labor market research; Sections III-IV
  provide the experimental design, logit estimation, and robustness checks including
  attention screens, probit alternatives, and common-baseline variants.
- Studying heterogeneity in labor market preferences; Tables 4-7 present results by
  gender, race, education, and age with cross-group p-values.

## Attribution and rights

This page is a distilled extract. The source paper is:

Maestas, Nicole, Kathleen J. Mullen, David Powell, Till von Wachter, and Jeffrey B.
Wenger. 2023. "The Value of Working Conditions in the United States and Implications
for the Structure of Wages." *American Economic Review* 113(7): 2007-2047.
https://doi.org/10.1257/aer.20190846

Copyright American Economic Association; reproduced with permission. Extract only;
full text available at [aeaweb.org](https://www.aeaweb.org/articles?id=10.1257/aer.20190846)
(paywalled; AEA 3-year embargo elapses July 2026). This summary is LLM-distilled
by IAR, not human-verified, and not reproduced.


==============================================================================
# Political Economy of International Regulatory Cooperation: Maggi & Ossa (2023)
# https://instituteforautomatedresearch.org/wiki/papers/aer/2023/maggi-ossa-political-economy-international-regulatory-2023/
# Distilled: cooperative agreements on product standards induce co-lobbying and lead to excessive deregulation when producer lobbies are strong, reducing welfare; agreements on process standards trigger counter-lobbying, tightening regulations and improving welfare when lobbies are powerful. American Economic Review 113(8) 2023, paywalled. Five core propositions with source locators, the lobbying-extended regulatory model, and the equilibrium characterization method.
# Tags: paper-summary, international-trade, trade-policy, political-economy
==============================================================================

**What this is.** This is a distilled skeleton of the paper by Giovanni Maggi and Ralph Ossa (2023). Read the [original article](https://doi.org/10.1257/aer.20200780) to replicate or extend the results.

## TL;DR

Maggi and Ossa build a political economy model of international regulatory cooperation under lobbying by producer groups. The central insight is that the welfare properties of regulatory agreements depend on whether producer interests across countries are aligned or in conflict. For product standards (restrictions on the characteristics of products sold locally, such as emissions caps for automobiles), deregulation in any country benefits producers worldwide by raising world prices; international negotiations therefore induce co-lobbying and lead to excessive deregulation when lobbies are strong, reducing global welfare. For process standards (restrictions on production methods on domestic soil, such as factory pollution limits), deregulation at home reduces world prices and hurts foreign producers; international negotiations induce counter-lobbying that moderates the influence of lobbies on regulatory outcomes, tightening regulations when lobbying is strong and improving welfare. The paper formalizes a concern raised by Rodrik (2018) for product standards while reaching the opposite conclusion for process standards, and extends the analysis to large countries where Bagwell and Staiger (1999) terms-of-trade motives also affect standards.

## Core results

| # | Result | Locator | Key condition / statement |
|---|--------|---------|--------------------------|
| R1 | Cooperative agreement loosens all product standards | Proposition 1, p. 2181 | Holds under mild regularity: demand semi-elasticities do not vary too much with price, or countries are not too asymmetric, or lobby strength gamma_g is sufficiently high |
| R2 | Cooperation on product standards increases global welfare iff lobbying is weak; decreases welfare when lobbying is sufficiently strong | Proposition 2, p. 2183 | Welfare change Delta_g > 0 for low gamma_g; Delta_g < 0 for high gamma_g; Delta_g monotone decreasing in gamma_g (Figure 1, p. 2182) |
| R3 | Cooperation on process standards loosens standards when lobbying is weak; tightens them when lobbying is strong | Proposition 3, p. 2190 | (i) loosens all process standards for sufficiently small gamma_g under mild regularity; (ii) tightens all process standards for sufficiently large gamma_g unconditionally |
| R4 | Cooperation on process standards increases welfare when lobbying is weak or strong; may decrease welfare at intermediate lobbying levels | Proposition 4, p. 2192 | Delta_g > 0 for very low or very high gamma_g; possible Delta_g < 0 for intermediate gamma_g; intermediate welfare loss is guaranteed to exist if countries are symmetric (Figure 2, p. 2191) |
| R5 | Main qualitative results extend to N large countries; asymmetric countries add a terms-of-trade motive for product-standard manipulation | Section III, pp. 2193-2197 | With N large countries, importers tend to tighten product standards to depress world prices while exporters tend to loosen them; with sufficiently strong lobbying, the cooperative agreement still loosens product standards and tightens process standards |

**Overall.** The paper's central lesson is that product-standard agreements are prone to excessive deregulation when lobbies are powerful, while process-standard agreements have a built-in counter-lobbying correction that tightens regulations and preserves welfare. The distinction between co-lobbying (aligned producer interests amplify lobby influence) and counter-lobbying (conflicting interests dilute it) drives both the positive and normative results.

## Theory / model

The model (Section I.A, pp. 2175-2177 for product standards; Section II.A, pp. 2186-2187 for process standards) considers a perfectly competitive world with a continuum of small countries (extended to N large countries in Section III). There are $$\mathcal{G} + 1$$ goods: a numeraire good produced one-for-one from labor, and $$G$$ non-numeraire goods freely traded. The two settings are analyzed separately: one focuses on product standards, the other on process standards.

**Preferences (product standards setting).** Each country $$i$$ has a unit mass of consumers with quasi-linear utility (eq. 1, p. 2176):

$$
U_i = c_{i0} + \sum_{g \in \mathcal{G}} \left[ u_{ig}(c_{ig}) - E_{ig} \right] \tag{1}
$$

where $$c_{i0}$$ is numeraire consumption, $$c_{ig}$$ is consumption of good $$g$$, and $$E_{ig} = -a_{ig} e_{ig} d_{ig}(p^c_{ig})$$ is the local consumption externality. The parameter $$a_{ig} > 0$$ measures how strongly country $$i$$ dislikes pollution; $$e_{ig} \in [0, \infty)$$ indexes the dirtiness of the variety sold (a product standard is a cap on $$e_{ig}$$). Cleaner varieties are more costly: producers incur abatement cost $$\phi_{ig}(e_{ig})$$ per unit (strictly positive, decreasing, convex). Consumer price is $$p^c_{ig} = p_g + \phi_{ig}(e_{ig})$$ where $$p_g$$ is the world price.

**Welfare and government objective.** Country $$i$$'s aggregate welfare across sectors is (eq. 2, p. 2177):

$$
W_i = \sum_{g} \left[ \pi_{ig}(p_g) + S_{ig}\!\left(p_g + \phi_i(e_{ig})\right) - a_{ig} e_{ig} d_{ig}\!\left(p_g + \phi_i(e_{ig})\right) \right] \tag{2}
$$

where $$\pi_{ig}$$ is producer surplus and $$S_{ig}$$ is consumer surplus. Following Grossman and Helpman (1994), governments face lobbying by specific-factor owners and attach extra weight $$\gamma_{ig} \geq 0$$ to sector $$g$$ producer surplus. Government $$i$$ maximizes (eq. 3, p. 2177):

$$
\Omega_i = W_i + \sum_{g \in \mathcal{G}} \gamma_{ig} \pi_{ig} \tag{3}
$$

where $$\gamma_{ig} = 0$$ for all $$g$$ gives a welfare-maximizing government.

**Market clearing.** With free trade and competitive markets, the world market clears for each good $$g$$ (eq. 5, p. 2178):

$$
\int_i y_{ig}(p_g) = \int_i d_{ig}\!\left(p_g + \phi_{ig}(e_{ig})\right) \tag{5}
$$

**Process standards setting.** Good $$g$$ is homogeneous but can be produced with technologies $$z_{ig} \in [0, \infty)$$ indexed by dirtiness; dirtier processes are cheaper. The per-unit abatement cost $$\varphi_{ig}(z_{ig})$$ is paid by producers, so the producer price net of abatement is $$p^P_{ig} = p_g - \varphi_{ig}(z_{ig})$$, and the associated local pollution is $$b_{ig} z_{ig} y_{ig}(p^P_{ig})$$ where $$b_{ig}$$ is the disutility per pollution unit. The incidence of process standards falls on domestic producers (not consumers), which is the source of the counter-lobbying mechanism. Government $$i$$'s objective in sector $$g$$ is:

$$
\Omega_{ig} = (1 + \gamma_{ig})\pi_{ig}\!\left(p_g - \varphi_{ig}(z_{ig})\right) + S_{ig}(p_g) - b_{ig} z_{ig} y_{ig}\!\left(p_g - \varphi_{ig}(z_{ig})\right) \tag{8}
$$

(eq. 8 derivation, p. 2187). Market clearing in process standards requires (eq. 10, p. 2187):

$$
\int_i y_{ig}\!\left(p_g - \varphi_{ig}(z_{ig})\right) = \int_i d_{ig}(p_g) \tag{10}
$$

## Method

The paper characterizes noncooperative Nash equilibria and a cooperative (joint-payoff-maximizing) equilibrium using first-order conditions and local perturbation arguments; formal proofs are in online Appendix B.

**Noncooperative product standards.** Each government maximizes $$\Omega_i$$ over $$e_{ig}$$ taking the world price and all other standards as given. Since countries are small and the problem is separable across industries, the first-order condition for good $$g$$ yields (eq. 4, p. 2178):

$$
e_{ig} = \frac{1}{\sigma_{ig}} \left( \frac{1}{a_{ig}} + \frac{1}{\phi'_{ig}} \right) \quad \text{for all } i \tag{4}
$$

where $$\sigma_{ig} \equiv -d'_{ig}/d_{ig} > 0$$ is the demand semi-elasticity. Lobby strength $$\gamma_{ig}$$ does not enter eq. (4) because the incidence of product standards falls entirely on consumers in the small-country case, leaving producer surplus unaffected at the margin.

**Cooperative product standards.** Governments jointly maximize $$\int_i \Omega_i$$ subject to market clearing (eq. 5). Applying a standard Lagrangian approach, the cooperative product standard satisfies (eq. 6, p. 2179):

$$
e_{ig} = \frac{1}{\sigma_{ig}} \left( \frac{1}{a_{ig}} + \frac{1}{\phi'_{ig}} \right) + \frac{\lambda_g}{a_{ig}} \quad \text{for all } i \tag{6}
$$

where the Lagrange multiplier is

$$
\lambda_g = \frac{\int_i \!\left(\gamma_{ig} y_{ig} + a_{ig} e_{ig} \sigma_{ig} d_{ig}\right)}{\int_i \!\left(\varepsilon_{ig} y_{ig} + \sigma_{ig} d_{ig}\right)} > 0
$$

with $$\varepsilon_{ig} \equiv y'_{ig}/y_{ig} > 0$$ the supply semi-elasticity. Since $$\lambda_g > 0$$ always, cooperative standards are always looser than noncooperative standards (Proposition 1).

**Local argument for product standards direction.** To confirm that $$\lambda_g > 0$$, the paper evaluates the derivative of the joint government payoff with respect to the world price at the noncooperative equilibrium (eq. 7, p. 2180):

$$
\left. \frac{\partial \Omega_g}{\partial p_g} \right|_{\text{NE}} = \int_i \left( \gamma_{ig} y_{ig} + a_{ig} e^N_{ig} \sigma_{ig} d_{ig} \right) > 0 \tag{7}
$$

Both terms are positive: the first captures a political externality (higher world price benefits producers worldwide, so co-lobbying applies) and the second captures an environmental externality (higher price reduces consumption and hence pollution). The agreement internalizes this positive externality by loosening standards to raise the world price.

**Noncooperative process standards.** The first-order condition for $$z_{ig}$$ yields (eq. 9, p. 2187):

$$
z_{ig} = \frac{1}{\varepsilon_{ig}} \left( \frac{1 + \gamma_{ig}}{b_{ig}} + \frac{1}{\varphi'_{ig}} \right) \quad \text{for all } i \tag{9}
$$

Unlike eq. (4), lobby strength $$\gamma_{ig}$$ directly enters the noncooperative process standard: stronger lobbying yields looser process standards unilaterally, since the incidence falls on domestic producers.

**Cooperative process standards.** Joint maximization yields (eq. 11, p. 2188):

$$
z_{ig} = \frac{1}{\varepsilon_{ig}} \left( \frac{1 + \gamma_{ig}}{b_{ig}} + \frac{1}{\varphi'_{ig}} \right) - \frac{\lambda_g}{b_{ig}} \quad \text{for all } i \tag{11}
$$

where

$$
\lambda_g = \frac{\int_i y_{ig}\!\left(\gamma_{ig} - b_{ig} z_{ig} \varepsilon_{ig}\right)}{\int_i \varepsilon_{ig} y_{ig} + \int_i \sigma_{ig} d_{ig}}
$$

The sign of $$\lambda_g$$ determines whether cooperation tightens ($$\lambda_g > 0$$) or loosens ($$\lambda_g < 0$$) process standards. Crucially, $$\lambda_g > 0$$ when $$\gamma_{ig} > b_{ig} z_{ig} \varepsilon_{ig}$$ for all $$i$$, i.e., when lobbying is sufficiently strong.

**Local argument for process standards direction.** At the noncooperative equilibrium, the derivative of the joint payoff with respect to the world price is (eq. 12, p. 2189):

$$
\left. \frac{\partial \Omega_g}{\partial p_g} \right|_{\text{NE}} = \int_i \left( \gamma_{ig} y_{ig} - b_{ig} z^N_{ig} \varepsilon_{ig} y_{ig} \right) \tag{12}
$$

The first term is the positive political externality (tighter standards raise world price, benefiting foreign producers) and the second is a negative environmental externality (higher world price stimulates supply and increases pollution). This is the counter-lobbying mechanism: each lobby prefers loose domestic regulations but tight foreign regulations, so their demands partially offset each other in the cooperative setting. The sign of eq. (12) changes with $$\gamma_g$$, driving Proposition 3.

**Welfare analysis (Propositions 2 and 4).** The paper scales lobby strength proportionally as $$\gamma_{ig} = \gamma_g \cdot \nu_{ig}$$ and tracks the welfare change $$\Delta_g = W^A_g - W^N_g$$ as $$\gamma_g$$ varies. For product standards (Figure 1, p. 2182), $$\Delta_g$$ is positive at $$\gamma_g = 0$$ (noncooperative standards are over-tight from the welfare standpoint) and decreasing in $$\gamma_g$$, turning negative once $$\gamma_g$$ crosses a threshold $$\bar{\gamma}_g$$. For process standards (Figure 2, p. 2191), $$\Delta_g$$ is positive at $$\gamma_g = 0$$, may turn negative for an intermediate range $$[\gamma^L_g, \gamma^H_g]$$, and becomes positive again for large $$\gamma_g$$ (since by then the agreement tightens standards enough to counteract the race to the bottom in the noncooperative equilibrium).

**Large-country extension.** Section III replaces the continuum of small countries with N large countries that each have market power over world prices. The noncooperative product standard becomes (eq. 13, p. 2193):

$$
e^N_{ig} = \frac{1}{\sigma_{ig}} \left( \frac{1}{a_{ig}} + \frac{1}{\phi'_{ig}} \right) + \frac{\lambda^N_{ig}}{a_{ig}} \quad \text{for all } i \tag{13}
$$

where $$\lambda^N_{ig} = (\gamma_{ig} y_{ig} + a_{ig} e^N_{ig} \sigma_{ig} d_{ig} - m_{ig}) / \sum_i (\varepsilon_{ig} y_{ig} + \sigma_{ig} d_{ig})$$ and $$m_{ig} \equiv d_{ig} - y_{ig}$$ is imports. The import term $$-m_{ig}$$ reflects each country's terms-of-trade incentive: importers tighten product standards to depress world prices, exporters loosen them. The cooperative product standard aggregates these effects (eq. 14, p. 2194):

$$
e^A_{ig} = \frac{1}{\sigma_{ig}} \left( \frac{1}{a_{ig}} + \frac{1}{\phi'_{ig}} \right) + \frac{\lambda^A_g}{a_{ig}} \quad \text{for all } i \tag{14}
$$

where $$\lambda^A_g = \sum_i (\gamma_{ig} y_{ig} + a_{ig} e^A_{ig} \sigma_{ig} d_{ig}) / \sum_i (\varepsilon_{ig} y_{ig} + \sigma_{ig} d_{ig})$$. The terms-of-trade motive vanishes in symmetric countries (no trade in equilibrium), and the main results from the small-country model carry through. With strong enough lobbying, the political externality dominates and Propositions 1-4 hold qualitatively for large countries too.

## Datasets used

This is a pure theory paper. No empirical datasets are used; all results are derived analytically.

| Dataset | Role | Wiki page |
|---------|------|-----------|
| (none) | | |

## When to read

Read the full paper when working on the political economy of "deep" trade agreements (non-tariff barriers, regulatory convergence, and CETA/TTIP-style regulatory cooperation councils); when you need the formal conditions under which international lobbying either amplifies or attenuates regulatory distortions; or when extending the Grossman and Helpman (1994) protection-for-sale framework to regulatory rather than tariff policy. The four core propositions (pp. 2181-2192) and Figures 1-2 (pp. 2182, 2191) are the main reusable reference points. Grossman, McCalman and Staiger (2021) provide a complementary analysis focused on harmonization vs. regulatory diversity under imperfect competition. Bagwell and Staiger (1999) supply the terms-of-trade foundation. The companion survey by Maggi and Ossa (2021) in the Annual Review of Economics provides additional context on the political economy of deep integration.

## Attribution and rights

Giovanni Maggi and Ralph Ossa, "The Political Economy of International Regulatory Cooperation," *American Economic Review* 113(8): 2168-2200, 2023. DOI: [10.1257/aer.20200780](https://doi.org/10.1257/aer.20200780).

This page is an LLM-distilled summary (extract-only). It is not human-verified and has not been reproduced. The original article is paywalled; see the [AEA website](https://www.aeaweb.org/articles?id=10.1257/aer.20200780) for access.


==============================================================================
# Choices and Outcomes in Assignment Mechanisms: Agarwal, Hodgson & Somaini (2025)
# https://instituteforautomatedresearch.org/wiki/papers/econometrica/2025/agarwal-et-al-choices-outcomes-assignment-mechanisms-2025/
# Distilled: Using quasi-experimental variation in deceased donor kidney offers and a scarcity instrument, this paper identifies a joint model of patient acceptance decisions and survival outcomes, finding the kidney waitlist mechanism achieves an average LYFT of 9.29 years (1.75 years above random assignment) while the maximum possible is 14.08 years, exposing a planner's dilemma between efficiency and prioritizing the sickest. Econometrica 2025, paywalled. Seven core results with source locators, the assignment-outcomes joint model, and the defining equations.
# Tags: paper-summary, mechanism-design, market-design, health-economics, organ-allocation
==============================================================================

**What this is.** The paper's core results, the structural model it estimates (joint decisions and survival outcomes), and the two-instrument identification strategy with the defining equations: enough to know what it found and how, without reading all 44 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.3982/ECTA20203).

## TL;DR

The paper evaluates the deceased donor kidney allocation mechanism by estimating patient Life-Years from Transplantation (LYFT), defined as the gain in median survival from receiving a transplant. The current observational standard (Wolfe et al. (2008)) relies on hazard-ratio comparisons that cannot account for selection on unobservables. This paper uses two quasi-experimental instruments: (1) random variation in the sequence of organ offers made to each patient, and (2) a scarcity measure (number of donors or offers in the patient's Donor Service Area) that shifts acceptance decisions but is excluded from survival outcomes. These instruments identify a joint structural model of accept/reject decisions and post-transplant and untransplanted survival.

The canonical mechanism design literature, following Roth and Sotomayor (1992), and its school choice applications, such as Abdulkadiroglu and Sonmez (2003), evaluate mechanisms based on revealed agent preferences. This paper takes a different approach: it evaluates the kidney allocation mechanism based on a downstream outcome (patient survival), because policymakers and transplantation communities focus on survival rather than preference satisfaction.

Applied to 175,640 patients on the US kidney waitlist (2000-2010), the preferred estimates find that the realized allocation achieves an average LYFT of 9.29 years among patients registered in 2005, 1.75 years above a random assignment benchmark. Most of this gain is driven by patient choice rather than priority rules: removing patient choice while keeping priority rules yields a LYFT of only 8.05 years. The maximum achievable LYFT is 14.08 years, but reaching it requires systematically transplanting healthier patients with longer expected untransplanted survival, creating a sharp conflict with prioritarianism for the sickest.

## Core results

Magnitudes are as reported; locators point into the source PDF. The 2005 cohort results (R1-R4) are from Figure 4 (p. 425); the preferred-spec LYFT (R5) is from Table VII (p. 421).

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Realized assignment achieves average LYFT of **9.29 years**, 1.75 years above random | Figure 4, p. 425 | Realized = 9.29; random = 7.54; difference = +1.75 years |
| R2 | **Patient choice drives 70.4% of the LYFT gain** over random assignment; priority rules alone (no-choice) achieve only 8.05 | Figure 4, p. 425 | No-choice LYFT = 8.05; realized = 9.29; random = 7.54; choice share = 70.4% (PDF text, p. 426; no-choice achieves 29.6% of the gain over random) |
| R3 | **Maximum possible LYFT = 14.08 years**, 4.8 years above realized; optimal rematching of transplanted yields only 9.93 (13.4% of the gap) | Figure 4, p. 425 | Optimal = 14.08; realized = 9.29; optimal rematching = 9.93; patient-selection dominates |
| R4 | Observables-only optimal assignment reaches **11.04 years**, 1.8 years above realized and 3.0 below full-information optimal | Figure 4, p. 425 | Observables optimum = 11.04; full optimal = 14.08; realized = 9.29 |
| R5 | Preferred specification (instruments + unobservables) LYFT = **8.93 years**; observational model (no instruments) = 8.25, a 0.68-year underestimate | Table VII, col 1-2, p. 421 | Preferred = 8.93 (s.e. 0.12); no-instruments = 8.25 (s.e. 0.07); positive selection on unobservables biases the observational estimate downward |
| R6 | **Positive selection on unobservables**: 1-SD rise in patient selectivity reduces acceptance by 4.0 pp and raises untransplanted survival by 0.316 SD | Table VI, p. 420; Figure 2b, p. 422 | Acceptance effect: -0.040 (s.e. 0.001); untransplanted survival: +0.316 SD (s.e. 0.062); transplanted patients' predicted LYFT distribution shifted ~1.1 years right vs full distribution |
| R7 | **Patient heterogeneity dominates LYFT variance**: patient-specific SD = 3.26 years; donor-specific = 0.99 years; match-specific = 0.41 years | p. 423 | Variance decomposition of LYFT: patient component = 3.26 yr SD vs donor = 0.99 yr vs match = 0.41 yr; rematching alone captures only 13.4% of maximum gain |

**Overall (paper's conclusion).** The mechanism outperforms random assignment primarily through patient choice and selection, not priority rules or patient-kidney matching. However, meaningful gains in average LYFT require changing which patients are transplanted, not only to whom they are matched, creating a dilemma for policymakers who also wish to prioritize the sickest patients. Observational methods that do not account for selection on unobservables underestimate both LYFT and the potential gains from improved allocation.

## Theory / model

The model generalizes the Roy selection framework to a sequential assignment setting in which agents (patients, indexed $$i$$) receive offers for heterogeneous objects (organs, indexed $$j$$) arriving sequentially and make accept-or-reject decisions.

**Unassigned outcome** (eq. 3.1, p. 404): survival if patient $$i$$ never receives a transplant,

$$
Y_{i,0} = g_0(x_i, \nu_{i,0}), \tag{3.1}
$$

where $$x_i \in \mathbb{R}^{d_x}$$ are patient observables and $$\nu_{i,0} \in \mathbb{R}$$ is a patient-specific unobservable (frailty).

**Assignment outcome** (eq. 3.2, p. 404): survival if patient $$i$$ is transplanted organ $$j$$,

$$
Y_{i,j} = g_1(q_j, x_i, \nu_{i,1}, \varepsilon_{i,j,1}), \tag{3.2}
$$

where $$q_j \in \mathbb{R}^{d_q}$$ are organ-type observables, $$\nu_{i,1}$$ is a patient-specific post-transplant unobservable, and $$\varepsilon_{i,j,1}$$ is a match-specific shock.

**Decision equation** (eq. 3.3, p. 405): patient accepts organ $$j$$ if

$$
D_{i,j} = g_D(q_j, x_i, z_i, \nu_{i,D}, \varepsilon_{i,j,D}) = 1, \tag{3.3}
$$

where $$z_i \in \mathbb{R}^{d_z}$$ is the scarcity instrument (excluded from outcome equations), $$\nu_{i,D}$$ is unobserved selectivity, and $$\varepsilon_{i,j,D}$$ is a match-specific preference shock.

Patient $$i$$ is transplanted organ $$j$$ if $$T_{i,j} = 1\{A_i \ge t_{i,j}\} \prod_{j' < j, j' \in J_i} (1 - D_{i,j'}) D_{i,j} = 1$$, i.e., she is alive when the organ arrives, rejects all prior offers, and accepts this one. The observed outcome is then

$$
Y_i = \sum_{j \in J_i} T_{i,j} Y_{i,j} + \Bigl(1 - \sum_{j \in J_i} T_{i,j}\Bigr) Y_{i,0}.
$$

**Key assumptions** (pp. 406-407):

Assumption 1: $$\varepsilon_i$$, $$\nu_i$$, and $$z_i$$ are mutually independent conditional on $$x_i$$ (the scarcity instrument is excludable).

Assumption 2: The potential offer sequence $$J_i$$ is conditionally independent of $$(\nu_i, \varepsilon_i)$$ given $$(x_i, z_i)$$ (organ arrivals are random conditional on patient priority type and geography).

**LYFT** (eq. 7.1, p. 421) for patient-organ pair $$(i, j)$$ conditional on covariates $$I_{i,j} = \{x_i, q_j, D_{i,j}, \eta_j, \nu_{i,D}, \nu_{i,f}\}$$:

$$
\text{LYFT}(I_{i,j}) = M(Y_{i,j} \mid I_{i,j},\, Y_{i,0} \ge t_{i,j}) - M(Y_{i,0} \mid I_{i,j},\, Y_{i,0} \ge t_{i,j}), \tag{7.1}
$$

where $$M(Y \mid X)$$ denotes the median of $$Y$$ given $$X$$ and $$t_{i,j}$$ is the time elapsed between patient registration and organ arrival. This conditions on the patient being alive at the time of the offer and accounts for selection on both patient observables and unobservables.

## Method

Identification proceeds in three steps building on Heckman and Navarro (2007) and Imbens and Angrist (1994).

**Lemma 1** (p. 411) uses variation in the offer sequence $$J_i$$. Let $$N_i = \min\{n : D_{i,j(i,n)} = 1\}$$ be the number of offers rejected before first acceptance. For a patient with priority type $$x_i$$ and scarcity $$z_i$$, comparing patients who received offer-type sequences $$(q_{j(i,1)}, \ldots, q_{j(i,n)})$$ vs $$(q_{j(i,1)}, \ldots, q_{j(i,n-1)})$$ identifies the marginal distributions of $$Y_{i,j(i,n)}$$ and $$Y_{i,0}$$ conditional on $$N_i = n$$. This is a standard LATE argument (Imbens and Angrist (1994)) extended to the sequential setting.

**Lemma 2** (p. 413) shows that offer-sequence variation identifies the choice function $$g_D(\cdot)$$ via its Fourier-Legendre approximation. The key quantity is the moment generating structure: for a sequence $$q_j^n$$ of $$n$$ identical organ-type offers and $$k \le n$$,

$$
P(N_i > k \mid q_j^n, z) = \int_0^1 \varepsilon_D^k \,\mathrm{d}v(\varepsilon_D;\, q_j, z), \tag{5.1}
$$

where $$v(\varepsilon_D; q_j, z)$$ is the CDF of rejection probabilities across patients given organ type $$q_j$$ and scarcity $$z$$. These moments identify the $$(n-1)$$-th order Fourier-Legendre approximation of $$v(\cdot; q_j, z)$$, which converges in Cesaro mean to the true CDF.

**Theorem 1** (p. 414) combines the offer instrument and the scarcity instrument $$z_i$$ to identify the expected outcomes conditional on the selection unobservables $$\nu_{i,D}$$ and $$\varepsilon_{i,j,D}$$. The scarcity instrument "traces out" the selectivity unobservable via (eq. 5.2, p. 414):

$$
E\!\left[Y_{i,0} \times 1\{T_i = 0\} \mid q_j^k, z_i\right] = \int_0^1 E\!\left[Y_{i,0} \mid \nu_D = v(\varepsilon_D;\, z_i, q_i)\right] \varepsilon_D^k \,\mathrm{d}v(\varepsilon_D;\, z_i, q_i). \tag{5.2}
$$

**Estimation** uses a parameterized Box-Cox version of equations (3.1)-(3.3), estimated by Gibbs sampling (McCulloch and Rossi (1994)) (eqs. 5.3-5.7, pp. 415-416). Let $$B(Y; \rho) = (Y^\rho - 1)/\rho$$ denote the Box-Cox transformation (Box and Cox (1964)):

$$
y_{i,0} = B(Y_{i,0};\, \rho_0) = x_i \beta_x + \nu_{i,0}, \tag{5.3}
$$

$$
y_{i,j} = B(Y_{i,j};\, \rho_1) = \chi(x_i, q_j)\alpha_{x,q} + \alpha_\eta \eta_j + \nu_{i,1} + \varepsilon_{i,j,1}, \tag{5.4}
$$

$$
D_{i,j} = 1\!\left\{\chi(x_i, q_j)\gamma_{x,q} + z_i \gamma_z + \eta_j - \nu_{i,D} + \varepsilon_{i,j,D} > 0\right\}, \tag{5.5}
$$

where $$\eta_j \sim N(0, \sigma_\eta^2)$$ captures unobserved organ-level quality and $$\chi(x_i, q_j)$$ is a flexible function of patient and donor characteristics. The unobservables follow the factor structure (eqs. 5.6-5.7, p. 415):

$$
\nu_{i,1} = \delta_{1,D}\, \nu_{i,D} + \nu_{i,f}, \quad \nu_{i,0} = \delta_{0,D}\, \nu_{i,D} + \delta_{0,f}\, \nu_{i,f} + \tilde{\nu}_{i,0}, \tag{5.6-5.7}
$$

where $$\nu_{i,D}$$, $$\nu_{i,f}$$, and $$\tilde{\nu}_{i,0}$$ are independently distributed mean-zero normals. This factor structure allows selectivity into transplantation ($$\nu_{i,D}$$) to be correlated with both post-transplant and untransplanted survival. The Gibbs sampler (Geweke, Gowrisankaran, and Town (2003)) draws sequentially from conditional posteriors; by the Bernstein-von Mises theorem this is interpreted as maximum likelihood. Monte Carlo simulations on 100 data sets with 10,000 patients and 2,500 donors confirm good coverage and convergence (footnote 23, p. 416).

## Empirical specifications

**First stage: offer instrument** (Table III, p. 409). A linear probability model for whether a transplant occurs and its type, as a function of the number of "desirable" donors available in the two years following registration:

$$
\text{Transplant}_{i} = \alpha \log(1 + \#\text{top-10 offers in 2 years}) + x_i \gamma + \text{DSA FE} + \text{year FE} + \text{blood-type FE} + e_i.
$$

Coefficients: 0.0479 (s.e. 0.0046) for KDPI $$\le 50\%$$ organs (column 1). F-statistics range from 142.6 to 162.7 across columns, far above the conventional threshold of 10. Sample: N = 132,507 non-pediatric patients registered 2000-2008.

**First stage: scarcity instruments** (Table IV, p. 410). A linear probability model for whether patient $$i$$ accepts an offer from donor $$j$$:

$$
\text{Accept}_{ij} = \alpha_1 \log(1 + \#\text{donors}) + \alpha_2 \log(1 + \#\text{offers}) + x_i \gamma + w_j \psi + m_{ij} \delta + \text{FE} + e_{ij}.
$$

The number of donors has coefficient $$-0.0434$$ (s.e. 0.00209) and the number of offers $$-0.039$$ (s.e. 0.00106) in columns 1-2. F-statistics range from 296.8 to 1,361.8 across specifications. The instruments remain significant and of similar magnitude after adding patient characteristics (columns 3-4), donor characteristics (columns 5-6), and match characteristics (columns 7-8). Sample: N = 851,753-863,073 offers from the first 100 donors per patient, registered 2000-2009.

Standard errors are clustered by DSA, registration year, and blood type in Table III; by DSA, offer year, years waited at offer, and blood type in Table IV.

**Survival and choice estimates** (Table V, p. 417-418). The structural model is estimated across three specifications: (1) observational (no instruments, $$\nu_{i,D}$$ independent of $$\nu_{i,0}$$ and $$\nu_{i,1}$$), (2) preferred (scarcity instrument = number of past donors), and (3) robustness (past offers instrument). Marginal half-life effects are reported for a 1-SD increase in continuous characteristics. Diabetic patients have a shorter half-life by 3.58 years with a transplant (Panel B) and 1.45 years without a transplant (Panel A; PDF p. 419). Positive tissue-type matching raises post-transplant half-life substantially (Panel B). Selectivity raises untransplanted survival by 0.316 SD per 1-SD increase in $$\nu_{i,D}$$ (Table VI, Panel A, p. 420).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| OPTN Potential Transplant Recipient (PTR) dataset | All organ offers made to each waitlisted patient and each patient's accept/reject decision; match characteristics including HLA mismatches | no page yet |
| OPTN Standard Transplantation Analysis and Research (STAR) dataset | Deceased donor characteristics, patient histories, transplant outcomes, annual follow-up data, patient survival (death dates merged from Social Security records) | no page yet |

Sample: 175,640 patients registered January 1, 2000 through December 31, 2010, excluding pediatric patients and those needing multiple organs or a living donor. Data on approximately 6,195 donors per year and 15,967 new patients per year. Survival tracked through February 29, 2020 (up to 20 years and 2 months from registration).

Both datasets are supplied by UNOS as contractor for OPTN. Access requires a data use agreement with OPTN (https://optn.transplant.hrsa.gov/data/request-data/); data are proprietary-confidential.

## When to read the full paper

Use the [original](https://doi.org/10.3982/ECTA20203) if you are:
- Extending the identification framework to other assignment settings (public housing, school choice, gig-economy jobs) where agents face sequentially arriving heterogeneous objects and outcomes depend on the match;
- Evaluating the distributional consequences of alternative kidney allocation policies, particularly the trade-off between LYFT maximization and prioritizing the sickest (Table VIII, p. 427);
- Replicating the Gibbs sampler estimation of the joint model (replication code available; footnote 23, p. 416);
- Building on the identification results (Lemmas 1-2, Theorem 1) for settings with multiple unobserved dimensions of heterogeneity in both choices and outcomes.

The variance decomposition (patient vs donor vs match, p. 423) and the planner's dilemma (Figure 4, p. 425; Table VIII, p. 427) are the most directly policy-relevant sections.

## Attribution and rights

Source: peer-reviewed, *Econometrica* 93(2), March 2025. This distillation was extracted by an LLM on 2026-06-26 and is **not human-verified or independently reproduced**. The paper is paywalled; only text extraction is permitted here.

> Agarwal, Nikhil, Charles Hodgson, and Paulo Somaini.
> "Choices and Outcomes in Assignment Mechanisms: The Allocation of Deceased Donor Kidneys."
> *Econometrica* 93, no. 2 (March 2025): 395-438.
> DOI: 10.3982/ECTA20203. (c) 2025 The Econometric Society.
> Extract-only; full text at the [Econometric Society](https://doi.org/10.3982/ECTA20203).


==============================================================================
# Location Sorting and Endogenous Amenities: Almagro & Dominguez-Iino (2025)
# https://instituteforautomatedresearch.org/wiki/papers/econometrica/2025/almagro-iino-location-sorting-endogenous-amenities-2025/
# Distilled: A dynamic spatial equilibrium model of Amsterdam shows that heterogeneous household preferences over endogenous consumption amenities increase residential sorting across neighborhoods but reduce welfare inequality, while short-term rental entry raises rents for all but redistributes welfare gains and losses across household types through the amenity channel. Econometrica 2025, CC BY-NC-ND 4.0. Five core results with source locators, datasets used, the model, and the method with its defining equations.
# Tags: paper-summary, urban-economics, residential-sorting, housing-markets, short-term-rentals
==============================================================================

**What this is.** The paper's core results, the structural equilibrium model (endogenous amenities + dynamic location choice), and the estimation equations: enough to understand what was found and how, without reading all 41 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.3982/ECTA21394).

## TL;DR

The paper builds and estimates a dynamic spatial equilibrium model of Amsterdam in which heterogeneous households make forward-looking residential choices and firms endogenously supply consumption amenities (restaurants, bars, nurseries, touristic venues, food and non-food stores) in response to the neighborhood's demographic composition. Using restricted Dutch administrative microdata (CBS) linked to neighborhood amenity counts (ACD BBGA) and short-term rental listings (Inside Airbnb), and exploiting tourist inflows as a demand shifter via a shift-share instrument, the paper shows: (1) short-term rental (STR) penetration raises rents 0.09-0.21% per 1% growth in listings (IV); (2) tourist presence increases touristic amenities and restaurants but leaves nurseries unchanged; (3) preference heterogeneity across household types increases residential sorting but reduces welfare inequality relative to the homogeneous-preference benchmark, because neighborhoods become horizontally differentiated; and (4) STR entry produces winner-loser welfare splits by household type once amenity adjustment is allowed: the highest-income group (Older Families) loses 4% of income while lower-income Singles and Younger Families gain 1-2%. The paper extends prior work by Guerrieri, Hartley, and Hurst (2013) and Diamond (2016) by microfounding how different amenity types respond to demographic heterogeneity.

## Core results

Magnitudes and significance are as reported; `\*` = 5%, `\*\*` = 1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **STR penetration raises neighborhood rents** (IV); OLS is downward-biased, consistent with tourist-attractive areas becoming locally less attractive to residents | Table I, p.1038 | IV (full controls + district-year FE): coeff = 0.205 (SE 0.093), F = 69.66; OLS = 0.109 (SE 0.018); range across specs: 0.091-0.205 |
| R2 | **STR penetration raises house sale prices** (IV); OLS severely underestimates the price effect | Table I, p.1038 | IV (full controls + district-year FE): coeff = 0.326 (SE 0.102), F = 65.9; OLS = 0.037 (SE 0.022); range: 0.149-0.326 |
| R3 | **Tourist presence drives supply of touristic amenities but not nurseries**; supply responses are sectorally differentiated by the demographic type driving demand | Table III, p.1054; text p.1053 | 10% more tourists: +2.3% touristic amenities, +0.5% restaurants, +2.3% bars, +0.9% food stores, +2.9% non-food stores, 0% nurseries |
| R4 | **Heterogeneous preferences increase residential sorting but reduce welfare inequality** relative to the homogeneous benchmark; horizontal neighborhood differentiation is the mechanism | Figure 9, p.1064; Table VII, p.1065 | Entropy index: 0.8 (heterogeneous) vs 0.4 (homogeneous); welfare gap (max/min consumer surplus): ~1 vs ~10; Gini indices rise for 5 of 6 amenity sectors under heterogeneous preferences |
| R5 | **STR entry produces welfare gains for younger/lower-income households and losses for older/higher-income households** once amenity endogeneity is accounted for; direction reverses vs the exogenous-amenity benchmark | Figure 10, p.1066 | Older Families: -4% income CE loss; Singles: +1-2% CE gain; Younger Families: +1-2% CE gain; under exogenous amenities all lose 1-2% (dark bars in Figure 10) |

**Overall (paper's conclusion).** Two-way heterogeneity, in household preferences and in amenity supply responses, determines both the degree of horizontal differentiation across neighborhoods and the distributional incidence of urban policies. Low-income households may gain rather than lose from STR entry if the amenities tourists bring align with their preferences, reversing naive predictions based on rent effects alone. The amenity channel matters for incidence qualitatively, not just quantitatively.

## Theory / model

The model has three blocks: endogenous amenity determination, housing supply, and household/tourist location demand. There are $$J+1$$ locations ($$J$$ inside the city plus an outside option) and $$K+1$$ household types ($$K$$ local types and a tourist type $$T$$). The population composition of location $$j$$ at time $$t$$ (equation 1, p.1042) is:

$$
M_{jt} \equiv [M^1_{jt}, \ldots, M^K_{jt}, M^T_{jt}]', \tag{1}
$$

and the amenity vector (equation 2, p.1042) is:

$$
a_{jt} \equiv [N_{1jt}, \ldots, N_{Sjt}]', \tag{2}
$$

where $$N_{sjt}$$ is the number of varieties in amenity sector $$s$$ at location $$j$$.

**Endogenous amenities.** Households have Cobb-Douglas preferences over housing and a composite amenity good, with $$\phi^k$$ the expenditure share on amenities for type $$k$$. Within each amenity sector, firms supply differentiated varieties under CES preferences (substitution elasticity $$\sigma_s > 1$$). Individual demand for variety $$i$$ in sector $$s$$ at location $$j$$ (equation 3, p.1042) is:

$$
q^k_{isjt} = \frac{\alpha^k_s \phi^k w^k_t}{p_{isjt}} \left(\frac{p_{isjt}}{P_{sjt}}\right)^{1-\sigma_s}, \tag{3}
$$

where $$\alpha^k_s$$ is type $$k$$'s sectoral budget share and $$P_{sjt}$$ is the sector-location price index. Firms engage in monopolistic competition with free entry, equating variable profits to a fixed cost $$F_{sjt}(N_{jt})$$ increasing in total firm count $$N_{jt}$$. The zero-profit condition yields equilibrium varieties (equation 7, p.1043):

$$
N_{sjt} = \frac{1}{\sigma_s F_{sjt}} \sum_k \alpha^k_s \phi^k w^k_t M^k_{jt}. \tag{7}
$$

This delivers a mapping $$a_{jt} = \mathcal{A}(M_{jt})$$ (equation 8, p.1043): equilibrium amenities are a function of population composition alone, encoding the preference-externality mechanism.

**Housing supply.** Absentee landlords choose between the long-term (LT) rental market (income $$r_{jt}$$ per floor-space unit) and the short-term (ST) market (income $$p_{jt}$$) subject to an operating cost wedge $$\kappa_{jt}$$. Under Type I EV shocks, long-term supply (equation 9, p.1044) and short-term supply (equation 10) are:

$$
\mathcal{H}^{LT,S}_{jt}(r_{jt}, p_{jt}) = \frac{\exp(\alpha r_{jt})}{\exp(\alpha r_{jt}) + \exp(\alpha p_{jt} - \kappa_{jt})} \mathcal{H}_{jt}, \tag{9}
$$

$$
\mathcal{H}^{ST,S}_{jt}(r_{jt}, p_{jt}) = \mathcal{H}_{jt} - \mathcal{H}^{LT,S}_{jt}(r_{jt}, p_{jt}). \tag{10}
$$

**Local household location choice.** At each period $$t$$, household $$i$$ of type $$k$$ chooses location $$j_{it}$$ to maximize discounted expected utility. The flow utility inside the city (p.1044-1045) is:

$$
u^k_t(j, x_{it}) = \bar{u}^k_t(j) + \delta^k_\tau \log \tau_{it} - MC^k(j_{it}, j_{it-1}), \tag{11}
$$

where $$\tau_{it}$$ is location tenure, $$MC^k$$ is a moving cost that combines a bilateral distance-adjusted component and a fixed component, and $$\bar{u}^k_t(j)$$ collects aggregate state-dependent payoffs. Under Type I EV preference shocks, location choice probabilities (equation 12, p.1045) are:

$$
\mathbb{P}^k_t(j|x_{it}) = \frac{\exp\!\left(u^k_t(j,x_{it}) + \beta \mathbb{E}_t[V^k_{t+1}(x_{it+1},\varepsilon_{it+1})|j,x_{it},\varepsilon_{it}]\right)}{\sum_{j'} \exp\!\left(u^k_t(j',x_{it}) + \beta \mathbb{E}_t[V^k_{t+1}(x_{it+1},\varepsilon_{it+1})|j',x_{it},\varepsilon_{it}]\right)}. \tag{12}
$$

**Stationary equilibrium.** A stationary equilibrium (Definition, p.1048) is a vector of long-term rental prices $$\mathbf{r}$$, short-term rental prices $$\mathbf{p}$$, amenities $$\mathbf{a}$$, and stationary population distributions $$\pi^k(\mathbf{r}, \mathbf{a})$$ for each type $$k$$, such that the long-term and short-term rental markets clear for every location and $$a_j = \mathcal{A}(M_j)$$ for every $$j$$. Population and amenities are thus jointly determined in equilibrium.

## Method

Estimation proceeds in three independent blocks.

**Amenity supply (GMM).** Taking logs of equation (7) and parameterizing the fixed cost as $$F_{sjt}(N_{jt}) = \Lambda_j \Lambda_t R(N_{jt}) \Omega_{sjt}$$ with $$R(N_{jt}) = N^\eta_{jt}$$, the estimating equation (equation 24, p.1051) is:

$$
\log N_{sjt} = \lambda_j + \lambda_t - \eta \log N_{jt} + \log\!\left(\sum_k \beta^k_s X^k_{jt}\right) + \omega_{sjt}, \tag{24}
$$

where $$X^k_{jt} \equiv \phi^k w^k_t M^k_{jt}$$ is total amenity expenditure by type $$k$$ in location $$j$$, $$\beta^k_s \equiv \alpha^k_s / \sigma^s$$ captures how type $$k$$'s expenditure converts to amenity supply in sector $$s$$, and $$\omega_{sjt}$$ is an unobservable supply shock. The key endogeneity concern is that $$\omega_{sjt}$$ shifts firm costs and thus residential composition $$M^k_{jt}$$ simultaneously. The instrument is $$Z^k_{jt} = w^k_t S^{\gamma(k)}_{jt}$$, the interaction of type $$k$$'s wages with the housing stock of its modal tenancy status $$\gamma(k)$$, exploiting the idea that neighborhoods composed primarily of social housing attract households qualifying for social housing assistance. GMM is implemented on a three-way panel of 22 districts for 2008-2018. The housing supply inverse elasticity $$\eta = 1.52$$ is calibrated from Saiz (2010).

**Housing demand from locals (ECCP).** The method builds on `eccp-estimator` (Aguirregabiria and Mira (2010), Scott (2013), Kalouptsidi, Scott, and Souza-Rodrigues (2021b)). The parametric flow utility (equation 27, p.1056) for type $$k$$ in location $$j \neq 0$$ is:

$$
\bar{u}^k_t(j) = \delta^k_j + \delta^k_t + \delta^k_r \log r_{jt} + \delta^k_a \log a_{jt} + \delta^k_b \log b_{jt} + \xi^k_{jt}, \tag{27}
$$

where $$\delta^k_a = [\delta^k_1, \ldots, \delta^k_s, \ldots, \delta^k_S]$$ is the vector of amenity preference parameters (one per sector). Exploiting renewal actions (household pairs who at $$t+1$$ choose the same new location $$\tilde{j}$$, so their continuation values cancel), the estimating equation (equation 29, p.1057) is:

$$
Y^k_{t,j,\tilde{j},x_{it}} = \delta^k_j + \delta^k_t + \delta^k_r \log r_{jt} + \delta^k_a \log a_{jt} + \delta^k_b \log b_{jt} + \delta^k_\tau \Delta\tau_{it} - \Delta MC^k_{it} + \tilde{\xi}^k_{t,j,\tilde{j},x_{it}}, \tag{29}
$$

where $$Y^k$$ is the log ratio of path likelihoods for the two paths diverging from state $$x_t$$ and converging at $$\tilde{j}$$ in period $$t+1$$, and $$\Delta\tau_{it}$$ is the change in location capital (tenure). The left-hand side is formed from conditional choice probabilities estimated via multinomial logit. Seven instruments are used to address endogeneity of rents and amenities: three post-2011 rental-market policy dummies (social-housing reclassification 2011, rent deregulation 2015, STR regulation 2017) interacted with lagged tenancy stock, plus removal of housing units inside and outside the precinct. First-stage F-stat = 169.8.

**Housing supply.** From equation (9)-(10), the log ratio of long- to short-term supply shares gives (p.1061):

$$
\log \mathcal{H}^{LT,S}_{jt} - \log \mathcal{H}^{ST,S}_{jt} = \alpha(r_{jt} - p_{jt}) + \kappa_j + \kappa_t + \nu_{jt}.
$$

The instrument for the relative price $$r_{jt} - p_{jt}$$ is predicted tourist demand from a shift-share following Barron, Kung, and Proserpio (2021): the "shift" is Airbnb worldwide search volume; the "share" is neighborhood-level exposure from the historic spatial distribution of touristic attractions. IV estimate (two-way FE preferred spec): $$\hat{\alpha} = 0.385$$ (Table VI, p.1062), implying a 1 SD increase in the STR-LT price gap (29%) raises the short-term market share by 13.6%.

**Household type classification.** Six types are identified using `k-means-clustering` on income, skill, household composition, and ethnicity from CBS tax returns (Table II, p.1050): Older Families, Singles, Younger Families (market-determined types used in structural estimation) and Students, Immigrant Families, Dutch Low Income (treated as exogenous allocation in social-housing/university assignment). Discount factor $$\beta = 0.85$.

## Empirical specifications

**Reduced-form STR rent effects (R1, R2).** The estimating equation for Table I (p.1038) is:

$$
\ln Y_{jt} = \beta \ln(\text{Commercial Airbnb listings}_{jt}) + \gamma X_{jt} + \mu_{dt} + \varepsilon_{jt},
$$

where $$Y_{jt}$$ is either rent/m² or house sale price, $$X_{jt}$$ includes housing stock, average income, and high-skill population share, and $$\mu_{dt}$$ are district-year fixed effects. Standard errors are clustered at the wijk (neighborhood) level. The shift-share IV for Airbnb listings uses worldwide Airbnb search volume as the shift and the spatial density of historic monuments as the neighborhood share. First-stage F-stats exceed 65 in all IV specifications.

**Amenity supply estimation (R3).** GMM on equation (24) for 6 amenity sectors simultaneously, using the district-level panel 2008-2018. Parameter $$\beta^k_s$$ captures how a 1% increase in type $$k$$'s expenditure on location $$j$$ changes the number of sector-$$s$$ firms. The economic magnitude is translated to tourist effects in the text (p.1053): a 10% increase in city-wide tourists shifts amenity composition toward touristic and retail amenities and away from nurseries.

**Preference heterogeneity and sorting/inequality counterfactual (R4).** Two equilibria are compared: the baseline with heterogeneous amenity preferences (from Table IV estimates) against a homogeneous-preference counterfactual in which $$\delta^k_a$$ is replaced by its population-weighted average. Sorting is measured by an entropy index (Figure 9, p.1064); welfare inequality is the ratio of the highest to lowest consumer surplus in euros across household types.

**STR welfare decomposition (R5).** Three equilibria are compared step by step: (i) pre-STR equilibrium $$(\mathbf{r}_0, \mathbf{a}_0)$$; (ii) post-STR with exogenous amenities $$(\mathbf{r}_1, \mathbf{a}_0)$$; (iii) post-STR with endogenous amenities $$(\mathbf{r}_1, \mathbf{a}_1)$$. Welfare is measured in consumption equivalent (CE) terms: how much extra income a household in the pre-STR equilibrium must receive to be as well off as in the counterfactual. Positive CE values indicate welfare gains. Homeowners receive back landlord income from rent increases; renters do not (Supplemental Appendix A.5, p.1065).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CBS residential cadaster (Centraal Bureau voor de Statistiek, Netherlands) | Individual-level annual residential histories for universe of Dutch residents; key panel for location choice estimation | no page yet |
| CBS tax return data | Household income, educational attainment, employment status, ethnic background; source for household type classification | no page yet |
| CBS housing unit tax appraisal panel 2006-2020 | Property values, tenancy status, geo-coordinates, quality measures for universe of Dutch residential units | no page yet |
| CBS national rent survey 2006-2019 | Rental prices per neighborhood; imputed via random forest and CBS valuations (Mullainathan and Spiess 2017) | no page yet |
| Amsterdam City Data BBGA (ACD) | Annual neighborhood-level demographics, amenity establishment counts, tourist inflows; 95 wijk / 22 districts, 2008-2018; publicly available at ACD BBGA | no page yet |
| ACD Tourism data | City-level tourist overnight stays and hotel room counts; public via ACD Tourism portal | no page yet |
| Inside Airbnb | Monthly web-scraped listing-level STR data for Amsterdam (prices per night, calendar availability, reviews); used to construct commercial listings time series | no page yet |

Sample period: 2008-2018 (annual). Household type classification uses the CBS panel of 672,093 households. Amenity supply estimated on 22 districts. Housing demand estimated on 22 districts with 46 individual states per type per year.

## When to read the full paper

Read the [original](https://doi.org/10.3982/ECTA21394) if you are: (i) building a structural spatial equilibrium model with endogenous amenities and need the full equilibrium existence/uniqueness arguments (Supplemental Appendix A.4); (ii) running welfare counterfactuals for STR regulation in a city with heterogeneous amenity demand and need the CE calculation formulas (Supplemental Appendix A.5); (iii) implementing the ECCP estimator for dynamic location choice and need the finite-dependence / renewal-action derivation (Supplemental Appendix A.6); or (iv) using CBS microdata or ACD BBGA for Amsterdam and need the exact variable construction (Supplemental Appendix A.2). Tables I, III, IV, VI are the main empirical anchors; Figures 9, 10, 12 are the main counterfactual exhibits.

## Attribution and rights

Source: peer-reviewed, *Econometrica* 93(3) (May 2025). This distillation was extracted by an LLM on 2026-06-26 and is **not human-verified or independently reproduced**. The CC BY-NC-ND 4.0 licence permits non-commercial use with attribution and no modifications; the verbatim PDF is not hosted here.

> **Attribution (CC BY-NC-ND 4.0).** Almagro, Milena, and Tomás Domínguez-Iino.
> "Location Sorting and Endogenous Amenities: Evidence From Amsterdam."
> *Econometrica* 93, no. 3 (May 2025): 1031-1071.
> DOI: 10.3982/ECTA21394. © 2025 The Authors.
> Licensed under [CC BY-NC-ND 4.0](https://creativecommons.org/licenses/by-nc-nd/4.0/).
> This page is a distillation by the Institute for Automated Research: core results extracted and re-expressed. The licence prohibits modifications and commercial use; this extract is used for non-commercial research reference only.


==============================================================================
# Competitive Capture of Public Opinion: Alonso & Padró i Miquel (2025)
# https://instituteforautomatedresearch.org/wiki/papers/econometrica/2025/alonso-miquel-competitive-capture-public-opinion-2025/
# Distilled: Two opposed interested parties compete to capture news coverage; rational citizens discount informative messages and sort into aligned sources, so competition compounds rather than cancels harm to social learning. Econometrica 2025, CC BY 4.0. Six core propositions with locators, the capture-and-communication game model, and equilibrium characterization with equations.
# Tags: paper-summary, political-economy, media-bias, information-economics, public-opinion
==============================================================================

**What this is.** The propositions, model, and equilibrium characterization of this paper on competitive capture of public opinion: enough to know what was proved and how, without reading the full 33 pages. To replicate, extend, or verify any result, read the original at [https://doi.org/10.3982/ecta22072](https://doi.org/10.3982/ecta22072). Page references are to the accepted author manuscript (AAM).

## TL;DR

Two interested parties (IPs), right (R) and left (L), compete to capture news items produced by multiple information sources that reach citizens with heterogeneous prior beliefs over a binary state of the world. When an IP captures a news item it can publish any message it likes, genuine disinformation with no commitment and no restriction. Citizens rationally discount suspicious coverage. The paper characterizes the Perfect Bayesian Equilibrium of this capture-and-communication game and obtains four results: (i) each IP mixes over an interval of favorable messages, equalizing effective informational content within its support (Proposition 1); (ii) published coverage is more polarized than honest coverage and rational skepticism makes it less informative than face value suggests (Section 3.2); (iii) competing IPs do not cancel each other but compound harm to social learning because capture efforts are strategic substitutes (Proposition 3); (iv) citizens sort into ideologically aligned sources despite knowing the bias (Proposition 6). These results match documented empirical patterns in the item-level distribution of media slant.

## Core results

Page locators refer to the AAM; equation numbers are identical across AAM and VOR.

| # | Result | Locator | Magnitude as stated |
|---|---|---|---|
| R1 | In the unique communication equilibrium, R randomizes over messages with $$\lambda_H(m) \ge \overline{\lambda}$$ and L over messages with $$\lambda_H(m) \le \underline{\lambda}$$; citizens treat every message in each IP's support as conveying the same constant effective likelihood ratio | Proposition 1, eq. (2), AAM p. 13 | $$\lambda^*(m) = \overline{\lambda}$$ for all $$m \in \text{supp}(\tau_R^*)$$; $$\lambda^*(m) = \underline{\lambda}$$ for all $$m \in \text{supp}(\tau_L^*)$$; moderate messages $$m \in (\underline{m}^*, \overline{m}^*)$$ taken at honest face value |
| R2 | Capture shifts the published coverage distribution to the tails: extreme messages become more frequent, moderate messages less frequent, than under honest coverage | Proposition 1, Figure 1, AAM pp. 13-15 | Equilibrium density has higher mass at both tails relative to the honest distribution $$F_H(\lambda)$$; model accommodates Budak, Goel, and Rao (2016) and Kim, Lelkes, and McCrain (2022) empirical patterns of within-outlet slant variation |
| R3 | Capture uniformly reduces Blackwell-informativeness of the source: the equilibrium message distribution SOSD-dominates honest coverage; higher effort by either IP compresses citizen posteriors further | Lemma 1, eq. (6), Section 3.2, AAM pp. 16-17 | $$F(\lambda; p) = \pi_L(r,l) + \pi_H(r,l) F_H(\lambda; p)$$ for $$\lambda \in [\underline{\lambda}, \overline{\lambda}]$$; $$\overline{\lambda}$$ is decreasing and $$\underline{\lambda}$$ is increasing in each IP's own effort, so more capture always makes the source less informative |
| R4 | Competing capture efforts are strategic substitutes at the item level: one IP's higher effort reduces the other IP's marginal return to capture | Proposition 3, AAM p. 21 | Under Assumptions I-II: $$\partial B^R / \partial l < 0$$ and $$\partial B^L / \partial r < 0$$ along the best-response locus; equilibrium $$r^*$$ and $$l^*$$ move in opposite directions to each other's effort |
| R5 | A horizontal source attribute (favoring one IP) unambiguously increases that IP's capture and decreases the opponent's; strategic substitution amplifies differentiation | Proposition 5, AAM p. 26 | Under Assumptions I-II: if horizontal attribute $$\zeta$$ favoring R increases, there exists an equilibrium $$(\bar{r}, \bar{l})$$ with $$\bar{r}_j \ge r_j^*$$ and $$\bar{l}_j \le l_j^*$$ |
| R6 | Citizens sort ideologically: those with rightist priors choose the source mostly captured by R; those with leftist priors choose the source mostly captured by L | Proposition 6, AAM pp. 28-29 | With symmetric sources and $$\pi_R^1/\pi_R^2 > \pi_H^1/\pi_H^2 > \pi_L^1/\pi_L^2$$: there exist $$\underline{p} \le \bar{p}$$ such that citizens with $$p < \underline{p}$$ choose source 2 and $$p > \bar{p}$$ choose source 1; when $$\pi_H^1 = \pi_H^2$$, sorting is monotone in $$p$$ |

**Overall (paper's conclusion).** Competition between IPs does not restore informational balance: opposing capture efforts are strategic substitutes at each item so they do not cancel, they compound harm to social learning. Horizontal differentiation between sources is amplified by competition, not dampened. Citizens rationally sort into ideologically aligned sources, consistent with recent experimental evidence on demand for biased news, not because they prefer bias but because the lies they fear most come from the ideologically opposed source.

## Theory / model

**State and citizens.** The unknown binary state is $$\theta \in \Theta = \{-1, 1\}$$. A mass $$M$$ of citizens have heterogeneous prior beliefs $$p = \Pr[\theta = 1]$$ distributed with CDF $$F_p(p)$$. IP R wants citizens to hold the highest possible posterior on $$\theta = 1$$; IP L wants the lowest. Their indirect utilities over citizen posteriors are $$v_R(\mu)$$ (strictly increasing) and $$v_L(\mu)$$ (strictly decreasing), differentiable on $$[0,1]$$ with bounded derivatives.

**Honest news.** Each of $$n$$ sources produces one news item $$j$$. If uncaptured (honest), item $$j$$ conveys an informative signal $$m^j \in \mathcal{M} \subset \mathbb{R}$$ with state-dependent density $$\Pr[m^j = m \mid \theta] = q_\theta^j(m)$$, conditionally independent across items. The honest posterior of a $$p$$-citizen who observes message $$m$$ is (§2, eq. (1), AAM p. 8):

$$\mu_H^j(m; p) = \Pr\!\bigl[\theta = 1 \mid m^j = m,\, \text{honest},\, p\bigr] = \frac{q_1^j(m)\, p}{q_1^j(m)\, p + q_{-1}^j(m)(1-p)} \tag{1}$$

Messages are ordered by the likelihood ratio $$\lambda_H(m) = q_1^j(m)/q_{-1}^j(m)$$: higher $$\lambda_H(m)$$ means a message more favorable to $$\theta = 1$$.

**Capture and timing.** IPs simultaneously and covertly choose efforts $$r_j \in [0, \bar{x}_R^j]$$ and $$l_j \in [0, \bar{x}_L^j]$$ for each item $$j$$. Nature draws the state of capture $$S^j \in \{R, L, H\}$$ with probabilities $$\pi_R^j(r_j, l_j)$$, $$\pi_L^j(r_j, l_j)$$, $$\pi_H^j = 1 - \pi_R^j - \pi_L^j$$. If IP $$i$$ wins, it publishes any $$m \in \mathcal{M}$$ regardless of the true state (genuine disinformation). Citizens observe the published message and update beliefs without observing whether capture occurred. Each IP's cost of capture effort across items is $$C_R(r) = \sum_j C_{Rj}(r_j)$$ and $$C_L(l) = \sum_j C_{Lj}(l_j)$$ with $$C_{ij}$$ increasing and strictly convex.

The equilibrium concept is Perfect Bayesian Equilibrium (PBE). Citizens hold assessments $$(\tilde{r}, \tilde{l}, \tilde{\tau}_R, \tilde{\tau}_L)$$ of IPs' efforts and reporting strategies; in any PBE these assessments are correct on the equilibrium path.

The paper advances on prior work by allowing two opposing IPs (unlike Besley and Prat (2006), which has one), using a continuous message space (unlike binary disclosure models), and imposing no commitment to an editorial line (unlike Gentzkow and Kamenica (2017), where the sender commits to an information structure). Prat (2018) gives upper bounds on IP influence in a multiple-media setting; this paper endogenizes the capture incentives.

## Method

**Characterizing the communication equilibrium.** For fixed efforts $$(r, l)$$, Proposition 1 (AAM p. 13) characterizes the unique communication equilibrium. R mixes over messages whose honest likelihood ratio $$\lambda_H(m) \ge \overline{\lambda}$$; L mixes over messages with $$\lambda_H(m) \le \underline{\lambda}$$. The equilibrium likelihood ratio $$\lambda^*(m) \equiv \Pr[m \mid \theta=1]/\Pr[m \mid \theta=-1]$$ takes the censored form (eq. (2)):

$$\lambda^*(m) = \begin{cases} \underline{\lambda} & \text{if } m \le \underline{m}^* \\ \lambda_H(m) & \text{if } \underline{m}^* < m < \overline{m}^* \\ \overline{\lambda} & \text{if } m \ge \overline{m}^* \end{cases} \tag{2}$$

The key step is that each IP's indirect payoff $$V_i(m) = M \int_0^1 v_i(\mu^*(\lambda; p))\, dF_p(p)$$ is strictly monotone in $$\lambda^*(m)$$, so IP optimality requires equalizing $$\lambda^*(m)$$ across all messages in the support of $$\tau_i^*$$ (the mixing condition). The thresholds $$\overline{\lambda}$$ and $$\underline{\lambda}$$ are pinned down by the mass conditions (Proposition 1, Part 3, eqs. (3)-(4)):

$$\int_{\overline{\lambda}}^{\infty} \!\!(\lambda - \overline{\lambda})\, dF_{H,-1}(\lambda) = \frac{\pi_R(r,l)}{\pi_H(r,l)}\,(\overline{\lambda} - 1) \tag{3}$$

$$\int_0^{\underline{\lambda}} (\underline{\lambda} - \lambda)\, dF_{H,-1}(\lambda) = \frac{\pi_L(r,l)}{\pi_H(r,l)}\,(1 - \underline{\lambda}) \tag{4}$$

where $$F_{H,-1}(\lambda)$$ is the distribution of honest likelihood ratios in state $$\theta = -1$$. Because the right-hand side of (3) is strictly increasing in $$\overline{\lambda}$$ and the left-hand side is strictly decreasing, the solution is unique.

**Full-game equilibrium.** IP $$i$$'s payoff given efforts $$(r, l)$$ and citizens' assessment $$(\tilde{r}, \tilde{l})$$ is (eq. (7), AAM p. 19):

$$W_i(r,l;\tilde{r},\tilde{l}) = \pi_L(r,l)\,V_i\!\bigl(\underline{\lambda}(\tilde{r},\tilde{l})\bigr) + \pi_H(r,l)\,\mathbb{E}_H\!\bigl[V_i(\lambda); p_i\bigr] + \pi_R(r,l)\,V_i\!\bigl(\overline{\lambda}(\tilde{r},\tilde{l})\bigr) \tag{7}$$

where $$\mathbb{E}_H[V_i(\lambda); p_i]$$ integrates $$V_i(\lambda)$$ over the honest message distribution. Proposition 2 (AAM p. 19) establishes existence of a pure-strategy equilibrium $$(r^*, l^*)$$ satisfying the first-order conditions:

$$B_R(r^*, l^*;\, r^*, l^*) = C_R'(r^*) \tag{11}$$
$$B_L(r^*, l^*;\, r^*, l^*) = C_L'(l^*) \tag{12}$$

combined with the communication-equilibrium conditions (3)-(4), where $$B_i$$ is the marginal benefit from capture (the integral of $$V_i'(\lambda)$$ weighted by the effect of a marginal increase in capture probability over the contested message range, eqs. (9)-(10) in Proposition 2).

**Strategic substitutability.** Proposition 3 (AAM p. 21) is proved by differentiating $$B_R$$ with respect to $$l$$ (Assumption I rules out cross-partials in the contest function; Assumption II requires $$\pi_R/\pi_H$$ increasing in $$l$$, ensuring that higher left effort increases the perceived odds that honest coverage is crowded out rather than that the right is crowded in). Both effects reduce R's marginal return, giving (eq. (15), AAM p. 21):

$$\frac{\partial B_R}{\partial l} + \frac{\partial B_R}{\partial \tilde{l}}\bigg|_{l=\tilde{l}} < 0$$

and symmetrically $$\partial B_L / \partial r < 0$$ along the best-response locus.

## Empirical specifications

This is a pure theory paper. There are no estimating equations, regression specifications, or structural estimation exercises. The model's equilibrium predictions on the shape of the published coverage distribution are compared qualitatively to three empirical studies of item-level slant (§3.1, AAM pp. 14-15). Budak, Goel, and Rao (2016) measure ideological content of articles at top US news outlets using crowdsourced scoring and find that within-outlet variation in slant far exceeds across-outlet variation and that a large share of items is centrist. Kim, Lelkes, and McCrain (2022) study dynamic media bias in cable news and find large week-by-week variation within program. Braghieri, Eichmeyer, Levy, Mobius, Steinhardt, and Zhong (2024) document article-level slant on social media and find only about 35% of article-level variance is explained by outlet differences. All three patterns are consistent with the model: equilibrium coverage puts mass at both tails and in the center because each IP mixes over an interval of favorable messages, not a single extreme one.

Suen (2004) is discussed in Section 6 as a contrast: in that model media filters rather than lies, and bias can create value for aligned citizens. Here disinformation without commitment destroys value for rational citizens, generating sorting for a different reason.

In Section 7, robustness is established analytically for: (i) a mixed population including naive citizens who take coverage at face value; (ii) multi-homing (citizens observe more than one news item); and (iii) ideology reflecting heterogeneous preferences rather than heterogeneous beliefs. Shapiro (2016) is discussed there as related work on multiple IPs but a single outlet.

## Datasets used

This is a pure theory paper. No empirical datasets are used in the analysis. The model is motivated by and compared qualitatively to published empirical studies of media slant; no proprietary or public microdata are analyzed directly.

## When to read the full paper

Read the source at [https://doi.org/10.3982/ecta22072](https://doi.org/10.3982/ecta22072) if you are:
- modeling how competing interest groups influence information intermediaries (media, social media platforms, scientific discourse)
- studying information transmission under strategic manipulation without commitment to an editorial or publishing rule
- extending the framework to allow sources to be strategic (profit-maximizing, reputation-seeking) rather than passive
- applying the model to social media bot campaigns (Section 2 explicitly discusses this interpretation), public health campaigns, or regulatory communications
- replicating the proofs: the online appendix contains extensions including non-separable cost functions (§OA-13), naive citizens (§OA-15), and preference heterogeneity (§OA-16)

## Attribution and rights

Source: peer-reviewed, *Econometrica* 93(4), 2025. The accepted author manuscript is available under CC BY 4.0 from LSE Research Online. This distillation was extracted by an LLM on 2026-06-26 and is **not human-verified or independently reproduced**. The VOR licence could not be confirmed via Crossref (no license block); the AAM licence and the VOR licence may differ.

> **Attribution (CC BY 4.0 - AAM).** Alonso, Ricardo, and Gerard Padró i Miquel.
> "Competitive Capture of Public Opinion."
> *Econometrica* 93, no. 4 (2025): 1265-1297.
> DOI: 10.3982/ecta22072.
> Accepted author manuscript available at LSE Research Online (eprint/127777)
> under Creative Commons Attribution 4.0 International (CC BY 4.0).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Adaptive Maximization of Social Welfare: Cesa-Bianchi, Colomboni & Kasy (2025)
# https://instituteforautomatedresearch.org/wiki/papers/econometrica/2025/bianchi-et-al-adaptive-maximization-social-welfare-2025/
# Distilled: A policymaker repeatedly setting a tax rate to maximize social welfare (weighted sum of public revenue and private consumer surplus) cannot observe welfare directly, only demand outcomes; cumulative regret must grow at rate T^{2/3} (vs T^{1/2} for standard bandits), and Tempered Exp3 achieves this bound while Dyadic Search recovers T^{1/2} under concavity. Econometrica 2025, CC BY 4.0. Six core results with source locators, the setup model, and both algorithms with equations.
# Tags: paper-summary, optimal-taxation, online-learning, public-economics, social-welfare
==============================================================================

**What this is.** The paper's core results, the social welfare setup, and the two algorithms (Tempered Exp3 for Social Welfare and Dyadic Search) with their defining equations: enough to understand what was proved and how, without reading all 32 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.3982/ECTA22351).

## TL;DR

A policymaker repeatedly sets a tax rate $$x_i \in [0,1]$$ to maximize social welfare (a weighted sum of public revenue and private consumer surplus). Welfare is not observed: after setting $$x_i$$, only the binary demand decision $$y_i = \mathbf{1}(x_i \leq v_i)$$ is revealed, where $$v_i$$ is the individual's unobserved willingness to pay. The paper shows this is harder than monopoly pricing or standard multiarmed bandits: any algorithm must suffer cumulative regret of order $$T^{2/3}$$, because welfare depends on integrated demand over counterfactual (suboptimal) tax rates that must be actively explored. The paper proposes Tempered Exp3 for Social Welfare, achieving $$c \cdot \log(T)^{1/3} T^{2/3}$$ adversarial regret (matching the lower bound up to a log factor), and Dyadic Search, achieving $$T^{1/2}$$ in the stochastic setting when welfare is concave. Extensions to nonlinear income taxation (Section 5) and commodity taxation (Section 6) are developed.

## Core results

Rates hold up to logarithmic factors unless stated. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Lower bound on regret**: any randomized algorithm must suffer cumulative regret of order $$T^{2/3}$$ for both stochastic and adversarial preference sequences | Theorem 1, p. 1081 | $$\mathcal{R}_T \geq C \cdot T^{2/3}$$ for a constant $$C > 0$$; holds for finite and continuous policy sets, unlike standard bandits ($$T^{1/2}$$) |
| R2 | **Upper bound for Tempered Exp3**: with optimal tuning, adversarial (and stochastic) regret is bounded by $$c_4 \log(T)^{1/3} T^{2/3}$$ | Theorem 2, Corollary 1, Eq. 11, p. 1084 | $$\mathcal{R}_T \leq c_4 \cdot \log(T)^{1/3} T^{2/3}$$; matches lower bound up to $$\log(T)^{1/3}$$ |
| R3 | **Concave case lower bound**: stochastic regret cannot grow slower than $$T^{1/2}$$ when social welfare is concave | Theorem 3, p. 1086 | $$\mathcal{R}_T(\mathbf{G}) \geq C \cdot T^{1/2}$$; matches the multiarmed bandit rate |
| R4 | **Dyadic Search achieves $$T^{1/2}$$** for stochastic concave welfare via active interval narrowing | Theorem 4, p. 1086 | Order $$T^{1/2}$$ (up to log terms); rate-optimal for stochastic concave case |
| R5 | **Income taxation extension**: Tempered Exp3 adapted to $$H$$ wage brackets achieves adversarial regret of order $$H^{1/3} \log(T)^{1/3} T^{2/3}$$ | Theorem 5, Eq. 24, p. 1092 | $$\mathcal{R}_T \leq c_4 \cdot H^{1/3} \log(T)^{1/3} T^{2/3}$$ for $$H$$ tax brackets |
| R6 | **Welfare is harder than monopoly pricing**: for finite policy sets, monopoly pricing achieves $$T^{1/2}$$ while optimal taxation requires $$T^{2/3}$$ | Table I, p. 1080 | $$T^{1/2}$$ (monopoly pricing, finite) vs $$T^{2/3}$$ (optimal taxation, finite) vs $$T^{2/3}$$ (bilateral trade, finite; continuous rate is $$T$$) |

**Overall (paper's conclusion).** Welfare maximization is a strictly harder adaptive learning problem than reward maximization in standard bandits, because welfare depends on the integral of demand over counterfactual policies. The $$T^{2/3}$$ rate is sharp (lower and upper bounds coincide up to $$\log(T)^{1/3}$$). Concavity of welfare restores the $$T^{1/2}$$ bandit rate. The algorithms are adversarially robust and apply to public policy settings where behavioral responses (demand, labor supply) are observable but utility is not.

## Theory / model

**Setup.** Each period $$i = 1, 2, \dots, T$$, one individual arrives with unknown willingness to pay $$v_i \in [0, 1]$$. The policymaker sets a tax rate $$x_i \in [0, 1]$$ and observes the binary demand decision $$y_i = \mathbf{1}(x_i \leq v_i)$$. Define $$G_i(x) = \mathbf{1}(v_i \geq x)$$. Social welfare in period $$i$$ at the chosen tax rate $$x_i$$ is (Eq. 1, p. 1078):

$$
U_i(x_i) = \underbrace{x_i \cdot \mathbf{1}(x_i \leq v_i)}_{\text{Public revenue}} + \lambda \cdot \underbrace{\max(v_i - x_i,\, 0)}_{\text{Private welfare}} \tag{1}
$$

with $$\lambda \in (0, 1)$$ the welfare weight on private utility. Private welfare equals consumer surplus (integrated demand), so social welfare rewrites as (Eq. 2, p. 1078):

$$
U_i(x) = \underbrace{x \cdot G_i(x)}_{\text{Public revenue}} + \lambda \cdot \underbrace{\int_x^1 G_i(x')\,\mathrm{d}x'}_{\text{Private welfare}} \tag{2}
$$

Evaluating welfare at a counterfactual policy $$x \neq x_i$$ requires knowing integrated demand over $$[x, 1]$$, which must be learned by sampling suboptimal tax rates. This integral term makes the problem harder than monopoly pricing, where the objective $$U_i^{\text{MP}}(x) = x \cdot G_i(x)$$ requires only pointwise demand (Kleinberg and Leighton (2003)).

**Regret.** In the adversarial case, cumulative expected regret is (Eq. 3, p. 1078):

$$
\mathcal{R}_T\!\left(\{v_i\}_{i=1}^T\right) = \sup_{x}\, E\!\left[\mathbb{U}_T(x) - \mathbb{U}_T \,\big|\, \{v_i\}_{i=1}^T\right] \tag{3}
$$

where $$\mathbb{U}_T(x) = \sum_{i \leq T} U_i(x)$$ is cumulative welfare for fixed policy $$x$$, and $$\mathbb{U}_T = \sum_{i \leq T} U_i(x_i)$$ is welfare achieved by the algorithm. In the stochastic case, $$v_i \overset{\text{i.i.d.}}{\sim} \mu$$, the demand function is $$\mathbf{G}(x) = P(v_i \geq x) = E[y_i | x_i = x]$$, and cumulative regret is (Eq. 4, p. 1079):

$$
\mathcal{R}_T(\mathbf{G}) = T \cdot \sup_x \mathbf{U}(x) - E\!\left[\sum_{i \leq T} U(x_i)\right] \tag{4}
$$

**Lower bound (Theorem 1, p. 1081).** The proof constructs a family of distributions $$\{\mu^\epsilon\}_{\epsilon \in [-1,1]}$$ for $$v_i$$ with four-point support $$\{1/4, 1/2, 3/4, 1\}$$. Depending on the sign of $$\epsilon$$, the optimal policy is $$x^* = 1$$ or $$x^* = 1/4$$. Distinguishing the two requires sampling from the intermediate interval $$[1/2, 3/4]$$, which incurs regret of order $$|\epsilon|^{-2}$$. Minimizing over $$\epsilon \propto T^{-1/3}$$ gives the $$T^{2/3}$$ lower bound. The adversarial lower bound follows since worst-case regret dominates average-case regret.

**Comparison of learning problems (Table I, p. 1080).** The rate hierarchy is: monopoly pricing (Kleinberg and Leighton (2003)) achieves $$T^{1/2}$$ because its objective is one-sided Lipschitz and requires only pointwise demand; optimal taxation requires $$T^{2/3}$$ because welfare also depends on integrated demand for counterfactual policies; bilateral trade (Cesa-Bianchi, Cesari, Colomboni, Fusco, and Leonardi (2024a)) has rate $$T$$ because its objective is not one-sided Lipschitz. Mirrlees (1971) and Saez (2001) treat income tax design as a static structural problem; this paper adapts it to adaptive online learning. Ramsey (1927) commodity taxation is the basis for the Section 6 extension. Chetty (2009) establishes the sufficiency of the envelope-theorem welfare representation for arbitrary preference structures beyond the binary baseline. Lattimore and Szepesvari (2020) provide the $$T^{1/2}$$ multiarmed bandit baseline.

## Method

The paper proposes two algorithms. Algorithm 1 handles the general (possibly adversarial, nonconcave) case; Algorithm 2 exploits concavity in the stochastic case.

**Algorithm 1: Tempered Exp3 for Social Welfare.** The continuous policy space $$[0,1]$$ is discretized to $$K+1$$ evenly spaced grid points $$\tilde{x}_k = (k-1)/K$$. Three modifications are made relative to standard Exp3 (Auer, Cesa-Bianchi, Freund and Schapire (2002)) for the nonstochastic bandit:

1. Discretize $$[0,1]$$ to $$K+1$$ arms.
2. Estimate welfare indirectly via inverse-probability-weighted demand, since $$U_i(x)$$ is not directly observed.
3. Use a larger uniform exploration weight $$\gamma$$ than standard Exp3 to ensure sufficient sampling of suboptimal policies needed to estimate the integral term in $$U_i(x)$$.

Assignment probability for arm $$k$$ at period $$i$$ (Eq. 7, p. 1083):

$$
p_{ik} = (1 - \gamma) \cdot \frac{\exp(\eta \cdot \hat{\mathbb{U}}_{ik})}{\displaystyle\sum_{k'} \exp(\eta \cdot \hat{\mathbb{U}}_{ik'})} + \frac{\gamma}{K+1} \tag{7}
$$

where $$\eta$$ is the learning rate and $$\hat{\mathbb{U}}_{ik}$$ is the running cumulative welfare estimate at arm $$k$$. Demand at grid point $$k$$ is updated via inverse probability weighting after observing $$y_i$$ (Eq. 8, p. 1083):

$$
\hat{\mathbb{G}}_{i+1,k} = \hat{\mathbb{G}}_{i,k} + y_i \cdot \frac{\mathbf{1}(k_i = k)}{p_{ik}} \tag{8}
$$

Welfare at arm $$k$$ is estimated using a step-function approximation to the integral term (Eq. 9, p. 1083):

$$
\hat{\mathbb{U}}_{i+1,k} = \tilde{x}_k \cdot \hat{\mathbb{G}}_{i+1,k} + \frac{\lambda}{K} \cdot \sum_{k' > k} \hat{\mathbb{G}}_{i+1,k'} \tag{9}
$$

**Theorem 2 + Corollary 1 (Adversarial and stochastic upper bound, p. 1083-1084).** With optimal tuning $$\gamma = c_1 (\log(T)/T)^{1/3}$$, $$\eta = c_2 \gamma^2$$, $$K = \lfloor c_3/\gamma \rfloor$$, expected regret is bounded above by (Eq. 11, p. 1084):

$$
\mathcal{R}_T \leq c_4 \cdot \log(T)^{1/3} T^{2/3} \tag{11}
$$

The same bound applies to stochastic regret (Corollary 1). Combined with Theorem 1, the bound is rate-optimal up to $$\log(T)^{1/3}$$.

**Algorithm 2: Dyadic Search for Social Welfare.** For the stochastic concave case (Section 4, p. 1086), the algorithm maintains an active interval $$I_\tau \subseteq [0,1]$$ that contains the optimal policy with probability at least $$1 - \delta$$. In each epoch $$\tau$$, it selects three points $$l, c, r$$ (left, center, right) from a dyadic grid (points of the form $$k/2^m$$) inside $$I_\tau$$, and forms estimates of welfare differences. The welfare difference between policies $$x < x'$$ is:

$$
\Delta(x, x') = \mathbf{U}(x') - \mathbf{U}(x) = x' \cdot \mathbf{G}(x') - x \cdot \mathbf{G}(x) - \lambda \int_x^{x'} \mathbf{G}(x'')\,\mathrm{d}x'' \tag{12}
$$

The sample estimator is (Eq. 13, p. 1088):

$$
\hat{\Delta}_t(x, x') = x' \cdot \hat{G}_t(x') - x \cdot \hat{G}_t(x) - \lambda \cdot (x' - x) \cdot \hat{G}_t(x, x') \tag{13}
$$

where $$\hat{G}_t(x, x')$$ is the average of $$y_i$$ for observations $$x_i \in (x, x')$$. The confidence interval is (Eq. 15, p. 1088):

$$
J_t(x, x') = \hat{\Delta}_t(x, x') \pm \bigl(\Gamma_t(x') + \Gamma_t(x) + \Gamma_t(x, x')\bigr) \tag{15}
$$

with half-lengths $$\Gamma_t(x) = x \cdot \sqrt{\frac{1}{2 n_t(x)} \log(2/\delta)}$$ for the revenue component and $$\Gamma_t(x, x') = \lambda (x'-x)\bigl(\sqrt{\frac{\log(2/\delta)}{2(n_t(x,x')+1)}} + \frac{2}{n_t(x,x')+1}\bigr)$$ for the integral component. If the confidence interval $$J_t(l, c)$$ or $$J_t(l, r)$$ lies entirely above zero, the optimal policy cannot be to the left of $$l$$, so $$I_\tau$$ is trimmed accordingly. Concavity ensures the trimming is valid and yields $$T^{1/2}$$ regret (Theorem 4).

## Empirical specifications

The paper is entirely theoretical. The one numerical illustration (Figure 2, p. 1085) uses simulated data: $$v_i \overset{\text{i.i.d.}}{\sim} U[0,1]$$, $$\lambda = 0.7$$, $$K = 20$$, $$\eta = 0.025$$, $$\gamma = 0.1$$, $$T = 1000$$ periods, averaged across 4000 Monte Carlo replications. Tempered Exp3 reduces average cumulative regret to below half the uniform-random baseline by period 1000 (Figure 2, left panel). All theoretical claims rest on formal proofs in Appendix A and the Online Supplement (Cesa-Bianchi, Colomboni, and Kasy (2025)); no real-world data are used.

## Datasets used

No datasets are used. This is a theory paper with a single numerical simulation using artificially generated preference draws.

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Simulated uniform draws ($$v_i \sim U[0,1]$$) | Numerical illustration only (Figure 2, p. 1085) | Not applicable (simulated) |

## When to read the full paper

Read the [original](https://doi.org/10.3982/ECTA22351) if you are:
designing a policy-learning algorithm where welfare (utility) is not observed but demand or participation is;
proving regret bounds for bandit problems with integral-valued objectives;
extending to commodity taxation (Section 6, analysis left to future work), general preference heterogeneity, or Thompson sampling variants (Section 7);
or seeking the proofs of Theorems 1 and 2 in Appendix A, or the proofs of Theorems 3 and 4 in the Online Supplement.

The locators above point to the exact theorems, algorithms, and figures in the source PDF.

## Attribution and rights

Source: peer-reviewed, *Econometrica* 93(3). This distillation was extracted by an LLM on 2026-06-26 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Cesa-Bianchi, Nicolò, Roberto Colomboni, and Maximilian Kasy.
> "Adaptive Maximization of Social Welfare."
> *Econometrica* 93, no. 3 (May 2025): 1073-1104.
> DOI: 10.3982/ECTA22351. © 2025 The Authors.
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Insurance and Inequality With Persistent Private Information: Bloedel, Krishna & Leukhina (2025)
# https://instituteforautomatedresearch.org/wiki/papers/econometrica/2025/bloedel-et-al-insurance-inequality-persistent-private-2025/
# Distilled: Under any ergodic finite-state Markov type process, the optimal insurance contract always generates immiseration (Theorem 1), with backloaded high-powered incentives under positive serial correlation (Theorem 2). Econometrica 2025, paywalled. Five core results with source locators, the recursive contract model, the marginal cost martingale method, and numerical illustrations of speed of immiseration and short-run distortions.
# Tags: paper-summary, mechanism-design, optimal-contracting, dynamic-contracting
==============================================================================

**What this is.** The paper's core results, the recursive model, and the marginal cost martingale method with the defining equations: enough to know what was proved and how, without reading all 37 pages. To replicate or extend, read the original at [https://doi.org/10.3982/ECTA20404](https://doi.org/10.3982/ECTA20404).

## TL;DR

The paper studies long-run welfare and inequality in optimal insurance contracts when the agent's privately observed type follows an ergodic finite-state Markov chain, filling the gap between the i.i.d. benchmark of Thomas and Worrall (1990) and the permanent-shock benchmark of Williams (2011). A risk-neutral principal offers an infinite-horizon insurance contract to a risk-averse agent whose privately observed endowment evolves with arbitrary serial correlation bounded between these extremes.

**Theorem 1** (the central result) shows that immiseration is universal under ergodic persistence: the agent's promised utilities, flow utilities, and consumption all converge to their lower bounds in probability. **Theorem 2** strengthens this under positive serial correlation (FOSD): the spread in continuation utility across types and the conditional variance of promised utility both diverge to infinity, reflecting "backloaded high-powered incentives." The key insight is that ergodicity (mean-reversion) is the critical determinant: Williams (2011) shows bliss only at the knife-edge of zero mean-reversion (permanent shocks); any positive amount of mean-reversion restores immiseration.

The proofs construct a **marginal cost martingale**: a specific directional derivative of the principal's value function that is a strictly positive martingale under the optimal contract. The Martingale Convergence Theorem combined with a "renewal property" of the Markov process shows this martingale converges to zero, implying immiseration. Numerical simulations with CARA utility and two endowment types show that greater persistence accelerates immiseration in the medium run, generates over-insurance (negative insurance wedge) after consecutive low shocks, and introduces order-dependence absent in the i.i.d. CARA case.

## Core results

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Theorem 1 (Immiseration):** under any TVC-Regular environment with ergodic Markov types, the optimal contract generates immiseration | Theorem 1, p. 834 | $$v_i^{(t)} \to -\infty$$ in probability for all $$i \in S$$; $$u^{(t)} \to -\infty$$ in probability; $$c^{(t)} + \omega^{(t)} \to \underline{c}$$ in probability; no stationary distribution exists |
| R2 | **Theorem 2 (Backloaded Incentives):** under FOSD, the optimal contract exhibits growing cross-type utility spreads | Theorem 2, p. 836 | $$v_i^{(t)} - v_{i-1}^{(t)} \to +\infty$$ in probability for all $$i \geq 2$$; conditional variance $$\mathbf{V}(v_{s^{(t+1)}}^{(t)} \mid \mathbf{v}^{(t)}, s^{(t)}) \to +\infty$$ in probability |
| R3 | **Theorem 3 (Recursive Domain):** under MLRP or PPR type process, the implementable domain is characterized explicitly | Theorem 3, Appendix B, p. 853 | $$D = V_d = \{\mathbf{v} \in \mathcal{U}^d : v_d > v_{d-1} > \cdots > v_1\}$$ under MLRP or PPR; under CARA + MLRP/PPR, $$D = D^* = V_d$$; an open, convex cone independent of $$\alpha$$ and $$U$$ (within DARA class) |
| R4 | **Numerical (§5.1):** greater persistence yields faster immiseration in medium-run | Figure 3, p. 842 | 420,000 simulated paths of CARA/$$d=2$$ model: for $$q = 0.8$$ vs i.i.d. ($$q = 0.5$$), mean consumption $$\mu_{C,t}$$ declines faster and variance $$\sigma^2_{C,t}$$ grows faster in medium-run; patterns reverse in the first few periods |
| R5 | **Numerical (§5.2):** persistence induces over-insurance and large intertemporal wedges after consecutive low shocks | Figure 5, p. 845 | Insurance wedge turns negative (over-insurance) after strings of low shocks; intertemporal wedge grows to orders of magnitude larger than i.i.d. case; consumption depends on shock order (early bad luck penalized more than late bad luck) |

**Overall (paper's conclusion).** Immiseration is not an artifact of the i.i.d. assumption but is universal under ergodic persistence. The key determinant of long-run outcomes is mean-reversion in the type process. Any positive amount of mean-reversion is sufficient to generate immiseration, while the bliss result of Williams (2011) arises only at the knife-edge of zero mean-reversion. Persistence does affect the speed of immiseration and generates qualitatively new short-run distortions: over-insurance, large intertemporal wedges, and order-dependent consumption that are absent in the i.i.d. case.

## Theory / model

The environment is a discrete-time infinite-horizon insurance model (§2, p. 827). A risk-neutral principal with discount factor $$\alpha \in (0,1)$$ offers an insurance contract to a risk-averse agent (same discount factor). The agent's Bernoulli utility is $$U: (\underline{c}, \infty) \to \mathbb{R}$$ satisfying **Assumption DARA** (p. 827): strictly increasing, strictly concave, satisfying Inada conditions $$\lim_{c \to \underline{c}} U'(c) = +\infty$$ and $$\lim_{c \to \infty} U'(c) = 0$$, bounded above and unbounded below ($$\mathcal{U} = (-\infty, 0)$$), and with decreasing absolute risk aversion. Standard CARA and HARA utilities satisfy DARA.

The agent's type $$\omega^{(t)} \in S := \{\omega_1, \dots, \omega_d\}$$ ($$\omega_d > \cdots > \omega_1$$) evolves as a **fully connected, time-homogeneous, first-order Markov chain** (Assumption Markov, p. 828) with transition matrix $$\mathbf{F} = (f_{ij})$$, where $$f_{ij} = \mathbf{P}(\omega^{(t+1)} = \omega_j \mid \omega^{(t)} = \omega_i) > 0$$ for all $$i, j$$. This ensures the process is ergodic and bounded, allowing arbitrary serial correlation. **Assumption NHB** (p. 828) restricts the agent to under-reporting: he cannot report a type higher than his true type, capturing the idea that endowments are partially verifiable.

The principal's goal is to minimize lifetime cost of delivering a vector of promised utilities $$\mathbf{v}^{(0)} = (v_1, \dots, v_d) \in \mathcal{U}^d$$ subject to promise keeping and incentive compatibility.

**Recursive formulation (§3, p. 829).** Following Green (1987) and Fernandes and Phelan (2000), the state variable for the recursive problem is $$(\mathbf{v}, s) \in D \times S$$, where $$\mathbf{v} = (v_1, \dots, v_d) \in \mathcal{U}^d$$ is the vector of **interim promised utilities** (the agent's continuation utility $$v_i$$ conditional on his true current type being $$i$$, assuming truthful reporting in all future periods) and $$s \in S$$ is the agent's report in the previous period.

At state $$(\mathbf{v}, s)$$, the contract offers a menu $$(u_i, \mathbf{w}_i)_{i \in S}$$ of flow utility $$u_i$$ and continuation utility vector $$\mathbf{w}_i$$. The recursive constraints (p. 830) are:

$$
v_i = u_i + \alpha \mathbf{E}^{\mathbf{f}_i}[\mathbf{w}_i], \qquad \forall i \in S \tag{PK$_i$}
$$

$$
v_i \geq \psi(u_j, i, j) + \alpha \mathbf{E}^{\mathbf{f}_i}[\mathbf{w}_j], \qquad \forall i > j \in S \tag{IC$_{ij}$}
$$

where $$\mathbf{E}^{\mathbf{f}_i}[\mathbf{w}_j] := \sum_{k=1}^d f_{ik} w_{jk}$$ is the expected continuation utility that a type-$$i$$ agent obtains by reporting $$j$$, and $$\psi(u_j, i, j) := U(\omega_i + C(u_j, j))$$ is the flow utility a true type-$$i$$ agent receives when the contract delivers consumption $$C(u_j, j) := U^{-1}(u_j) - \omega_j$$.

Substituting (PK$$_j$$) into (IC$$_{ij}$$) yields the combined incentive constraint (p. 836; specific d=2 case at p. 830):

$$
v_i - v_j \geq \underbrace{\psi(u_j, i, j) - u_j}_{\text{i.i.d. info rent}} + \alpha\underbrace{\bigl[\mathbf{E}^{\mathbf{f}_i}[\mathbf{w}_j] - \mathbf{E}^{\mathbf{f}_j}[\mathbf{w}_j]\bigr]}_{\text{Markov info rent}} \tag{IC$_{ij}^*$}
$$

The Markov information rent term is new relative to the i.i.d. case: it arises because the agent's current type also determines his beliefs about future types, giving him intertemporal preferences over continuation contracts.

The principal's **recursive problem** (p. 831) is to minimize expected discounted cost:

$$
P(\mathbf{v}, s) := \inf_{\xi \in \Xi} \mathbf{E}\!\left[\sum_{t=0}^\infty \alpha^t C\!\left(u_\xi^{(t)}, s^{(t+1)}\right)\bigg|\,(\mathbf{v}^{(0)}, s^{(0)}) = (\mathbf{v}, s)\right] \tag{RP}
$$

where $$C(u, j) := U^{-1}(u) - \omega_j$$ is the consumption cost to the principal and $$\Xi$$ is the set of feasible recursive contracts. The **Bellman equation** characterizing $$P$$ is (Proposition 3.2, p. 834):

$$
P(\mathbf{v}, s) = \min_{(u_i, \mathbf{w}_i)_{i \in S} \in \Gamma(\mathbf{v})} \sum_{i \in S} f_{si}\bigl[C(u_i, i) + \alpha P(\mathbf{w}_i, i)\bigr] \tag{FE}
$$

where $$\Gamma(\mathbf{v})$$ is the constraint correspondence of all menus satisfying (PK$$_i$$)-(IC$$_{ij}$$) with $$\mathbf{w}_i \in D$$ for all $$i$$. Under (TVC)-Regularity, $$P(\cdot, s)$$ is convex, and there exists a unique optimal contract $$\xi^*$$ that is continuous on $$D \times S$$ (Proposition 3.2(b)).

The **conditional variance of continuation utility**, used in Theorem 2, is defined at (4.1, p. 835):

$$
\mathbf{V}\!\left(v_{s^{(t+1)}}^{(t)} \;\Big|\; \mathbf{v}^{(t)}, s^{(t)}\right) := \sum_{i=1}^d f_{s^{(t)},i} \left(v_i^{(t)} - \sum_{k=1}^d f_{s^{(t)},k}\, v_k^{(t)}\right)^2
$$

## Method

The core methodology is the **marginal cost martingale** (§4.3, pp. 837-840). It builds on the `mechanism-design` framework and `value-function-iteration` (Bellman equation) ideas, extending the martingale approach of Thomas and Worrall (1990) for i.i.d. types to the general Markovian setting via the Fernandes and Phelan (2000) recursive formulation. The paper's stated primary methodological contributions are (i) the recursive formulation using interim promised utilities (extending Fernandes and Phelan (2000)) and (ii) the marginal cost martingale for analyzing long-run outcomes.

Let $$DP(\mathbf{v}, s) = (P_1(\mathbf{v}, s), \dots, P_d(\mathbf{v}, s))$$ denote the gradient of $$P$$ with respect to $$\mathbf{v}$$. The **directional derivative in direction $$\mathbf{1} = (1, \dots, 1) \in \mathbb{R}^d$$** is:

$$
D_{\mathbf{1}} P(\mathbf{v}, s) := \sum_{i \in S} P_i(\mathbf{v}, s)
$$

This direction is unique in that increasing $$\mathbf{v}$$ along $$\mathbf{1}$$ raises every type's continuation utility by the same amount $$\varepsilon$$, leaving all downward incentive constraints (IC$$_{ij}^*$$) unchanged (because the left-hand side $$v_i - v_j$$ is unaffected). Consequently, $$D_{\mathbf{1}} P$$ captures the marginal cost of increasing the agent's ex ante promised utility without distorting his information rents.

**Proposition 4.4** (p. 837): Under (TVC)-Regularity, the process $$(D_{\mathbf{1}} P(\mathbf{v}^{(t)}, s^{(t)}))_{t=0}^\infty$$ induced by the optimal contract is a **strictly positive martingale**.

The martingale property follows from an envelope argument on (FE): at the optimum,

$$
D_{\mathbf{1}} P(\mathbf{v}, s) = \sum_{i=1}^d f_{si}\, D_{\mathbf{1}} P(\mathbf{w}_i, i)
$$

which is precisely the martingale condition $$\mathbf{E}[D_{\mathbf{1}} P(\mathbf{v}^{(t+1)}, s^{(t+1)}) \mid \mathbf{v}^{(t)}, s^{(t)}] = D_{\mathbf{1}} P(\mathbf{v}^{(t)}, s^{(t)})$$. Strict positivity holds because the cost function $$C(\cdot, j)$$ is convex and the cost-smoothing motive always pushes $$D_{\mathbf{1}} P > 0$.

**Proof sketch for Theorem 1 (§4.3):**

**Step 1 (Marginal cost martingale, p. 837).** By Proposition 4.4, $$D_{\mathbf{1}} P(\mathbf{v}^{(t)}, s^{(t)})$$ is a strictly positive martingale. By the Martingale Convergence Theorem, it converges a.s. to some non-negative limit $$Z \geq 0$$.

**Step 2 (Convergence to zero, p. 839).** The key step is showing $$Z = 0$$ in probability. Assumption Markov implies the agent's highest-type realization $$\omega_d$$ occurs infinitely often along any sample path. At such "renewal" histories the optimal contract is **efficient** (renegotiation-proof): the principal does not need to screen through continuation contracts, so the marginal cost martingale splits like in the i.i.d. case. At these histories, if the martingale were to converge to a strictly positive number, then $$\mathbf{v}^{(t)}$$ would converge to some interior point of $$D$$, implying the optimal contract perfectly stabilizes consumption, which contradicts incentive compatibility (Lemma C.18). Thus the martingale must converge to zero at renewal histories, and the Markov ergodicity extends this to all histories.

**Step 3 (Convergence of allocations, p. 840).** $$D_{\mathbf{1}} P(\mathbf{v}^{(t)}, s^{(t)}) \to 0$$ implies that the Lagrange multipliers on the incentive constraints converge to zero. This in turn implies that the agent's consumption converges to the level that the first-best contract would deliver if cost were zero, which is $$\underline{c}$$.

For **Theorem 2** (backloaded incentives), the argument uses Theorem 1(b) (flow utility $$u^{(t)} \to -\infty$$) combined with the incentive constraint (IC$$_{ij}^*$$): for FOSD type processes, the Markov information rent (the second bracketed term) is non-negative (Theorem 3 in Appendix B guarantees $$\mathbf{E}^{\mathbf{f}_i}[\mathbf{w}_i] \geq \mathbf{E}^{\mathbf{f}_j}[\mathbf{w}_i]$$ whenever $$i > j$$). Since the i.i.d. information rent grows without bound (from Theorem 1(b)) and the Markov rent is non-negative, the spread $$v_i^{(t)} - v_{i-1}^{(t)}$$ must also grow without bound.

## Empirical specifications

Section 5 presents numerical simulations for the **CARA / binary-type ($$d = 2$$) / symmetric-transitions** special case, using (p. 841):

$$
U(c + \omega) = -e^{-(c + \omega)}, \quad \omega_1 = -\log 5, \quad \omega_2 = \log 10, \quad \alpha = 0.5
$$

with symmetric transition matrix $$f_{11} = f_{22} = q$$. Three persistence levels plus a very high case are studied: $$q \in \{0.5 \text{ (i.i.d.)}, 0.65 \text{ (low)}, 0.8 \text{ (high)}, 0.95 \text{ (very high)}\}$$. Under CARA utility and FOSD ($$q \geq 0.5$$), Theorem 3 gives $$D = V_2 = \{(v_1, v_2) : v_2 > v_1\}$$, so optimal contracts are homogeneous of degree 1 in $$\mathbf{v}$$ (property HD1, p. 843).

**Speed of immiseration (§5.1, Figure 3, p. 842).** For each $$q$$, 420,000 sample paths of the optimal contract are simulated for 40 time periods, with 21 initial $$\mathbf{v}^{(0)}$$ points drawn from a grid on $$D = V_2$$. Mean consumption $$\mu_{C,t} := \mathbf{E}[c^{(t)} + \omega^{(t)}]$$ and variance $$\sigma^2_{C,t} := \mathbf{V}[c^{(t)} + \omega^{(t)}]$$ serve as proxies for open-economy aggregate consumption and cross-sectional inequality. Key patterns:

1. $$\mu_{C,t} \to -\infty$$ and $$\sigma^2_{C,t} \to +\infty$$ at all persistence levels (Theorem 1).
2. Medium-run: greater $$q$$ yields faster decline in $$\mu_{C,t}$$ and faster growth in $$\sigma^2_{C,t}$$.
3. Short-run (first few periods): greater $$q$$ initially slows the decline of $$\mu_{C,t}$$ and the growth of $$\sigma^2_{C,t}$$.

**Short-run distortions (§5.2, Figure 5, p. 845).** The HD1 property implies the contract's dynamics trace a countable set of rays in $$V_2$$: ray $$E_2$$ (continuation state after a high shock) and rays $$\{B_k\}_{k \geq 1}$$ (after $$k$$ consecutive low shocks, $$B_k$$ strictly below $$B_{k-1}$$). Along sequences of consecutive high and low shocks starting from $$E_2$$ (Figure 5):

- **Insurance wedge** $$\equiv U'(c^{(t)} + \omega_1)/U'(c^{(t)} + \omega_2) - 1$$: always positive in the i.i.d. case (under-insurance, consistent with Thomas and Worrall (1990)). Under persistence, it remains positive after high shocks but turns **negative** (over-insurance, $$u_1 > u_2$$) after consecutive low shocks.
- **Intertemporal wedge** $$\equiv \mathbf{E}[U'(c^{(t+1)} + \omega^{(t+1)}) \mid \omega^{(t)}]/U'(c^{(t)} + \omega^{(t)}) - 1$$: always positive in i.i.d. case (consumption drift downward). Under persistence, becomes orders of magnitude larger after consecutive low shocks.
- **Order-dependence**: unlike the i.i.d. CARA case (where Atkeson and Lucas (1992) show order-independence), consumption under persistence depends on the sequence of shocks, not just their frequency; early bad luck is penalized more than late bad luck.

## Datasets used

No empirical datasets are used. All results are analytical (Theorems 1-3) or based on numerical simulations of the theoretical model. The replication code is publicly available.

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Synthetic model simulation (CARA utility, d=2 types, symmetric Markov, q in {0.5, 0.65, 0.8, 0.95}) | Numerical illustrations of speed of immiseration and short-run distortions (Figures 3-5, §5) | No page (theoretical model; no external data source) |

Sample: 420,000 paths per persistence level, 40 time periods, 21 initial state points (§5.1). Replication code: [https://doi.org/10.5281/zenodo.14720557](https://doi.org/10.5281/zenodo.14720557).

## When to read the full paper

Read the original at [https://doi.org/10.3982/ECTA20404](https://doi.org/10.3982/ECTA20404) if you:
- need the formal proofs of Theorems 1, 2, or 3 (in Supplemental Appendices C, E, and I-J of the companion working paper Bloedel, Krishna, and Leukhina (2025b));
- are extending the recursive formulation to other environments (Appendix A covers equivalence between sequential and recursive contracts);
- want the closed-economy (Atkeson and Lucas (1992)) extension or the Zhang (2009) / Williams (2011) comparison (Section 6);
- are studying the general DARA + Markov setting beyond the CARA/d=2 numerical illustrations.

Core locators: Theorem 1 (immiseration, p. 834), Theorem 2 (backloaded incentives, p. 836), Theorem 3 / Appendix B (recursive domain, p. 853), Figure 3 (speed of immiseration, p. 842), Figure 5 (short-run wedges, p. 845).

## Attribution and rights

Source: peer-reviewed, *Econometrica* 93(3), May 2025. This distillation was extracted by an LLM on 2026-06-26 and is **not human-verified or independently reproduced**. The article is paywalled; no CC licence was found in Crossref metadata.

> Bloedel, Alexander W., R. Vijay Krishna, and Oksana Leukhina.
> "Insurance and Inequality With Persistent Private Information."
> *Econometrica* 93, no. 3 (May 2025): 821-857.
> DOI: 10.3982/ECTA20404.
> © 2025 The Econometric Society. All rights reserved.
> This page extracts core results only and does not reproduce the full text.


==============================================================================
# Double Robust Bayesian ATE Inference: Breunig, Liu & Yu (2025)
# https://instituteforautomatedresearch.org/wiki/papers/econometrica/2025/breunig-et-al-double-robust-bayesian-inference-2025/
# Proposes a doubly robust Bayesian procedure for ATE estimation under unconfoundedness that adjusts the conditional mean prior and corrects the posterior via the semiparametric efficient influence function, proving a new Bernstein-von Mises theorem with exact frequentist coverage under double robust smoothness. Simulations on Lalonde-Dehejia-Wahba data show near-nominal coverage (0.95-0.98) with shorter credible intervals than prior-adjusted Bayesian and doubly robust frequentist alternatives. Econometrica 2025, CC BY 4.0; LLM-distilled, not human-verified, not reproduced.
# Tags: paper-summary, causal-inference, treatment-effects, bayesian-inference, semiparametric
==============================================================================

**What this is.** A distilled skeleton of "Double Robust Bayesian Inference on Average Treatment Effects" by Breunig, Liu, and Yu (2025). Read the original at <https://doi.org/10.3982/ECTA21442> to replicate or extend.

## TL;DR

The paper proposes a double robust (DR) Bayesian procedure for estimating the average treatment effect (ATE) under unconfoundedness. The procedure makes two adjustments to standard Bayesian inference: (i) it adjusts the prior distribution of the conditional mean function using a pilot propensity score estimator, in the least favorable direction identified by Hahn (1998); and (ii) it corrects each posterior draw by subtracting a bias term constructed from the semiparametric efficient influence function. The authors prove a new Bernstein-von Mises (BvM) theorem showing the corrected posterior is asymptotically normal with the semiparametric efficient variance under a double robust smoothness condition (Proposition 4.1, p. 550): insufficient smoothness of the conditional mean can be compensated by higher smoothness of the propensity score and vice versa. Monte Carlo simulations using WGAN-generated samples from the Lalonde-Dehejia-Wahba data show DR Bayes achieves near-nominal coverage (95% to 98%) with credible interval lengths shorter than both prior-adjusted Bayesian (PA Bayes, following Ray and van der Vaart (2020)) and doubly robust frequentist alternatives including DML (Chernozhukov et al. (2017)). An empirical illustration on the National Supported Work Demonstration finds employment effects of 12% to 18%, close to the experimental benchmark of 11%.

## Core results

| # | Result | Locator | Magnitude as reported |
|---|--------|---------|----------------------|
| R1 | BvM theorem: corrected posterior is asymptotically normal with efficient variance V_0 under Assumptions 1-4 | Theorem 3.2, p. 548 | d_BL(posterior of sqrt(n)(tau - tau-hat - b-hat), N(0, V_0)) → 0 in probability; V_0 = semiparametric efficiency bound of Hahn (1998) |
| R2 | Exact frequentist coverage: the Bayesian credible set C_n(alpha) is also a valid (1-alpha) confidence interval | Corollary 3.1, p. 548 | P_0(tau_0 in C_n(alpha)) → 1-alpha; sqrt(n)(tau-bar minus tau_0) => N(0, V_0) |
| R3 | Simulation CP at t=0.10 (avg n=240): near-nominal coverage, shorter than PA Bayes | Table I, p. 553 | DR Bayes CP=0.983, CIL=0.223; PA Bayes CP=0.981, CIL=0.260; DML CP=0.927, CIL=0.524 |
| R4 | Simulation CP at t=0.05 (avg n=363): coverage maintained; PA Bayes begins to slip | Table I, p. 553 | DR Bayes CP=0.970, CIL=0.221; PA Bayes CP=0.949, CIL=0.254; DML CP=0.870, CIL=0.393 |
| R5 | Simulation CP at t=0.01 (avg n=664): DR Bayes stable; PA Bayes degrades sharply | Table I, p. 553 | DR Bayes CP=0.952, CIL=0.258; PA Bayes CP=0.897, CIL=0.308; DML CP=0.918, CIL=0.522 |
| R6 | Empirical ATE: NSW job training increases employment by 12% to 18%, near the experimental benchmark | Table II, p. 554 | DR Bayes ATE=0.178 [0.061, 0.293] at t=0.10; ATE=0.184 [0.064, 0.294] at t=0.05; NSW expt=0.111 [0.026, 0.196] |

**Overall.** The DR Bayesian procedure achieves near-nominal coverage across all trimming thresholds, including near-boundary overlap (t=0.01), where PA Bayes degrades to 89.7% and DML coverage ranges from 87% to 93%. Credible interval lengths are consistently shorter than PA Bayes because the posterior correction removes the bias that PA Bayes carries. In the empirical illustration using the Lalonde-Dehejia-Wahba data, DR Bayes produces employment effect estimates of 12% to 18%, bracketing the experimental benchmark of 11%, while DML (Chernozhukov et al. (2017)) produces a negative estimate at t=0.10 when all other methods are positive (Table II, p. 554).

## Theory / model

The setup (Section 2.1, p. 541) considers binary potential outcomes $$Y_i(0), Y_i(1) \in \{0,1\}$$ and binary treatment $$D_i \in \{0,1\}$$. The observed outcome is $$Y_i = D_i Y_i(1) + (1-D_i)Y_i(0)$$. Covariates $$X_i \in \mathbb{R}^p$$. The joint density of $$Z_i = (Y_i, D_i, X_i^\top)^\top$$ factors as (eq. 2.1, p. 541):

$$p_{\pi,m,f}(z) = \pi(x)^d(1-\pi(x))^{1-d}\, m(d,x)^y(1-m(d,x))^{1-y}\, f(x), \tag{2.1}$$

where $$\pi(x) = P_0(D_i=1 \mid X_i=x)$$ is the propensity score and $$m(d,x) = P_0(Y_i=1 \mid D_i=d, X_i=x)$$ is the conditional mean. The parameter of interest is the ATE $$\tau_0 = \mathbb{E}_0[Y_i(1) - Y_i(0)]$$.

**Assumption 1** (Unconfoundedness and Overlap, p. 541): (i) $$(Y(0), Y(1)) \perp D_i \mid X_i$$, and (ii) there exists $$\bar{\pi} > 0$$ such that $$\bar{\pi} < \pi_0(x) < 1 - \bar{\pi}$$ for all $$x$$ in the support of $$F_0$$.

For Bayesian inference the paper uses the logistic link $$\Psi(t) = 1/(1+e^{-t})$$ and defines the reparametrization (eq. 2.2, p. 542):

$$\eta^\pi = \Psi^{-1}(\pi), \quad \eta^m = \Psi^{-1}(m), \quad \eta^f = \log f. \tag{2.2}$$

The ATE in terms of $$\eta$$ is (eq. 2.3, p. 542):

$$\tau_\eta = \mathbb{E}_\eta[m_\eta(1,X) - m_\eta(0,X)]. \tag{2.3}$$

The efficient influence function for ATE estimation (Hahn (1998); Hirano, Imbens, and Ridder (2003)) is (eq. 2.4, p. 542):

$$\tilde{\varphi}_\eta(z) = m_\eta(1,x) - m_\eta(0,x) + \gamma_\eta(d,x)\bigl(y - m_\eta(d,x)\bigr) - \tau_\eta, \tag{2.4}$$

where the Riesz representer $$\gamma_\eta$$ is (eq. 2.5, p. 542):

$$\gamma_\eta(d,x) = \frac{d}{\pi_\eta(x)} - \frac{1-d}{1-\pi_\eta(x)}. \tag{2.5}$$

Lemma 3.1 (p. 545) shows the least favorable direction for estimating $$\tau_\eta$$ is:

$$\xi_\eta(d,x) = \bigl(0,\; \gamma_\eta(d,x),\; m_\eta(1,x) - m_\eta(0,x) - \tau_\eta\bigr). \tag{3.4}$$

This derivation extends Lemma 2.1 of Ray and van der Vaart (2020) from the one-arm (missing data) to the two-arm (treatment-control) ATE context. The prior adjustment is in the direction of $$\gamma_\eta$$ (the propensity score component of the least favorable direction) and the posterior correction aligns with the efficient influence function $$\tilde{\varphi}_\eta$$.

## Method

The Double Robust Bayesian Procedure is described in Algorithm 1 (p. 543) with implementation details in Section 4.2 (p. 550).

**Prior specification.** Place a centered Gaussian process prior $$W^m$$ with squared exponential covariance $$K((d,x),(d',x')) = \nu^2\exp(-a_{0n}^2(d-d')^2/2 - \sum_{l=1}^p a_{ln}^2(x_l-x_l')^2/2)$$ on $$\eta^m(d,X_i)$$. Adjust the prior by the pilot Riesz representer:

$$\eta^m(d,X_i) = W^m(d,X_i) + \lambda\hat{\gamma}(d,X_i),$$

where $$\hat{\gamma}(d,x) = d/\hat{\pi}(x) - (1-d)/(1-\hat{\pi}(x))$$ (eq. 2.6, p. 542) uses logistic Lasso propensity score estimates $$\hat{\pi}$$ (Friedman, Hastie, and Tibshirani (2010)), and $$\lambda \sim N(0,\sigma_n^2)$$ with $$\sigma_n = (\log n)/(\sqrt{n}\,\Gamma_n)$$.

**Posterior computation and correction.** For each draw $$s = 1,\ldots,S$$: draw Bayesian bootstrap weights $$M^s_i = e^s_i / \sum_j e^s_j$$ with $$e^s_i \stackrel{iid}{\sim} \text{Exp}(1)$$; generate the posterior draw $$\tau^s_\eta$$ and the correction term $$\hat{b}^s_\eta$$; form the corrected draw (eqs. 2.7-2.8, p. 543):

$$\tilde{\tau}^s_\eta = \tau^s_\eta - \hat{b}^s_\eta, \tag{2.7}$$

$$\tau^s_\eta = \sum_{i=1}^n M^s_i\bigl(m^s(1,X_i) - m^s(0,X_i)\bigr), \quad \hat{b}^s_\eta = \frac{1}{n}\sum_{i=1}^n \boldsymbol{\tau}[m^s - \hat{m}](Z_i), \tag{2.8}$$

where $$\boldsymbol{\tau}[m](z) := m(1,x) - m(0,x) + \hat{\gamma}(d,x)(y - m(d,x))$$ is the influence-function correction and $$\hat{m}$$ is the uncorrected Gaussian process posterior mean. The point estimator is $$\bar{\tau}_\eta = S^{-1}\sum_s \tilde{\tau}^s_\eta$$ and the $$100(1-\alpha)\%$$ credible set is $$\mathcal{C}_n(\alpha) = \{\tau : q_n(\alpha/2) \le \tau \le q_n(1-\alpha/2)\}$$.

**Main theoretical results.** Theorem 3.2 (p. 548) establishes, under Assumptions 1-4:

$$d_{\text{BL}}\!\left(\mathcal{L}_\Pi\!\left(\sqrt{n}(\tau_\eta - \hat{\tau} - \hat{b}_\eta) \,\middle|\, Z^{(n)}\right),\, N(0,V_0)\right) \to_{P_0} 0$$

where $$V_0 = \mathbb{E}_0[\tilde{\varphi}_0(Z_i)^2]$$ is the semiparametric efficiency bound of Hahn (1998). Corollary 3.1 (p. 548) gives exact frequentist coverage:

$$\sqrt{n}(\bar{\tau}_\eta - \tau_0) \Rightarrow N(0, V_0) \quad \text{and} \quad P_0(\tau_0 \in \mathcal{C}_n(\alpha)) \to 1 - \alpha. \tag{3.7}$$

**Double robust smoothness** (Proposition 4.1 and Remark 4.1, pp. 549-550): Theorem 3.2 holds under Holder smoothness classes $$m_0(d,\cdot) \in \mathcal{C}^{s_m}([0,1]^p)$$ and propensity scores estimated at rate $$r_n$$ provided:

$$\sqrt{s_\pi s_m} > p/2,$$

where $$s_\pi$$ and $$s_m$$ are the Holder smoothness indices of $$\pi_0$$ and $$m_0$$. Low smoothness of $$m_0$$ can be offset by higher smoothness of $$\pi_0$$ (and vice versa), which is the double robustness of the smoothness condition. When $$s_m > p/2$$ alone (single robustness), the bias term $$b_{0,\eta}$$ vanishes and no posterior correction is needed, recovering the Ray and van der Vaart (2020) case.

The frequentist doubly robust estimator that the Bayesian procedure is asymptotically equivalent to (eq. 3.8, p. 548) is:

$$\hat{\tau} = n^{-1}\!\sum_{i=1}^n \bigl(\hat{m}(1,X_i) - \hat{m}(0,X_i)\bigr) + n^{-1}\!\sum_{i=1}^n \hat{\gamma}(D_i,X_i)\bigl(Y_i - \hat{m}(D_i,X_i)\bigr). \tag{3.8}$$

This is the standard doubly robust or double machine learning (DML) estimator of Chernozhukov et al. (2017). The BvM theorem establishes that the corrected Bayesian posterior yields a Bayesian interpretation of this frequentist estimator.

## Empirical specifications

The paper applies DR Bayes to the Lalonde-Dehejia-Wahba data from the National Supported Work (NSW) Demonstration study (LaLonde (1986), Dehejia and Wahba (1999)). The data (publicly available at Dehejia's NBER website, footnote 4, p. 551) combine:

- **Treated group**: 185 men in the NSW experimental program
- **Control group**: 2490 men from the Panel Study of Income Dynamics (PSID)
- **Outcome $$Y$$**: binary employment indicator for 1978
- **Treatment $$D$$**: participation in the NSW program
- **Covariates (9)** following Abadie and Imbens (2011): age, education, Black, Hispanic, married, earnings 1974, earnings 1975, unemployed 1974, unemployed 1975

**Propensity score**: estimated by logistic Lasso with cross-validated penalty (Friedman, Hastie, and Tibshirani (2010)). Observations with estimated propensity score outside $$[t, 1-t]$$ are trimmed; $$t \in \{0.10, 0.05, 0.01\}$$. The optimal threshold of Crump, Hotz, Imbens, and Mitnik (2009) gives an average threshold of 0.073 for these samples (footnote 7, p. 553).

**Conditional mean $$m$$**: estimated using the uncorrected Gaussian process posterior mean with the squared exponential kernel, with the rescaling parameter $$a_n \sim n^{1/(2s_m+p)}(\log n)^{-(1+p)/(2s_m+p)}$$ (eq. 4.1, p. 549). Posterior computed via Laplace approximation (Section 4.2, p. 550). $$S = 5000$$ posterior draws.

**Simulation study** (Section 5.1, pp. 551-553): samples of $$n = 185$$ treated + 2490 controls drawn from a population generated by WGAN (Athey, Imbens, Metzger, and Munro (2024)) applied to the Lalonde-Dehejia-Wahba data, focusing on the binary employment outcome for 1978. 1000 Monte Carlo replications per trimming threshold. Competitors: unadjusted Bayes, PA Bayes (Ray and van der Vaart (2020)), Match and Match BC (Abadie and Imbens (2011)), DR TMLE (Benkeser et al. (2017)), DML (Chernozhukov et al. (2017)).

**Experimental benchmark** (body text, p. 553): The NSW experimental data (both NSW treatment and NSW control, $$n=445$$) yields ATE = 0.111, 95% CI [0.026, 0.196], which serves as the ground truth for evaluating observational methods.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---------|--------------|-----------|
| Lalonde-Dehejia-Wahba (NSW + PSID) | Treatment (185 NSW men) and non-experimental controls (2490 PSID men); binary employment outcome 1978; 9 covariates including earnings 1974-1975 | no page yet |

Sample: 185 treated + 2490 PSID controls; cross-sectional, covariates from 1974-1975, outcome from 1978. Publicly available at <http://users.nber.org/~rdehejia/nswdata2.html> (footnote 4, p. 551).

## When to read the full paper

Read this paper when you need a Bayesian credible set for an ATE under unconfoundedness that is also frequentist-valid: Theorem 3.2 and Corollary 3.1 (pp. 547-548) give the formal BvM and coverage results. Read Algorithm 1 (p. 543) and Section 4.2 (pp. 550-551) to implement the procedure with Gaussian process priors and logistic Lasso propensity scores. Read Remark 4.1 (p. 550) for the double robust smoothness condition and how to choose the Gaussian process rescaling parameter. Read Tables I-II (pp. 553-554) for simulation and empirical performance against PA Bayes, DML, and matching. Read Section 6 (pp. 554-556) for extensions to continuous, multinomial, and other causal parameters. Replication code is available at the Zenodo archive (https://doi.org/10.5281/zenodo.14015435).

## Attribution and rights

Christoph Breunig, Ruixuan Liu, and Zhengfei Yu, "Double Robust Bayesian Inference on Average Treatment Effects," *Econometrica*, Vol. 93, No. 2 (March 2025), pp. 539-568. DOI: [10.3982/ECTA21442](https://doi.org/10.3982/ECTA21442).

This article is published under the Creative Commons Attribution 4.0 International License (CC BY 4.0). You are free to share and adapt the material provided you give appropriate credit, provide a link to the license, and indicate if changes were made. License URL: <https://creativecommons.org/licenses/by/4.0/>.

This page is an LLM-distilled summary (claude-sonnet-4-6, 2026-06-26). It has not been human-verified and the results have not been independently reproduced. All magnitudes are extracted from the published article; consult the original for replication.


==============================================================================
# Making Subsidies Work: Cingano, Palomba, Pinotti & Rettore (2025)
# https://instituteforautomatedresearch.org/wiki/papers/econometrica/2025/cingano-et-al-making-subsidies-work-rules-2025/
# Distilled: Using a regression discontinuity design around Italy's L488/92 investment subsidy program (1996-2007), this paper finds that subsidies raised firm investment by 43% and employment by 17% over six years, at a cost per new job 3.5 times higher in Southern than Northern Italy. Eliminating political discretion from allocation would reduce cost per job by 11%, while relying solely on discretion would raise it by 42%. Econometrica 2025, CC BY 4.0. Eight core results with source locators, datasets used, the identification strategy, and the empirical specifications.
# Tags: paper-summary, public-economics, place-based-policy, industrial-policy
==============================================================================

**What this is.** The paper's core results, identification strategy, and estimating equations: enough to know what it found and how, without reading all 32 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.3982/ECTA21319).

## TL;DR

This paper evaluates Italy's Law 488/92 (L488/92), the country's largest public investment subsidy program, which financed 77,000 investment projects at a total cost of nearly EUR 26 billion between 1996 and 2007. Projects were ranked within each call-region-category cell by a composite score combining objective quality indicators ("rules," sub-score SR) and regional politicians' priorities ("discretion," sub-score SD), creating a sharp eligibility cutoff exploited here as a regression discontinuity design. Firms scoring just above the cutoff increased investment by 43% and employment by 11% over three years; employment gains persist and grow to 17% by year six, with no evidence of spillovers to non-subsidized competitors. Extending the analysis to the full distribution of inframarginal firms via Angrist and Rokkanen (2015), the paper documents that firms preferred by political discretion generate similar percent employment gains as firms ranked high on objective criteria, but at 3.5 times higher cost per job in Southern regions. Counterfactual simulations show that removing political discretion would reduce the cost per new job by 11%, while relying exclusively on discretion would raise it by 42%. Cerqua and Pellegrini (2014) evaluated L488/92 in six Southern regions and found positive employment effects; this paper extends that to all 26 calls and quantifies the cost of the rules-vs.-discretion trade-off. Bartik (2020) places these cost estimates in the context of US place-based policy evidence.

## Core results

Magnitudes and significance are as reported; heteroscedasticity-robust standard errors clustered by cell (call-region-category) are in brackets. All monetary amounts at constant 2010 prices.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Subsidy raises **cumulative investment by 43%** over the three-year subsidy period | Table III, Panel A, col 2, p. 764 | +0.360 log points [SE 0.055]; linear RDD with cell FE; Adj. R² = 0.229; n = 17,425 |
| R2 | Subsidy raises **employment by 11%** over three years | Table III, Panel B, col 2, p. 764 | +0.104 log-change [SE 0.020]; stable across all 8 specifications (linear/quadratic, uniform/triangular, with/without cell FE); n = 31,681 |
| R3 | **Employment effect persists** and grows to 17% over six years | Table III, Panel C, col 2, p. 764 | +0.153 log-change [SE 0.024]; effect continues after subsidy disbursement ends; n = 28,759 |
| R4 | **Firm survival probability rises by 3 pp** (+6% above baseline) over six years | Figure 5, last panel, p. 765 | +3 pp on a baseline survival rate of 87%; from the dynamic event-study specification with linear RDD and cell FE |
| R5 | **Cost per new job is EUR 178,000** (all regions), with a 3.5x North-South gap | Table IV, col 1, p. 768 | EUR 178,000 [133; 299] overall; EUR 241,000 [195; 332] South; EUR 68,000 [41; 211] North-Center; per worker-year EUR 54,000 overall |
| R6 | **No-discretion counterfactual reduces cost per job by 11%** | Table VI, Panel A, col 2, p. 773 | -11.1 pp [CI -14.8; -8.0] overall; -12.1 pp in South; -8.7 pp in North-Center |
| R7 | **Only-discretion counterfactual raises cost per job by 42%** | Table VI, Panel A, col 3, p. 773 | +41.7% [17.7; 64.3] overall; +37.8% in South; cost per EUR 1 of investment rises by 22% |
| R8 | **Treatment effects range from 10% to 19%** across SR-SD quintile cells, with cost per job varying by a factor of five | Figure 8, Panels A-B, p. 771 | 6-year employment growth log-change 0.10 (low SR, low SD) to 0.19 (high SR and SD); cost per job highest (~5x lower-bound) for high-SD, low-SR cells |

**Overall (paper's conclusion).** Both firms selected by objective criteria and those preferred by local politicians generate employment and investment growth, but politically favored firms do so at higher cost per job because they are smaller and demand larger subsidies per worker. The same percent employment increase corresponds to fewer absolute new jobs in small firms. Eliminating political discretion from allocation improves cost-effectiveness, particularly in Southern regions that received the largest share of L488/92 funds. An optimal allocation based on estimated treatment effects would reduce the cost per new job by more than half (Table VI, col 4: -54% [-60.2; -52.4]).

## Theory / model

The paper has no formal equilibrium model. The empirical strategy tests two hypotheses about the allocation and impact of public investment subsidies.

**Hypothesis 1 (treatment effect).** Subsidized firms near the eligibility cutoff invest and hire more than similar non-subsidized firms. This tests whether L488/92 generated genuine real effects or merely crowded out private investment (as Bronzini and de Blasio (2006) found using a DiD approach on earlier data).

**Hypothesis 2 (rules vs. discretion).** The cost-effectiveness of subsidies depends on which firms receive them: those scoring high on objective criteria (SR) versus those preferred by politicians (SD). The rules-versus-discretion dilemma (Persson and Tabellini (2002); Laffont (1996)) has empirical content if political priorities are systematically misaligned with cost-efficiency objectives.

The institutional setting provides the identification lever. L488/92 ranked applicant firms within each call-region-category cell by a composite score S. During 1996-1997, the score combined three objective indicators (I1: investment-to-subsidy ratio "skin in the game"; I2: planned job creation; I3: no-waste ratio). Starting in 1998, regional governments gained discretionary authority to assign points to municipalities and project types, creating the sub-score SD (I4). The aggregate of standardized I1-I3 is denoted SR (objective sub-score) and the standardized I4 is SD (discretionary sub-score). The composite score is their standardized sum (eq. 1, p. 753):

$$
S_{ir} = \sum_{j=1}^{3} \frac{I^j_{ir} - \mu^j_r}{\sigma^j_r} \tag{1}
$$

where $$I^j_{ir}$$ is the value of indicator $$j$$ for project $$i$$ in call-region $$r$$, $$\mu^j_r$$ is the within-cell mean, and $$\sigma^j_r$$ is the within-cell standard deviation. Projects were funded in descending order of S until the call-region budget was exhausted, yielding a rationing cutoff that varies by cell.

The key identifying assumption is that applicants just above and below the cutoff are otherwise identical. Balancing tests on pre-application characteristics (Figure A7 in Cingano et al. (2025a)) show no discontinuity at the cutoff. The density test of McCrary (2008) also does not reject continuity (p-value 0.2; Figure A6 in Cingano et al. (2025a)), ruling out strategic sorting. One-sided non-compliance (about 20% of above-cutoff firms are not funded for exogenous reasons) means $$\tau$$ identifies an ITT effect; the LATE is approximately $$\tau / 0.8$$.

## Method

The analysis has two parts: a parametric RDD for firms near the cutoff, and the Angrist and Rokkanen (2015) conditional independence approach to characterize treatment effects across the full distribution of inframarginal firms.

**Characterizing the sub-scores with LASSO.** To understand which firm characteristics drive objective versus political allocation, the paper regresses SR and SD on a rich covariate vector $$Z_i$$ using the LASSO estimator (eq. 2, p. 757):

$$
\hat{\theta}^{\text{LASSO}} := \arg\min_{\theta \in \mathbb{R}^k} \left\{ \sum_{i=1}^{n} \left(Y_i - Z_i'\theta\right)^2 + \lambda \sum_{j=1}^{k} |\theta_j| \right\} \tag{2}
$$

where $$Y_i$$ is SR or SD, and $$\lambda \geq 0$$ is selected by the one-standard-deviation rule (James, Witten, Hastie, and Tibshirani (2013)). Key findings (Figure 1, p. 758): firm size is the strongest predictor of SR (positive) and SD (negative); the same is true for the subsidy amount requested (negatively for SR, positively for SD). Political discretion therefore systematically favors smaller firms demanding larger per-worker subsidies, which is the root cause of its lower cost-effectiveness.

**Angrist-Rokkanen extrapolation.** Following Angrist and Rokkanen (2015), the conditional independence assumption (CIA) states that potential outcomes are mean-independent of the running variable S conditional on a vector of pre-treatment firm characteristics $$X$$ (eq. 4, p. 760):

$$
\mathbb{E}[Y(d) \mid S, X] = \mathbb{E}[Y(d) \mid X], \quad d \in \{0, 1\} \tag{4}
$$

Combined with common support (eq. 5, p. 761):

$$
0 < \mathbb{P}(D = 1 \mid X) < 1 \quad \text{a.s.} \tag{5}
$$

the CIA permits identification of the ATE at any score value $$s'$$ (eq. 6, p. 761):

$$
\mathbb{E}[Y(1) - Y(0) \mid S = s'] = \mathbb{E}\!\left[\mathbb{E}[Y \mid X, D=1] - \mathbb{E}[Y \mid X, D=0] \,\Big|\, S = s'\right] \tag{6}
$$

The CIA is partially testable: if X absorbs all confounding, then conditional on X, outcomes should be mean-independent of S within treated and control groups. Table V (p. 770) confirms this for the chosen covariate vector $$X^\star$$: once $$X^\star$$ is included, coefficients on SR and SD in the conditional mean regression become insignificant (columns with $$X^\star$$). Panel B of Figure 6 (p. 767) confirms substantial common support in the estimated propensity score distribution.

## Empirical specifications

**Baseline RDD estimating equation.** Firm outcomes are regressed on the treatment dummy D (scoring above the cutoff = 1), a polynomial in the centered score S, its interaction with D, and cell fixed effects $$FE_c$$ (eq. 3, p. 760):

$$
Y = \tau D + \sum_{\ell=1}^{p} \gamma_\ell S^\ell + \sum_{\ell=1}^{p} \delta_\ell D \cdot S^\ell + FE_c + \varepsilon \tag{3}
$$

The coefficient $$\tau$$ identifies the ITT effect for firms near the cutoff (bandwidth $$S \in [-5, 5]$$, covering 82% of the sample). Specifications use $$p = 1$$ (linear) and $$p = 2$$ (quadratic) polynomials, uniform and triangular kernels, with and without cell fixed effects. Outcomes cover: log-cumulated investment over 3 years (Panel A of Table III); log-change in employment over 3 years (Panel B) and 6 years (Panel C); log-revenues and log-value-added (Figure 5); and survival probability (Figure 5, last panel). Standard errors are clustered by cell; results are stable across all eight specifications.

**Linear reweighting estimator for inframarginal effects.** The conditional mean is parametrized following Kline (2011) (eq. 7, p. 761):

$$
\mathbb{E}[Y \mid S, X, D = d] = \sum_{\ell=0}^{q} \alpha_{d,\ell} S^\ell + X'\beta_d, \quad d \in \{0, 1\} \tag{7}
$$

Restricting $$\alpha_{0,\ell} = \alpha_{1,\ell} = 0$$ for all $$\ell$$ is the testable implication of the CIA (Table V). Under the CIA, the ATE at any score value reduces to eq. 8 (p. 761):

$$
\mathbb{E}[Y(1) - Y(0) \mid S = s'] = (\beta_1 - \beta_0)' \mathbb{E}[X \mid S = s'] \tag{8}
$$

The covariate vector $$X^\star$$ includes: firm age (inversely related to growth per Evans (1987)); lagged employment growth of similar firms in the same local labor market (LLM) and 3-digit sector; average wage of white-collar workers; dummies for managers or apprentices in payroll; and investment project size relative to initial employment interacted with cell fixed effects.

**Conditional treatment effects by sub-score quintile.** Extending the CIA to both sub-scores (eq. 9, p. 761):

$$
\mathbb{E}[Y(d) \mid SR, SD, X] = \mathbb{E}[Y(d) \mid X], \quad d \in \{0, 1\} \tag{9}
$$

yields conditional ATEs for any point in the SR-SD distribution (eq. 10, p. 761):

$$
\mathbb{E}[Y(1) - Y(0) \mid SR = r, SD = d] = (\beta_1 - \beta_0)' \mathbb{E}[X \mid SR = r, SD = d] \tag{10}
$$

These are estimated for the 25 cells defined by the 5-by-5 quintiles of SR and SD (Figure 8, p. 771). Cost per new job in each cell is computed by scaling the subsidy by the treatment effect times average firm size in the cell.

**Counterfactual allocation rules.** Three counterfactual policies are simulated (Table VI, p. 773): (1) no-discretion (SD = 0 for all applicants, re-ranked by SR only); (2) only-discretion (rank by SD only); (3) cost-minimizing (rank by estimated treatment effects). The policy invariance assumption (eq. 11, p. 762) holds that applicant characteristics and project quality do not respond to the selection rule. This is validated by comparing applicant characteristics and objective sub-scores in regions that adopted versus did not adopt discretion, before and after the 1998 reform (Table A4 in Cingano et al. (2025a): means are not significantly different).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| L488/92 administrative data (Ministero dello Sviluppo Economico) | Project applications, numerical scores (I1-I5 and SR/SD sub-scores), subsidy amounts, and funding outcomes for 75,584 projects across 26 calls, 1996-2007 | No page yet |
| INPS Italian social security archives | Monthly employment records for all Italian firms with at least one employee (~1.6 million firms annually); firm start and closure dates; basis for employment outcome variables | No page yet |
| Cerved Group balance sheet database | Investment, revenues, value-added, and total assets for approximately 17,226 L488/92 applicant firms (all Italian limited liability companies, 1993-2015) | [Cerved](/wiki/commercial/cerved/) (licensed) |
| ISTAT census and administrative statistics | Municipality-level socioeconomic characteristics (labor force participation, NEET rates, employment composition, population); price deflators (IPCA) | No page yet |
| Gazzetta Ufficiale della Repubblica Italiana | Official gazette entries used to identify competition "cells" (call-region-category) for each applicant | No page yet |

Sample: 40,366 projects from 27,084 firms for the employment outcome regressions (n = 31,681 for 3-year and n = 28,759 for 6-year employment). Balance sheet outcomes are available for 17,226 companies. Period: 26 L488/92 calls, 1996-2007 (5 calls missing for data reasons). About 45% of applicants scored at or above the cutoff; roughly 80% of those received the subsidy (complier share approximately 0.8, Table II, p. 757).

## When to read the full paper

Read the [original](https://doi.org/10.3982/ECTA21319) if you are: evaluating the cost-effectiveness of firm subsidy programs and comparing rules-based versus discretionary allocation criteria (Tables IV and VI); applying the Angrist and Rokkanen (2015) CIA extrapolation to characterize treatment effects across the full running-variable distribution, not just around the cutoff (Section 4, eqs. 4-10); studying treatment effect heterogeneity by firm type under a place-based policy (Section 6.3, Figure 8); or interested in Italian regional economic disparities and the North-South productivity divide (Sections 2 and 6.2). The companion supplement (Cingano et al. (2025b)) covers extensive robustness checks, including non-parametric bandwidth variation, alternative propensity-score restrictions, and a fully flexible relationship between SD and SR.

## Attribution and rights

Source: peer-reviewed, *Econometrica*, Vol. 93, No. 3 (May, 2025). This distillation was extracted by an LLM on 2026-06-26 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Cingano, Federico, Filippo Palomba, Paolo Pinotti, and Enrico Rettore.
> "Making Subsidies Work: Rules versus Discretion."
> *Econometrica* 93, no. 3 (May 2025): 747-778.
> DOI: 10.3982/ECTA21319. © 2025 The Authors.
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Comparative Statics With Adjustment Costs: Dekel, Quah & Sinander (2025)
# https://instituteforautomatedresearch.org/wiki/papers/econometrica/2025/dekel-et-al-comparative-statics-adjustment-costs-2025/
# Distilled: Develops a general theory of monotone comparative statics for models with adjustment costs, showing that ordinal complementarity on the objective and minimal monotonicity of the cost function suffice for comparative-statics conclusions and a Le Chatelier principle. Applied to saving, factor demand, pricing, labor supply, and capital investment. Econometrica 2025, CC BY 4.0. Six core theorems with proof locators and formal equations.
# Tags: paper-summary, comparative-statics, adjustment-costs, le-chatelier
==============================================================================

**What this is.** The paper's core theorems, the model structure, and the proof approach: a distilled skeleton sufficient to understand what was established and how, without reading all 34 pages. To replicate or extend the results, read the original at [https://doi.org/10.3982/ECTA22841](https://doi.org/10.3982/ECTA22841).

## TL;DR

The paper develops a general theory of monotone comparative statics for models with costly adjustment. The key insight is that the standard comparative-statics conclusion (an increase in a parameter leads to a higher optimal action) holds under the ordinal complementarity conditions of quasi-supermodularity and single-crossing differences on the objective, with only a minimal monotonicity condition on the cost function: it must be weakly less costly to adjust less. This is used to prove a general Le Chatelier principle, originating with Samuelson (1947): under adjustment costs the short-run response to a shock is bounded by the long-run response. Both results are extended to a fully dynamic model with long-lived forward-looking agents (and, separately, short-lived agents). Applications include saving by wishful thinkers, factor demand following Milgrom and Roberts (1996), pricing, labor supply, and capital investment. A key feature is that convex and nonconvex adjustment costs are handled in a unified framework.

## Core results

Theorems are qualitative; the "magnitude" column gives the precise conclusion. All results require quasi-supermodularity of F in x and single-crossing differences in (x, θ).

| # | Result | Locator | Conclusion |
|---|---|---|---|
| R1 | Theorem 1: comparative statics with adjustment costs | Theorem 1, p. 666 | If C is minimally monotone and θ̄ ≥ θ̲, then x̂ ≥ x̲ for some x̂ ∈ arg max G(x, θ̄) (provided the argmax is nonempty) |
| R2 | Theorem 2: Le Chatelier principle | Theorem 2, §4, p. 671 | If C is monotone and x̄ ∈ arg max F(x, θ̄) satisfies x̄ ≥ x̲, then x̄ ≥ x̂ ≥ x̲ for some short-run x̂; if x̄ is the largest long-run optimum then x̄ ≥ x̂ for every short-run x̂ |
| R3 | Theorem 3: dynamic Le Chatelier (long-lived agents) | Theorem 3, §5.2, p. 675 | Under monotone C_t and θ̲ ≤ θ_t ≤ θ̄ for every t, there is a solution (x_t) with x̲ ≤ x_t ≤ x̄ for every period t |
| R4 | Theorem 4: strong dynamic Le Chatelier | Theorem 4, §5.2, p. 675 | Under supermodularity, BCS, and additive separability of a time-invariant C, there is a solution with x̲ ≤ x_t ≤ x_{t+1} ≤ x̄ for every t (monotone upward adjustment over time) |
| R5 | Theorem 5: short-lived dynamic Le Chatelier | Theorem 5, §6, p. 679 | Short-lived agents' equilibrium satisfies x̲ ≤ x_t ≤ x̄ in the same direction as in Theorem 3 |
| R6 | Theorem 6: short- vs. long-lived agents | Theorem 6, §6, pp. 679-680 | Under additional convexity and equi-BCS conditions, short-lived agents adjust more sluggishly: x̲ ≤ x̃_t ≤ x_t ≤ x̄ for some equilibrium pair |

**Overall (paper's conclusion).** Comparative statics and the Le Chatelier principle are robust to adjustment costs under ordinal (not cardinal) complementarity conditions on the objective and minimal monotonicity on the cost. The prior literature, including Milgrom and Roberts (1996), required that short-run adjustment be completely infeasible in some dimensions. Section 7 establishes converses showing that minimal monotonicity (Theorem 1) and weak monotonicity (Theorem 2) are necessary as well as sufficient.

## Theory / model

### Setting

An agent chooses an action x from a sublattice $$L \subseteq \mathbb{R}^n$$. Her objective $$F(x, \theta)$$ depends on a parameter $$\theta \in \Theta$$. At the initial parameter $$\theta = \underline{\theta}$$, the optimal choice is (p. 664):

$$
\underline{x} \in \arg\max_{x \in L} F(x, \underline{\theta}).
$$

When the parameter rises to $$\bar{\theta} \geq \underline{\theta}$$, adjusting from $$\underline{x}$$ to $$x$$ costs $$C(x - \underline{x}) \geq 0$$, where $$C : \Delta L \to [0, \infty]$$ and $$\Delta L = \{x - y : x, y \in L\}$$. The agent's new choice maximizes (p. 664):

$$
G(x, \bar{\theta}) = F(x, \bar{\theta}) - C(x - \underline{x}).
$$

Infinite cost captures infeasibility. Since $$\underline{x}$$ is held fixed, the cost depends on the adjustment vector $$\varepsilon = x - \underline{x}$$ only.

### Monotonicity conditions on C

Two conditions on C appear in the results (pp. 664-665):

**Monotone:** Shifting any one dimension's adjustment closer to zero reduces cost:

$$
C(\varepsilon_1, \ldots, \varepsilon_{i-1}, \varepsilon'_i, \varepsilon_{i+1}, \ldots, \varepsilon_n) \leq C(\varepsilon)
\quad \text{whenever } 0 \leq \varepsilon'_i \leq \varepsilon_i \text{ or } 0 \geq \varepsilon'_i \geq \varepsilon_i. \tag{M}
$$

An additively separable $$C(\varepsilon) = \sum_{i=1}^n C_i(\varepsilon_i)$$ is monotone if and only if each $$C_i$$ is single-dipped at zero. Monotonicity is an ordinal property, preserved by strictly increasing transformations.

**Minimally monotone:** Simultaneously canceling all upward (or all downward) adjustments reduces cost:

$$
C(\varepsilon \wedge 0) \leq C(\varepsilon) \geq C(\varepsilon \vee 0)
\quad \text{for any adjustment vector } \varepsilon \in \Delta L, \tag{MM}
$$

where $$\varepsilon \wedge 0$$ replaces all positive entries of $$\varepsilon$$ with zero and $$\varepsilon \vee 0$$ replaces all negative entries with zero. Monotonicity implies minimal monotonicity; the converse fails. In the additively separable case, minimal monotonicity requires that each $$C_i$$ is minimized at zero.

### Ordinal complementarity conditions on F

Throughout, F satisfies (pp. 666-667, following Milgrom and Shannon (1994)):

- **Single-crossing differences in (x, θ):** $$F(y, \theta') - F(x, \theta') \geq 0$$ implies $$F(y, \theta'') - F(x, \theta'') \geq 0$$ whenever $$x \leq y$$ and $$\theta' \leq \theta''$$.
- **Quasi-supermodularity in x:** $$F(x, \theta) - F(x \wedge y, \theta) \geq(>) 0$$ implies $$F(x \vee y, \theta) - F(y, \theta) \geq(>) 0$$ for all $$x, y \in L$$.

These are ordinal properties, strictly weaker than the cardinal conditions of increasing differences and supermodularity.

### Dynamic model

In Section 5, the agent is long-lived and forward-looking. In each period $$t \in \mathbb{N} = \{1, 2, 3, \ldots\}$$, she takes action $$x_t \in L$$ and earns $$F(x_t, \theta_t)$$. Adjusting from $$x_{t-1}$$ to $$x_t$$ costs $$C_t(x_t - x_{t-1})$$. Given initial choice $$x_0 = \underline{x}$$, the agent maximizes (p. 674):

$$
\mathcal{G}\!\left((x_t)_{t=1}^\infty, x_0\right) = \mathcal{F}\!\left((x_t)_{t=1}^\infty\right) - \mathcal{C}\!\left(x_0, (x_t)_{t=1}^\infty\right),
$$

where

$$
\mathcal{F}\!\left((x_t)_{t=1}^\infty\right) = \sum_{t=1}^\infty \delta^{t-1} F(x_t, \theta_t)
\qquad \text{and} \qquad
\mathcal{C}\!\left(x_0, (x_t)_{t=1}^\infty\right) = \sum_{t=1}^\infty \delta^{t-1} C_t(x_t - x_{t-1}).
$$

Section 6 considers the alternative in which each period's action is chosen by a short-lived (or myopic) agent who takes $$x_{t-1}$$ as given and maximizes the period-t payoff $$G_t(x, x_{t-1}) = F(x, \theta_t) - C_t(x - x_{t-1})$$.

## Method

The paper uses lattice-theoretic methods, building on the `lattice-comparative-statics` framework of Topkis, and Milgrom and Shannon (1994). The key tools are:

**Sublattice operations.** For $$x, y \in L$$, the meet $$x \wedge y = (\min\{x_1, y_1\}, \ldots, \min\{x_n, y_n\})$$ and join $$x \vee y = (\max\{x_1, y_1\}, \ldots, \max\{x_n, y_n\})$$ both belong to L. The proof of Theorem 1 (p. 667) constructs $$\hat{x} = \underline{x} \vee x'$$ for any $$x' \in \arg\max G(x, \bar{\theta})$$. The key step uses minimal monotonicity: $$C(\underline{x} \vee x' - \underline{x}) = C((x' - \underline{x}) \vee 0) \leq C(x' - \underline{x})$$, so $$G(\hat{x}, \bar{\theta}) \geq G(x', \bar{\theta})$$, confirming that $$\hat{x}$$ also maximizes G and satisfies $$\hat{x} \geq \underline{x}$$.

**Monotonization argument (Theorem 4, Appendix I, pp. 686-688).** The strong dynamic result is proved by showing that any solution $$(x_t)_{t=1}^\infty$$ to the forward-looking problem can be replaced by the running-maximum sequence $$X_t = x_1 \vee x_2 \vee \cdots \vee x_t$$ without reducing optimality. The key inequality (eq. (5) in Appendix I) is, for each dimension $$i$$:

$$
C_i(y \vee z - x \vee y) + C_i(y \wedge z - x \wedge y) \leq C_i(y - x) + C_i(z - y)
\quad \text{for all } x, y, z, \tag{5}
$$

which holds because each $$C_i$$ is single-dipped at zero. Combined with supermodularity of $$F(\cdot, \bar{\theta})$$, this ensures monotonization preserves optimality.

**Necessity results (Section 7, Theorems 1†, 2†, 3†, pp. 680-682).** The paper proves converses by explicit counterexample construction. For example, Theorem 1† (p. 680) shows that minimal monotonicity is equivalent to: for all quasi-supermodular F with single-crossing differences, $$\theta \geq \underline{\theta}$$ implies $$\hat{x} \geq \underline{x}$$ for some $$\hat{x} \in \arg\max G(x, \bar{\theta})$$, and $$\theta \leq \underline{\theta}$$ implies $$\hat{x} \leq \underline{x}$$. The counterexample when C fails minimal monotonicity is constructed on a sublattice $$X = \{\underline{x} \wedge \hat{x}, \underline{x}, \hat{x}, \underline{x} \vee \hat{x}\}$$ with explicitly specified F values (pp. 680-681).

## Empirical specifications

This paper contains no empirical analysis. The formal results are applied to five standard economic models; these applications demonstrate that the theory delivers sharp conclusions without the auxiliary functional-form assumptions that each literature has typically imposed.

**Saving by wishful thinkers (§3.3, pp. 669-671, Proposition 2).** Following Caplin and Leahy (2019), an agent consumes $$c \in [0, w]$$ and chooses a belief $$G$$ (a CDF over future income) from a set $$\mathcal{G}$$ ordered by first-order stochastic dominance. The lifetime payoff is (p. 669):

$$
U(c, G) = u_1(c) + \int_{\mathcal{Y}} u_2\!\left((1+r)(w - c) + y\right) G(\mathrm{d}y),
$$

where $$u_1, u_2$$ are continuous, concave, and strictly increasing. A wishful thinker chooses:

$$
(\hat{c}, \hat{G}) \in \arg\max_{(c, G) \in [0, w] \times \mathcal{G}} \left[U(c, G) - C(G - G_0)\right],
$$

where $$C$$ is minimally monotone and $$G_0$$ is the realist's belief. Proposition 2 (p. 670) establishes $$\hat{c} \geq c_0$$ (wishful thinkers over-consume) and $$\hat{G} \geq_1 G_0$$ (wishful thinkers adopt more optimistic beliefs). The proof applies Theorem 1*, the constraint-shift variant of Theorem 1, since $$[0, w] \times \mathcal{G}$$ is a sublattice. Notably, Caplin and Leahy (2019) assumed the Kullback-Leibler functional form for C; this assumption is not needed here.

**Factor demand (§4.2, p. 673).** A firm uses capital k and labor $$\ell$$ to produce output $$f(k, \ell)$$. Profit at factor prices $$(r, w)$$ is $$F(k, \ell, -w) = f(k, \ell) - rk - w\ell$$. By Theorem 2, a drop in the wage w precipitates a short-run increase in both k and $$\ell$$ when $$f$$ is supermodular (complements), with a further increase in the long run. When $$f$$ is submodular (substitutes), rewriting the choice variable as $$(x_1, x_2) = (-k, \ell)$$ restores supermodularity; $$\ell$$ still increases in both runs while k now decreases.

**Pricing (§4.3 and §5.3, pp. 673-674, 677).** A monopolist with constant marginal cost $$c$$ faces demand $$D(p, \eta)$$ where $$\eta$$ is an elasticity shifter. Profit $$F(p, (c, -\eta)) = (p - c)D(p, \eta)$$ has single-crossing differences in $$(p, (c, -\eta))$$ (using the "log increasing differences" condition, p. 673) and is quasi-supermodular since $$p \in \mathbb{R}$$. By Theorems 1 and 2, the monopolist raises her price in both the short and long run whenever marginal cost rises or demand becomes less elastic ($$\eta$$ falls), without any assumptions on the adjustment cost C beyond minimization at zero. Theorem 4 further implies that prices adjust monotonically upward over time in the dynamic version.

**Labor supply (§5.4, pp. 677-678).** A worker chooses labor supply $$x \in L \subseteq \mathbb{R}_+$$ with per-period utility $$F(x, T) = wx - T(wx) - \kappa(x)$$, where T is the tax schedule and $$\kappa$$ is effort disutility. A tax reform from T to $$\tilde{T}$$ with lower marginal rates ($$\tilde{T} \geq_{\text{flat}} T$$) yields F with single-crossing differences in $$(x, T)$$. Theorems 1-4 imply that labor supply rises at every horizon and adjusts monotonically upward over time under a one-off permanent rate cut.

**Capital investment (§5.5, pp. 678-679).** A firm adjusts capital $$k_t \in \mathbb{R}_+$$ with per-period profit $$F(k, (p, \eta, -r)) = pf(k, \eta) - rk$$ where $$f$$ has increasing differences (so F has increasing differences in $$(k, \theta)$$ for $$\theta = (p, \eta, -r)$$). Investing $$i_t = k_t - k_{t-1}$$ costs $$C(i_t) \geq 0$$, assumed only to be single-dipped at zero. Theorem 4 delivers monotone upward adjustment of capital over time whenever the marginal product of capital rises (fall in r, rise in p, or rise in $$\eta$$). The analysis covers both convex and nonconvex (lumpy) investment; when adjustment costs are not even single-dipped (e.g., a minimum investment threshold creates a region $$C(\varepsilon) = \infty$$), Theorem 1 still applies since C remains minimally monotone.

## Datasets used

This is a pure-theory paper. No datasets are used.

## When to read the full paper

Read the original at [https://doi.org/10.3982/ECTA22841](https://doi.org/10.3982/ECTA22841) if you:
- are applying these results to a new economic model and need the precise conditions and all proof details (Appendices A-N, pp. 682-693);
- are working on dynamic-adjustment models with nonconvex or general cost functions and need the full statements of Theorems 3-6;
- need the extension to uncertain adjustment costs (Appendix B, pp. 683-684, Theorems 1'-3');
- need the constraint-shift variant (Theorem 1*, §3.1, pp. 668-669) or the necessity results of Section 7 to understand when the conditions can be relaxed.

## Attribution and rights

Source: peer-reviewed, *Econometrica* 93(2), March 2025. This distillation was extracted by an LLM on 2026-06-26 and is **not human-verified or independently reproduced**.

> **Attribution (CC BY 4.0).** Dekel, Eddie, John K.-H. Quah, and Ludvig Sinander.
> "Comparative Statics With Adjustment Costs and the Le Chatelier Principle."
> *Econometrica* 93, no. 2 (March 2025): 661-694.
> DOI: 10.3982/ECTA22841. © 2025 The Authors.
> Licensed under the [Creative Commons Attribution License (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Feedback Design in Dynamic Moral Hazard: Ely, Georgiadis & Rayo (2025)
# https://instituteforautomatedresearch.org/wiki/papers/econometrica/2025/ely-et-al-feedback-design-dynamic-moral-2025/
# Distilled: In a dynamic moral hazard setting with a binary success signal, the jointly optimal performance feedback and reward contract takes a two-phase bang-bang form: an initial silent phase (agent kept in the dark) followed by a full-transparency pronto phase, driven by a backward compounding effect that makes front-loading ignorance uniquely optimal. Econometrica 2025, CC BY-NC 4.0. Five core theoretical results with source locators, the model equations, and the solution method; LLM-distilled, not reproduced.
# Tags: paper-summary, contract-theory, moral-hazard, information-design, dynamic-contracting
==============================================================================

**What this is.** The paper's core theoretical results, the model equations (principal-agent setup with binary success signal), and the solution method (minimal reward characterization, bang-bang relaxed problem): enough to know what it found and how, without reading all 25 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.3982/ECTA21871).

## TL;DR

The paper studies the optimal joint design of performance feedback and monetary rewards when the only available performance signal is coarse: a binary "success" that arrives stochastically as a function of the agent's accumulated effort. A "backward compounding effect" governs the analysis: promising greater future rewards to prevent pausing raises the cost of all earlier rewards as well, and this cost is larger the further into the future the information is hidden. As a result, it is always optimal to front-load the agent's ignorance. The optimal contract takes a two-phase bang-bang form (Theorem 1, p. 606): a fully silent phase in which the principal says nothing and the agent works regardless of success, followed by a full-transparency pronto phase in which the agent is immediately informed and stops upon success. The silent phase exists precisely when the flow cost satisfies c < 1/2. The two-phase structure is robust to continuation payoffs, a more informed agent, learning-by-doing, and costly direct monitoring.

## Core results

All results are theoretical; locators point into the source PDF.

| # | Result | Locator | Stated form |
|---|---|---|---|
| R1 | Local incentive constraint (LIC): necessary condition for incentive compatibility via deterring instantaneous pauses | Prop. 1, p. 602 | r(t)R(t)f(t) - cp(t) ≥ cr(t)Q(t)f(t) + ∫ r(s)[R(s)-cQ(s)]\|f'(s)\| ds - c∫ r(s)f(s) ds |
| R2 | Minimal implementing reward schedule: unique least-cost reward satisfying LIC with equality at every t | Prop. 2, p. 603 | r(t)R(t) = c[p(t)/f(t) - ∫ f'(s)/f(s)² p(s) ds - ∫ (r(s)-r(t)q(s\|t)) ds] (eq. 4) |
| R3 | Principal's objective simplifies to (Obj); relaxed problem (P) admits bang-bang solution with cutoffs t\* and T\* | Lemma 1 + Prop. 3, pp. 605-606 | t\* = min{t : 1-2c-cΦ(t) ≤ 0}; p = 1 on [0,t\*], 1-F(t) on (t\*,T\*], 0 thereafter; r = 1 on [0,T\*] |
| R4 | Every optimal policy takes exactly two phases: silent (no disclosure) then pronto (immediate disclosure) | Thm. 1, p. 606 | Silent reward: c/λ(T\*) + cF(t\*)/f(t\*); pronto reward: c/λ(T\*); silent phase iff c < 1/2 |
| R5 | With direct monitoring cost m > c, early silence dominates always pronto when the silent phase is sufficiently short | Prop. 4, p. 612 | Expected cost ratio Δ(t̃) < 1 iff m > c and t̃ sufficiently small; m > c is the exact threshold |

**Overall (paper's conclusion).** The backward compounding effect unifies the analysis: silence grants the agent rents, and these rents compound the longer they are deferred, making it uniquely optimal to front-load ignorance. The two-phase bang-bang structure holds under every extension considered, and the key quantity governing it is the function Φ(t) = F(t)(d/dt)(1/f(t)), which measures how severely the backward compounding escalates over time.

## Theory / model

The baseline model (Section 2, pp. 599-601) is a continuous-time game on [0, T] with T large and finite. At each instant, the agent privately chooses to work (incurring flow cost c < 1 per unit time) or wait. A success occurs at most once and is observed only by the principal. The agent's cumulative effort e maps to success via a CDF $$F : [0, T] \to [0, 1]$$ with F(0) = 0 and F(T) ≤ 1. Both the density $$f = F'$$ and the hazard rate $$\lambda(e) = f(e)/[1-F(e)]$$ are weakly decreasing in e. The key auxiliary function (p. 600, eq. 1) is:

$$
\Phi(t) := F(t)\,\frac{d}{dt}\frac{1}{f(t)}, \quad \Phi(0) = 0, \quad \Phi \text{ weakly increasing.} \tag{1}
$$

A larger $$\Phi(t)$$ means the backward compounding effect is more severe: the hazard rate is falling faster (large $$-f'/f^2$$) and there is more cumulative mass of past rewards to compound (large $$F(t)$$).

The principal designs (i) a reward schedule $$R(t) \geq 0$$ paid upon success at t and (ii) a feedback policy. By a direct-mechanism argument it suffices to consider direct policies: $$q(s|t)$$, the probability the agent is asked to continue working at s conditional on having succeeded at $$t \leq s$$, and $$r(s)$$, the probability the agent is asked to continue conditional on no success yet. The total probability the agent is asked to work at least until s is (p. 600, eq. 2):

$$
p(s) = r(s)\bigl[1 - F(s)\bigr] + \int_0^s r(u)\,f(u)\,q(s|u)\,du. \tag{2}
$$

The agent's continuation payoff from obeying recommendations from t onward is (p. 601, eq. 3):

$$
U(t) := \int_t^T r(s)\,R(s)\,f(s)\,ds - c\int_t^T p(s)\,ds. \tag{3}
$$

The principal maximizes total expected effort net of rewards: $$\int_0^T p(s)\,ds - \int_0^T r(s)R(s)f(s)\,ds$$, subject to incentive compatibility. The agent is risk neutral and cash-constrained (no punishment below zero). Both players are assumed not to discount time, though the results are qualitatively unchanged with a common discount rate.

The paper builds on the continuous-time principal-agent approach of Sannikov (2008) and the information-design frameworks of Kamenica and Gentzkow (2011) and Rayo and Segal (2010), extending them to a setting where the principal jointly controls both the feedback policy and the monetary rewards.

## Method

**Step 1 - Local incentive constraint (Proposition 1, p. 602).** It suffices to focus on instantaneous deviations (brief pauses). Dividing the no-pause condition by $$\Delta t$$ and letting $$\Delta t \to 0$$ establishes the local incentive constraint (LIC): for all t,

$$
r(t)R(t)f(t) - c\,p(t) \geq c\,r(t)Q(t)f(t) + \int_t^T r(s)\bigl[R(s) - cQ(s)\bigr]\lvert f'(s)\rvert\,ds - c\int_t^T r(s)\,f(s)\,ds, \tag{LIC}
$$

where $$Q(t) = \int_t^T q(u|t)\,du$$ is the agent's expected future work conditional on succeeding at t. The three right-hand terms represent: (i) the effort savings (the agent avoids cost c for $$Q(t)$$ future periods), (ii) the higher density of future success because accumulated effort is lower (the backward compounding integral), and (iii) the higher probability that future recommendation-to-work periods are reached.

**Step 2 - Minimal reward schedule (Proposition 2, p. 603).** For any given recommendation policy, there is a unique least-expensive reward schedule satisfying (LIC) with equality at every t. Equation (4) gives it:

$$
r(t)R(t) = c\!\left[\frac{p(t)}{f(t)} - \int_t^T \frac{f'(s)}{f(s)^2}\,p(s)\,ds - \int_t^T \bigl(r(s) - r(t)\,q(s|t)\bigr)\,ds\right]. \tag{4}
$$

The first term is the zero-rent reward (just enough to cover current expected cost). The second term is the backward compounding penalty: since $$f' \leq 0$$, this term is positive and requires raising current rewards to compensate for high future rewards. The third term is the information rebate: when the principal reveals a success at t (low $$q(s|t)$$), the agent works less thereafter, reducing future reward obligations, so current rewards can be lower.

**Step 3 - Simplified objective and bang-bang relaxed problem (Lemma 1 + Proposition 3, pp. 605-606).** Substituting (4) into the objective, Lemma 1 shows it simplifies to:

$$
\int_0^T p(t)\,dt - c\int_0^T \bigl[p(t)(1 + \Phi(t)) - (r(t) - p(t))\bigr]\,dt. \tag{Obj}
$$

The integrand $$r(t) - p(t)$$ is the information rebate (positive only when a successful agent is still asked to work). Rearranging and collecting terms, the relaxed problem (P) that ignores global deviations is:

$$
\sup_{r(\cdot),\,p(\cdot)} \int_0^T p(t)\{1 - 2c - c\Phi(t)\}\,dt + c\int_0^T r(t)\,dt \tag{P}
$$

$$
\text{s.t.} \quad r(t)[1-F(t)] \leq p(t) \leq 1, \quad 0 \leq r(t) \leq 1 \text{ non-increasing.} \tag{Feas}
$$

Since $$\Phi$$ is weakly increasing, the coefficient $$1-2c-c\Phi(t)$$ is weakly decreasing in t: positive at t = 0 when c < 1/2 and eventually non-positive. Because the objective is linear in $$p(t)$$ and $$r(t)$$, Proposition 3 establishes a bang-bang solution: define $$t^* = \min\{t : 1-2c-c\Phi(t) \leq 0\}$$ and let $$T^* \in (t^*, T]$$ be the optimal terminal date. Then:

$$
p(t) = \begin{cases} 1 & t \in [0, t^*], \\ 1-F(t) & t \in (t^*, T^*], \\ 0 & t > T^*, \end{cases} \qquad r(t) = \begin{cases} 1 & t \leq T^*, \\ 0 & t > T^*. \end{cases}
$$

**Step 4 - Two-phase optimal contract (Theorem 1, p. 606).** The bang-bang solution is implemented by a deterministic-deadline contract with exactly two phases. In Phase 1 (silent phase, $$t \in [0, t^*]$$): the principal is silent; $$q(t|s) \equiv r(t) \equiv 1$$; the agent works continuously regardless of success. In Phase 2 (pronto phase, $$t \in (t^*, T^*]$$): $$r(t) = 1$$ and $$q(s|t) \equiv 0$$; the agent quits immediately upon success. Substituting into (4) and simplifying using $$\lambda(T^*) = f(T^*)/[1-F(T^*)]$$, the reward for each phase is (Appendix A.5, pp. 618-619):

$$
R(t) = \frac{c}{\lambda(T^*)} + \frac{cF(t^*)}{f(t^*)} \quad \text{for } t \in [0, t^*], \qquad R(t) = \frac{c}{\lambda(T^*)} \quad \text{for } t \in (t^*, T^*].
$$

The phase 1 reward exceeds the phase 2 reward by $$cF(t^*)/f(t^*)$$: the "silent premium" that compensates for the rents the agent would earn during the pronto phase, compounded backward into the silent phase. The cutoff $$t^* > 0$$ (positive silent phase) if and only if c < 1/2. Global incentive compatibility is verified by showing that the non-increasing reward profile makes any global deviation unprofitable (Appendix A.5). The result contrasts with Ely and Szydlowski (2020), where quitting is irreversible and rewards are exogenous, and with Halac, Kartik, and Liu (2016) and Mason and Välimäki (2015), neither of which allows the principal to strategically withhold performance feedback.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| None | Pure theory paper; all results are analytical propositions and theorems | n/a |

## When to read the full paper

Read the [original](https://doi.org/10.3982/ECTA21871) if you are: designing performance evaluation or promotion systems with coarse binary signals in employment or professional services settings; extending the two-phase framework to multiple agents or contest environments (the paper's own suggested direction); studying how information design interacts with monetary incentives under dynamic moral hazard; or working through the formal proofs, which are detailed and constructive (Appendix A, pp. 613-620). Section 5 (pp. 608-613) covers four extensions: continuation payoffs, a more informed agent, an upfront learning-by-doing investment phase, and costly direct monitoring.

## Attribution and rights

Source: peer-reviewed, *Econometrica* 93(2), March 2025. This distillation was extracted by an LLM on 2026-06-26 and is **not human-verified or independently reproduced**. The CC BY-NC 4.0 licence permits noncommercial sharing; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY-NC 4.0).** Ely, Jeffrey C., George Georgiadis, and Luis Rayo.
> "Feedback Design in Dynamic Moral Hazard."
> *Econometrica* 93, no. 2 (March 2025): 597-621.
> DOI: 10.3982/ECTA21871. © 2025 The Authors. The Econometric Society.
> Licensed under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# How Well Does Bargaining Work: Freyberger & Larsen (2025)
# https://instituteforautomatedresearch.org/wiki/papers/econometrica/2025/freyberger-larsen-well-bargaining-work-consumer-2025/
# Distilled: Freyberger and Larsen (2025) derive sharp nonparametric bounds on buyer and seller private value distributions and on the first-best trade probability from eBay Best Offer bargaining data, using a hierarchy of behavioral assumptions without specifying a complete equilibrium model. Under preferred assumptions (stochastic monotonicity and positive correlation), at least 37% of failed trades are cases where gains from trade existed. Econometrica 2025, paywalled. Seven core results with source locators, the bounds framework with equations, and the estimation approach.
# Tags: paper-summary, bargaining, partial-identification, market-microstructure
==============================================================================

**What this is.** This is a distilled skeleton of Freyberger and Larsen (2025), *Econometrica*. It records the paper's core bounds results, framework equations, and dataset with locators to specific tables, figures, and equations. Read the original at https://doi.org/10.3982/ECTA20125 to replicate or extend.

## TL;DR

Freyberger and Larsen (2025) use eBay Best Offer platform data to measure how efficiently buyers and sellers in consumer markets reach agreement. Rather than estimating a structural bargaining model, they propose an incomplete-model (partial identification) approach: they derive sharp nonparametric bounds on buyer and seller private value distributions ($$F_B$$, $$F_S$$) and on the counterfactual first-best trade probability $$P(B \geq S)$$ under a hierarchy of behavioral assumptions. The weakest assumption (Assumption A1, revealed preferences only) gives wide bounds. The strongest assumptions (seller monotonicity A2, buyer independence A3, as in Perry (1986) and Cramton (1992)) cross for most products, indicating they are too strong for inexperienced consumer negotiators and fail in the presence of unobserved game-level heterogeneity. The preferred assumptions, stochastic monotonicity (A4) and positive correlation (A5), are consistent with the data for all 36 products and yield informative non-crossing bounds. Under these, the paper finds that for the median product at least 37% of failed trades are inefficient: the buyer genuinely valued the good more than the seller but the parties failed to agree. The auto accept/decline feature and new product status are each associated with lower inefficient impasse, while buyer experience appears to worsen it, consistent with information-rent extraction motives noted by Myerson and Satterthwaite (1983). The approach builds on the partial identification tradition of Manski (1989) and the incomplete-model auction bounds of Haile and Tamer (2003), extending both to a two-sided sequential bargaining setting. Keniston (2017) and Larsen (2021) are the closest related structural empirical studies; this paper extends beyond them by weakening the behavioral assumptions required.

## Core results

| # | Result | Locator | Magnitude as reported |
|---|---|---|---|
| R1 | Seller monotonicity bounds (A2) cross for all products, indicating the assumption is violated | Table III, p. 182 | Frac. Cross = 1.00 across 36 products; IVE = 0.23 |
| R2 | Buyer independence bounds (A3) cross for 42% of products; 11% statistically significant | Table III, p. 182 | Frac. Cross = 0.42; Frac. Reject = 0.11; IVE = 0.006 |
| R3 | Stochastic monotonicity + positive correlation (A4+A5) do not cross for any product | Table III, p. 182 | Frac. Cross = 0; IVE = 0 (seller and buyer bounds) |
| R4 | Cell phone product: first-best trade probability lower bound = 0.508 vs. P(sale) = 0.276 | Table V, p. 186 | Implied inefficient impasse = 45.6% (= 1 - 0.276/0.508); 95% CI [0.450, 0.540] |
| R5 | Median product: inefficient impasse lower bound = 37.3%; range 18.0% to 54.2% | Fig. 6B, pp. 187-188 | All 36 products have lower bounds above P(sale) under preferred assumptions |
| R6 | Auto accept/decline: inefficient impasse lower bound 5.8 pp lower for users vs. non-users | Table VI Panel A, p. 189 | Diff = -0.058, S.E. = 0.0250 (statistically significant) |
| R7 | New products: inefficient impasse lower bound 11.9 pp lower than used products | Table VI Panel C, p. 189 | Diff = -0.119, S.E. = 0.0610; t = 1.95 (nearly significant at 5%) |

**Overall (paper's conclusion).** Seller monotonicity, while satisfied in theoretical equilibria such as Cramton (1992) and Perry (1986), is rejected for all 36 products, most likely because unobserved game-level heterogeneity (e.g., aspects of the item's condition known to both parties but not the econometrician) induces nonmonotonicities between the seller's value and first offer. Stochastic monotonicity and positive correlation are consistent with the data and yield the tightest non-crossing bounds. Under these preferred assumptions, real-world eBay consumer bargaining exhibits substantial inefficient impasse: at least 37.3% of failed trades (median product) are cases where the buyer values the good above the seller. Automation tools (auto accept/decline) and new product status are associated with lower impasse; increased buyer experience appears linked to higher impasse, consistent with experienced agents extracting information rents at the cost of reducing total surplus.

## Theory / model

The paper has no formal theoretical model. It proposes an incomplete-model (partial identification) framework whose theoretical content lies in the bargaining game setup, the revealed-preference restrictions, and the sharpness proofs.

**Bargaining game setup (Section 3.1, p. 167).** A seller with private value $$S \sim F_S$$ and a buyer with private value $$B \sim F_B$$ negotiate over the eBay Best Offer protocol. The seller posts a list price as the first offer ($$P_1^S$$); the buyer responds with a first offer ($$P_2^B$$); each party then alternates accepting, countering, or quitting, up to three offers per side. Values represent net willingness to accept (seller) and willingness to pay (buyer) inclusive of outside options. The paper allows $$B$$ and $$S$$ to be correlated across instances through unobserved game-level heterogeneity $$W$$ known to both agents but not the econometrician.

Key sequence-level statistics (p. 167):
- $$X^S_{AC}$$: smallest offer the seller makes or accepts/counters (prices at which she is willing to trade)
- $$X^S_Q$$: largest price at which the seller quits
- $$X^B_{AC}$$: largest price the buyer accepts or offers
- $$X^B_Q$$: smallest price at which the buyer quits

Assumption A1 (revealed preferences, p. 170) implies $$X^S_Q \leq S \leq X^S_{AC}$$ and $$X^B_{AC} \leq B \leq X^B_Q$$ in every realization.

**Representation lemma (p. 167).** Applying the law of iterated expectations:

$$P(S \leq x) = \int P\!\left(S \leq x \mid P_1^S = y\right) dF_{P_1^S}(y), \tag{1}$$

$$P(B \leq x) = \int P\!\left(B \leq x \mid P_1^S = y,\, P_2^B = z\right) dF_{P_1^S, P_2^B}(y, z). \tag{2}$$

These representations are the foundation for all bounds: each assumption restricts the unobserved conditional $$P(S \leq x | P_1^S = y)$$ (or its buyer analogue), which is then bracketed by observed empirical quantities from the sequence of offers, acceptances, and quits.

**Identification logic.** The central objects $$F_S$$, $$F_B$$, and $$P(B \geq S)$$ are not directly observable. The paper asks what can be inferred from observable bargaining actions under progressively stronger behavioral restrictions, without selecting a specific equilibrium. The answer is sharp bounds: for every assumption set, the paper proves that any CDF between the lower and upper bound is consistent with the data and the assumptions (Theorems 1-7, pp. 170-185). Sharpness means there exists a data-generating process satisfying the assumptions under which the true distribution exactly equals the bound.

## Method

The paper derives a hierarchy of sharp bounds on $$F_S$$, $$F_B$$, and $$P(B - S \geq x)$$ under five assumption sets (A1 through A5 for marginal distributions, A6-A7 for the surplus object).

**Unconditional bounds from A1 alone (Theorem 1, p. 170).** Revealed preferences directly imply:

$$P(X^S_{AC} \leq x) \leq F_S(x) \leq P(X^S_Q \leq x), \tag{3}$$

$$P(X^B_Q \leq x) \leq F_B(x) \leq P(X^B_{AC} \leq x). \tag{4}$$

These are the weakest bounds. The seller upper bound is often near 1 because seller quit prices are unobserved when sequences end in agreement or buyer quit.

**Monotonicity bounds from A1+A2 (Theorem 2, p. 172).** Assumption A2 states that $$\overline{\text{supp}}(S | P_1^S = y)$$ is weakly increasing in $$y$$ (sellers with higher first offers have stochastically higher values), and analogously for buyers. Defining $$X^{S*}_{AC}(y) \equiv \overline{\text{supp}}(X^S_{AC} | P_1^S \geq y)$$:

$$\int \mathbf{1}\!\left(X^{S*}_{AC}(y) \leq x\right) dF_{P_1^S}(y) \leq F_S(x) \leq \int \mathbf{1}\!\left(X^{S*}_Q(y) \leq x\right) dF_{P_1^S}(y), \tag{5}$$

with analogous buyer bounds (eq. 6, p. 172). Seller monotonicity bounds cross for all 36 products (R1), and the auto-accept/decline validation confirms the rejection (Section 5.1.1, p. 178-179).

**Independence bounds from A1+A3 (Theorem 3, p. 173).** Assumption A3 states (i) $$S$$ is independent of $$P_2^B$$ conditional on $$P_1^S$$, and (ii) $$B$$ is independent of $$P_1^S$$. With $$m^S_{AC}(x, y, z) = P(X^S_{AC} \leq x | P_1^S = y, P_2^B = z)$$:

$$\int \max_z m^S_{AC}(x, y, z)\, dF_{P_1^S}(y) \leq F_S(x) \leq \int \min_z m^S_Q(x, y, z)\, dF_{P_1^S}(y), \tag{7}$$

$$\max_{y'} P(X^B_Q \leq x \mid P_1^S = y') \leq F_B(x) \leq \min_{y'} P(X^B_{AC} \leq x \mid P_1^S = y'). \tag{8}$$

Buyer independence bounds cross for 42% of products (R2). The paper demonstrates that additive or multiplicative unobserved heterogeneity violates A3 even within Perry (1986) and Cramton (1992) equilibria (Supplemental Appendix G, p. 173-174).

**Stochastic monotonicity bounds from A1+A4 (Theorem 4, pp. 174-175).** Assumption A4 weakens A2 to require only that $$P(S \leq x | P_1^S = y)$$ is weakly decreasing in $$y$$ for all $$x$$. The bounds are:

$$\int \max_{y' \geq y} P(X^S_{AC} \leq x \mid P_1^S = y')\, dF_{P_1^S}(y) \leq F_S(x) \leq \int \min_{y' \leq y} P(X^S_Q \leq x \mid P_1^S = y')\, dF_{P_1^S}(y), \tag{9}$$

with analogous buyer bounds (eq. 10). These are implied by A2 but do not cross.

**Positive correlation bounds from A1+A5 (Theorem 5, p. 175).** Assumption A5 states that $$P(S \leq x | P_1^S = y, P_2^B = z)$$ is weakly decreasing in $$z$$ (one agent's value is stochastically increasing in the other's first offer). Combined with A4:

$$\int \max_{z' \geq z} m^S_{AC}(x, y, z')\, dF_{P_1^S, P_2^B}(y, z) \leq F_S(x) \leq \int \min_{z' \leq z} m^S_Q(x, y, z')\, dF_{P_1^S, P_2^B}(y, z). \tag{11}$$

Combined A4+A5 bounds do not cross for any of the 36 products (R3, Table III), making these the preferred "Goldilocks" assumptions.

**Surplus bounds for $$P(B \geq S)$$ (Theorems 6-7, p. 185).** To bound the first-best trade probability directly, the paper adds Assumption A6 (surplus stochastic monotonicity: $$P(B - S \geq x | P_1^S = y, P_2^B = z)$$ increasing in $$z$$) and A7 (surplus weak monotonicity: $$\overline{\text{supp}}(B - S | P_1^S = y, P_2^B = z)$$ increasing in $$z$$). Under A1, buyer monotonicity A2.ii, and A7:

$$P(B - S \geq x) \geq \int \mathbf{1}\!\left(X^{B*-S}_{AC}(y, z) \geq x\right) dF_{P_1^S, P_2^B}(y, z), \tag{15}$$

where $$X^{B*-S}_{AC}(y, z) \equiv \overline{\text{supp}}(X^B_{AC}(y, z) - X^S_{AC} : P_2^B \geq z, P_1^S = y)$$. Evaluating at $$x = 0$$ gives a lower bound on $$P(B \geq S)$$. The inefficient impasse lower bound is then $$1 - P(\text{sale})/\widehat{P}(B \geq S)^{LB}$$.

**Estimation (Section 4, pp. 176-177).** Conditional probabilities such as $$P(X^S_{AC} \leq x | P_1^S = y)$$ are estimated using the Nadaraya-Watson kernel estimator with an Epanechnikov kernel and bandwidth $$n^{-1/4}$$ for one-dimensional conditioning. For two-dimensional conditioning on $$(P_1^S, P_2^B)$$ the bandwidth is $$n^{-1/5}$$. Because some plug-in estimators are inward biased (artificially tight), the paper modifies them to be half-median-unbiased following Chernozhukov, Lee, and Rosen (2013) (p. 176). All estimation is done separately by product; prices are normalized by the product's reference price.

## Empirical specifications

The estimation sample requires at least 200 bargaining sequences per product after restrictions (nonoverlapping buyer/seller time windows, first-seller-per-buyer limit, exclusion of extreme offers). This yields 12,012 sequences for 36 products (Table A1, Supplemental Appendix, p. 166).

**Bounds validation via auto-accept/decline prices (Section 5.1.1, pp. 178-180).** For the 363 negotiations where sellers reported secret auto-accept and auto-decline thresholds, the paper uses these as known bounds on $$S$$ (auto-accept price is a weak upper bound; auto-decline price is a weak lower bound) to cross-check the estimated $$F_S$$ bounds without using these prices in estimation. Under combined independence + stochastic monotonicity, the estimated bounds are consistent: the auto-accept CDF lies above the $$F_S$$ lower bound and the auto-decline CDF lies below the $$F_S$$ upper bound (Figure 3, right panel, p. 179). Seller monotonicity bounds are rejected by this exercise.

**Bound crossing tests (Section 5.2, Table III, pp. 181-182).** For each assumption set and each product, the paper tests whether the estimated lower bound significantly exceeds the estimated upper bound at any price point on a grid from 0 to 2.5 (increments of 0.1 units of reference price), using 95% one-sided subsampling confidence bands. The integrated violation error (IVE) measures the average excess where lower exceeds upper:

$$\text{IVE} = \int \max\!\left(F^L(x) - F^U(x),\, 0\right) dG(x),$$

where $$G$$ is the unconditional lower bound (sellers) or upper bound (buyers). Table III reports crossing fraction, rejection fraction, and IVE across all 36 products and all assumption sets.

**Inefficient impasse heterogeneity (Section 6.4, Table VI, pp. 188-191).** For each subsample condition (message exchanged, eBay store seller, U.S. buyer, auto accept/decline prices reported, number of photos relative to median, seller rating, seller/buyer experience level, new vs. used product, reference price relative to median), the paper computes the inefficient impasse lower bound $$1 - P(\text{sale})/\widehat{P}(B \geq S)^{LB}$$ separately for observations satisfying and not satisfying the condition, requiring at least 100 qualifying observations per group per product. Within-product differences are averaged across products with standard errors via the delta method. All bounds use surplus weak monotonicity (A7) combined with buyer monotonicity (A2.ii).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| eBay Best Offer bargaining sequences (Backus et al. 2020) | 12,012 bargaining sequences for 36 consumer products; list prices, buyer and seller counteroffers, acceptance and quit decisions, auto-accept/decline thresholds for a subset; U.S. eBay site, June 2012 to May 2013 | no page yet |

Sample: 36 products (bar-code + condition pairs), 12,012 sequences, June 2012 to May 2013 (Table I, p. 166). Reference prices are averages over non-Best-Offer posted-price sales of the same product during the sample period; all offers are expressed as fractions of the reference price.

## When to read the full paper

Read Freyberger and Larsen (2025) if you:
- Are designing or evaluating a bargaining or negotiation mechanism and want empirical benchmarks on inefficiency without imposing Nash bargaining or a specific equilibrium.
- Want to apply partial identification bounds to game-theoretic settings with incomplete information, especially where standard structural assumptions may be violated by unobserved heterogeneity.
- Are studying the eBay Best Offer marketplace or similar consumer negotiation platforms and need a validated nonparametric approach for bounding private value distributions.
- Need to understand which game-theoretic assumptions (monotonicity, independence, stochastic monotonicity, positive correlation) are empirically falsifiable from sequential-offer data and where they fail (Table III cross-check; Figure 3 auto-accept/decline validation, p. 179).

The Supplemental Appendix (Freyberger and Larsen (2024), https://doi.org/10.3982/ECTA20125) contains sharpness proofs (Appendix C), Monte Carlo simulations comparing bias-corrected and uncorrected estimators (Appendix F), and the theoretical analysis of Perry (1986) and Cramton (1992) equilibria under unobserved heterogeneity (Appendix G).

## Attribution and rights

Freyberger, Joachim, and Bradley J. Larsen. "How Well Does Bargaining Work in Consumer Markets? A Robust Bounds Approach." *Econometrica* 93, no. 1 (January 2025): 161-194. https://doi.org/10.3982/ECTA20125

Copyright 2025 The Econometric Society. All rights reserved. No open-access license found in Crossref metadata. Extract-only: this page reproduces no figures or tables verbatim; all results are cited with their original locators.

LLM-distilled by paper-distiller (claude-sonnet-4-6); not human-verified; not reproduced.


==============================================================================
# Estimating Candidate Valence: Kawai & Sunada (2025)
# https://instituteforautomatedresearch.org/wiki/papers/econometrica/2025/kawai-sunada-estimating-candidate-valence-2025/
# Distilled: Kawai and Sunada estimate valence measures for U.S. House candidates by adapting the Olley and Pakes (1996) production function control function approach to an election game, controlling for endogenous campaign spending and selection from challenger entry. Incumbents have about 3.5 percentage-point higher valence than challengers on average, accounting for about 21 percent of the incumbency advantage. Econometrica Vol. 93(2), 2025, paywalled. Eight core results with source locators, the dynamic game model equations, and the structural estimation strategy; LLM-distilled, not human-verified.
# Tags: paper-summary, political-economy, elections, incumbency-advantage, candidate-valence
==============================================================================

**What this is.** This page is a distilled skeleton of Kawai and Sunada (2025). Read the original at https://doi.org/10.3982/ECTA20496 to replicate or extend the results.

## TL;DR

The paper develops a structural method for estimating candidate valence (unobservable quality that affects vote share) from data on vote shares, campaign spending, savings, and strategic entry in U.S. House elections from 1984 to 2008. Adapting the control function approach of Olley and Pakes (1996) from production function estimation, the authors embed vote shares in a dynamic election game and use the injectivity of uncontested incumbents' policy functions to construct a control function for incumbent valence. Challenger valence is identified from the first-order conditions of the candidates' spending and saving decisions, treating them as a GMM moment system. Results show incumbents have roughly 3.5 percentage-point higher valence than challengers on average, with challengers exhibiting wider dispersion (IQR 9.2 pp vs. 3.8 pp for incumbents). Equalizing challenger and incumbent valence increases the average challenger winning probability from 6.5% to 12.1%. A regression discontinuity decomposition following Lee (2008) finds that the total 10.2 pp incumbency advantage in vote share decomposes into about 21% from valence, 43% from spending, and 19% from policy positions. The valence measure is validated against the observable seriousness dummies of Maestas and Rugeley (2008), showing positive and statistically significant Spearman rank correlations (Table VI, p. 491).

## Core results

| # | Result | Locator | Magnitude as reported |
|---|--------|---------|----------------------|
| R1 | Average valence of incumbents exceeds challengers | Fig. 4, §5.3, p. 488 | ~3.5 pp vote share advantage; incumbents 0.035 units higher |
| R2 | Dispersion of valence measures | Fig. 4, §5.3, pp. 488-489 | IQR incumbents 3.8 pp; IQR challengers 9.2 pp |
| R3 | Counterfactual challenger winning probability (equal valence, no spending adj.) | §7, Fig. 8, p. 493 | Rises from 6.5% to 12.1% |
| R4 | Counterfactual challenger winning probability (equal valence, with spending adj.) | §7, Fig. 8, p. 493 | 11.0% vs. baseline 6.5% |
| R5 | Total incumbency advantage (RD estimate) | Table VII col. (i), p. 495 | 10.2 pp (SE 0.012) |
| R6 | Valence component of incumbency advantage | Table VII cols. (ii)-(iii), p. 495-496 | 2.1 pp combined (~21% of total) |
| R7 | Spending component of incumbency advantage | Table VII cols. (iv)-(v), p. 496 | 4.3 pp (~43% of total) |
| R8 | Policy position component of incumbency advantage | Table VII cols. (vi)-(vii), p. 497 | 1.9 pp (~19% of total) |

**Overall (paper's conclusion).** Incumbents hold a persistent and quantitatively meaningful valence advantage over challengers that extends beyond spending capacity and more centrist policy positions. The 10.2 pp total incumbency advantage decomposes into roughly 21% from valence, 43% from spending, and 19% from policy positions, implying that spending-focused interventions such as subsidizing challengers' campaigns will be only partially effective. Open-seat candidates' valence distribution resembles that of incumbents in its upper tail, but has a larger mass of low-valence candidates.

## Theory / model

The paper embeds vote shares in a dynamic Markov Perfect Equilibrium model of U.S. House elections (solution concept: Maskin and Tirole (1988)). In each period $$t = 1, 2, \ldots, \infty$$, a stage game is either an election with an incumbent or an open-seat election. State variables for contested elections are $$\mathbf{s} = \{q_I, w_I, \text{ten}_I, p_I, pt, dn \times D_I, ue \times D_I \times D_P, \mathbf{1}\{\text{First Term}\}, \mathbf{1}\{\text{Midterm}\}\}$$.

**Vote share equation.** The incumbent's vote share is (p. 467, eq. 1):

$$
\text{vote}_I = \beta_I \ln d_I + \beta_C \ln d_C + \beta_P(p_I - p^*)^2 - \beta_P(p_C - p^*)^2 + \beta_{\text{ten}} \text{ten}_I + \beta_X X + q_I - q_C + \varepsilon, \tag{1}
$$

where $$d_I, d_C$$ are spending (disbursements) of the incumbent and challenger; $$p_I, p_C$$ are their policy positions; $$p^*$$ is the district's ideal policy position; $$\text{ten}_I$$ is the incumbent's tenure; $$X$$ is a vector of district controls; $$q_I, q_C$$ are the unobservable valence terms (candidate fixed effects in vote share units); and $$\varepsilon \sim \mathcal{N}(0.5, \sigma_\varepsilon^2)$$. The winning probability follows (p. 468, eq. 2):

$$
\Pr(\text{vote}_I > 0.5) = \Phi\!\left(\frac{1}{\sigma_\varepsilon}\!\left(\beta_I \ln d_I + \beta_C \ln d_C + \beta_P(p_I - p^*)^2 - \beta_P(p_C - p^*)^2 + \beta_{\text{ten}} \text{ten}_I + \beta_X X + q_I - q_C\right)\right). \tag{2}
$$

**Incumbent's dynamic program.** Facing a challenger with known valence $$q_C$$ and policy $$p_C$$, the incumbent chooses spending $$d_I$$ and savings $$w'_I$$ to solve (p. 468, eq. 3):

$$
v_I(\mathbf{s}, q_C, p_C) = \max_{d_I \geq 0,\, w'_I \geq 0} u_I + \delta \Pr(\text{vote}_I > 0.5)\, \mathbb{E}_{s'|s}\!\left[(1 - \lambda(s')) V_I(s')\right], \tag{3}
$$

where $$u_I = B \cdot \Pr(\text{vote}_I > 0.5) - C_I(w'_I + d_I - w_I;\, q_I) + H_I(d_I)$$. Here $$B = 1$$ is the normalized utility from winning, $$C_I(\cdot; q_I)$$ is the fund-raising cost (strictly decreasing in $$q_I$$, so higher-valence incumbents face lower marginal cost), $$H_I(\cdot)$$ is the consumption value of spending, $$\delta = 0.9$$, and $$\lambda(s')$$ is the endogenous retirement probability. The ex ante value function before the challenger's entry decision realizes is (p. 470, eq. 4):

$$
V_I(\mathbf{s}) = (1 - P_e(\mathbf{s}))\, \bar{v}_I(\mathbf{s}) + P_e(\mathbf{s}) \int_{q_C, p_C} v_I(\mathbf{s}, q_C, p_C)\, dG_{q_C, p_C}(q_C, p_C \mid \mathbf{s}), \tag{4}
$$

where $$P_e(\mathbf{s})$$ is the equilibrium entry probability and $$G_{q_C, p_C}(\cdot \mid \mathbf{s})$$ is the joint distribution of the entering challenger's valence and policy.

**Challenger's problem.** The general election challenger solves (p. 470, eq. 5):

$$
v_C(\mathbf{s}, q_C, p_C) = \max_{d_C \geq 0,\, w'_C \geq 0} B \cdot \Pr(\text{vote}_I < 0.5) - C_C(w'_C + d_C, q_C) + H_C(d_C) + \delta \Pr(\text{vote}_I < 0.5)\, \mathbb{E}_{s'|s}\!\left[(1-\lambda(s'))V_I(s')\right]. \tag{5}
$$

A potential challenger enters if and only if $$q_C > \bar{q}_C(\mathbf{s}, p_C)$$, where the entry threshold is defined implicitly by $$p(\mathbf{s}, \cdot, p_C)\, v_C(\mathbf{s}, \cdot, p_C) = \kappa$$ (entry cost $$\kappa$$; p. 471). Challengers with higher valence are more likely to enter.

**Two propositions that drive identification.** Proposition 1 (Injectivity, p. 473): If the marginal cost of fund-raising $$\frac{\partial}{\partial x} \tilde{C}_I(x, q_I)$$ is strictly decreasing in $$q_I$$, then the policy functions of uncontested incumbents $$\{d_I(\mathbf{s}), w'_I(\mathbf{s})\}$$ are one-to-one from $$q_I$$ to $$(d_I, w'_I)$$, holding other state variables fixed. This mirrors the invertibility of the investment function in Olley and Pakes (1996) and allows expressing $$q_I = q_I(\bar{\mathbf{s}}_U)$$ as a function of observables. Proposition 2 (Sufficient statistic, p. 473): $$m(\mathbf{s}) \equiv \{P_e(\mathbf{s}),\, F_{p_C}(p_C \mid \mathbf{s},\, \chi = 1)\}$$ is a sufficient statistic for the distribution of the general-election challenger's valence $$G_{q_C}(\cdot \mid \mathbf{s})$$. This parallels the propensity score in Olley and Pakes (1996) and allows conditioning out the challenger selection bias.

## Method

The four-step estimation adapts the control function strategy of Olley and Pakes (1996) to handle two unobservables ($$q_I$$ and $$q_C$$) and a dynamic game structure.

**Step 1: Vote share equation and incumbent valence.** By Proposition 1, substitute $$q_I = q_I(\bar{\mathbf{s}}_U)$$ into the vote share equation. Decompose $$q_C = \mathbb{E}[q_C \mid \mathbf{s}] + (q_C - \mathbb{E}[q_C \mid \mathbf{s}])$$ and use Proposition 2 to write $$\mathbb{E}[q_C \mid \mathbf{s}] = g(m(\mathbf{s}))$$. The endogeneity-corrected vote share equation becomes (p. 476, eq. 1'):

$$
\text{vote}_I = \beta_I \mathbb{E}[\ln d_I \mid \mathbf{s}] + \beta_C \mathbb{E}[\ln d_C \mid \mathbf{s}] + \beta_P(p_I^2 - \mathbb{E}[p_C^2 \mid \mathbf{s}]) - 2\beta_P p^*(p_I - \mathbb{E}[p_C \mid \mathbf{s}]) + \beta_{\text{ten}} \text{ten}_I + \beta_X X + q_I(\bar{\mathbf{s}}_U) - g(m(\mathbf{s})) + \epsilon, \tag{1'}
$$

where $$\epsilon = \text{vote}_I - \mathbb{E}[\text{vote}_I \mid \mathbf{s}]$$ is orthogonal to $$\mathbf{s}$$ by construction, so $$\mathbb{E}[\ln d_I \mid \mathbf{s}]$$ and $$\mathbb{E}[\ln d_C \mid \mathbf{s}]$$ are valid instruments. The coefficients $$\beta_I$$ and $$\beta_C$$ are identified by variation in $$\mathbf{s}$$ holding $$m(\mathbf{s})$$ constant. The sieve minimum distance estimator of Ai and Chen (2003) is applied to the semiparametric equation.

**Step 2: Challenger valence and structural parameters.** The first-order conditions of the contested incumbent's spending and saving decisions jointly identify challenger valence $$q_C$$ and structural parameters $$\theta = (c_1, c_2, \eta_I, \eta_C, \alpha, \gamma)$$. The spending and saving FOCs are (p. 478, eqs. 12-13):

$$
\frac{\partial C_I}{\partial d_I}(w'_I + d_I - w_I,\, q_I;\, \theta) = \frac{\beta_I}{\sigma_\varepsilon d_I}\, \phi(K)\, \bigl(B + \delta\, \mathbb{E}_{s'|s}[V_I(s')]\bigr) + \frac{\partial H_I}{\partial d_I}(d_I;\, \theta), \tag{12}
$$

$$
\frac{\partial C_I}{\partial w'_I}(w'_I + d_I - w_I,\, q_I;\, \theta) = \delta\, \Phi(K)\, \frac{\partial}{\partial w'_I}\, \mathbb{E}_{s'|s}[V_I(s')], \tag{13}
$$

where $$K$$ is the standardized expected vote margin (p. 478, eq. 14):

$$
K = \frac{1}{\sigma_\varepsilon}\bigl(\beta_I \ln d_I + \beta_C \ln d_C + \beta_P(p_I - p^*)^2 - \beta_P(p_C - p^*)^2 + \beta_{\text{ten}} \text{ten}_I + \beta_X X + q_I - q_C\bigr). \tag{14}
$$

GMM treats the FOCs as moment conditions and identifies $$\theta$$ by requiring that the two expressions for $$K$$ obtained from eqs. (12) and (13) coincide at the true parameter values.

**Forward simulation of continuation values.** The continuation value $$\mathbb{E}_{s'|s}[V_I(s')]$$ and its derivative $$\frac{\partial}{\partial w'_I}\mathbb{E}_{s'|s}[V_I(s')]$$ are computed by forward simulation using the methods of Hotz, Miller, Sanders, and Smith (1994) and Bajari, Benkard, and Levin (2007), estimating the distribution of actions and outcomes nonparametrically without solving for an equilibrium at each candidate parameter value.

**Functional form specifications** (p. 484):

$$
\tilde{C}_I(fr_I;\, q_I) = c(q_I)(\ln fr_I)^2, \quad \tilde{H}_I(d_I) = \gamma_U \sqrt{\ln d_I},
$$

$$
C_I(fr_I;\, q_I) = \eta_I \times c(q_I)(\ln fr_I)^\alpha, \quad C_C(fr_C;\, q_C) = \eta_C \times c(q_C)(\ln fr_C)^\alpha, \quad H(d) = \gamma\sqrt{\ln d},
$$

where $$c(q) = c_1 + c_2 \exp(-q)$$ ensures $$c(\cdot)$$ is positive and strictly decreasing in $$q$$.

**Steps 3-4.** Open-seat election parameters (including $$\beta_O$$) are identified by analogous GMM from open-seat candidates' FOCs. Valence for incumbents who never appear in uncontested elections is recovered by solving all four FOCs jointly as a system of equations in $$(q_I, q_C)$$, stacked as GMM moments.

## Empirical specifications

**Vote share specification.** The full parameterization estimated in Section 5 is (p. 483):

$$
\text{vote}_I = \beta_I \ln d_I + \beta_C \ln d_C + \beta_P(p_I - p^*)^2 - \beta_P(p_C - p^*)^2 + \beta_{\text{ten}} \text{ten}_I + D_I(\beta_d + \beta_{dn}\, dn) + \beta_{ue}(ue \times D_I \times D_P) + \text{Election cycle FE} + q_I - q_C + \varepsilon,
$$

where $$p^* = \beta_{\text{ID},0} + \beta_{\text{ID},1}\, pt$$ is estimated as a linear function of the Republican partisanship index $$pt$$; $$dn$$ is log population density (interacted with incumbent party $$D_I$$) to capture differential urban vs. rural electoral strength; $$ue \times D_I \times D_P$$ captures retrospective voting through unemployment interacted with whether the incumbent is of the same party as the President; and election cycle FE include midterm, first-term President, and their interaction. Identification uses elections in which the incumbent has previously been uncontested, and requires that $$m(\mathbf{s})$$ varies across elections holding fixed $$\bar{\mathbf{s}}_U$$ (the control for $$q_I$$).

Key parameter estimates from the control function approach (Table IV, p. 486): $$\hat\beta_I = 0.039$$ (SE 0.020), $$\hat\beta_C = -0.039$$ (SE 0.011), $$\hat\beta_P = -0.031$$ (SE 0.021), $$\hat\sigma_\varepsilon = 0.069$$ (SE 0.003). Standard errors for $$\beta_I$$ and $$\beta_C$$ are from 500 bootstrap samples. A standard deviation increase in incumbent spending raises incumbent vote share by about 2.7 pp; the same for challenger spending decreases it by about 6.9 pp. OLS estimates of $$\beta_I$$ are negative and significant (Table IV, col. 2), reflecting omitted-variable bias from the positive correlation between challenger strength and incumbent spending.

**Counterfactual analysis (Section 7, p. 493, Figure 8).** To assess the role of valence differences, each challenger's $$q_C$$ is replaced by the corresponding percentile of the incumbent valence distribution. The baseline mean challenger winning probability is 6.5%. Equalizing valence without allowing spending to adjust raises this to 12.1%. Allowing candidates to adjust spending to their new equilibrium levels (using the estimated policy functions) yields 11.0%. The moderation comes primarily from increased incumbent spending (log spending increases by about 0.30 points, or roughly $144,600).

**Incumbency advantage decomposition (Section 8, p. 494-497, Table VII).** Following Lee (2008), the incumbency advantage is defined via the regression discontinuity limit (p. 494, eq. 15):

$$
\text{IA} = \lim_{\varepsilon \to +0} \mathbb{E}[\text{vote}_{\text{Dem},\, t+1} \mid \text{vote}_{\text{Dem},\, t} = 0.5 + \varepsilon] - \lim_{\varepsilon \to +0} \mathbb{E}[\text{vote}_{\text{Dem},\, t+1} \mid \text{vote}_{\text{Dem},\, t} = 0.5 - \varepsilon]. \tag{15}
$$

The same RD regression is estimated replacing the outcome (period $$t+1$$ vote share) with candidate valence, log spending, and policy position in turn. Using the bias-corrected RD estimator of Calonico, Cattaneo, and Titiunik (2014), the total incumbency advantage is 10.2 pp (SE 0.012, Table VII col. i, bandwidth 0.092). The valence component (combined Democratic and Republican RD estimates multiplied by the vote share effect) is 2.1 pp. The spending component (Democratic $$+0.526$$ and Republican $$-0.650$$ log spending RD estimates, Table VII cols. iv-v, converted via $$\hat\beta_I$$ and $$\hat\beta_C$$) is 4.3 pp. The policy component (Democratic $$+0.179$$ and Republican $$+0.169$$ policy position RD estimates, Table VII cols. vi-vii) is 1.9 pp. Sample for the RD: all election pairs $$(t, t+1)$$ in which neither period is uncontested (N = 2,320 per column).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---------|--------------|-----------|
| FEC campaign finance data (2011) | Spending, fund-raising, and savings for all U.S. House candidates, 1984-2008 | no page yet |
| CQ Press electoral database | Electoral outcomes and candidate characteristics | no page yet |
| U.S. Census Bureau (2015) | Congressional district demographics (population density) | [Census](/wiki/datasets/census/) |
| Bureau of Labor Statistics (BLS, 2011) | Local area unemployment statistics (retrospective voting controls) | [BLS](/wiki/datasets/bls/) |
| POLIDATA (2015) | Presidential vote shares by district (source for partisanship index) | no page yet |
| Bonica (2023) DIME database | Incumbent and challenger policy positions (ideology scores from campaign contributions) | no page yet |

Sample scope: 3,065 contested elections with incumbents, 787 uncontested elections, 445 open-seat elections, all from the 1984-2008 U.S. House election cycle (biennial). Dollar values normalized to 1984 dollars and reported in units of $1,000. Dropped observations include elections in Louisiana and Texas 1996 (affected by Supreme Court redistricting rulings), elections involving major scandals, and elections in which candidates' spending or savings are near zero, or a policy position is missing.

## When to read the full paper

Read Kawai and Sunada (2025) to (i) replicate or extend the structural valence estimation procedure, in particular the forward simulation of continuation values and the GMM system from first-order conditions (Supplemental Appendices 10.5-10.6); (ii) examine the model fit in detail (Figures 6-7, p. 491-492), which compares predicted vs. realized vote shares and predicted vs. actual candidate actions; (iii) study the full incumbency advantage decomposition with binned scatter plots of valence, spending, and policy position at the 50% vote share threshold (Figures 10-13, pp. 495-498); or (iv) see the cross-validation against the Maestas and Rugeley (2008) seriousness measure (Table VI, p. 491). The replication code and non-restricted data are available at https://doi.org/10.5281/zenodo.14172367; restricted data (CQ Press, POLIDATA) are subject to an exemption and were shared separately with the journal.

## Attribution and rights

Kawai, Kei, and Takeaki Sunada. "Estimating Candidate Valence." *Econometrica*, Vol. 93, No. 2 (March, 2025), pp. 463-501. DOI: 10.3982/ECTA20496. © 2025 The Econometric Society. All rights reserved; no Creative Commons license; standard copyright. This page is an extract-only distillation: it reproduces a structured summary of the paper's methods, equations, and findings for research reference under fair-use conventions for scholarly excerpts. LLM-distilled, not human-verified; results have not been independently reproduced.


==============================================================================
# Auctioning Control and Cash-Flow Rights Separately: Liu & Bernhardt (2025)
# https://instituteforautomatedresearch.org/wiki/papers/econometrica/2025/liu-bernhardt-auctioning-control-cash-flow-2025/
# A seller increases expected revenue by sometimes allocating control and cash-flow rights to different bidders: separation reduces a controller's information rent because project payoffs are most sensitive to his signal when he runs the project. Two ex post incentive-compatible separation mechanisms always strictly dominate no-separation English auctions in expected revenue for any minimum stake requirement. Econometrica 2025, CC BY 4.0. Six core results with source locators, the model equations, and the mechanism designs.
# Tags: paper-summary, mechanism-design, auction-theory, information-economics, open-access, cc-by, peer-reviewed, unreplicated
==============================================================================

**What this is.** Core results, the model, and the mechanism designs from Liu and Bernhardt (2025), distilled from the source PDF. To replicate or extend, read the original at [https://doi.org/10.3982/ecta21343](https://doi.org/10.3982/ecta21343).

## TL;DR

The paper studies a classical auction setting where a seller sells a single asset or project to risk-neutral bidders who privately observe signals about the project's future cash flows. The key departure from the standard literature is allowing the seller to allocate control rights and cash-flow rights to different bidders. Because project payoffs are more sensitive to a bidder's signal when he controls the project than when a rival does (single-crossing condition), awarding cash flows to a bidder who does not control the project reduces that bidder's informational advantage, lowering the seller's cost of rent extraction. The paper proposes two families of separation mechanisms (Mechanism A with inefficient control allocation and Mechanism B with efficient control allocation) and proves that each can always be designed to yield strictly higher expected seller revenues than any ex post incentive-compatible no-separation English auction. The gains from separation are largest when the two highest signals are close: the cost of potentially misallocating control is then small, but the benefit of reduced information rents remains positive.

## Core results

Locators refer to the source PDF (Econometrica 93(3), pp. 859–889).

| # | Result | Locator | Statement |
|---|---|---|---|
| R1 | Proposition 1: In the symmetric equilibrium of the two-stage separation auction, bidding strategies are identical to those of a no-separation English auction; the first-stage winner acquires control if and only if $$\Delta(t_1, t_2;\, t_3,\ldots,t_n) \geq p_{\text{extra}}$$; and the equilibrium is ex post incentive compatible | Proposition 1, p. 867 | Equilibrium characterization; ex post IC for any symmetric $$p_{\text{extra}}(\cdot)$$ |
| R2 | Result 1: There exists $$p^* > 0$$ such that for all $$p_{\text{extra}} \in (0, p^*)$$, the two-stage separation auction generates strictly higher expected revenues than the no-separation English auction | Result 1, p. 868 | Strict revenue dominance for sufficiently small constant price offer |
| R3 | Result 2: The revenue-maximizing price offer is the monopoly price conditional on the highest signal being at least $$t_2$$; in the two-bidder i.i.d. uniform-[1,2] linear example: $$p_{\text{extra}}^{\text{optimal}} = (2-t_s)/6$$, accepted with probability 0.5, yielding an expected revenue gain of $$1/18$$ above the no-separation English auction | Result 2, p. 868-869; eq. (17) | Optimal price offer formula; $$1/18$$ revenue gain in the tractable example |
| R4 | Proposition 2: Mechanism A (second-highest bidder receives control plus share $$q$$ of cash flows; highest bidder receives share $$1-q$$) is globally ex post incentive compatible for any separation function $$S$$ whenever $$\rho_{\min} \geq q/(1-q)$$, where $$\rho_{\min}$$ measures the minimum sensitivity of cash flows to the non-controller's signal | Proposition 2, p. 872 | IC condition for Mechanism A (inefficient control allocation) |
| R5 | Proposition 3: Mechanism B (highest bidder receives control plus share $$q$$ of cash flows; second-highest bidder receives share $$1-q$$) is globally ex post incentive compatible for any separation function $$S$$ whenever $$\rho_{\max} \leq q/(1-q)$$; a stake $$q \geq 0.5$$ always satisfies this condition | Proposition 3, p. 876 | IC condition for Mechanism B (efficient control allocation) |
| R6 | Proposition 4: Given the respective IC conditions for Mechanism A or B, separation functions $$S$$ exist for which each mechanism generates strictly higher expected seller revenues than any ex post IC no-separation English auction, for any minimum stake requirement $$q \in [0,1)$$ and any number $$n \geq 2$$ of bidders with weakly affiliated signals | Proposition 4, p. 877 | Revenue dominance in the general setting |

**Overall (paper's conclusion).** The mechanism design literature has focused on settings where the bidder who controls the project receives all cash flows. This paper shows a seller can always do better by designing mechanisms that sometimes allocate control to one bidder and cash flows to another. Separation lowers a controller's information rent because a project's payoff is most sensitive to his signal when he runs it; when signals are close, the cost of assigning control inefficiently is small but the gain from reduced sensitivity is strictly positive, so separation strictly raises revenue over any ex post IC no-separation mechanism.

## Theory / model

There are $$n > 1$$ ex ante identical, risk-neutral bidders for a single asset or project. Each bidder $$i$$ receives a private signal $$t_i \in [\underline{t}, \bar{t}]$$. The bidders and the seller are risk-neutral. Signals are weakly affiliated with a joint density $$f(\mathbf{t})$$ that is symmetric and uniformly continuous and strictly positive on $$[\underline{t}, \bar{t}]^n$$.

**Valuations.** Expected future cash flows from the project under bidder $$i$$'s control are (eq. 1, p. 863):

$$
v_i(t_1,\ldots,t_n) = u(t_i;\,\mathbf{t}_{-i}), \qquad \text{for all } i,
\tag{1}
$$

where $$u$$ is the same for each bidder and symmetric in its last $$n-1$$ arguments. Valuations are interdependent: they depend on all bidders' signals, not just the controller's.

**Single-crossing condition.** A bidder's signal has greater influence on cash flows when he controls the project than when another bidder does (eq. 2, p. 863):

$$
\frac{\partial v_i}{\partial t_i}(\mathbf{t}) \geq \frac{\partial v_j}{\partial t_i}(\mathbf{t}), \qquad \text{for all } i \text{ and all } j \neq i.
\tag{2}
$$

By symmetry, this reduces to $$u_1(t_1;\, t_2,\ldots,t_n) \geq u_2(t_2;\, t_1,\ldots,t_n)$$ (eq. 3, p. 863), where $$u_i$$ denotes the derivative of $$u$$ with respect to its $$i$$th argument. The paper also imposes the strict inequality when $$t_1 = t_2$$ (eq. 4, p. 863), which is the key assumption enabling separation gains. A leading specialization used throughout the examples is the linear valuation function (eq. 5, p. 863):

$$
u(t_i,\,\mathbf{t}_{-i}) = A_n\!\left(t_i + \rho \sum_{j \neq i} t_j\right), \qquad A_n \equiv \frac{1}{1+(n-1)\rho},
\tag{5}
$$

where $$\rho \in (0,1)$$ measures the degree of common values. Higher $$\rho$$ means control assignment matters more for realized cash flows, so the single-crossing difference $$u_1 - u_2$$ is larger.

**Mechanism design variables.** The paper allows direct-revelation mechanisms that allocate control and cash-flow rights separately. Let $$R_j(\mathbf{t}) \in [0,1]$$ be the probability bidder $$j$$ is assigned control and $$Q_{ji}(\mathbf{t}) \in [0,1]$$ be the share of cash flows that bidder $$i$$ receives when bidder $$j$$ controls. Feasibility requires (pp. 863-864):

$$
\sum_j R_j(\mathbf{t}) \leq 1, \qquad \sum_i Q_{ji}(\mathbf{t}) = 1 \;\text{ for all } j, \qquad Q_{jj}(\mathbf{t}) \geq q,
\tag{6--8}
$$

where $$q \in [0,1)$$ is the minimum cash-flow stake the controller must retain. Standard no-separation mechanisms correspond to $$Q_{jj}(\mathbf{t}) = 1$$ and $$Q_{ji}(\mathbf{t}) = 0$$ for all $$i \neq j$$.

**Bidder payoffs and seller revenue.** Bidder $$i$$'s expected profit when his true type is $$t_i$$ but he reports $$t_i'$$ is (eq. 9, p. 864):

$$
U_i(t_i,\, t_i';\,\mathbf{t}_{-i}) \equiv \sum_j R_j(t_i';\mathbf{t}_{-i})\, Q_{ji}(t_i';\mathbf{t}_{-i})\, v_j(\mathbf{t}) - M_i(t_i';\mathbf{t}_{-i}).
\tag{9}
$$

The first term is the expected value of cash flows awarded to bidder $$i$$ (which may be generated under another bidder $$j$$'s control). The seller's expected revenue is (eq. 13, p. 865):

$$
\pi_s = \sum_{i=1}^n \int M_i(\mathbf{t})\, f(\mathbf{t})\, d\mathbf{t}.
\tag{13}
$$

**Intuition from the envelope theorem.** Applying the envelope theorem to the equilibrium profit yields (eq. 14, p. 865):

$$
\frac{d\tilde{U}_i(t_i,t_i)}{dt_i} = \int_{\Omega_{n-1}} \sum_j R_j(\mathbf{t})\, Q_{ji}(\mathbf{t})\, \frac{\partial v_j(\mathbf{t})}{\partial t_i}\, f_{-i}(\mathbf{t}_{-i}|t_i)\, d\mathbf{t}_{-i} + \text{(correlation term)}.
\tag{14}
$$

In no-separation mechanisms, bidder $$i$$ receives cash flows only when he controls, so the relevant sensitivity is $$\partial v_i/\partial t_i$$ (own influence, higher by single-crossing). With separation, bidder $$i$$ may receive cash flows $$Q_{ji}$$ when a rival $$j$$ controls, so the sensitivity is $$\partial v_j/\partial t_i$$ (rival's influence, lower by single-crossing). This reduced sensitivity lowers information rents and raises seller revenue. The gain is zero when signals differ a lot (separation is costly), but strictly positive when signals are close.

**Efficiency gain.** Define the efficiency gain from assigning control to the higher bidder rather than the lower (p. 867):

$$
\Delta(t_1, t_2;\, t_3,\ldots,t_n) \equiv u(t_1;\, t_2,t_3,\ldots,t_n) - u(t_2;\, t_1,t_3,\ldots,t_n).
$$

This gain is nonnegative when $$t_1 \geq t_2$$ and weakly increases in $$t_1$$.

**Relationship to prior work.** The analysis builds on the optimal auction design of Myerson (1981), which considers no-separation mechanisms where the highest bidder always receives both control and cash flows. Mezzetti (2003) studies two-stage mechanisms with interdependent valuations but focuses on implementing efficiency; this paper focuses on revenue. Ekmekci, Kos, and Vohra (2016) consider a related setting where a single buyer splits cash flows with the seller; the paper instead distributes cash flows among multiple bidders, which is the channel for rent reduction that the separation framework adds. Full surplus extraction is achievable with correlated signals per Cremer and McLean (1988), but requires large side bets that create large regrets. Separation yields revenue gains without exploiting correlation and applies to both i.i.d. and affiliated signals.

## Method

The paper proposes and analyzes two families of ex post incentive-compatible separation mechanisms. Both build on `mechanism-design` principles and extend the English auction.

**Two-stage separation auction (Definition 1, p. 867; $$q = 0$$ case).** The first stage is a standard English ascending auction. Losers exit at prices revealing their signals; when the next-to-last bidder exits, the seller offers the first-stage winner a second-stage choice: (a) accept cash flows with control going to the second-highest bidder, paying only the exit price, or (b) pay an additional fee $$p_{\text{extra}}(\cdot) \geq 0$$ to acquire both control and cash flows. The first-stage bidding strategy has the same form as in the no-separation English auction (eq. 15, p. 866):

$$
\beta^k(t_i,\, t_{k+1},\ldots,t_N) = u(t_i;\, t_i,\ldots,t_i,\, t_{k+1},\ldots,t_N),
\tag{15}
$$

the expected cash flows when all $$k$$ active bidders have signal $$t_i$$ and the revealed losing types are $$t_{k+1},\ldots,t_N$$. Proposition 1 shows the winner acquires control if and only if $$\Delta(t_1, t_2;\, t_3,\ldots,t_n) \geq p_{\text{extra}}$$ and this equilibrium satisfies ex post incentive compatibility (the Bergemann and Morris (2008) criterion: no bidder regrets his strategy after observing all signals). Revenue-maximizing price offer (Result 2, p. 869):

$$
p_{\text{extra}}^{\text{optimal}}(t_2,\ldots,t_n) = \Delta\!\left(t^{\text{opt}},\, t_2;\, t_3,\ldots,t_n\right),
$$

where $$t^{\text{opt}} \equiv \arg\max_t \Delta(t, t_2;\, t_3,\ldots,t_n) \int_t^{\bar{t}} f_1(x|\mathbf{t}_{-1})\, dx$$. In the two-bidder i.i.d. uniform-$$[1,2]$$ linear example with $$v_i = \frac{2}{3}t_i + \frac{1}{3}t_{-i}$$, the optimal price offer is (eq. 17, p. 869):

$$
p_{\text{extra}}^{\text{optimal}} = \frac{2 - t_s}{6},
\tag{17}
$$

accepted with probability 0.5, yielding revenue gain $$1/18$$ above the no-separation English auction, despite an expected social welfare loss of $$1/36$$ from occasionally assigning control inefficiently.

**Separation functions (Definitions 2-3, pp. 870-871).** For the general $$q > 0$$ case, both Mechanism A and B are parameterized by a "separation function" $$S(s_1,\ldots,s_{n-1})$$: a symmetric function of the $$n-1$$ reported signals weakly increasing in the highest report $$s_h$$. Bidder $$i$$ with the highest report receives all rights when $$t_i' \geq S(\mathbf{t}_{-i}')$$; otherwise separation occurs. The "quasi-inverse" $$S^{QI}(\mathbf{t}_{-i}')$$ gives the threshold below which bidder $$i$$'s report is low enough to receive neither control nor cash flows.

**Mechanism A: inefficient splitting (Definition 4, p. 871).** When the highest report $$t_1' \geq S(\mathbf{t}_{-1}')$$, bidder 1 receives control and all cash flows and pays (eq. 18, p. 871):

$$
M_1 = u(S(\mathbf{t}_{-1}');\, t_2',\ldots,t_n') - (1-q)\,u(t_2';\, S(\mathbf{t}_{-1}'),\ldots,t_n') + (1-2q)\,u(t_2';\, t_2',\ldots,t_n') + q\,u(S^{QI}(\mathbf{t}_{-1}');\, t_2',\ldots,t_n').
\tag{18}
$$

When $$t_1' < S(\mathbf{t}_{-1}')$$, bidder 2 receives control and share $$q$$ of cash flows; bidder 1 receives share $$1-q$$ and pays (eqs. 19-20):

$$
M_1 = (1-2q)\,u(t_2';\, t_2',\ldots,t_n') + q\,u(S^{QI}(\mathbf{t}_{-1}');\, t_2',\ldots,t_n'),
\tag{19}
$$

$$
M_2 = q\,u(S^{QI}(\mathbf{t}_{-2}');\, t_1', t_3',\ldots,t_n').
\tag{20}
$$

Mechanism A is globally ex post IC if the single-crossing condition holds in the weighted form (Proposition 2, p. 872):

$$
\rho_{\min} \equiv \min_{\mathbf{t}} \frac{\partial v_2(\mathbf{t})/\partial t_1}{\partial v_1(\mathbf{t})/\partial t_1} \geq \frac{q}{1-q}.
$$

For linear valuations $$\rho_{\min} = \rho$$, so the condition is $$\rho \geq q/(1-q)$$: when common values are high (large $$\rho$$) and the minimum stake $$q$$ is small, Mechanism A applies.

**Mechanism B: efficient splitting (Definition 6, p. 875).** When $$t_1' \geq S(\mathbf{t}_{-1}')$$, bidder 1 receives control and all cash flows and pays (eq. 22, p. 875):

$$
M_1 = (1-q)\,u(S(\mathbf{t}_{-1}');\, t_2',\ldots,t_n') + (2q-1)\,u(t_2';\, t_2',\ldots,t_n') + (1-q)\,u(t_2';\, S^{QI}(\mathbf{t}_{-1}'),\ldots,t_n').
\tag{22}
$$

When $$t_1' < S(\mathbf{t}_{-1}')$$, bidder 1 receives control and share $$q$$; bidder 2 receives share $$1-q$$ (eqs. 23-24):

$$
M_1 = (2q-1)\,u(t_2';\, t_2',\ldots,t_n') + (1-q)\,u(t_2';\, S^{QI}(\mathbf{t}_{-1}'),\ldots,t_n'),
\tag{23}
$$

$$
M_2 = (1-q)\,u(t_1';\, S^{QI}(\mathbf{t}_{-2}'),\, t_3',\ldots,t_n').
\tag{24}
$$

Mechanism B is globally ex post IC whenever (Proposition 3, p. 876):

$$
\rho_{\max} \equiv \max_{\mathbf{t}} \frac{\partial v_2(\mathbf{t})/\partial t_1}{\partial v_1(\mathbf{t})/\partial t_1} \leq \frac{q}{1-q}.
$$

A stake $$q \geq 0.5$$ always satisfies $$\rho_{\max} \leq q/(1-q)$$, so Mechanism B can be designed to satisfy any minimum stake requirement and strictly dominate no-separation auctions (Proposition 4).

## Empirical specifications

The paper is a pure theory contribution; all results follow from formal proofs. The primary technique for revenue dominance is a "delta-separation" construction (Appendix, pp. 881-888): for any target signal $$s^*$$, define a separation function $$S_\delta$$ that induces separation only when the highest $$n-1$$ signals are within a small $$\delta$$-interval around $$s^*$$. Expected revenue difference $$E[D]$$ between the separation and no-separation mechanisms decomposes by the law of iterated expectations into two cases:

- **Case 1** (both highest signals in the $$\delta$$-interval): probability shrinks at rate $$\delta^2$$, and the revenue deficit per realization is bounded above by a term linear in $$\delta$$, so $$E[D|\text{Case 1}] \cdot \text{prob}$$ goes to zero at rate $$\delta^3$$.
- **Case 2** (highest signal exceeds the interval, lower signal inside): probability shrinks at rate $$\delta$$, and the revenue surplus per realization is bounded below by a term proportional to $$\omega(1-q) > 0$$ (from the strict single-crossing inequality), so $$E[D|\text{Case 2}] \cdot \text{prob}$$ goes to zero at rate $$\delta^2$$.

For $$\delta$$ small, Case 2 dominates Case 1 and the total expected revenue difference is strictly positive. The argument applies to all $$n \geq 2$$ bidders, any weakly affiliated signal distributions, and any continuous valuation functions satisfying the single-crossing condition, with or without signal correlation.

## Datasets used

This paper is a pure theory contribution with no empirical data.

| Dataset | Role in paper | Wiki page |
|---|---|---|
| None | Theoretical model only | N/A |

## When to read the full paper

Read the source at [https://doi.org/10.3982/ecta21343](https://doi.org/10.3982/ecta21343) if you are: designing auction mechanisms for assets where control and cash-flow rights can be split (venture capital exits, bankruptcy resolution, corporate takeovers); studying mechanism design with interdependent valuations and the role of the single-crossing condition in determining how rent-reducing separation is; extending Bergemann and Morris (2008) ex post IC requirements to settings with multi-dimensional allocation; or working on comparative statics of Mechanisms A vs. B with respect to the minimum stake $$q$$. Propositions 2 and 3 give the exact IC thresholds on $$\rho_{\min}$$ and $$\rho_{\max}$$; Proposition 4 and the Appendix contain the revenue-dominance proof and the delta-separation construction for arbitrary bidder counts and signal distributions.

## Attribution and rights

Source: peer-reviewed, *Econometrica* 93, no. 3 (May 2025): 859-889. This distillation was extracted by an LLM on 2026-06-26 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Liu, Tingjun, and Dan Bernhardt.
> "Auctioning Control and Cash-Flow Rights Separately."
> *Econometrica* 93, no. 3 (May 2025): 859-889.
> DOI: 10.3982/ecta21343. (c) 2025 The Authors. Econometrica published by John Wiley and Sons Ltd on behalf of The Econometric Society.
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Active Fund Management when ESG Matters: Avramov, Cheng & Tarelli (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jbf/2026/avramov-active-fund-management-esg-2026/
# Distilled: This paper develops and tests an equilibrium model of active fund management with ESG considerations, showing that heterogeneous fund ESG preferences intensify information acquisition across the ESG spectrum, improving price informativeness and lowering the cost of capital for green firms through a concave, amplified ESG-return relation. Journal of Banking and Finance vol. 182 (2026), CC BY-NC-ND 4.0. Six core results with source locators, datasets used, the model equations, and the method.
# Tags: paper-summary, esg, asset-pricing, mutual-funds, information-acquisition
==============================================================================

**What this is.** The paper's core results, the equilibrium model it builds on (multi-asset noisy rational expectations with ESG preferences), and the method (information acquisition optimality conditions and two Fama-MacBeth regression designs) with defining equations: enough to understand what it found and how, without reading all 16 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1016/j.jbankfin.2025.107597).

## TL;DR

Avramov, Cheng, and Tarelli develop a noisy rational-expectations equilibrium model of active fund management in which agents have heterogeneous ESG preferences and can acquire costly private signals about asset payoffs. Building on the information acquisition framework of Grossman and Stiglitz (1980), the model shows that in equilibrium ESG-perceptive fund managers intensify information acquisition for assets that deviate from ESG neutrality, especially green (high-ESG) assets, which broadens the scope of active management. The enhanced signal precision lowers the posterior variance of green asset payoffs, reducing their implied cost of equity capital (ICC) and making the ESG-ICC relation negative and concave. The paper relates to the evidence in Hartzmark and Sussman (2019) that sustainability ratings drive fund flows, and tests the equilibrium ESG-return predictions of Pastor, Stambaugh & Taylor (2021) by documenting a concave (not merely linear) ESG-ICC relation amplified by an information channel absent from prior theory. Applied to monthly data on U.S. equity mutual funds and common stocks from 2007 to 2021, the model predictions are confirmed: stocks held by funds with heterogeneous ESG preferences display higher price informativeness, green stocks held by green funds have significantly lower ICC than brown stocks, and green funds earn significantly positive abnormal returns when investing in green stocks.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Stock price informativeness increases with departure from green neutrality (ESGDev) and with fund ESG preference heterogeneity (ESGDisp) | Table 2, Models 2-4, p. 9 | β₂ on Log(M/A) × ESGDev = 0.039\*\*\* (t=5.51) in Model 2; 1-SD increase in ESGDev → 43.9% (h=1) higher price informativeness; 1-SD increase in ESGDisp → 71.4% (h=1); β₃ on Log(M/A) × ESGDisp = 0.098\*\*\* (h=1) |
| R2 | ESG-ICC relation is negative and concave: the difference in DGTW-adjusted ICC widens sharply from low-to-mid to mid-to-high ESG quintiles | Figure 1, p. 13; Table 3, p. 12 | DGTW-adj. ICC difference between ESG quintiles Q1 and Q3: 0.012%/month; between Q3 and Q5: 0.072%/month; all differences significant at 1% |
| R3 | When Green IO is high, green stocks display significantly lower ICC than brown stocks; the effect is absent at low Green IO | Table 3, Panel A, p. 12 | High Green IO group: green vs. brown ICC = -0.115\*\*\* (-0.066\*\*\* DGTW-adj.) per month; HML-R ICC spread across high/low Green-IO portfolios = -0.094% (-0.046%), significant |
| R4 | When Brown IO is low, green stocks display significantly lower ICC than brown stocks; at high Brown IO the spread is insignificant | Table 3, Panel B, p. 12 | Low Brown IO group: green vs. brown ICC = -0.138\*\*\* (-0.079\*\*\* DGTW-adj.) per month; ICC spread across high/low Brown-IO portfolios = 0.134% (0.077%), significant at 0.134% (0.077%) |
| R5 | Fama-MacBeth ICC regression confirms: Green IO and its interaction with high-ESG stock significantly reduce ICC | Table 4, Models 1-4, p. 14 | Green IO = -0.824\*\*\* (t=-7.84); High Green IO = -0.035\*\*\* (t=-9.09); High ESG × High Green IO = -0.020\*\*\* (t=-3.03); Low ESG × Brown IO = -0.352\*\*\* (t=-3.17); Low ESG × High Brown IO = -0.014\*\* (t=-2.24) |
| R6 | Green funds earn significantly positive abnormal returns when investing in green stocks; green stocks held by brown funds yield insignificant returns | Online Appendix Table A.2, referenced p. 13 | High-Green-IO, high-ESG stocks: CAPM alpha = -0.304%/month; HML-R (high vs. low ESG, high Green IO) = -0.588%/month (Panel A1); Low-Brown-IO, low-ESG stocks: CAPM alpha = +0.455%/month, outperform high-ESG stocks by 0.579%/month (Panel A3) |

**Overall (paper's conclusion).** ESG considerations play a central role in shaping mutual funds' information decisions, portfolio choices, and the cross-section of asset prices. Information acquisition driven by ESG motives not only provides capital to green firms at a lower cost but also improves overall financial market efficiency by incorporating more private information into equilibrium prices. The concave ESG-ICC relation and the asymmetric performance of green and brown funds in their preferred ESG domains provide corroborating evidence for the model's information channel.

## Theory / model

The economy has $$N$$ risky assets. For $$i \in \{1, \ldots, N-1\}$$, asset payoffs load on both idiosyncratic and an aggregate risk factor; the $$N$$th asset is a pure aggregate asset (p. 3, Eq. 1):

$$
\begin{cases}
f_i = \mu_i + b_i z_N + z_i, & i \in \{1, \ldots, N-1\} \\
f_N = \mu_N + z_N
\end{cases}
\tag{1}
$$

where $$\mu_i$$ is the expected payoff, $$b_i$$ is the asset's exposure to the aggregate factor $$z_N$$, and $$z_i \sim \mathcal{N}(0, \sigma_i)$$ is an idiosyncratic shock. Shocks are uncorrelated. The risk factor supply for asset $$i$$ is $$\bar{x}_i + x_i$$ where $$\bar{x}_i$$ is mean supply and $$x_i \sim \mathcal{N}(0, \sigma_N^2(1+b_i^2))$$ is a random component. Random supply introduces noise that prevents prices from fully revealing private signals.

Agents are indexed by $$j$$ on a continuum. Each agent-asset pair $$(i,j)$$ can acquire a private signal (p. 3, Eq. 2):

$$
\eta_{ij} = z_i + \varepsilon_{ij}, \qquad \varepsilon_{ij} \sim \mathcal{N}\!\left(0, S_{ij}^{-1}\right)
\tag{2}
$$

where $$S_{ij} \geq 0$$ is the signal precision chosen by agent $$j$$ at cost $$c_{ij}(S_{ij})$$, a continuous, increasing, convex function with $$c_{ij}(0)=0$$. Agents are heterogeneous in their cost functions (stock-picking skill) and in their ESG preference parameter $$\delta_j \geq 0$$ (nonnegative for all agents, strictly positive for ESG-perceptive agents).

In period 2, after signals are realized, agents choose portfolios to maximize mean-variance utility with an ESG preference term (p. 4, Eq. 3):

$$
U_{2j} = \text{E}_{2j}\!\left[W_j\right] - \frac{\rho}{2}\,\text{Var}_{2j}\!\left[W_j\right] + \delta_j G_j
\tag{3}
$$

where $$\rho > 0$$ is the common risk-aversion coefficient, $$W_j$$ is terminal wealth, and $$G_j = \sum_{i=1}^N q_{ij} g_i$$ is the portfolio ESG score ($$q_{ij}$$ = holding of asset $$i$$, $$g_i$$ = ESG score, mean-zero with $$g_N = 0$$). A higher $$\delta_j$$ implies stronger ESG-driven preferences; $$\delta_j = 0$$ yields standard mean-variance.

The key pricing implication follows from Proposition 4 (p. 5). The expected net payoffs are (Eqs. 9-10):

$$
\text{E}\!\left[f_N - p_N\right] = \rho\bar{\sigma}_N
\tag{9}
$$

$$
\text{E}\!\left[f_i - p_i\right] = b_i\,\text{E}\!\left[f_N - p_N\right] + \rho\bar{\sigma}_i - \bar{\delta}_i g_i
\tag{10}
$$

where $$\bar{\sigma}_i$$ is the cross-agent average posterior payoff variance (the inverse of $$\bar{\sigma}_i^{-1} = \sigma_i^{-1} + \bar{S}_i + \sigma_{p_i}^{-1}$$) and $$\bar{\delta}_i = \bar{\sigma}_i \int \hat{\sigma}_{ij}^{-1}\,\delta_{ij}\,dj$$ is the aggregate, posterior-precision-weighted ESG preference for asset $$i$$ (where $$\hat{\sigma}_{ij}^{-1} = \sigma_i^{-1} + S_{ij} + \sigma_{p_i}^{-1}$$ is agent $$j$$'s posterior precision for asset $$i$$; Proposition 1, p. 4). Equation (10) shows that the expected return on asset $$i$$ is reduced by $$\bar{\delta}_i g_i$$, all else equal: green assets ($$g_i > 0$$) have lower expected returns, and brown assets ($$g_i < 0$$) have higher expected returns. The negative ESG-ICC relation is therefore a direct equilibrium prediction, amplified by the information channel through $$\bar{\sigma}_i$$ (which falls when signals are more precise).

**Concavity.** For green assets the ESG preference motive (nonpecuniary benefit $$\delta_j g_i > 0$$) and the information-acquisition motive (more precise signals reduce posterior variance) both push expected returns down. For brown assets the two forces partially offset: the nonpecuniary motive lowers expected returns, but better signals increase them. Hence the negative ESG-return relation is more pronounced for green assets, making the curve concave.

## Method

In period 1, each agent chooses signal precision $$S_{ij}$$ for each asset to maximize expected utility (p. 4, Eq. 5). Proposition 2 characterizes the optimum (p. 5, Eq. 6):

$$
\hat{S}_{ij} = \max\!\left[0,\; s \;\middle|\; c'_{ij}(s) = \psi_{ij}\right]
\tag{6}
$$

The pre-cost marginal benefit of information for asset-agent pair $$(i,j)$$ is (p. 5, Eq. 7):

$$
\psi_{ij} = \frac{1}{2\rho}\!\left(\bar{\sigma}_i + \left(\rho^2\sigma_X + \bar{S}_i\right)\bar{\sigma}_i^2 + \left(\rho\bar{\sigma}_i + \left(\bar{\delta}_j - \bar{\delta}_i\right)g_i\right)^{\!2}\right)
\tag{7}
$$

where $$\bar{S}_i = \int S_{ij}\,dj$$ is the cross-agent average signal precision and $$\sigma_X$$ is the variance of the aggregate risk factor supply. The term $$(\bar{\delta}_j - \bar{\delta}_i)g_i$$ captures the ESG motive: funds whose ESG preference $$\bar{\delta}_j$$ is above (below) the aggregate $$\bar{\delta}_i$$ have a higher (lower) marginal benefit of acquiring information about asset $$i$$ when that asset's ESG score $$g_i$$ is nonzero.

Proposition 3 establishes that the cross-agent average signal precision increases with the absolute departure from green neutrality (p. 5, Eq. 8):

$$
\frac{\partial \bar{S}_i}{\partial |g_i|} = \xi_{Ai}\,\sigma_\delta\,|g_i| > 0
\tag{8}
$$

where $$\sigma_\delta$$ is the cross-agent dispersion in ESG preferences and $$\xi_{Ai} > 0$$ is a positive scalar. This proves that aggregated information acquisition rises for both green and brown assets as their ESG scores depart from zero, because the marginal benefit of information acquisition is quadratic in ESG preferences. As a consequence, the informational efficiency of asset prices (price informativeness) increases for stocks with more extreme ESG profiles and for stocks held by funds with more dispersed ESG preferences.

The equilibrium is solved by a fixed-point problem on $$\bar{S}_i$$: agents choose optimal signal precisions given aggregate precision, and aggregate precision is consistent with individual choices. The model builds on the information acquisition framework of Breugem and Buss (2019) for institutional investors, extending it to incorporate heterogeneous ESG preferences. The fund performance measure follows Kacperczyk, Van Nieuwerburgh & Veldkamp (2016): the expected excess net payoff (EENP) decomposes into an ESG-based portfolio tilt component and a skill (private signal precision) component, with both building on the `noisy-rational-expectations` and `fama-macbeth` primitives listed above.

## Empirical specifications

**Price informativeness (Eq. 14, Table 2, p. 9).** Price informativeness is measured following Bai et al. (2016) as the ability of the current market-to-book ratio to predict future earnings-to-assets. The monthly Fama and MacBeth (1973) regression tests the model predictions about ESGDev and ESGDisp:

$$
\frac{E_{i,y+h}}{A_{i,y}} = \alpha + \beta_1\log\!\frac{M_{i,y}}{A_{i,y}} + \beta_2\log\!\frac{M_{i,y}}{A_{i,y}} \times \text{ESGDev}_{i,y} + \beta_3\log\!\frac{M_{i,y}}{A_{i,y}} \times \text{ESGDisp}_{i,y}
+ \beta_4\,\text{ESGDev}_{i,y} + \beta_5\,\text{ESGDisp}_{i,y} + \beta_6\frac{E_{i,y}}{A_{i,y}} + c N_{i,y} + \varepsilon_{i,y+h}
\tag{14}
$$

where $$E_{i,y+h}/A_{i,y}$$ is earnings-before-interest-and-taxes over total assets for stock $$i$$ in year $$y+h$$, $$M_{i,y}/A_{i,y}$$ is the market-to-book ratio, $$\text{ESGDev}_{i,y}$$ is the absolute departure from green neutrality (from LASSO residual of MSCI ESG on 94 non-ESG characteristics), $$\text{ESGDisp}_{i,y}$$ is the stock-level dispersion in fund ESG preferences, and $$N_{i,y}$$ stacks all other stock-level controls. Standard errors follow Newey and West (1987). Forecasting horizons are $$h=1$$ year (Models 1-5) and $$h=5$$ years (Models 6-10).

**ESG-ICC regression (Eq. 15, Table 4, p. 14).** The implied cost of capital is estimated following Hou et al. (2012) and Pastor, Stambaugh & Taylor (2022). The main Fama and MacBeth (1973) regression is:

$$
\text{ICC}_{i,t} = \alpha + \beta_1\,\text{ESG}_{i,t-1} + \beta_2\,\text{IO}_{i,t-1} + \beta_3\,\text{ESG}_{i,t-1} \times \text{IO}_{i,t-1} + c N_{i,t-1} + \varepsilon_{i,t}
\tag{15}
$$

where $$\text{ICC}_{i,t}$$ is the monthly implied cost of capital for stock $$i$$, $$\text{ESG}_{i,t-1}$$ is a high/low ESG indicator (top/bottom quintile), $$\text{IO}_{i,t-1}$$ is a vector of fund-ownership indicators (Green IO, Brown IO, high/low variants), and $$N_{i,t-1}$$ stacks stock-level controls (Log(Size), Log(BM), ROE, I/A, 1M Return, 12M Return). Standard errors follow Newey and West (1987).

**Portfolio double sorts (Table 3).** At the end of each month $$t$$, stocks are first sorted into terciles by Green IO (or Brown IO) and then within each tercile into quintiles by ESG rating, yielding 15 ($$3 \times 5$$) portfolios. Value-weighted ICC is computed in month $$t+1$$ and rebalanced monthly. The HML-R spread (high ESG minus low ESG within each ownership tercile) measures the ESG-ICC relation. The HML-G spread (high minus low ownership within each ESG quintile) measures the effect of fund ESG preference alignment. ICCs are additionally adjusted for the CAPM market factor, the Fama-French six-factor model (FF6), and the characteristic-adjusted DGTW model.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CRSP common stocks (daily/monthly returns, prices, shares) | Stock-level returns, market cap, turnover, idiosyncratic volatility, short-term reversal | [WRDS](/wiki/commercial/wrds/) (licensed) |
| Compustat annual fundamentals | Book-to-market, profitability, investment, leverage, sales, tangibility, earnings | [WRDS](/wiki/commercial/wrds/) (licensed) |
| MSCI ESG Ratings (INDUSTRY_ADJUSTED_SCORE) | Monthly ESG scores, residualized against 94 non-ESG characteristics via LASSO to produce Stock ESG and Stock ESGDev | [KLD / MSCI ESG](/wiki/commercial/kld/) (licensed) |
| CRSP mutual fund database (via WRDS MFLINKS) | Monthly net-of-fee fund returns, TNAs, turnover, expense ratio, fund flows, multiple share classes consolidated | [CRSP Mutual Funds](/wiki/commercial/crsp-mutual-funds/) (licensed) |
| Thomson-Reuters 13F institutional holdings | Quarterly fund equity holdings, used to compute fund-level ESG preference and stock-level fund ownership (Green IO, Brown IO, ESGDisp) | [Thomson 13F](/wiki/commercial/thomson-13f/) (licensed) |
| I/B/E/S analyst forecasts | Analyst coverage and forecast dispersion as stock-level controls; earnings forecasts for ICC computation via Hou et al. (2012) | [I/B/E/S](/wiki/commercial/ibes/) (licensed) |

Sample: January 2007 to December 2021 (15 years, monthly). Full sample contains 4031 unique equity funds and 3422 unique stocks; average 1777 funds and 1374 stocks per month. Equity funds are restricted to those with TNA of at least $15 million, identified as active via CRSP objective codes.

## When to read the full paper

Read the [original](https://doi.org/10.1016/j.jbankfin.2025.107597) if you are: building or testing equilibrium models of ESG-driven information acquisition; studying the cross-section of expected returns under ESG-heterogeneous investors; examining the scope-of-active-management implications of sustainable investing; or replicating the ICC double-sort or price-informativeness Fama-MacBeth designs. The Online Appendix contains calibration exercises (Appendix B), model extensions for ESG rating disagreement and heterogeneous information costs (Appendix C), and additional empirical robustness checks (Appendix D). The locators above point to the exact tables and figures for each headline result.

## Attribution and rights

Source: peer-reviewed, *Journal of Banking and Finance* vol. 182 (2026). This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**.

The article is published under CC BY-NC-ND 4.0, which permits noncommercial redistribution with attribution in unmodified form but does not permit derivative works. This page is an IAR adaptation prepared for noncommercial research and educational purposes.

> **Citation.** Avramov, Doron, Si Cheng, and Andrea Tarelli.
> "Active fund management when ESG matters."
> *Journal of Banking and Finance* 182 (2026): 107597.
> DOI: [10.1016/j.jbankfin.2025.107597](https://doi.org/10.1016/j.jbankfin.2025.107597).
> © 2025 The Authors. Published by Elsevier B.V. under
> [CC BY-NC-ND 4.0](https://creativecommons.org/licenses/by-nc-nd/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Illegal Insider Trading Profitability and the Legal Environment: Batten, Liu & Sha (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jbf/2026/batten-illegal-insider-trading-profitability-2026/
# Distilled: Using 521 hand-collected adjudicated insider-trading cases from China (2006-2018), this paper finds that stronger provincial legal environments are associated with significantly higher per-trade abnormal returns, consistent with a risk-compensation mechanism in which stricter enforcement screens out low-return trades and leaves only high-return ones. Journal of Banking and Finance 185 (2026) 107609, CC BY 4.0. Six core results with source locators, datasets, and regression specifications. LLM-distilled, not human-verified.
# Tags: paper-summary, insider-trading, legal-environment, china, market-regulation
==============================================================================

**What this is.** The paper's core results, the hypotheses it tests (risk-compensation vs. deterrence), the regression specifications, and the datasets used: enough to know what it found and how, without reading all 17 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1016/j.jbankfin.2025.107609).

## TL;DR

This paper asks whether legal risk is priced in illegal insider trading in China. Using 521 adjudicated insider-trading cases hand-collected from court judgments and China Securities Regulatory Commission (CSRC) sanction documents (2006-2018), the authors measure each insider's buy-and-hold abnormal return (BHAR) and regress it on three proxies for provincial legal environment quality, combined with firm-level ex ante litigation risk. Across all specifications, stronger legal environments are associated with significantly higher per-trade profitability, consistent with a risk-compensation mechanism: stricter enforcement screens out low-return trades, leaving only those with sufficiently high expected gains to justify elevated detection risk. This counterintuitive pattern rules out the simple deterrence story (stricter enforcement reduces profits) in favor of a selective-deterrence story in which observed returns rise because low-return opportunities are filtered away. Firm-level litigation exposure (lnRISK), constructed following Kim and Skinner (2012), further raises BHAR, suggesting insiders incorporate both provincial and firm-specific legal risk into their trading decisions. The findings also rule out M&A rumors, financial literacy, political connections, and corporate governance quality as alternative channels (Sections 5.2-5.5), and survive selection-correction and a range of robustness tests. Kacperczyk and Pagnotta (2024) show a related legal-risk channel for legal insider trading in the US; this paper extends that logic to illegal trading in an emerging-market setting with rich within-country legal variation. The evidence aligns with the rational-crime model of Becker (1968): insiders behave as rational agents weighing expected gains against expected penalties. It also extends work by Ahern (2020) on the determinants of illegal insider trading profitability and by Sha et al. (2020) on the puzzle of low average returns in China's insider-trading cases.

## Core results

Magnitudes and significance are as reported; `\*\*\*`/`\*\*`/`\*` = 1%/5%/10%. Locators point to the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Provincial market development index (LAW^Institution) positively and significantly predicts BHAR from illegal insider trading; industry and year fixed effects included | Table 4, col 1, p. 7 | coeff 0.011\*\*\* (SE 0.002); adj. R² = 0.155; N = 478 |
| R2 | Provincial legal environment index (LAW^Environment) positively and significantly predicts BHAR, confirming the pattern across a second legal quality proxy | Table 4, col 2, p. 7 | coeff 0.013\*\*\* (SE 0.002); adj. R² = 0.162; N = 478 |
| R3 | Economic magnitude: 1-SD improvement in legal environment quality predicts 2.77-5.78 percentage-point increase in insider-trading abnormal returns (all three proxies) | Table 4, p. 7 | 2.77 pp (LAW^Institution) to 5.78 pp (LAW^Environment) per 1-SD increase |
| R4 | Ex ante litigation risk (lnRISK) positively predicts BHAR, incremental to provincial legal environment: both firm-level and provincial risk are priced in illegal insider trades | Table 6, col 1, p. 8 | lnRISK coeff 0.181\*\*\* (SE 0.049); a 1-pp increase in lnRISK → 18.1 bp higher BHAR |
| R5 | Univariate test: high-legal-environment provinces yield significantly higher insider-trading BHAR than low-legal-environment provinces | Table 3, Panel A, p. 6 | mean BHAR difference 0.085\*\*\* (high vs. low LAW^Institution); median difference 0.044\*\*\* |
| R6 | Heckman selection correction confirms the legal environment effect persists after accounting for potential selection bias from undetected cases | Table 7, Panel A, col 1, p. 10 | LAW^Institution coeff 0.011\*\*\* (SE 0.003) in Heckman model; same sign and significance as OLS baseline |

**Overall (paper's conclusion).** Provincial legal quality plays a decisive role in shaping insider-trading outcomes. Insiders weigh expected gains against enforcement risk and trade only when the anticipated return exceeds the expected penalty, consistent with the rational-crime framework of Becker (1968) and extending the law and finance literature of La Porta et al. (1998) to the enforcement of securities law. Stricter legal environments produce higher conditional profitability because only high-return trades survive deterrence. Firm-level litigation exposure reinforces this relationship. Results are robust to selection-correction procedures, alternative return measures, dummy-variable legal proxies, and geographic heterogeneity tests. Political connections, M&A rumors, and financial literacy of the insider do not explain the premium. Insiders who are closer to the CSRC in Beijing face the strongest legal environment effects (Table 10), consistent with tighter central oversight raising the required risk premium.

## Theory / model

The paper has no formal structural model; the theoretical frame is organized around two competing hypotheses derived from the rational-crime model of Becker (1968) and extended by the law-and-finance literature of La Porta et al. (1998).

**Deterrence hypothesis (H1a):** Stronger legal environments raise expected penalties, reducing the profitability of illegal insider trades by making even high-information trades unattractive.

**Risk-compensation hypothesis (H1b):** Stronger legal environments deter low-return trades altogether. The trades that still occur are a selective right-tail subset with unusually high expected gains. The conditional average observed return rises even as the unconditional volume falls. The selection effect dominates the deterrence effect.

**Firm-level litigation hypothesis (H2):** Firm-level ex ante litigation risk (probability of regulatory sanction, conditional on trading at a particular company) increases BHAR because insiders at high-litigation-risk firms require higher compensation for elevated firm-specific enforcement exposure. This is incremental to the provincial-level effect.

The hypotheses predict opposite signs on the legal environment coefficient in the BHAR regression: H1a predicts a negative coefficient (deterrence reduces profits), H1b predicts a positive coefficient (risk compensation raises the conditional mean). The data strongly support H1b and H2.

The intuition for H1b follows from the rational-crime trade-off. Define the insider's decision as: trade if and only if expected gain exceeds expected penalty:

$$\mathbb{E}[\text{gain}] > p(\text{detection}) \times \text{penalty}$$

In a stronger legal environment, $$p(\text{detection})$$ rises, so the threshold gain required to justify trading also rises. The observed (adjudicated) trades are draws from the right tail of the gain distribution. As the threshold rises with legal quality, the observed conditional mean rises even if the full distribution of potential gains is unchanged. The paper's data capture only adjudicated cases, so the composition of observed returns shifts upward in high-enforcement provinces.

## Method

The paper applies `panel-regression` as the primary estimator and `probit-regression` as a first-stage tool for constructing the ex ante litigation-risk mediator following the methodology of Kim and Skinner (2012).

**Baseline OLS with fixed effects (Models 3-4, pp. 5-6, Table 4).** BHAR is regressed on the provincial legal environment proxy, optionally with firm characteristics, and with industry and year fixed effects. Standard errors are clustered at both the firm and year levels. Three legal environment proxies are used in separate columns: LAW^Institution (the "market intermediaries and legal environment" sub-index of the Wang et al. (2017) Marketization Index), LAW^Environment (the overall legal environment sub-index), and LAW^Resources (the provincial judicial resources index of Gao et al. (2016)).

**Mediation analysis (Models 5-6, pp. 7-8, Table 6).** To assess whether ex ante litigation risk is the channel through which the legal environment affects BHAR, a two-equation mediation system is estimated. First, litigation risk (lnRISK) is regressed on the legal environment proxies and firm characteristics (eq. 5, p. 8):

$$\text{Med} = \beta_0 + \beta_1 \text{Law} + \beta_2 \text{Firm Characteristics} + \text{Ind} + \text{Year} + \varepsilon \tag{5}$$

Second, BHAR is regressed on both the mediator and the legal environment proxies (eq. 6, p. 8):

$$\text{BHAR} = \gamma_0 + \gamma_1 \text{Med} + \gamma_2 \text{Law} + \gamma_3 \text{Firm Characteristics} + \text{Ind} + \text{Year} + \varepsilon \tag{6}$$

If $$\gamma_1$$ is significant and $$\gamma_2$$ remains significant, the litigation risk channel is a partial (not complete) mediator of the legal environment effect. Results in Table 6 confirm both coefficients are significant at the 5% level across all three legal environment proxies.

**Heckman selection correction (Models 7-13, pp. 8-10, Table 7).** Since only detected insider-trading cases are observable, the observed BHAR may be a biased estimate of the full population BHAR. The paper addresses this via Heckman's two-step procedure. Two probit first-stage models identify the probability of appearing in the sample: (i) detection likelihood based on company characteristics (eq. 11), and (ii) top-30% profitability rank within the sample (eq. 12):

$$\text{Prob}(S_{\text{litigation}} = 1) = a + b_1 \text{Firm Characteristics} + \eta \tag{11}$$

$$\text{Prob}(S_{\text{profit}} = 1) = a + b_1 \ln\text{ME} + b_2 \ln\text{BE/ME} + b_3 \text{MOM} + b_4 \text{TURNOVER} \tag{12}$$

The Inverse Mills Ratio (IMR) from each first stage is included as an additional control in the BHAR regression (eq. 13):

$$\text{BHAR}_{i,j,t} = \alpha_0 + \alpha_1 \text{LAW}_{j,t} + \alpha_2 \text{Firm Characteristics}_{i,j,t} + \rho\sigma\, \text{IMR}_{i,j,t} + \text{Ind}_{j,t} + \text{Year}_t + \varepsilon_{i,j,t} \tag{13}$$

The legal environment coefficient remains positive and significant at the 1% level after the IMR correction in both panels of Table 7, ruling out selection bias as the driver of the main result.

## Empirical specifications

**Dependent variable construction (eqs. 1-2, p. 4, Table 1).** For each adjudicated insider-trading case $$i$$, the raw holding-period return is computed from the legal documents:

$$\text{ret}_{\text{Raw}_i} = \frac{\text{Amount of illegal income}_i}{\text{Trading volume}_i \times \text{Closing price}_{i,\text{PurchaseDay}}} \tag{1}$$

The market-adjusted benchmark return for the same holding period is:

$$\text{ret}_{\text{benchmark}_i} = \frac{\text{Closing price}_{i,\text{SellDay}} - \text{Closing price}_{i,\text{PurchaseDay}}}{\text{Closing price}_{i,\text{PurchaseDay}}} \tag{2}$$

BHAR equals the difference between the raw insider trading return and the benchmark return. For cases where trading occurs over multiple days, the average daily closing price during the trading dates is used. The amount of illegal income, trading volumes, and dates are read directly from the court judgment or CSRC sanction document for each case.

**Baseline regression specifications (eqs. 3-4, p. 5, Table 4).** The two baseline models are:

$$\text{BHAR}_{i,j,t} = \alpha_0 + \alpha_1 \text{LAW}_{j,t} + \text{Ind}_{j,t} + \text{Year}_t + \varepsilon_{i,j,t} \tag{3}$$

$$\text{BHAR}_{i,j,t} = \alpha_0 + \alpha_1 \text{LAW}_{j,t} + \alpha_2 \text{Firm Characteristics}_{j,t} + \text{Ind}_{j,t} + \text{Year}_t + \varepsilon_{i,j,t} \tag{4}$$

where $$i$$ indexes the insider-trading case, $$j$$ the company, and $$t$$ the year. $$\text{LAW}_{j,t}$$ is one of the three provincial legal environment proxies for the province where company $$j$$ is registered. Firm characteristics (lagged two months to align with public availability) include: size (lnME), book-to-market ratio (lnBE/ME), momentum (MOM), turnover ratio, leverage (DEBT/ASSET), return on equity (ROE), cash/assets ratio, firm age, institutional ownership (FUND), and state-ownership dummy (DSOE). Industry and year fixed effects are included in all specifications; standard errors are clustered at both the firm and year levels.

Table 4 columns 1-3 report model (3) for the three legal environment proxies; columns 4-6 report model (4). All six coefficients on the legal environment measures are positive and significant (1% for LAW^Institution and LAW^Environment; 1% and 5% for LAW^Resources). The adjusted R² ranges from 0.139 to 0.207.

**Robustness.** Alternative dependent variable: BHAR_High, computed using the highest stock price during the holding period (Table 8). Alternative legal proxies: binary dummies based on the national median of each index (Table 9). Selection correction: Heckman two-step with two alternative first stages (Table 7, Panels A and B). Geographic heterogeneity: subsamples of cases near the CSRC (Beijing-Tianjin-Hebei region) vs. far provinces (Table 10). Market conditions: bear vs. bull market subsamples (Table 11). Corporate governance: ESG score, managerial ownership, CEO duality, and G-index added to model (4) (Table 12); legal environment coefficients remain positive and significant.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Hand-collected court judgments and CSRC sanction documents (PKU-LAW, Lawyee databases) | Primary dataset: 521 insider-trading cases with trading dates, volumes, execution prices, illicit gains, and case characteristics; 312 unique companies, 2006-2018 | No page yet |
| China Stock Market and Accounting Research (CSMAR) | Firm-level control variables: market capitalization, book-to-market ratio, past returns, turnover, leverage, return on equity, cash/assets, institutional ownership, state ownership, firm age | [CSMAR](/wiki/commercial/csmar/) (licensed) |
| Wang, Fan and Yu (2017) Marketization Index of China's Provinces (NERI) | Two provincial legal environment proxies: LAW^Institution (market intermediaries and legal sub-index) and LAW^Environment (overall provincial legal environment sub-index); updated biannually | No page yet |
| Gao et al. (2016) provincial judicial resources index | LAW^Resources proxy: provincial count of lawyers and legal service offices; measures availability of non-public judicial resources | No page yet |
| Bloomberg (appendix only) | M&A event verification for alternative-channel tests in Section 5.2 (dummy DRINFO) | No page yet |

Sample: 521 insider-trading cases involving 312 companies, 2006-2018. Regression sample N = 478 for most specifications (limited by legal environment data coverage); N = 491 for specifications using LAW^Resources. All continuous independent variables winsorized at the 1st and 99th percentiles. A two-month lag is applied between firm fundamentals and the insider-trading date.

## When to read the full paper

Read the full [original](https://doi.org/10.1016/j.jbankfin.2025.107609) if you are: studying the determinants of illegal insider trading profitability in emerging markets; modeling risk-return trade-offs in illicit market activity; extending the rational-crime or law-and-finance framework to securities law enforcement; working on empirical cross-regional legal variation using the Chinese provincial institutional setting; or building on the hand-collected dataset of 521 Chinese insider-trading cases (data available upon request per the paper's data-availability statement). Table 10 is particularly useful for understanding how geographic proximity to the central regulator moderates the legal environment effect.

## Attribution and rights

Source: peer-reviewed, *Journal of Banking and Finance* 185 (2026) 107609. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Batten, Jonathan A., Lanlan Liu, and Yezhou Sha.
> "Illegal insider trading profitability and the legal environment."
> *Journal of Banking and Finance* 185 (2026): 107609.
> DOI: 10.1016/j.jbankfin.2025.107609. © 2025 The Author(s). Published by Elsevier B.V.
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Crowded Spaces and Anomalies: Chincarini, Lazo-Paz & Moneta (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jbf/2026/chincarini-crowded-spaces-anomalies-2026/
# Distilled: This paper shows that crowded equity positions in well-known stock market anomalies earn significantly higher risk-adjusted returns (FF3 monthly alpha of 1.44% for the most vs. least crowded stocks) and that crowding increases institutional exposure to crash risk. The anomaly alpha is concentrated among the most crowded stocks and persists after publication dates. Journal of Banking and Finance 182 (2026) 107579, CC BY-NC-ND 4.0. Six core results with source locators, datasets used, the crowding measures, and the empirical specifications.
# Tags: paper-summary, asset-pricing, anomalies, factors, cross-section, crowding
==============================================================================

**What this is.** The paper's core results, the crowding measures it constructs, and the
empirical specifications behind the main findings: enough to know what it found and how,
without reading all 17 pages. To replicate or extend it, read the full source at the
[original](https://doi.org/10.1016/j.jbankfin.2025.107579).

## TL;DR

This paper investigates whether crowded equity positions, those in which many institutional
investors hold the same stocks and thereby exhaust the liquidity for normal exits, are
associated with higher future returns and greater crash risk. Using Thomson/Refinitiv 13F
institutional holdings from 1980 to 2021, the authors construct a Days-ADV crowding
measure (the days of average daily trading volume needed for all institutions to exit a
position). They find that more crowded anomaly stocks deliver significantly higher
risk-adjusted returns across all 11 anomalies studied by Stambaugh, Yu, and Yuan (2012),
that the anomaly alpha is entirely absent among non-crowded stocks, that the result persists
after the anomaly publication dates identified by Mclean and Pontiff (2016), and that
crowding increases institutional exposure to stock price crash risk. The paper extends the
hedge-fund crowding-return result of Brown, Howard, and Lundblad (2021) to all 13F
institutions and contradicts the mutual-fund finding of Zhong, Ding, and Tay (2017), and
frames crowding as an additional channel within limits-to-arbitrage theory.

## Core results

Magnitudes and significance are as reported; `\*\*`/`\*\*\*` = 5%/1%. Locators point into
the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Most-crowded (Q5) stocks earn higher FF3 alphas than least-crowded (Q1); the VW spread is 1.44%/month | Table 3 Panels A-B, p. 8; Table 4, p. 9 | VW Q5: FF3 alpha = 0.54%/month (t=8.87); Q1: -0.90%/month (t=-7.86); Q5-Q1 spread: 1.44%/month (t=9.67); EW spread: 1.57%/month (t=12.23) |
| R2 | Bivariate double-sort aggregate anomaly portfolio (long crowded long-leg, short least-crowded short-leg) earns large EW alpha | Table 6 Panel B, p. 10-11 | EW FF3 alpha = 1.69%/month (t=11.09) full sample; 1.96%/month in-sample; 1.61%/month post-publication (t=7.67) |
| R3 | Non-crowded anomaly stocks earn near-zero or insignificant alphas across all factor models and periods | Table 7, p. 11 | EW non-crowded portfolio FF3 alpha = 0.009%/month (t=0.18); VW = 0.008%/month (t=0.09); near zero across FF3, FF5P, FF5A, FF5AM |
| R4 | Fama-MacBeth cross-sectional regressions confirm positive LADV-return association; stronger among anomaly stocks and post-publication | Table 9, p. 12 | LADV coef = 0.546 (t=4.31) full sample (col 1); Long x LADV = 0.287 (t=3.12) full sample with interactions (col 4); Short x LADV = 0.485 (t=4.86) (col 4); post-pub spec (col 5): Long x LADV = 0.206 (t=3.01), Short x LADV = 0.308 (t=3.83) |
| R5 | Crowding positively predicts future stock price crash risk, measured by NCSKEW and DUVOL | Table 10, p. 14 | NCSKEW: LADV coef = 0.011 (t=3.29) full sample; DUVOL: LADV coef = 0.018 (t=5.13) full sample; robust to both subperiods |
| R6 | Crowded anomaly portfolios declined significantly more than uncrowded portfolios during the 2007-2009 and COVID-19 crises | Figure 3, p. 13 | CAR differences statistically significant for financial crisis (t=1.97) and COVID-19 crisis (t=2.03) |

**Overall (paper's conclusion).** Crowding is positively associated with future abnormal
returns across all 11 stock market anomalies studied, and the anomaly alpha is generated
almost entirely by the most crowded stocks. This result is robust to different factor model
specifications (FF3, FF5, FF5 augmented with liquidity and momentum), persists after
publication dates, and is stronger for transient and short-horizon institutions. Crowding
also increases institutional exposure to crash risk, consistent with the idea that crowded
positions impose additional risk for which investors require compensation and that crowding
adds a new consideration to limits-to-arbitrage arguments.

## Theory / model

The paper has no formal economic model. It tests two empirical hypotheses derived from
limits-to-arbitrage theory.

**Hypothesis 1 (returns).** Crowded equity positions impose additional risks on investors,
because correlated exit decisions can cause large price declines (coordination risk) and
because the presence of many similarly positioned investors makes liquidity scarce when all
try to exit simultaneously. Following the limits-to-arbitrage literature (Shleifer and
Vishny (1997); De Long et al. (1990); Lam et al. (2011)), arbitrageurs must be compensated
for this extra risk, so the long (short) leg of anomalies should earn positive (negative)
abnormal returns specifically among the most crowded stocks.

**Hypothesis 2 (crash risk).** When too many institutional investors hold the same stock
they create exposure to a correlated crash: if information or margin calls trigger
simultaneous exit decisions, prices decline sharply. This is captured by the negative
conditional skewness (NCSKEW) and down-to-up volatility (DUVOL) of firm-specific weekly
returns, following Hutton et al. (2009) and Callen and Fang (2015).

**Identification.** The paper makes no causal claim. The empirical work is descriptive and
correlational: it documents a positive association between lagged Days-ADV (institutional
crowding) and subsequent returns or crash risk. Fama-MacBeth regressions control for known
determinants of institutional demand (size, book-to-market, turnover, cumulative return), but
there is no instrument, discontinuity, or difference-in-differences design.

## Method

**Days-ADV crowding measure (equation 3, p. 5).** The main crowding proxy is the total
dollar value of institutional holdings in a stock relative to its average daily dollar
trading volume over the same quarter:

$$
\text{Days-ADV}_{i,t} = \frac{\sum_{j=1}^{N} \text{InstHold}_{i,j,t-1}}{\text{ADV}_{i,t-1}} \tag{3}
$$

where $$\text{InstHold}_{i,j,t-1}$$ is the total dollar value invested in security $$i$$ by
institutional investor $$j$$ in quarter $$t-1$$, and $$\text{ADV}_{i,t-1}$$ is the average
daily dollar trading volume of security $$i$$ over quarter $$t-1$$. Higher Days-ADV means
institutions would need more trading days to fully unwind the position at normal volume.
The paper uses the log of Days-ADV (LADV) in regressions to reduce the influence of
outliers. A complementary portfolio-level similarity measure based on cosine similarity
between institutional portfolio weight vectors is also constructed (equations 1-2, p. 5),
but Days-ADV is the primary measure throughout because it directly links ownership
magnitude to the liquidity of the individual security.

**Activity Ratio crowding measure (equation 4, p. 5).** As an alternative, the Activity
Ratio (ActRatio) of Zhong et al. (2017) is used as a cross-check:

$$
\text{ActRatio}_{i,t} = \frac{\sum_{j=1}^{N} \text{Shares}_{i,t-2}}{\text{AvgTurn}_{i,t-1}} \tag{4}
$$

where the numerator is the percentage of shares held by active investors at $$t-2$$ and the
denominator is the average share turnover of stock $$i$$ at $$t-1$$. The correlation between
Days-ADV and ActRatio is 0.99, confirming they proxy the same construct.

**Crash risk metrics (equations 5-7, p. 7-8).** Firm-specific residual returns $$R_{i,t}$$
are first obtained by stripping market and industry effects from weekly returns via a
market-model regression that includes lead and lag terms (equation 5, p. 7). Negative
conditional skewness (NCSKEW) is then:

$$
\text{NCSKEW}_{i,t} = -\frac{n(n-1)^{3/2} \sum R_{i,t}^3}{(n-1)(n-2)\left(\sum R_{i,t}^2\right)^{3/2}} \tag{6}
$$

Down-to-up volatility (DUVOL) is:

$$
\text{DUVOL}_{i,t} = \log\left(\frac{(n_u - 1)\sum_{\text{DOWN}} R_{i,t}^2}{(n_d - 1)\sum_{\text{UP}} R_{i,t}^2}\right) \tag{7}
$$

where $$n_u$$ ($$n_d$$) is the count of up (down) weeks in the year and DOWN (UP) is the
subsample of weeks with returns below (above) the mean. Higher values of both measures
indicate greater crash risk.

**Days-to-Cover (DTC, equation 8, p. 15).** To test the short-leg alpha against a
competing explanation (Hong et al. (2016)), DTC is computed as:

$$
DTC = \frac{SR}{\text{Average Daily Turnover}} \tag{8}
$$

where $$SR$$ is the short ratio (short interest divided by shares outstanding). DTC
approximates the number of days required for all short sellers to cover at normal trading
volume, capturing marginal-cost crowding for short positions.

## Empirical specifications

**Portfolio sorts (Section 3.1, p. 8).** Each calendar quarter, stocks are ranked into
quintiles by each crowding measure (Days-ADV, ActRatio, NI, PSO). Value- and equal-weighted
quintile portfolio returns are computed for the following quarter and the time-series alpha
is estimated via:

$$
R^e_{p,t} = \alpha_p + \beta_1 \text{MktRf}_t + \beta_2 \text{SMB}_t + \beta_3 \text{HML}_t + \varepsilon_{p,t}
$$

with variants that also include the Fama and French (1993) five-factor model (FF5), the
Pastor and Stambaugh (2003) traded liquidity factor (FF5P), the Amihud (2019)
illiquid-minus-liquid factor (FF5A), and further the Carhart (1997) momentum factor (FF5AM).
Newey-West standard errors are used throughout.

For the bivariate anomaly analysis (Section 3.2, p. 10), stocks are sorted first on the
anomaly variable into quintiles and then within the long (short) leg the top (bottom) 30%
by Days-ADV are selected. The aggregate portfolio across anomalies takes an equally weighted
average of the resulting anomaly-specific portfolio returns each month. Non-crowded sorted
portfolios use the middle 40% of Days-ADV within each anomaly leg (Table 7).

**Fama-MacBeth regressions (Section 3.3, p. 12).** Each quarter, cumulative monthly
returns over the following quarter are regressed cross-sectionally on LADV and a vector of
controls $$X_{i,t}$$ (log size, age, return standard deviation, book-to-market, dividend
yield, average monthly turnover, and cumulative returns over the past three and nine months).
To test whether the anomaly channel amplifies the crowding-return link, indicator dummies
for long-leg and short-leg anomaly membership and their interactions with LADV are added:

$$
r_{i,t+1} = \alpha + \beta_1 \text{LADV}_{i,t} + \beta_2 \text{Long}_{i,t} + \beta_3 (\text{Long} \times \text{LADV})_{i,t} + \beta_4 \text{Short}_{i,t} + \beta_5 (\text{Short} \times \text{LADV})_{i,t} + \gamma' X_{i,t} + \varepsilon_{i,t}
$$

The time-series average of quarterly cross-sectional coefficients gives the estimates.
Standard errors are Newey-West with four lags.

**Crash risk panel regression (Section 3.4.1, Table 10, p. 14).** One-year-ahead crash risk
is regressed on log Days-ADV and controls, with firm and year fixed effects. Standard
errors are clustered by firm:

$$
\text{CrashRisk}_{i,t+1} = \alpha + \beta \, \text{LADV}_{i,t} + \gamma' \text{Controls}_{i,t} + \text{FirmFE} + \text{YearFE} + \varepsilon_{i,t}
$$

Controls include cumulative firm-specific daily returns, kurtosis and standard deviation of firm-specific daily returns, market-to-book ratio, book value of liabilities to total assets, ROA, log market cap, average monthly share turnover, number of analysts, and the lag of the crash risk variable. Anomaly-leg dummies and post-publication
indicators are interacted with LADV in extended specifications.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Thomson/Refinitiv (TR) 13F institutional holdings | Primary crowding measure (Days-ADV, cosine similarity); institution-type classification into transient, dedicated, quasi-indexer, hedge fund, mutual fund | [WRDS](/wiki/commercial/wrds/) (licensed) |
| CRSP monthly stock data | Stock returns, prices, trading volume, shares outstanding; anomaly variable construction | [WRDS](/wiki/commercial/wrds/) (licensed) |
| Compustat annual fundamentals | Accounting-based anomaly variables (accruals, NOA, asset growth, profitability, etc.) and short interest data (2003-2021) | [WRDS](/wiki/commercial/wrds/) (licensed) |
| I/B/E/S analyst data | Number of analysts following each stock (control variable in FM regressions) | [WRDS](/wiki/commercial/wrds/) (licensed) |
| Kenneth French Data Library | FF3, FF5, and momentum factor returns for risk adjustment | [Ken French library](/wiki/datasets/ken-french/) |
| Brian Bushee institution classification | Transient, dedicated, quasi-indexer institution type labels | no page yet |

Sample: US common stocks on NYSE, AMEX, and Nasdaq with price above $5, excluding utilities
and financial firms. Main sample: 1980:Q1 to 2021:Q4 (quarterly rebalancing). Exception:
momentum anomaly portfolios are rebalanced quarterly (not annually). DTC analysis restricted
to 2003-2021, when Nasdaq short interest data becomes available.

## When to read the full paper

Use the [original](https://doi.org/10.1016/j.jbankfin.2025.107579) if you are:
investigating the relationship between institutional crowding and anomaly returns (Table 6
provides anomaly-by-anomaly bivariate-sort alphas for all 11 anomalies across five factor
models); studying crash risk as a channel linking institutional crowding to limits-to-
arbitrage; extending the results to the 97 anomalies of Mclean and Pontiff (2016) (Table 8
and the Internet Appendix); or assessing whether the DTC measure of Hong et al. (2016)
accounts for the short-leg alpha in crowded spaces.

## Attribution and rights

Source: peer-reviewed, *Journal of Banking and Finance* 182 (2026) 107579. This distillation
was extracted by an LLM on 2026-06-25 and is **not human-verified or independently
reproduced**. The CC BY-NC-ND 4.0 licence permits verbatim sharing but not derivative works;
the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY-NC-ND 4.0).** Chincarini, Ludwig B., Renato Lazo-Paz, and Fabio Moneta.
> "Crowded spaces and anomalies." *Journal of Banking and Finance* 182 (2026) 107579.
> DOI: 10.1016/j.jbankfin.2025.107579. © 2025 The Authors. Published by Elsevier B.V.
> Licensed under [CC BY-NC-ND 4.0](https://creativecommons.org/licenses/by-nc-nd/4.0/).
> This page is a distillation by the Institute for Automated Research:
> core results extracted and re-expressed as structured text. The licence does not permit
> derivative works; this page constitutes extract-only fair use documentation.


==============================================================================
# Fed Put in the Equity Options Markets: Dahiya, Kamrad, Poti & Siddique (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jbf/2026/dahiya-fed-put-equity-options-2026/
# Distilled: Documents the Fed Put (Greenspan Put) in S&P 500 and S&P 100 equity index option markets. Put implied volatility is 3 to 5 percentage points lower during accommodative monetary policy, strongest when investor risk aversion is high, and concentrated in the pre-2008 period; the effect largely vanishes after the Global Financial Crisis. Journal of Banking and Finance 188 (2026), paywalled. Seven core results with source locators, the Taylor Rule identification design, and IV-GMM estimation.
# Tags: paper-summary, monetary-policy, options-markets, implied-volatility, asset-pricing
==============================================================================

**What this is.** The paper's core results, the monetary policy identification design (Taylor Rule deviation combined with the Wu-Xia shadow rate), and the regression specifications with enough detail to know what it found and how, without reading all 14 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1016/j.jbankfin.2026.107697).

## TL;DR

The paper tests whether the "Fed Put" (Greenspan Put) is detectable in equity options prices. The premise: if investors believe the Federal Reserve will support markets during downturns, exchange-traded put options are partially substituted by the Fed's implicit backstop, so out-of-the-money put implied volatility should be lower during accommodative monetary policy periods. Using 1,342 weekly observations of S&P 500 (SPX) and S&P 100 (OEX) index option implied volatilities from OptionMetrics (January 1996 to December 2021), with monetary policy stance measured as deviations from the Taylor (1993) Rule extended to the zero lower bound via the Wu and Xia (2016) shadow fed funds rate, the paper finds robust evidence of a Fed Put in the pre-2008 period: put implied volatility is 3 to 5 percentage points lower during accommodative periods (controlling for option characteristics), with the effect amplified in high-risk-aversion regimes per Bekaert, Engstrom, and Xu (2021) and robust to IV-GMM estimation. Post-2008, the effect largely disappears, consistent with a structural break induced by the Global Financial Crisis and the subsequent shift to unconventional monetary policy.

## Core results

Magnitudes and significance are as reported. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Accommodative Fed stance (FedSupport) lowers S&P 500 put implied volatility across all 11 moneyness levels (univariate)** | Table 1, Panel A, p. 8 | β₁ = -3.32 (t=-20.48) at moneyness=50; -5.33 (t=-37.48) at moneyness=100; all 11 categories 1\% significant |
| R2 | **Effect is concentrated in high-risk-aversion regimes; negligible when risk aversion is low** | Table 2, Panel A, p. 9 | Quintile 5 (highest RA): β₁ = -6.27 (t=-14.44); Quintile 1 (lowest RA): β₁ = 0.09 (t=0.50, not significant); at moneyness=50 |
| R3 | **Pre-crisis (1996-2007) multivariate OLS effect is nearly 3x the full-sample estimate** | Table 5, p. 11 | Pre-crisis OLS: β₁ = -0.659 (t=-6.746); full-sample OLS: β₁ = -0.223 (t=-2.668) |
| R4 | **Post-crisis (2009-2021) Fed Put effect disappears in the multivariate regression** | Table 5, p. 11 | Post-crisis OLS: β₁ = -0.102 (t=-0.759, not significant); post-crisis IV-GMM: not significant |
| R5 | **Pre-crisis IV-GMM causal estimate is approximately 2x the OLS estimate** | Table 5, p. 11 | Pre-crisis IV-GMM: β₁ = -1.205 (t=-8.152), 1\% significant |
| R6 | **Binary FedSupport measure corroborates OLS; pre-crisis IV-GMM also strongly negative** | Table 6, p. 13 | Full-sample OLS: β₁ = -1.827 (t=-6.172); pre-crisis OLS: β₁ = -2.674 (t=-9.122); pre-crisis IV-GMM: β₁ = -3.969 (t=-8.571) |
| R7 | **Alternative monetary policy proxies (money-market-futures changes) confirm the negative relation** | Table 7, p. 13 | ΔMP1 OLS: β₁ = -9.427 (t=-51.566); ΔMP1 IV-GMM: β₁ = -80.292 (t=-1.813, 10\% significant) |

**Overall (paper's conclusion).** There is robust evidence of a Fed Put in the pre-2008 period: accommodative monetary policy is associated with substantially lower implied volatility of equity index put options, consistent with investors treating the Fed's implicit backstop as a partial substitute for put protection. The effect is amplified in high-risk-aversion regimes, aligning with the meta moral hazard mechanism of Miller, Weller, and Zhang (2002). Post-2008, the relationship largely vanishes, suggesting the Global Financial Crisis permanently altered market expectations of Fed intervention. IV-GMM confirms the pre-crisis effect is not driven by reverse causality or endogenous risk aversion.

## Theory / model

The paper has no formal model of its own. It tests the theoretical framework of Miller, Weller, and Zhang (2002), who distinguish two cases for the Fed Put's effect on option prices. In the first (complete-credibility) case, investors fully believe the Fed will prevent large market declines; return distributions exhibit a truncated downside and a fatter upside, and variation in policy would have no impact on beliefs or option prices. In the second (partial-credibility) case, monetary policy stance shifts investor beliefs about the probability and magnitude of Fed intervention, generating a negative cross-price effect between accommodative policy and put implied volatility. The paper tests the partial-credibility case empirically. While Miller, Weller, and Zhang (2002) and Drechsler, Savov, and Schnabl (2018) discuss theoretical mechanisms through which central bank intervention distorts asset prices, prior empirical evidence had been scarce.

**Identification via the Taylor Rule** (§2.1, p. 3): Monetary policy stance is identified through deviations of the effective fed funds rate from the Taylor (1993) benchmark rate. The original Taylor Rule specification is (Eq. 1, p. 3):

$$
i = r^* + \pi + w_1(\pi - \pi^*) + w_2(y - y^*) \tag{1}
$$

where $$i$$ is the nominal federal funds rate, $$r^* = 2\%$$ is the target real FFR, $$\pi$$ is the inflation rate, $$\pi^* = 2\%$$ is the target inflation rate, $$y$$ is log real output, $$y^*$$ is log potential output, and $$w_1 = w_2 = 0.5$$ in the original specification. The paper also tests variants labeled Final Taylor Rule 1_1, 2_1, 2_2, and 2_3 that modify the inflation measure and output-gap weight.

For ZLB periods when the observed FFR cannot capture the full accommodation of unconventional policy, the paper substitutes the Wu and Xia (2016) shadow fed funds rate. This rate is derived from a factor-augmented VAR (FAVAR) model with three latent factors; the Atlanta Fed maintains updated estimates. The shadow rate equals the observed FFR when the observed rate exceeds 0.25%, and extends below zero at the ZLB.

**FedSupport measure** (§2.1, pp. 3-4): Two variants are constructed:
- Continuous: (Wu-Xia shadow rate) minus (Taylor Rule implied rate); more negative = more accommodative
- Binary: indicator equal to 1 when the shadow rate falls below the Taylor Rule benchmark (Fed supports markets), and 0 otherwise

The key prediction is β₁ < 0: when the Fed provides implicit downside protection, investors demand less explicit insurance from put options. The results of Adrian et al. (2020), who show accommodative interest rate policy shifts rather than eliminates macroprudential risk over time, motivate the expectation of a structural break around the Global Financial Crisis.

## Method

The paper applies OLS panel regression and IV-GMM to weekly options data. The baseline specification (Eq. 2, p. 4) is:

$$
\sigma(\text{Put})_{i,t} = \beta_0 + \beta_1 \times \text{FedSupport}_t + \phi' \text{Controls}_{i,t} + u_t \tag{2}
$$

where $$\sigma(\text{Put})_{i,t}$$ is the implied volatility of put option $$i$$ in week $$t$$, expressed as a percentage, and $$\text{FedSupport}_t$$ captures the monetary policy stance. It builds on `panel-regression` as the baseline estimator.

Endogeneity is addressed via `instrumental-variables` estimation combined with `gmm`. Both FedSupport and RiskAversion are potentially endogenous: unobserved macroeconomic or financial variables could influence both the Fed's policy stance and investors' risk preferences simultaneously. The IV-GMM estimator (Appendix A, Eq. 9, p. 13) is:

$$
\hat{\beta}_{\text{GMM}} = (\mathbf{X}'\mathbf{Z}\mathbf{W}^{-1}\mathbf{Z}'\mathbf{X})^{-1} \mathbf{X}'\mathbf{Z}\mathbf{W}^{-1}\mathbf{Z}'\mathbf{y} \tag{9}
$$

where $$\mathbf{X}$$ is the matrix of endogenous regressors (FedSupport and RiskAversion), $$\mathbf{Z}$$ is the instrument matrix, and $$\mathbf{W}$$ is a HAC-consistent covariance matrix of moment conditions, following Baum, Schaffer, and Stillman (2003) and building on Griliches and Hausman (1986). The two-step feasible GMM handles heteroskedasticity and autocorrelation common in derivatives panel data.

## Empirical specifications

**Data construction** (§3, p. 5): Daily option implied volatility data come from the OptionMetrics Ivy DB, compiled from 3:59 PM EST closing prices for S&P 500 (CBOE ticker: SPX) and S&P 100 (CBOE ticker: OEX) index options. Black-Scholes implied volatility is computed for each option using the underlying index price, dividend yield, risk-free rate, time to maturity, and strike price. The volatility surface is reconstructed via cubic spline interpolation. Weekly observations are taken every Wednesday. Moneyness is defined as (strike price / current index price) × 100, so an out-of-the-money put has moneyness < 100.

**Univariate** (Eq. 3, p. 7; produces R1, R2):

$$
\sigma(\text{Put})_{i,t} = \beta_0 + \beta_1 \times \text{FedSupport}_t + u_t \tag{3}
$$

Estimated separately for each of 11 moneyness levels (50, 55, ..., 100) and 9 maturity buckets (10 to 90 days). The binary FedSupport indicator is used. Panel A of Table 1 (p. 8) aggregates across maturities for each moneyness level; Panel B aggregates across moneyness for each maturity. Table 2 (p. 9) stratifies by quintile of the Bekaert, Engstrom, and Xu (2021) relative risk aversion index to test the heterogeneity prediction (R2).

**Multivariate OLS** (Eq. 4, p. 10; produces R3, R4):

$$
\sigma(\text{Put})_{i,t} = \beta_0 + \beta_1 \times \text{FedSupport}_t + \beta_2 \times \text{Moneyness}_{i,t} + \beta_3 \times \text{Expiration}_{i,t} + \beta_4 \times \text{RiskAversion}_t + \varepsilon_{i,t} \tag{4}
$$

The continuous measure of FedSupport (shadow rate minus Taylor Rule rate) is used in Table 5 (p. 11); the binary measure in Table 6 (p. 13). $$\text{RiskAversion}_t$$ is the Bekaert, Engstrom, and Xu (2021) relative risk aversion index, sampled at weekly frequency. Standard errors are robust using the HAC procedure of Chang and McAleer (2015). The sample splits at December 31, 2007 (pre-crisis: 1996-2007) and January 1, 2009 (post-crisis: 2009-2021).

**IV-GMM structural** (Eq. 5, p. 11; produces R5):

$$
\sigma(\text{Put})_{i,t} = \beta_0 + \beta_1 \times \widehat{\text{FedSupport}}_t + \beta_2 \times \text{Moneyness}_{i,t} + \beta_3 \times \text{Expiration}_{i,t} + \beta_4 \times \widehat{\text{RiskAversion}}_t + \varepsilon_{i,t} \tag{5}
$$

Both FedSupport and RiskAversion are instrumented. The first-stage equations (Eqs. 6-7, p. 12) use four lagged macro indicators as instruments:

$$
\text{FedSupport}_t = \kappa_0 + \kappa_1 \text{LaborMkt}_{t-1} + \kappa_2 \text{SP500}_{t-1} + \kappa_3 \text{FinStab}_{t-1} + \kappa_4 \text{PE}_{t-1} + u_t \tag{6}
$$

$$
\text{RiskAversion}_t = \gamma_0 + \gamma_1 \text{LaborMkt}_{t-1} + \gamma_2 \text{SP500}_{t-1} + \gamma_3 \text{FinStab}_{t-1} + \gamma_4 \text{PE}_{t-1} + \eta_t \tag{7}
$$

where LaborMkt is the Kansas City Fed Labor Market Conditions Index, SP500 is the lagged S&P 500 return, FinStab is the FRB St. Louis Financial Stress Index, and PE is the lagged S&P 500 Price-Earnings Ratio. The system is exactly identified (4 instruments, 2 endogenous variables). Instrument relevance is confirmed: Kleibergen-Paap LM test chi² = 76.557 (rejects under-identification, p < 0.01); weak instruments null rejected at chi² = 24.462 (p < 0.01) (footnote 15, p. 11).

**Alternative MP proxies** (Eq. 8, p. 12; produces R7):

$$
\Delta\sigma(\text{Put})_{i,t} = \beta_0 + \beta_1 \times \Delta MP_t + \beta_2 \times \text{Moneyness}_{i,t} + \beta_3 \times \text{Maturity}_{i,t} + \beta_4 \times \text{RiskAversion}_t + \varepsilon_{i,t} \tag{8}
$$

where $$\Delta MP_t$$ is the weekly change in a money-market futures price index: MP1 is the 8-quarter-ahead 3-month Eurodollar futures level (100 minus yield); MP2 is the 3-month 30-day Fed Funds futures level. An increase in MP1 or MP2 corresponds to expected monetary easing; a negative β₁ confirms the Fed Put. Table 7 (p. 13) reports OLS and IV-GMM estimates for both proxies.

**Robustness**: Markov Switching Dynamic Regression (MSDR) and Hidden Markov Models (HMM) are used as alternative identification strategies for Fed support regimes (Online Appendix). Results are replicated on S&P 100 (OEX) index options (Table 4, p. 10).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| OptionMetrics Ivy DB (S&P 500 SPX, S&P 100 OEX) | Primary data: daily index option implied volatility surface, 1996-2021 | no page yet |
| Wu and Xia (2016) shadow Fed Funds rate (Atlanta Fed) | Monetary policy stance during ZLB; extends observed FFR below zero | [FRED](/wiki/datasets/fred/) (related public source) |
| Taylor Rule implied rate (Bernanke 2015 blog spreadsheet) | Benchmark rate for identifying FedSupport vs non-support periods | no page yet |
| CRSP S&P 500 and S&P 100 index levels | Underlying prices for moneyness construction | [WRDS](/wiki/commercial/wrds/) (licensed) |
| Federal Reserve H15 series (effective Fed Funds rate) | Observed FFR for non-ZLB periods | [FRED](/wiki/datasets/fred/) |
| Bekaert, Engstrom and Xu (2021) relative risk aversion index (nancyxu.net) | Time-varying risk aversion control and endogenous regressor | no page yet |
| Kansas City Fed Labor Market Conditions Index | Instrument for FedSupport and RiskAversion | [FRED](/wiki/datasets/fred/) |
| FRB St. Louis Financial Stress Index | Instrument for FedSupport and RiskAversion | [FRED](/wiki/datasets/fred/) |
| S&P 500 Price-Earnings Ratio | Instrument for FedSupport and RiskAversion | no page yet |

Sample: 1,342 weekly observations, January 4, 1996 to December 31, 2021. Primary results on S&P 500 (SPX). Replicated on S&P 100 (OEX). 132,858 option-week observations in the multivariate regressions (Table 5, p. 11). Authors state they do not have permission to share data (Appendix B, p. 13).

## When to read the full paper

Use the [original](https://doi.org/10.1016/j.jbankfin.2026.107697) if you are: testing whether the Fed Put effect extends to other derivatives markets or international indices; examining how unconventional monetary policy (QE, forward guidance) affects investor beliefs about central bank intervention; replicating the IV-GMM instrument set for options-market panel regressions with endogenous risk aversion; or extending the analysis to post-2021 data when the Fed raised rates rapidly. Tables 1-2 (pp. 8-9) document the univariate pattern and risk aversion heterogeneity; Tables 5-6 (pp. 11, 13) contain the multivariate OLS and IV-GMM estimates for pre- and post-crisis subperiods.

## Attribution and rights

Source: peer-reviewed, *Journal of Banking and Finance* 188 (2026). This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. The article is paywalled (Elsevier, all rights reserved); only core results are extracted here.

> Dahiya, Sandeep, Bardia Kamrad, Valerio Poti, and Akhtar Siddique.
> "Fed put in the equity options markets."
> *Journal of Banking and Finance* 188 (2026) 107697.
> DOI: [10.1016/j.jbankfin.2026.107697](https://doi.org/10.1016/j.jbankfin.2026.107697).
> (c) 2026 Elsevier B.V. All rights reserved.
> Extract only; redistribution not permitted.


==============================================================================
# Bank Market Power and Monetary Policy Transmission: Enkhbold (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jbf/2026/enkhbold-monetary-policy-bank-market-power-2026/
# Distilled: Using US bank- and loan-level data from 2000 to 2019, the paper shows that a 100 bps monetary policy shock transmits 34 bps to mortgage rates in competitive banking markets but near-zero in concentrated markets; wholesale funding reliance amplifies the gap in competitive markets and dampens it in concentrated ones. Journal of Banking and Finance 187 (2026), paywalled. Six core results with source locators, datasets used, and the estimating regression specification.
# Tags: paper-summary, monetary-policy, macro, banking, market-power, mortgage-markets
==============================================================================

**What this is.** The paper's core results, the empirical specification, and the identification strategy: enough to know what it found and how, without reading the full paper. To replicate or extend, read the original at the [DOI link](https://doi.org/10.1016/j.jbankfin.2026.107690).

## TL;DR

The paper studies how the interaction between local deposit market concentration and bank wholesale funding reliance (WFR) shapes the transmission of monetary policy surprises to mortgage rates. Using US bank- and loan-level data from 2000 to 2019, and building on the deposit channel of Drechsler et al. (2017) and the wholesale funding analysis of Choi and Choi (2021), the paper shows that in competitive markets (low HHI) a 100 bps policy shock raises mortgage rates by 34 bps for banks with high WFR; wholesale funding amplifies pass-through because it ties funding costs directly to market rates. In concentrated markets (high HHI), the same shock produces near-zero pass-through: banks use market power to hold deposit rates steady and absorb the cost change in margins rather than passing it to borrowers. The 35 bps differential translates to approximately $67 per month on a $300,000 mortgage ($24,000 over the loan life). Wang et al. (2022) show that market power dampens transmission in a structural model; this paper adds WFR as an interacting channel and traces heterogeneity across the interest rate cycle. These contrasts sharpen at the zero lower bound and during contractionary episodes, and hold across alternative market power measures including the Lerner index and branch market share.

## Core results

Magnitudes and significance as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%. Locators point to the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Competitive markets: 100 bps shock transmits 34.1 bps to mortgage rates for high-WFR banks | Table 4, p. 9 | Direct effect 18.47\*\*\* bps; WFR interaction adds 15.68\* bps; total ~34.1 bps for high-WFR competitive banks |
| R2 | Concentrated markets: near-zero pass-through; WFR interaction is negative | Table 4, p. 9 | Baseline 9.10 bps (insig., t=0.83); WFR interaction = -0.728\*\* bps/log-unit; net pass-through near zero |
| R3 | Zero lower bound amplifies competitive pass-through to 38.8 bps | Table 5, p. 10 | Competitive baseline at ZLB = 38.80\*\*\* bps (vs 23.14\*\*\* non-ZLB); WFR interaction = 10.05\*\*\* bps |
| R4 | ZLB reverses concentrated-market pass-through to -21.2 bps | Table 5, p. 10 | Concentrated ZLB = -21.18\*\*\* bps (vs -1.81\* non-ZLB); sign reversal vs competitive ZLB (+38.8 bps) |
| R5 | Asymmetric transmission: contractionary shocks pass through 4x more than expansionary | Table 7, p. 13 | Contractionary: competitive 29.66\*\*\* bps, concentrated 8.73\*\*\*; expansionary: competitive 6.86\*\*\* bps, concentrated -3.17 (insig.) |
| R6 | Robustness: Lerner index and branch share confirm market power dampening | Table 9, pp. 15-16 | Lerner x shock = -9.40\*\*\* bps; branch share x log(WFR) x shock (triple interaction) = +60.79\*\*\* bps; lagged HHI as IV yields -9.14\*\* bps interaction |

**Overall (paper's conclusion).** Market concentration dampens monetary policy transmission by allowing banks to absorb policy shocks in margins. Wholesale funding reliance amplifies this gap: in competitive markets it ties funding costs to market rates and forces pass-through; in concentrated markets it provides a substitute for unraised deposit rates, further cushioning borrowers from the policy change. The Federal Reserve's ability to affect household mortgage costs depends on local banking market structure throughout the interest rate cycle.

## Theory / model

The paper has no formal theoretical model. It builds on two prior mechanisms and documents a new interaction between them.

**Deposit channel (Drechsler et al. (2017)).** When the policy rate rises, banks with market power over local depositors need not match the rate increase: they hold the deposit spread wide and allow some depositors to leave without triggering mass outflows. This contracts aggregate deposits but does not immediately force pass-through to lending rates.

**Wholesale funding substitution (Choi and Choi (2021)).** As policy tightens, banks can replace contracting retail deposits with wholesale funding (repos, federal funds purchased, brokered deposits, time deposits from institutional investors). In concentrated markets, this substitution is cost-effective because raising deposit rates on a large base is expensive relative to borrowing wholesale at the margin. In competitive markets, thin markups force deposit repricing anyway, so wholesale funding adds cost pressure rather than providing insulation.

**Central hypothesis.** The two mechanisms interact: wholesale funding amplifies pass-through in competitive markets (funding costs track the policy rate closely) but dampens it in concentrated markets (market power absorbs wholesale cost increases into margins). The paper documents this interaction empirically across multiple policy-rate regimes.

**Identification.** The paper exploits cross-bank variation in HHI (measured from the FDIC Summary of Deposits) and WFR (from Call Reports), holding constant time trends via the monetary shock variable and location trends via MSA fixed effects. Bauer and Swanson (2023) high-frequency monetary surprises serve as the exogenous policy variable (see Method). Within each MSA, the identifying variation is cross-bank heterogeneity in market concentration and funding mix.

## Method

The estimating equation (equation (1), p. 8) is a panel regression of loan-level mortgage rate changes on the monetary shock, the log wholesale funding ratio, and their interaction, with bank and MSA fixed effects:

$$
\Delta r_{mbt} = \alpha_b + \alpha_m + \beta_1 \Delta_t + \beta_2 \log(\text{WFR}_{bt-1}) + \beta_3 \log(\text{WFR}_{bt-1}) \times \Delta_t \tag{1}
$$

$$
+ \;\Gamma \text{HH Controls}_{mbt-1} + \Xi \text{HH Controls}_{mbt-1} \times \Delta_t
$$

$$
+ \;\Pi \text{Bank Controls}_{mbt-1} + \Lambda \text{Bank Controls}_{mbt-1} \times \Delta_t
$$

$$
+ \;\Psi \text{Macro Controls}_{mt-1} + \Omega \text{Macro Controls}_{mt-1} \times \Delta_t + \epsilon_{mbt}
$$

where $$\Delta r_{mbt}$$ is the change in the loan-level mortgage rate at MSA $$m$$, bank $$b$$, quarter $$t$$; $$\alpha_b$$ is a bank fixed effect; $$\alpha_m$$ is an MSA fixed effect; $$\Delta_t$$ is the Bauer and Swanson (2023) monetary shock normalized to a 100 bps impact; and $$\text{WFR}_{bt-1} = \text{wholesale funding}_b / \text{retail deposits}_b$$ at quarter $$t-1$$. HH controls (interacted with the shock) include the borrower's credit score, LTV, and debt-to-income ratio. Bank controls include number of branches, liquidity asset ratio, duration mismatch, liability interest rate, real estate loans ratio, commercial and industrial loans ratio, and MBS-to-asset ratio. Macro controls include the unemployment rate, house price index, and personal income per capita. Standard errors are clustered at the bank and quarter levels.

**Policy shocks.** The shock $$\Delta_t$$ aggregates changes in financial variables in a 30-minute window around FOMC announcements (10 minutes before to 20 minutes after), orthogonalized to Fed information effects following Bauer and Swanson (2023). An alternative sign-based decomposition from Jarocinski and Karadi (2020) is used in robustness checks (Table 8).

**Market concentration.** Local deposit market concentration in MSA $$m$$ at quarter $$t$$ is the Herfindahl-Hirschman Index constructed from the FDIC Summary of Deposits (p. 3):

$$
\text{HHI}_{mt} = \sum_{b \in m} \left( \frac{dep_{mbt}}{\sum_{b' \in m} dep_{mb't}} \right)^2
$$

where $$dep_{mbt}$$ is deposits of bank $$b$$ in MSA $$m$$ in year $$t$$. A lower HHI indicates a competitive market; a higher HHI indicates a concentrated market. The main results split the sample at the median HHI into competitive (HHI = 0) and concentrated (HHI = 1) subsamples.

The method builds on `panel-regression` with two-way fixed effects and `instrumental-variables` for the monetary shock. Robustness checks instrument for potential endogeneity of market structure using lagged HHI and WFR as instruments (Table 9, col 2; interaction coefficient -9.14** bps).

## Empirical specifications

All headline results come from equation (1), estimated separately on competitive and concentrated subsamples. The focal coefficient is $$\beta_3$$, the interaction of log WFR and the monetary shock, capturing how wholesale funding reliance moderates policy pass-through.

**Main pass-through (R1, R2; Table 4).** The competitive subsample (HHI = 0, N = 24,539) yields a direct shock effect of 18.47\*\*\* bps (SE = 6.14) and a WFR interaction of 15.68\* bps (SE = 8.69). The concentrated subsample (HHI = 1, N = 30,527) yields a baseline of 9.10 bps (SE = 11.05, insignificant) and a WFR interaction of -0.728\*\* bps (SE = 0.318). For a $300,000 mortgage the 35 bps gap translates to approximately $67 per month.

**ZLB vs non-ZLB (R3, R4; Table 5).** The same regression is estimated separately for the ZLB period (2009-2015, cols 3-4) and non-ZLB periods (cols 1-2). The deposit floor constraint pins deposit rates at zero in competitive markets, removing the margin buffer and raising competitive pass-through to 38.80\*\*\* bps. In concentrated markets, forward guidance and quantitative easing compress long-term premia while market power prevents pass-through of cheap funding to borrowers, producing a large negative coefficient of -21.18\*\*\* bps.

**Low vs high policy rates (Table 6).** Estimated separately for low and high rate periods (split at the median federal funds rate). In competitive markets, the WFR interaction is largest at low rates (25.77\*\*\* bps) and remains large at high rates (18.37\*\*\* bps). In concentrated markets the baseline is positive at low rates (32.07\*\*\* bps) but deeply negative at high rates (-16.99\*\*\* bps), reflecting that banks at low rates lower mortgage rates to expand lending volume while using their pricing power to avoid fully passing the cost saving to borrowers at high rates.

**Contractionary vs expansionary shocks (R5; Table 7).** The sample is split into expansionary and contractionary periods. Within each, the model estimates separate coefficients for positive shocks ($$\Delta_t > 0$$, rate hikes) and negative shocks ($$\Delta_t < 0$$, rate cuts). Contractionary shocks produce 29.66\*\*\* bps pass-through in competitive markets (concentrated: 8.73\*\*\*) with WFR adding 16.03\*\*\* bps in competitive markets. Expansionary shocks produce only 6.86\*\*\* bps in competitive markets and an insignificant -3.17 bps in concentrated markets, consistent with banks in concentrated markets rebuilding margins during easing.

**Alternative market power measures (R6; Tables 8-9).** The specification is replicated with (i) liability-side wholesale funding share, (ii) branch market share, and (iii) the Lerner index as market power proxies. The Lerner index specification (Table 9, col 4) yields a direct interaction of -9.40\*\*\* bps, providing the most direct evidence that price-setting ability, not just market concentration, dampens transmission. Branch share produces a positive triple interaction of +60.79\*\*\* bps (col 3; this is the Δt × Branch share × log(WFR) coefficient; the simpler Δt × Branch share = 73.35\*\* bps), showing that physical presence amplifies rather than dampens transmission, distinct from market power measured by HHI or Lerner.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Fannie Mae Single-Family Loan Performance Data | Loan-level mortgage rates, borrower characteristics (FICO, LTV, loan purpose, MSA) | No page yet |
| Freddie Mac Single-Family Loan-Level Dataset | Loan-level mortgage rates, originator identity | No page yet |
| FDIC Summary of Deposits (SOD) | Branch-level deposits; used to construct HHI for each bank-MSA-quarter | No page yet |
| Federal Reserve Call Reports | Bank-level wholesale funding, assets, liabilities, branch count (quarterly) | No page yet |
| HMDA (Home Mortgage Disclosure Act) | Loan-level origination data; used to construct mortgage market HHI for robustness | No page yet |

Sample: 2000Q1 to 2019Q4. Working sample: 27 largest US banks with assets over $1 billion, approximately 40,000 bank-MSA-quarter observations. Mortgages restricted to 30-year fixed-rate single-family loans acquired by Fannie Mae and Freddie Mac (30-year, fully amortizing, full documentation, conventional).

## When to read the full paper

Read the [original](https://doi.org/10.1016/j.jbankfin.2026.107690) if you are: studying heterogeneous monetary policy transmission across bank types (Tables 4-7 carry the full coefficient estimates); designing monetary policy that accounts for local banking market structure (the ZLB and asymmetric results in Tables 5-7 are particularly relevant to forward guidance); extending the baseline specification with additional controls or alternative periods; comparing with shadow bank transmission (see Enkhbold (2024) for the companion paper on traditional vs shadow banks); or replicating the deposit-market concentration and wholesale funding interaction documented by the paper.

## Attribution and rights

Source: peer-reviewed, *Journal of Banking and Finance* 187 (2026) 107690. Crown Copyright 2026, published by Elsevier B.V. All rights reserved, including text and data mining. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. No CC licence is in effect; redistribution requires publisher permission. Reproduced here as extract-only commentary.

> Enkhbold, Amina. "Monetary policy transmission, bank market power, and wholesale funding reliance." *Journal of Banking and Finance* 187 (2026): 107690. DOI: 10.1016/j.jbankfin.2026.107690.


==============================================================================
# Options Trading and Price Stability: Kim (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jbf/2026/kim-options-trading-stabilize-prices-2026/
# Using the SEC Penny Pilot Program as a natural experiment, Kim (2026) provides causal evidence that options trading reduces stock price volatility: a one-standard-deviation increase in options volume lowers total volatility by 1.21 percentage points via a liquidity buffer channel and a mispricing correction channel. Journal of Banking and Finance 185 (2026), paywalled. Six core results with source locators, datasets used, the identification strategy, and the regression specifications. LLM-distilled, not human-verified.
# Tags: paper-summary, options, volatility, equities, market-microstructure, price-stability
==============================================================================

**What this is.** A distilled summary of Da-Hea Kim, "Does options trading stabilize stock prices? Evidence from a natural experiment," *Journal of Banking and Finance* 185 (2026) 107612. It covers the six headline results with exact source locators, the identification strategy (Penny Pilot Program as a quasi-natural experiment), and the regression specifications. Read the original at [doi.org/10.1016/j.jbankfin.2025.107612](https://doi.org/10.1016/j.jbankfin.2025.107612) to replicate or extend.

## TL;DR

Kim (2026) revisits whether options trading stabilizes or destabilizes underlying stock prices. Pooled OLS regressions find a positive association between options volume and stock price volatility, but this reflects reverse causality: high-volatility stocks attract more options activity. Using the SEC's Penny Pilot Program (PPP), which reduced options tick sizes for roughly 500 underlying securities in a staggered fashion from 2007 to 2020, as an exogenous shock to options trading volume, instrumental variable regressions show that a one-standard-deviation increase in options volume reduces total volatility by 1.21 percentage points (46% of the mean). Difference-in-differences (DiD) regressions using a propensity-score-matched control sample confirm that pilot-stock total volatility falls by 0.21 percentage points (7.5% of pre-treatment average) relative to never-included controls. Two mechanisms drive the result: (1) the options market absorbs liquidity shocks to the underlying stock, reducing excessive trade concentration and extreme daily returns; and (2) options trading corrects mispricing by anchoring prices to intrinsic values through enhanced price discovery, as evidenced by declines in the Stambaugh, Yu, and Yuan (2015) mispricing score and in the Bogousslavsky and Muravyev (2024) informed trading intensity measures.

## Core results

Significance stars: `\*` 10%, `\*\*` 5%, `\*\*\*` 1%. Locators cite the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Pooled OLS finds a positive (endogenous) association between options volume and total stock volatility | Table 3, col. (1), p. 5 | Ln(OPTVOLM) on TVOL = +0.223\*\*\* (t = 60.55); 1-SD increase associated with +34% SD change in TVOL |
| R2 | IV (PPP as instrument) reverses the sign: increased options trading causally reduces total volatility | Table 4 Panel A, col. (2), p. 7 | Instrumented Ln(OPTVOLM) on TVOL = -0.435\*\*\* (t = -4.15); 1-SD increase in options volume -> -1.21 pp TVOL = 46% of mean = 67% of SD |
| R3 | DiD (matched sample): pilot firms' total volatility declines relative to controls after program inclusion | Table 6, col. (3), p. 9 | TREAT x POST on TVOL = -0.205\*\*\* (t = -2.73); 7.5% of pre-treatment mean; robust to additional controls and lagged TVOL |
| R4 | DiD: idiosyncratic volatility and extreme return range also decline for pilot firms | Table 6, cols. (7) and (9), p. 9 | TREAT x POST on IVOL = -0.134\*\* (t = -2.28); TREAT x POST on MAXMIN = -0.871\*\*\* (t = -2.65) |
| R5 | Liquidity buffer mechanism: options inclusion reduces excessive stock trade concentration and extreme daily price moves | Table 9, cols. (1) and (7), p. 14 | TREAT x POST on Ln(MAXVOLM/MEDVOLM) = -0.038\*\*\* (t = -2.95); TREAT x POST on MAXRET = -0.427\* (t = -2.26); 5-11% of pre-treatment mean |
| R6 | Mispricing correction mechanism: options trading reduces stock mispricing and informed trading intensity | Tables 10-11, cols. (1), p. 15 | TREAT x POST on MISP = -3.180\*\*\* (t = -3.87; 11-13% of pre-treatment); TREAT x POST on ITI_13D = -0.020\*\*\* (t = -4.71) |

**Overall (paper's conclusion).** The positive OLS correlation between options volume and stock price volatility reflects reverse causality. Once endogeneity is addressed via the Penny Pilot Program, the causal effect reverses: expanded options trading reduces stock price volatility. The stabilizing effect operates through both a liquidity buffer channel and a mispricing correction channel, and it holds across nine alternative matching methods, alternative sample periods (including the pre-2011 concentrated-addition period and the post-2009 cohort), and shorter event windows of plus or minus three or six months. The findings support the beneficial role of options markets in enhancing underlying equity price stability.

## Theory / model

The paper proposes no formal structural model. It tests two competing hypotheses drawn from the prior theoretical literature and identifies two empirical channels for its main result.

**Stabilization hypothesis.** Options markets facilitate information transmission, provide hedging opportunities, and offer an alternative avenue for liquidity demand, all of which may reduce volatility in the underlying stock (Grossman (1988); Figlewski and Webb (1993); Cao (1999)). On this view, options absorb excess supply or demand that would otherwise move stock prices, and their informational role anchors prices to fundamental values.

**Destabilization hypothesis.** Options attract noise traders and speculators, undermine stock market liquidity by diverting it toward derivatives, and enable new sources of volatility through leveraged positions (Stein (1987); Gorton and Pennacchi (1993)).

Prior empirical work gives mixed results; the central identification challenge is endogeneity (higher-volatility stocks endogenously attract more options trading). Cao et al. (2024) use the same PPP to show that expanded options trading improves price informativeness, but leave open whether the net effect on stock return volatility is positive or negative. This paper addresses that question directly.

**Two operative channels:**

1. **Liquidity buffer.** By providing hedging alternatives and an additional venue for investor liquidity demand, the options market can absorb excess supply or demand in the underlying stock, mitigating extreme price movements. Empirically tested via concentrated-trading metrics: Ln(MAXVOLM/MEDVOLM) (max-to-median daily volume within a month), the maximum daily turnover deviation MAXTURN - MEDTURN, and extreme return measures MAXRET and |MINRET| (Table 9).

2. **Mispricing correction.** Options trading enhances price discovery by incorporating diverse information into asset prices and enabling arbitrage between options and the underlying. As prices adjust more quickly to fundamental values, the scope for profitable informed trading diminishes and mispricing is reduced. Empirically tested via the Stambaugh, Yu, and Yuan (2015) composite mispricing score SYY_SCORE (based on 11 anomaly variables) and the Bogousslavsky and Muravyev (2024) informed trading intensity measures (Table 10, Table 11).

**Identification logic.** From January 2007 to June 2020 the SEC's Penny Pilot Program reduced minimum options tick sizes for selected securities (from nickel-and-dime to penny-and-nickel increments), lowering options trading costs and increasing volume. The selection of securities was designed to represent diverse trading characteristics rather than target volatility levels, supporting the exclusion restriction that inclusion affects stock volatility only through the channel of options trading volume. The program's staggered, phased implementation (securities added at eight distinct dates over 13 years) helps disentangle the effect of options trading from contemporaneous trends. Hao and Li (2022) and Anagnostopoulou et al. (2023) use the same program in related settings.

## Method

Two complementary strategies identify the causal effect of options trading on stock price volatility.

**Strategy 1: Instrumental variable (IV).** The PPP inclusion status (TREAT, time-invariant) interacted with post-inclusion months (POST) instruments for options trading volume. The first-stage regression (Table 4, Panel A, col. (1)) is:

$$
\text{OptionsTrading}_{it} = \alpha + \beta \bigl(\text{TREAT}_i \times \text{POST}_{it}\bigr) + \gamma \, \text{Controls}_{i,t-1} + \varepsilon_{it} \tag{1}
$$

where TREAT equals 1 if firm $$i$$'s options are included in the PPP (0 otherwise) and POST equals 1 for months after inclusion. The first-stage coefficient $$\hat{\beta}$$ on TREAT $$\times$$ POST is 0.492 (t = 11.74), confirming a strong instrument. The second-stage regression (Table 4, Panel A, cols. (2)-(4)) is:

$$
\text{VolatilityProxy}_{it} = \delta + \zeta \, \widehat{\text{OptionsTrading}}_{it} + \lambda \, \text{Controls}_{i,t-1} + \varepsilon_{it} \tag{2}
$$

Two interchangeable options activity measures are used: the natural logarithm of total monthly options contracts traded, Ln(OPTVOLM), and the option-to-stock volume ratio O/S = OPTVOLM / (STKVOL x 100). Both panels of Table 4 yield consistently negative second-stage coefficients. All regressions include firm and month fixed effects; standard errors are clustered by firm.

**Strategy 2: Difference-in-differences (DiD) with propensity-score-matched controls.** Treated firms (pilot options included in the PPP) are matched 1:1 to never-treated controls using nearest-neighbor logit propensity-score matching without replacement (caliper 0.2), yielding 264 matched pairs. Matching covariates are TVOL, STKRET, SIZE, STKVOL, IOR, MOM, BM, and ROA measured in the month before inclusion. The baseline DiD regression (Table 6) uses a 25-month event window (12 months before and after inclusion):

$$
\text{Volatility}_{it} = \alpha + \beta \bigl(\text{TREAT}_i \times \text{POST}_{it}\bigr) + \gamma \, \text{Controls}_{i,t-1} + \varepsilon_{it} \tag{3}
$$

The coefficient $$\beta$$ on TREAT $$\times$$ POST identifies the within-matched-pair change in volatility attributable to program inclusion. Placebo tests (Table 7) assign pseudo-treatment dates six months before actual inclusion and confirm that TREAT $$\times$$ PseudoPOST coefficients are statistically indistinguishable from zero for all volatility outcomes, supporting the parallel-trends assumption. The dynamic DiD specification (Fig. 3) replaces POST with event-window dummies $$D_{ik}$$ ($$k = -11, \ldots, 12$$):

$$
\text{TVOL}_{it} = \alpha + \sum_{k=-11}^{12} \beta_{1k} \,\text{TREAT}_i \times D_{ik} + \beta_2 \,\text{TREAT}_i + \sum_{k=-11}^{12} \beta_{3k} D_{ik} + \gamma \,\text{Controls}_{i,t-1} + \varepsilon_{it} \tag{4}
$$

Pre-addition $$\beta_{1k}$$ estimates are flat and statistically insignificant; post-addition estimates turn significantly negative, tracing the onset of the stabilization effect.

## Empirical specifications

**Pooled OLS baseline (Table 3).** Sample: 380,064 firm-month observations, 2006-2021. Dependent variables are TVOL (cols. 1-4), IVOL (cols. 5-8), MAXMIN (cols. 9-12). Options activity is Ln(OPTVOLM) or O/S. Controls (lagged one month): STKRET, Ln(SIZE), Ln(STKVOL), IOR, MOM, BM, ROA. Columns (2), (4), (6), (8), (10), (12) also include the lagged dependent variable. Firm and month fixed effects; standard errors clustered by firm. Result R1 comes from col. (1).

**IV specifications (Table 4).** Same sample and controls. Panel A uses Ln(OPTVOLM) as the endogenous variable, Panel B uses O/S. Second-stage dependent variables: TVOL (col. 2 or 6), IVOL (col. 3 or 7), MAXMIN (col. 4 or 8). Columns (5)-(8) in each panel add the lagged volatility as an additional control. Result R2 comes from Panel A, col. (2). The standard deviation of Ln(OPTVOLM) is 2.78 (footnote 12), so the 1-SD partial effect is $$-0.435 \times 2.78 = -1.21$$ pp TVOL.

**DiD baseline (Table 6).** Sample: up to 13,020 firm-month observations from 264 matched pairs; 25-month event window (12,667 where additional controls require options trading data). Cols. (1)-(2) use options trading as the dependent variable (confirming the first-stage effect of PPP inclusion). Cols. (3)-(5) use TVOL with progressively richer controls (col. 5 adds lagged TVOL). Cols. (6)-(8) use IVOL; cols. (9)-(11) use MAXMIN. Results R3 and R4 come from cols. (3) and (7)/(9), respectively.

**Liquidity buffer channel (Table 9).** Same matched sample and DiD specification. Dependent variables: Ln(MAXVOLM/MEDVOLM) (log max-to-median daily volume ratio, cols. 1-2), (MAXVOLM - MEDVOLM)/MEDVOLM (percentage deviation, cols. 3-4), MAXTURN - MEDTURN (daily turnover range, cols. 5-6), MAXRET (maximum daily return, cols. 7-8), |MINRET| (absolute minimum daily return, cols. 9-10). Result R5 comes from cols. (1) and (7).

**Mispricing correction channel (Table 10).** Sample of mispriced stocks (SYY_SCORE > 70 or < 30). Dependent variables: MISP = |SYY_SCORE - 50| (cols. 1-2), SYY_SCORE for underpriced stocks (cols. 3-4), SYY_SCORE for overpriced stocks (cols. 5-6). Additional robustness in Table 11 uses five Bogousslavsky and Muravyev (2024) ITI measures (ITI_13D, ITI_Impatient, ITI_Patient, ITI_Insider, ITI_Short) as dependent variables. Result R6 comes from Table 10, col. (1) and Table 11, col. (1).

**Robustness (Tables 12-14).** Nine alternative matching methods (Mahalanobis distance and eight covariate-cell-based procedures following Cao et al. (2024)), alternative sample periods (2006-2011 only and excluding the Global Financial Crisis through June 2009), and shorter event windows of plus or minus three and six months around the inclusion month. Results are consistent in sign and significance across all specifications.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Ivy OptionMetrics | Options trading volume, open interest, implied volatility, bid-ask quotes for all exchange-listed options (2006-2021) | [OptionMetrics](/wiki/commercial/optionmetrics/) |
| CRSP | Stock returns, share prices, trading volume, shares outstanding for underlying stocks | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| Compustat | General accounting data: book-to-market ratio, return on assets | [WRDS / Compustat](/wiki/commercial/wrds/) (licensed) |
| Thomson Reuters 13F | Institutional ownership ratio (IOR) for matching and controls | [WRDS / 13F](/wiki/commercial/wrds/) (licensed) |
| Stambaugh, Yu, and Yuan (2015) SYY_SCORE | Composite mispricing score (percentile rank on 11 anomalies); downloaded from R. F. Stambaugh's website | No page yet |
| Bogousslavsky, Fos, and Muravyev (2024) ITI | Informed Trading Intensity measures (ITI_13D, ITI_Impatient, ITI_Patient, ITI_Insider, ITI_Short); downloaded from D. Muravyev's website | No page yet |
| CBOE SEC filings (hand-collected) | Penny Pilot Program inclusion schedule: 460 securities, 8 addition events, 2007-2020 | No page yet |

Sample: January 2006 to December 2021 (192 months). The main matched-sample analysis covers a 25-month window around each inclusion event; the IV analysis uses the full 380,219 firm-month panel. Options and stock data are merged at the underlying-security level; accounting data are lagged one year.

## When to read the full paper

Read the source at [doi.org/10.1016/j.jbankfin.2025.107612](https://doi.org/10.1016/j.jbankfin.2025.107612) if you are:
- assessing whether options markets stabilize or destabilize underlying equity prices (the causal IV and DiD evidence is in Tables 4 and 6);
- studying the liquidity buffer or mispricing correction channels of derivatives markets (Tables 9-11);
- using the Penny Pilot Program as a quasi-natural experiment for options trading (compare with Cao et al. (2024) for price informativeness; Hao and Li (2022) for earnings management implications);
- evaluating the robustness methodology for staggered DiD with many alternative matching procedures (Table 12) or shorter event windows (Table 14);
- benchmarking results against the recent retail and zero-day-to-expiration options literature.

## Attribution and rights

Source: peer-reviewed, *Journal of Banking and Finance* 185 (2026) 107612. DOI: 10.1016/j.jbankfin.2025.107612. Published by Elsevier B.V. All rights reserved. This distillation was extracted by an LLM (claude-sonnet-4-6) on 2026-06-25 and is **not human-verified or independently reproduced**. The article is paywalled; only textual extracts are permitted here.

> Kim, Da-Hea. "Does options trading stabilize stock prices? Evidence from a natural experiment." *Journal of Banking and Finance* 185 (2026): 107612. DOI: 10.1016/j.jbankfin.2025.107612. Copyright 2025 Elsevier B.V. All rights reserved.


==============================================================================
# Election Cycles and Systemic Risk: Kladakis & Skouralis (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jbf/2026/kladakis-election-cycles-systemic-risk-2026/
# Distilled: Election years are associated with significantly higher bank systemic risk across 22 OECD economies (2000-2023), with ΔCoVaR rising 3.57% above the overall average in the election year, while the pre-election period shows a decline. The effect is stronger for snap elections, new-government outcomes, and common-law countries; macroprudential tightening mitigates it. Journal of Banking and Finance 2026, CC BY 4.0. Eight core results with source locators, datasets used, the ΔCoVaR estimation method, and the panel regression specification.
# Tags: paper-summary, systemic-risk, elections, political-economy, banking
==============================================================================

**What this is.** The paper's core results, the ΔCoVaR methodology it applies (from Adrian
and Brunnermeier 2016), and the panel regression specification: enough to know what it found
and how, without reading all 26 pages. To replicate or extend it, read the full source at the
[original](https://doi.org/10.1016/j.jbankfin.2026.107676).

## TL;DR

Kladakis and Skouralis examine whether national elections are associated with higher bank
systemic risk using a panel of 193 banks from 22 OECD economies, covering 147 elections
over 2000-2023. Systemic risk is measured by ΔCoVaR (the additional tail risk to the
financial system when an institution is in distress), following Adrian and Brunnermeier
(2016). The central finding is a robust, time-varying relationship: bank ΔCoVaR rises by
approximately 3.57% above the overall mean in the election year, but the effect is
heterogeneous across the electoral cycle. In the pre-election period, suppressed negative
information and expansionary fiscal policies push systemic risk downward (-2.19%); the
surge occurs at election time and in the post-election period. Snap elections drive larger
increases than scheduled end-of-term elections, incumbent turnover amplifies the effect while
re-election dampens it, common-law countries show a stronger response than civil-law
jurisdictions, and macroprudential policy tightening can partially offset the election-driven
rise in systemic risk. Results are robust to alternative systemic risk measures (MES, SRISK),
instrumental-variable estimation (term limits; Google Trends uncertainty index), exclusion of
banking-crisis years, and monthly data.

## Core results

Magnitudes and significance are as reported; `\*\*`/`\*\*\*` = 5%/1%. Locators point into the
source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Elections raise bank systemic risk** in the election year | Table 3 Model (3), p. 11 | ELECTIONS: 0.062\*\*\* (SE 0.012); the election year is associated with ΔCoVaR 3.57% above the overall mean (mean = 1.737%) |
| R2 | **Pre-election period: systemic risk falls; post-election period: systemic risk rises** | Table 4 Models (1)(4), p. 12 | PRE (year before election): -0.032\*\*\* (SE 0.009); POST (year after election): +0.037\*\*\* (SE 0.010) |
| R3 | **Snap elections drive larger systemic risk increases than end-of-term elections** | Table 3 Models (4)(5), p. 9-11 | SNAP: 0.084\*\*\* (SE 0.023); END-OF-TERM: 0.053\*\*\* (SE 0.013); snap elections increase ΔCoVaR by 4.83% vs 3.05% for end-of-term (relative to mean) |
| R4 | **Incumbent turnover amplifies effect; re-election dampens it** | Table 3 Model (6), p. 10-11 | NEW GOV: 0.084\*\*\* (SE 0.017); RE-ELECTED: 0.043\*\* (SE 0.018); new government coefficient is roughly twice that under re-election |
| R5 | **Effect extends to all financial institutions** (banks + insurance + investment trusts) | Table 6 Model (1), p. 14 | ELECTIONS: 0.043\*\*\* (SE 0.006); sample of 697 institutions; snap: 0.056\*\*\*, end-of-term: 0.042\*\*\* |
| R6 | **Common-law countries show a stronger elections-systemic risk link** than civil-law countries | Table 9 Model (1), p. 17 | LEGAL ORIGIN × ELECTIONS: 0.049\*\*\* (SE 0.022); ELECTIONS main coefficient: 0.043\*\*\* (SE 0.014) |
| R7 | **Macroprudential policy tightening mitigates election-related systemic risk** | Table 10 Model (2), p. 17 | MP TIGHTENING × ELECTIONS: -0.121\*\*\* (SE 0.026); effect remains positive in tightening years but is significantly reduced |
| R8 | **2SLS with term limits IV confirms positive election effect**, addressing reverse causality | Table 13 Models (2)(4), pp. 18-19 | 2SLS ELECTIONS: 0.045\*\*\* (SE 0.011); first stage: ELECTIONS on TERM LIMITS 0.685\*\*\* (SE 0.007); consistent with OLS |

**Overall (paper's conclusion).** Elections are associated with a robust and time-varying
increase in bank systemic risk. The hump-shaped trajectory (lower pre-election, peak around
election time and first two post-election quarters, gradual decline over 12 months) is driven
primarily by turnover episodes: in re-elected cases the post-election rise is smaller and
shorter-lived. Multiple transmission channels are evidenced: stock market volatility (VIX
interaction), suppressed stock price informativeness (reduced transparency), expansionary
fiscal policies before elections, and declining trust in government. Macroprudential
tightening, strong economic growth, and high public trust in government partially buffer the
effect.

## Theory / model

The paper proposes no formal economic model. It develops six testable hypotheses grounded in
the prior literature on political uncertainty and financial markets.

**Competing hypotheses on election-period systemic risk.**

- H1a: Election periods are associated with increased systemic risk (via heightened policy
  uncertainty, reduced information disclosure, and amplified market volatility).
- H1b: Election periods are associated with reduced systemic risk (via uncertainty resolution,
  improved investor confidence when a competent government is expected, and credible policy
  commitments).
- H2: Snap elections are associated with increased systemic risk relative to scheduled
  elections (owing to their unexpected nature and greater uncertainty about outcomes).
- H3: Re-election of the incumbent is associated with reduced systemic risk (via continuity
  and reduced policy risk).
- H4a/H4b: Systemic risk is increased/reduced in the pre-election period (depending on
  whether anticipatory political risk or information suppression and fiscal stimulus dominate).
- H5: The impact of elections on systemic risk is stronger in common-law countries (where
  more market-based financial systems transmit political shocks more directly through asset
  prices and intermediaries).
- H6: Macroprudential policy can mitigate election-related systemic risk (by strengthening
  system resilience against political-economic shocks).

**Identification logic.** The core identification challenge is that systemic risk may itself
influence election timing (reverse causality: distressed governments may call early elections,
or delay elections to avoid political punishment). The prior literature documents that
Bialkowski et al. (2008) find country-specific stock market volatility roughly doubles in the
week around a national election, and Matousek et al. (2020) show that policy uncertainty
exerts a significant and persistent impact on bank capital shortfall, peaking around
11 months after elections, providing direct motivation for the systemic-risk focus here. The paper addresses this via (i) the
argument that national election schedules in parliamentary democracies are largely exogenous
to individual bank risk (particularly for scheduled end-of-term elections); (ii) a 2SLS
approach using term limits (Jens 2017) as an instrument for election occurrence, which is
predetermined and uncorrelated with contemporaneous financial conditions; and (iii) an
additional instrument based on Google Trends election search intensity. The authors also
exclude years in which banking crises occurred (Harvard Global Crisis Data; Metrick and
Schmelzing 2021) and run a separate sub-sample restricted to US presidential elections, which
occur at fixed intervals.

## Method

The paper applies the ΔCoVaR methodology of Adrian and Brunnermeier (2016) to construct
the systemic risk measure. The estimation has three steps (Eqs. 4-8, p. 6).

**Step 1: institution-level VaR.** For each financial institution $$i$$, run a quantile
regression of weekly returns $$R^i_t$$ on state variables $$S_{t-1}$$ (stock market returns,
short-term government bond yield change, and the 10Y-to-short-term yield spread) at the
distress quantile $$q = 0.05$$ (Eq. 4, p. 6):

$$
R^i_t = a_q + \beta_q S_{t-1} + \varepsilon_{q,t} \tag{4}
$$

$$
\widehat{\text{VaR}}^i_{q,t} = \hat{a}_q + \hat{\beta}_q S_{t-1} \tag{5}
$$

**Step 2: system CoVaR.** For the country-level financial system index (returns
$$R^{\text{system}}_t$$), run a second quantile regression conditioning on institution $$i$$'s
return (Eq. 6, p. 6):

$$
R^{\text{system}}_t = a_q^{\text{system}} + \beta_q^{\text{system}} S_{t-1} + \gamma_q^{\text{system}} R^i_t + \varepsilon_{q,t} \tag{6}
$$

$$
\widehat{\text{CoVaR}}^{si}_{q,t} = \hat{a}_q^{\text{system}} + \hat{\beta}_q^{\text{system}} S_{t-1} + \hat{\gamma}_q^{\text{system}} \widehat{\text{VaR}}^i_t \tag{7}
$$

**Step 3: ΔCoVaR.** The systemic importance measure is the difference between the system's
CoVaR when institution $$i$$ is at its distress level ($$q = 0.05$$) and when it is at its
median ($$q = 0.5$$) (Eq. 8, p. 6):

$$
\Delta\text{CoVaR}^{si} = \widehat{\text{CoVaR}}^{si}_{q=0.05} - \widehat{\text{CoVaR}}^{si}_{q=0.5} \tag{8}
$$

The system index $$R^{\text{system}}_t$$ is the return of the DS Financials country index
from Thomson Reuters EIKON Datastream, which includes large listed financial institutions in
each country. All data are weekly. The resulting annual average of ΔCoVaR is the dependent
variable in the panel regressions.

The paper also uses two alternative systemic risk measures for robustness (Table 11, p. 18):
Marginal Expected Shortfall (MES) from Acharya et al. (2017), and SRISK from
Brownlees and Engle (2017). MES is the expected equity loss of institution $$i$$ when the
market experiences an extreme loss. SRISK is the expected capital shortfall conditional on a
systemic event:

$$
\text{SRISK}_{i,t} = k \cdot \text{DEBT}_{i,t} - (1 - k) \cdot W_{i,t} \cdot (1 - \text{LRMES}_{i,t}) \tag{15}
$$

where $$k = 0.08$$ is the prudential capital fraction, $$W_{i,t}$$ is market capitalization,
and $$\text{LRMES}_{i,t}$$ is the long-run marginal expected shortfall (p. 15, Eq. 15).

## Empirical specifications

**Benchmark panel regression.** The headline specification (Eq. 9, p. 7) is a panel fixed-
effects regression of annual ΔCoVaR on an election dummy and controls:

$$
\Delta\text{CoVaR}^{ci}_t = \beta_0 + \beta_1 \text{ELECTIONS}_{c,t} + \beta_2 X_{L,t-1} + \beta_3 M_{c,t-1} + \alpha_i + \alpha_t + \varepsilon_{i,t} \tag{9}
$$

where subscripts $$t$$, $$i$$, $$c$$, and $$s$$ refer to year, firm, country, and financial-
system index. $$\text{ELECTIONS}_{c,t} = 1$$ in years when national elections occurred in
country $$c$$. $$X_{L,t-1}$$ is a vector of lagged firm controls: log total assets (size),
VaR (idiosyncratic risk), leverage (total debt to market-cap ratio), and ROE (profitability).
$$M_{c,t-1}$$ is a vector of lagged country-level controls: GDP growth, inflation, real house
price growth, and credit growth to non-financials. $$\alpha_i$$ and $$\alpha_t$$ are firm and
year fixed effects. Standard errors are clustered at the firm level.

The sample is 193 banks from 22 OECD countries, yielding 3,827 firm-year observations in the
full-control specification (Table 3 Model 3, p. 11). The ELECTIONS dummy is split into SNAP
(elections called before end of term) and END-OF-TERM (within six months of term limit) to
test H2, and into RE-ELECTED and NEW GOV based on the electoral outcome to test H3 (Table 3,
Models 4-7). Pre- and post-election dynamics are examined by replacing ELECTIONS with PRE
(year before elections) and POST (year after) in Table 4.

**2SLS instrumental-variable specification.** To address reverse causality (Table 13, p. 19),
ELECTIONS is instrumented by two variables. The first is TERM LIMITS: a dummy equal to one
if the country's constitution or law prohibits the incumbent government from seeking
re-election; Jens (2017) shows term limits are strongly correlated with election timing but
unrelated to financial conditions. The second is a Google Trends political uncertainty index
(GT Political Uncertainty dummy): the equally-weighted sum of standardized search volumes for
election-related terms, equal to one if the index exceeds the upper quartile of its country
distribution. First-stage coefficient on TERM LIMITS: 0.685\*\*\* (SE 0.007); on
GT Political Uncertainty dummy: 0.356\*\*\*. Second-stage ELECTIONS coefficient: 0.045\*\*\*
(SE 0.011, Table 13 Model 2), close to the OLS estimate of 0.062.

**Transmission-channel tests.** Table 8 (p. 16) introduces four transmission-channel
interaction terms one at a time, each interacted with ELECTIONS: (1) VIX (stock market
volatility index): positive and significant interaction (VIX × ELECTIONS: 0.073\*\*\*,
SE 0.025), confirming that market-sentiment amplifies election effects; (2) PRICE\_INFO (stock
price informativeness, measured as the country average bid-ask spread, lower = more
informative): negative interaction (PRICE\_INFO × ELECTIONS: -0.001\*\*\*, SE 0.000),
consistent with reduced informativeness dampening the election-risk link; (3) GDP growth ×
ELECTIONS: -0.040\*\*\* (SE 0.005), confirming strong economic growth mitigates the effect;
(4) GOV.EXP × ELECTIONS: -0.002\* (SE 0.001), fiscal expansion partially buffers.
These results tie back to the hypotheses in Section 2 and establish that multiple channels
operate simultaneously.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Thomson Reuters EIKON Datastream (DS Financials index) | Weekly stock return data for all financial institutions; country financial system index; firm-level total assets, ROE, leverage, VaR | [no page yet](/wiki/commercial/) |
| OECD database | Country-level GDP growth and inflation (year-on-year) | no page yet |
| BIS | Real residential property price growth; credit growth to non-financial sector | no page yet |
| iMaPP (IMF / Alam et al. 2019) | Macroprudential policy indicators: countercyclical capital buffer, LTV, LTD, DSTI, stress tests, SIFI measures; annual aggregate at country level | no page yet |
| National election databases (22 OECD countries) | Date, type (snap vs. end-of-term), and outcome (re-elected vs. new government) of 147 national elections, 2000-2023; collected by the authors from national sources | no data: tag (hand-collected) |
| Harvard Global Crisis Data (Reinhart-Rogoff 2014) | Banking crisis dates for banking-crisis robustness exclusion | no page yet |
| Metrick-Schmelzing (2021) banking crisis dataset | Alternative banking crisis dates for robustness | no page yet |
| Baker et al. (2016) EPU index | Economic Policy Uncertainty index; robustness subsample of 12 OECD countries | no page yet |

Sample: 2000-2023 (annual), 22 OECD countries, 193 banks in the main sample (3,827 firm-years).
ΔCoVaR estimation uses weekly returns. Extended sample includes 697 financial institutions
(banks, insurance companies, financial services companies, investment trusts). Macroeconomic
controls are from OECD (GDP, inflation) and BIS (house prices, credit).

## When to read the full paper

Use the [original](https://doi.org/10.1016/j.jbankfin.2026.107676) if you are:
studying how political events transmit into tail risk measures (Tables 3-5 give the full
decomposition by election type and outcome); building a stress-testing framework that
incorporates election cycles (Section 4.7 / Table 10 on the macroprudential buffer channel
is the most policy-relevant section); extending the ΔCoVaR approach to other political
events; or investigating legal-origin heterogeneity in political-financial transmission
(Table 9 and Section 4.6). Figure 5 (p. 22) shows the monthly impulse-response of systemic
risk around elections and Figure 6 replicates it on US data alone.

## Attribution and rights

Source: peer-reviewed, *Journal of Banking and Finance* 187 (2026) 107676. This distillation
was extracted by an LLM on 2026-06-25 and is **not human-verified or independently
reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in
this batch.

> **Attribution (CC BY 4.0).** Kladakis, George, and Alexandros Skouralis.
> "Election cycles and systemic risk."
> *Journal of Banking and Finance* 187 (2026) 107676.
> DOI: 10.1016/j.jbankfin.2026.107676. © 2026 The Author(s).
> Published by Elsevier B.V. Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Air Pollution and Bank Loan Pricing: Li et al. (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jbf/2026/li-air-pollution-bank-loan-pricing-2026/
# Distilled: Using proprietary loan data from a Chinese state-owned commercial bank linked to firm-level ESR emissions, Li et al. find that higher air pollutant intensity significantly raises bank loan spreads via labor risk and environmental transition risk channels, confirmed causal by a PSM-DID design around China's 2013 Air Pollution Control Action Plan. Journal of Banking and Finance 185 (2026), paywalled. Eight core results with source locators, datasets, and estimating specifications.
# Tags: paper-summary, environmental-finance, bank-lending, air-pollution, credit-pricing
==============================================================================

**What this is.** The paper's core results, identification strategy, estimating equations, and key datasets: enough to understand what it found and how, without reading the full paper. To replicate or extend, read the original at [https://doi.org/10.1016/j.jbankfin.2026.107655](https://doi.org/10.1016/j.jbankfin.2026.107655).

## TL;DR

Using a proprietary loan dataset from a nationwide Chinese state-owned commercial bank (55,483 loan observations, 745 borrowing firms in 175 cities, 2007-2019) linked to firm-level air pollutant emissions from China's Environmental Survey and Reporting (ESR) database, this paper documents that firms with higher air pollution intensity pay significantly higher bank loan spreads. In the baseline specification, a one-standard-deviation increase in air pollutant emissions is associated with a 17.85% increase relative to the mean loan spread (Table 2, Col. 4, p. 5). The effect is confirmed causal via propensity-score matching combined with a difference-in-differences design around China's 2013 Air Pollution Prevention and Control Action Plan: after the plan, treated high-pollution firms saw significantly lower loan pricing (Table 3). Two economic channels explain the premium: (1) a labor risk channel, where pollution undermines retention of skilled employees and raises default probability; and (2) an environmental transition risk channel, where high-pollution firms face elevated regulatory compliance costs. Both channels are confirmed through interaction tests and survive horse-racing together. The premium is concentrated in politically unconnected firms, in provinces with weak environmental governance, and for non-headquarter borrowers, and extends to non-price terms: pollution also reduces total credit volume extended to borrowers and increases mortgage-related contract requirements.

## Core results

Magnitudes as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%. Locators point to the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Air pollutant emissions positively predict bank loan spread in the full baseline specification with firm, year, borrower-category, and risk-type fixed effects | Table 2, Col. (4), p. 5 | Airpollution = 0.047\*\*\* (t=3.20); 1-SD increase associated with 17.85% of mean loan spread; N=55,483 |
| R2 | PSM-OLS confirms the baseline: positive and significant after controlling for self-selection of high-polluting firms | Table 3, Panel B, Col. (1), p. 6 | Airpollution = 0.062\*\*\* (t=4.58); N=42,530 |
| R3 | PSM-DID: the 2013 Air Pollution Control Action Plan significantly reduces loan pricing for treated high-pollution firms; parallel trends confirmed pre-shock | Table 3, Panel B, Col. (2), p. 6; Table 4, p. 7 | Treatpost = -0.124\*\*\* (t=-3.07); pre-period interactions Pre3\*Treat and Pre2\*Treat insignificant |
| R4 | Labor risk channel: pollution premium concentrated in firms lacking high-skill workers or with higher bankruptcy risk | Table 5, Cols. (1)-(3), p. 8 | Airpollution × Educated\_dummy = 0.031\*\* (t=2.33); × Technical\_dummy = 0.028\*\* (t=2.00); × Bankruptcy\_dummy = 0.039\*\*\* (t=3.34) |
| R5 | Transition risk channel: premium concentrated in firms with weak green innovation, rich environmental violation records, or low green subsidy support | Table 6, Cols. (1)-(3), p. 8 | Airpollution × Greeninno\_dummy = 0.024\*\* (t=1.97); × Violation\_dummy = 0.065\*\* (t=2.40); × Gsubsidy\_dummy = 0.021\*\* (t=2.00) |
| R6 | Cross-section: premium significant only for politically unconnected firms; banks discount pollution risk for firms with implicit political insurance | Table 8, Cols. (1)-(2), p. 10 | Political=0: 0.052\*\*\* (t=3.16); Political=1: 0.018 (t=0.97); subsample difference p=0.000 |
| R7 | Cross-section: premium significant only in provinces with weak environmental governance; strong governance regions show no effect | Table 8, Cols. (3)-(4), p. 10 | Weak governance: 0.079\*\*\* (t=4.67); strong governance: -0.002 (t=-0.11); difference p=0.000 |
| R8 | Air pollution tightens non-price contract terms: reduces total credit volume and increases mortgage requirements | Table 9, Cols. (1)-(2), p. 11 | Sum\_volume: Airpollution = -0.017\* (t=-1.67); Mortgage\_terms: 0.009\*\* (t=2.08) |

**Overall (paper's conclusion).** Firm-level air pollution intensity, measured by annual ESR emissions, is a significant and causal predictor of bank loan spreads in China. The effect operates through two complementary channels (labor risk and transition risk) that are both statistically independent and economically distinct. The premium is not uniform: it is largely absent for politically connected firms and in provinces with stringent environmental enforcement, consistent with banks pricing the residual uninsured risk. Beyond price, pollution also constrains the non-price terms of credit, reducing total loan availability and tightening collateral requirements. The findings extend to multiple emission types (SO2, NOx, smoke dust), with smoke emissions having the strongest effect (Table A1, p. 14). Bolton and Kacperczyk (2021) established pollution premiums in equity markets; this paper establishes the same channel in bank debt markets using firm-level Chinese data, extending the evidence beyond the US equity context of Hsu et al. (2023).

## Theory / model

This paper has no formal theoretical model. The two tested hypotheses are derived from prior literature and stated as verbal mechanisms:

**Hypothesis 1 (main).** *Ceteris paribus*, firms with higher air pollutant emissions intensity are charged higher bank loan prices than those with lower emissions intensity (pp. 2-3).

**Channel 1: Labor risk.** Air pollutant emissions have an adverse impact on employee health and retention, especially for skilled workers with higher education or technical background. This results in lower operational efficiency and greater default risk for the borrower. Following Liu and Yu (2020) and Xue et al. (2021), employees at high-pollution firms "vote with their feet," depressing productivity and raising expected credit losses for the bank lender. The mechanism is tested by interacting emissions with borrower-level proxies for the proportion of educated staff (Educated\_dummy), technical staff (Technical\_dummy), and bankruptcy risk (Bankruptcy\_dummy).

**Channel 2: Transition risk.** Schneider (2011) and Hsu et al. (2023) document that environmental transition risk, the cost of shifting toward a low-carbon economy, is elevated for high-pollution firms. Funding green production, meeting regulatory obligations, and managing the risk of factory shutdowns or contract termination raises the cash-flow concerns that lenders must price. The mechanism is tested via interactions with green innovation experience (Greeninno\_dummy), environmental violation records (Violation\_dummy), and government green subsidies (Gsubsidy\_dummy).

**Identification logic.** Because air pollution and loan pricing may share unobserved determinants, the paper exploits the Air Pollution Prevention and Control Action Plan promulgated by China's State Council in 2013 as an exogenous shock. Firms above the industry-year median emission intensity are designated as treated; propensity-score matching (PSM) balances pre-treatment covariates. The DID design compares treated and control firms before and after 2013, isolating the causal effect of the regulatory reduction in pollution on loan pricing. The parallel trend assumption is validated by insignificant pre-period coefficients in Table 4 (p. 7).

## Method

The paper applies standard OLS panel regression for the baseline, and a two-step PSM-DID for causal identification. It builds on `panel-regression` (multi-way fixed effects), `difference-in-differences` (pre/post policy shock), and `matching` (propensity-score matching on firm characteristics).

**Baseline OLS.** The estimating equation (Eq. 1, p. 3) is:

$$
\text{Loan\_spread}_{ijt} = \alpha + \beta \, \text{Airpollution}_{it-1} + \delta \, \text{Controls}_{t-1} + \text{FirmFE} + \text{LoanAttributesFE} + \text{YearFE} + \varepsilon_{ijt}
$$

where $$\text{Loan\_spread}_{ijt}$$ is the actual loan rate on loan $$j$$ of firm $$i$$ in year $$t$$ minus the PBOC benchmark rate, $$\text{Airpollution}_{it-1}$$ is the year-specific air pollutant emission volume (tons) lagged one year to mitigate endogeneity, and $$\text{Controls}_{t-1}$$ includes loan attributes (Volume, Maturity, Collateral) and firm attributes (Firmsize, Lev, Soe, Roa, Growth, Tangible, Age, BM, Cash\_holding). Fixed effects span firm (FirmFE), year (YearFE), borrower category (Borrower\_category\_FE), and loan risk type (Risk\_type\_FE). Standard errors are clustered at the firm level (p. 3-4).

**PSM-DID.** After PSM matching on 2012 (pre-policy) firm characteristics, the DID equation (Eq. 2, p. 4) is:

$$
\text{Loan\_spread}_{ijt} = \alpha + \beta \, \text{Treatpost}_{it-1} + \delta \, \text{Controls} + \text{FirmFE} + \text{LoanAttributesFE} + \text{YearFE} + \varepsilon_{ijt}
$$

where $$\text{Treatpost}$$ equals one for firms above the industry-year median emission intensity in the post-2013 period. The coefficient $$\beta$$ identifies the policy-induced change in loan pricing for treated firms relative to matched controls. Matching variables are Firmsize, Lev, Growth, Roa, Age, BM, and Cash\_holding; 1:1 nearest-neighbor matching without replacement is used. Balance tests are in Table 3, Panel A (p. 6).

## Empirical specifications

**Baseline (R1).** Table 2 (p. 5) reports four progressively richer columns: (1) air pollution and explained variable only, no FE; (2) firm and year FE; (3) adds firm-level controls; (4) adds loan-level controls plus borrower-category and risk-type FE. The headline result from Col. (4): Airpollution = 0.047\*\*\* (t=3.20), implying a 17.85% of mean loan spread increase per standard deviation of emissions (calculation: 0.047 × 4.783 / 1.259, where 4.783 = SD of Airpollution and 1.259 = mean of Loan\_spread, from Table 1).

**PSM-DID (R2, R3).** Table 3, Panel B (p. 6), Col. (1) is PSM-OLS on the matched sample: Airpollution = 0.062\*\*\* (t=4.58). Col. (2) is PSM-DID: Treatpost = -0.124\*\*\* (t=-3.07), confirming that the 2013 regulation causally reduces spreads for high-pollution firms. All columns include firm, year, borrower-category, and risk-type FE. The Graham et al. (2008) framework for loan-pricing determinants motivates the choice of control variables.

**Parallel trend test (R3 supplement).** Table 4 (p. 7) replaces Treatpost with interactions of Treat with year dummies. Pre-period interactions (Pre3\*Treat, Pre2\*Treat) are insignificant (t=0.71, -0.57). Post-period interactions (Post1\*Treat, Post2\*Treat) are significantly negative (t=-3.58, -1.92), consistent with causal post-policy reduction. Figure 1 (p. 7) plots the coefficient path with 95% confidence intervals.

**Mechanism tests (R4, R5).** Tables 5-6 (pp. 8) add triple-interaction terms Airpollution × mechanism\_dummy to Eq. (1). Subgroup dummies (Educated\_dummy, Technical\_dummy, Bankruptcy\_dummy for labor risk; Greeninno\_dummy, Violation\_dummy, Gsubsidy\_dummy for transition risk) take the value one for firms in the riskier half of each proxy distribution. All interaction coefficients are positive and significant at 5-10%, confirming both channels. Table A3 (p. 15) horse-races all six interactions jointly; Bankruptcy\_dummy (t=3.58) and Violation\_dummy (t=2.24) and Gsubsidy\_dummy (t=2.85) remain significant, supporting channel independence.

**Cross-sectional heterogeneity (R6, R7).** Table 8 (p. 10) splits the sample by political connection (Political = 0 vs. 1), regional environmental governance (Envir\_governance = 0 vs. 1 based on province-level pollution treatment cost above/below median), and borrower headquarters status. Bootstrap p-values test coefficient equality across subsamples. For Houston et al. (2014), politically connected borrowers face an implicit government guarantee that attenuates the pollution premium; the cross-section confirms this by showing zero effect for Political=1 firms. Barwick et al. (2024) find environmental regulation eases creditor concerns; this is consistent with the PSM-DID result and the strong-governance subsample showing no pollution premium.

**Non-price terms (R8).** Table 9 (p. 11) re-estimates Eq. (1) with two non-price outcomes: Sum\_volume (total annual loan volume from the bank to the borrowing firm) and Mortgage\_terms (change in mortgage-related contract terms). The regression is at the firm-year level. Both Airpollution coefficients have the expected sign and are statistically significant, indicating that pollution tightens credit quantity and collateral requirements, not only the interest rate.

**Robustness.** Alternative explanatory variables (Airpollution\_revenue: emission intensity scaled by revenue; separate SO2, NOx, smoke-dust regressions in Table A1), alternative outcome variables (Actual\_rate, Loan\_spread2, Loan\_spread3 in Table A4), province-year and industry-year FE replacing year FE (Tables A5), branch-bank FE (Table A5, Col. 4), and exclusion of the 2008 financial crisis period and non-SOE borrowers (Table A6) all confirm positive and significant Airpollution coefficients.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Environmental Survey and Reporting (ESR) database, China | Main explanatory variable: firm-year air pollutant emissions (tons); jointly managed by Ministry of Environmental Protection and National Bureau of Statistics; self-reported and government-audited | No page yet |
| China Stock Market Accounting Research (CSMAR) | Firm-level financial and governance controls (size, leverage, ROA, BM, tangibility, state ownership, etc.); lagged one year | [CSMAR](/wiki/commercial/csmar/) (licensed) |
| Proprietary state-owned commercial bank loan data | Main dataset: 61,718 raw loan observations (55,483 after exclusions) from 745 firms in 175 cities; annual credit exceeding RMB 50 million; nationwide deposit and loan network; covers 2008-2019 | No page yet (proprietary) |
| Washington University satellite PM2.5 data | City-level PM2.5 concentrations (ug/m³) from the Atmospheric Composition Analysis Group; used for city-level robustness analysis (not the main explanatory variable) | No page yet |

Sample period: air pollution controls 2007-2018 (lagged to match loan data 2008-2019); final loan-level sample 55,483 observations, annual frequency.

## When to read the full paper

Read the [original](https://doi.org/10.1016/j.jbankfin.2026.107655) if you are: studying how environmental risk gets priced into bank loan markets (Tables 2-4 contain the headline regressions); designing a study that uses China's ESR emission database or proprietary bank loan data (Sections 3.1 and Appendix B contain the variable construction); working on the mechanisms linking pollution to firm credit risk (Tables 5-7 and Appendix C break down emission types and loan types); or extending this evidence to other countries, private banks, or post-2019 periods (the paper's own institutional scope is limited to one Chinese state-owned bank, 2007-2019).

## Attribution and rights

Source: peer-reviewed, *Journal of Banking and Finance* 185 (2026), article 107655. © 2026 Elsevier B.V. All rights reserved. This paper is paywalled; no open-access or CC licence was found in Crossref metadata as of 2026-06-25. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. Extract-only; the verbatim PDF is not hosted here.

> Li, Donghui, Jian Sun, Rui Xu, Chun Yuan, and Liyi Zhu. "Air pollution and bank loan pricing."
> *Journal of Banking and Finance* 185 (2026): 107655.
> DOI: [10.1016/j.jbankfin.2026.107655](https://doi.org/10.1016/j.jbankfin.2026.107655)


==============================================================================
# Hidden Cost of ETF Investing: Liu, T. Zhang & Y. Zhang (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jbf/2026/liu-hidden-cost-etf-investing-2026/
# Distilled: ETFs earn significantly positive overnight returns and negative intraday returns; the gap is driven by retail demand near the market open and arbitrage constraints that prevent immediate price correction. Journal of Banking and Finance 2026, CC BY 4.0. Seven core results with source locators, datasets used, the three tested hypotheses, and the estimating equations.
# Tags: paper-summary, asset-pricing, equities, portfolio-sort, fama-macbeth, panel-regression, open-access, cc-by, peer-reviewed, unreplicated, data:crsp-mutual-funds, data:wrds, data:taq, data:morningstar, data:thomson-13f
==============================================================================

**What this is.** The paper's core results (the overnight vs intraday return differential in the US ETF market, its magnitude, and its mechanism), the three tested hypotheses, and the key estimating equations: enough to know what it found and how, without reading all 15 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1016/j.jbankfin.2025.107621).

## TL;DR

Decomposing ETF close-to-close mid-quote returns into overnight and intraday components (January 2004 to December 2021, 2,916 US ETFs), the paper documents that overnight returns are significantly positive on average (0.78% per month), while intraday returns are not significantly different from zero (-0.15%), producing a persistent overnight-intraday gap of 0.93% per month. This gap is ubiquitous across asset types (equity, fixed income, and other ETFs) and exchanges (NYSE and NASDAQ). Three candidate explanations are tested and the first two are rejected: the differential is not explained by overnight risk being higher than intraday risk (H1 rejected), nor by information asymmetry driving informed traders to exit at the close (H2 rejected). Instead, the evidence supports H3: the gap is driven by excess retail demand near the market open and by arbitrage constraints that slow correction. ETFs with the highest retail demand and highest arbitrage constraints show a monthly differential nearly six times larger than ETFs in the lowest demand and constraint group (1.99% vs 0.31% per month). Using COVID-19 Economic Impact Payments (EIPs) as an exogenous shock to retail demand, the paper shows that EIP months raise the overnight-intraday return difference by 2.37% per month, providing causal evidence for the retail demand channel.

Prior work by Lou, Polk and Skouras (2019) documents a tug-of-war between overnight and intraday returns for individual stocks; this paper establishes the same pattern in ETFs and identifies the underlying mechanism. Lachance (2021) focuses on ETFs' high overnight returns from a microstructure perspective; this paper complements her work by focusing on the full overnight-intraday differential and by decomposing the sources. Bogousslavsky (2021) documents the cross-section of intraday and overnight stock returns; this paper extends those findings to ETFs and links them to retail demand and arbitrage supply. Berkman et al. (2012) link retail investor attention and bid-ask bounce to inflated open prices; this paper strips out the bid-ask effect via mid-quote returns and shows the return pattern survives. Boehmer et al. (2021) propose an algorithm to identify retail orders in TAQ; the paper uses their method to measure retail order imbalances near the open.

## Core results

Magnitudes and significance are as reported; `\*\*`/`\*\*\*` = 5%/1%. All returns are in percentages. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | The all-ETF market portfolio has a **significantly positive overnight return and an insignificant intraday return**, producing a 0.93%/month overnight-intraday differential | Table 3, Panel A, p. 7 | Overnight = 0.784%\*\*\* (SE 0.165%), Intraday = -0.150% (SE 0.186%), Overnight-Intraday = 0.933%\*\*\* (SE 0.207%) |
| R2 | **ETFs with the highest retail demand have a significantly larger overnight-intraday differential** than low-retail-demand ETFs; the composite retail demand spread is 1.162%/month | Table 4, Panel B (bottom row), p. 8 | High-minus-low composite retail demand = 1.162%\*\*\* (SE 0.137%); individual proxies: max return 0.926%\*\*\* (SE 0.188%), retail ownership 0.625%\*\*\* (SE 0.078%), retail flow 0.360%\*\*\* (SE 0.084%) |
| R3 | **ETFs with the tightest arbitrage constraints have a larger overnight-intraday differential**; the composite arbitrage constraint spread is 0.875%/month | Table 4, Panel C (bottom row), p. 8 | High-minus-low composite arbitrage constraint = 0.875%\*\*\* (SE 0.126%); proxies: AP Concentration 0.424%\*\*\* (SE 0.095%), IVol 1.012%\*\*\* (SE 0.191%), Bid-Ask Spread 0.634%\*\*\* (SE 0.121%), Amihud Illiquidity 0.373%\*\*\* (SE 0.115%) |
| R4 | **Jointly, high retail demand and high arbitrage constraints produce a differential nearly six times larger** than the low/low group (1.99% vs 0.31% per month) | Table 6, Panel A, p. 10 | High retail demand / high arbitrage constraint = 1.988%\*\*\* (SE 0.281%); low / low = 0.312%\*\* (SE 0.152%); difference = 1.143%\*\*\* (SE 0.157%) |
| R5 | **Fama-MacBeth regressions confirm that retail investor ownership positively predicts the overnight-intraday differential** after controlling for risk measures | Table 5, col. 3 and col. 6, p. 9 | Retail Ownership coefficient = 0.898%\*\*\* (SE 0.102) in col. 3 (univariate with controls); 0.833%\*\*\* (SE 0.092) in col. 6 (full specification); N = 200,118 |
| R6 | **COVID-19 Economic Impact Payments (EIPs) increase the overnight-intraday differential by 2.37%/month**, confirming causal role of retail demand | Table 8, col. 2-3, p. 11 | EIP months: +2.372%\*\*\* (SE 0.076); Retail_demand x EIP interaction = 1.370%\*\*\* (SE 0.086); sample: January 2020 to December 2021, N = 45,492 |
| R7 | **Retail order imbalances near the market open (not the close) drive the overnight-intraday differential**, measured directly from TAQ | Table 9, col. 1-2, p. 11 | Market open retail imbalance coefficient = 0.082%\*\*\* (SE 0.006) on NDiff; market close retail imbalance coefficient = 0.001 (SE 0.002, insignificant); sample: January 2010 to December 2021, N = 3,975,301 |

**Overall (paper's conclusion).** The convenience of buying ETFs during intraday trading hours comes at a cost: retail investors bid up the opening price and arbitrageurs cannot fully correct this by the end of the day. This hidden cost is economically large (4.2 basis points per day on average, or 0.93% per month), ubiquitous across ETF types and exchanges, and causal: exogenous increases in retail demand during EIP months raise the differential. Investors can reduce the cost by purchasing near the market close, when ETF prices are more efficient due to the AP creation and redemption mechanism restoring pricing accuracy.

## Theory / model

The paper tests three competing hypotheses about why overnight returns exceed intraday returns for ETFs; there is no formal structural model.

**Hypothesis 1 (H1) -- Risk-return trade-off.** If holding assets overnight entails greater risk than intraday trading, overnight returns should compensate for that risk. Prediction: the overnight-intraday return difference is positively correlated with overnight risk and negatively correlated with intraday risk.

**Hypothesis 2 (H2) -- Information asymmetry.** Following Slezak (1994) and Hong and Wang (2000), if informed investors trade near the close to realize overnight information advantages, closing prices are discounted, raising overnight returns. Prediction: the differential is positively correlated with the proportion of informed (institutional) investors, so ETFs with more retail investors (less informed) should have a smaller differential.

**Hypothesis 3 (H3) -- Retail demand and arbitrage constraints.** Retail investors, who prefer to trade near the market open (Lou, Polk and Skouras (2019)), create excess demand that temporarily inflates opening prices. Arbitrageurs facing inventory limits, execution costs, and concentration constraints cannot immediately correct this mispricing; it unwinds gradually through the trading day, reducing intraday returns. Prediction: the differential is positively correlated with both retail demand and arbitrage constraints.

The identification strategy for H3 exploits the three rounds of COVID-19 Economic Impact Payments (EIPs) distributed in April/May 2020, December 2020/January 2021, and March/April 2021. EIPs are exogenous government transfer payments that increase household cash and retail participation in ETF markets (Divakaruni and Zimmerman (2024)), creating plausibly exogenous variation in retail demand.

## Method

The core methodological contribution is the decomposition of ETF close-to-close returns into overnight and intraday components using average NBBO mid-quotes from the first and last 5-minute intervals of the trading day to minimize microstructure noise (bid-ask bounce, as in Berkman et al. (2012)).

**Return construction (pp. 4-5).** For ETF $i$ on day $t$, letting $$P^{i}_{\text{close},t}$$ be the average NBBO mid-quote during the last five minutes and $$P^{i}_{\text{open},t}$$ the average mid-quote during the first five minutes, the daily intraday return (equation 1) is:

$$
r^{i}_{\text{intraday},t} = \frac{P^{i}_{\text{close},t}}{P^{i}_{\text{open},t}} - 1 \tag{1}
$$

The daily close-to-close return adjusting for dividends $$\text{Div}^i_t$$ and cumulative factors $$\text{CFACPR}^i_t$$ (equation 2) is:

$$
r^{i}_{\text{close-to-close},t} = \frac{P^{i}_{\text{close},t}/\text{CFACPR}^{i}_t + \text{Div}^{i}_t/\text{CFACPR}^{i}_t}{P^{i}_{\text{close},t-1}/\text{CFACPR}^{i}_{t-1}} - 1 \tag{2}
$$

The daily overnight return (equation 3) is:

$$
r^{i}_{\text{overnight},t} = \frac{1 + r^{i}_{\text{close-to-close},t}}{1 + r^{i}_{\text{intraday},t}} - 1 \tag{3}
$$

Monthly returns are standardized to 21 trading days to make observations comparable across months with different trading-day counts (equations 4 and 5):

$$
r^{i}_{\text{intraday},m} = \left[\prod_{t \in m}(1 + r^{i}_{\text{intraday},t})\right]^{21/n} - 1 \tag{4}
$$

$$
r^{i}_{\text{overnight},m} = \left[\prod_{t \in m}(1 + r^{i}_{\text{overnight},t})\right]^{21/n} - 1 \tag{5}
$$

where $$n$$ is the number of trading days in month $$m$$. Equal-weighted portfolio overnight-intraday return difference (equation 9, denoted $$\text{NDdiff}$$) is:

$$
r^{p}_{\text{NDdiff},m} = r^{p}_{\text{overnight},m} - r^{p}_{\text{intraday},m} = \sum_{i \in p} w^{i}_{m-1} \left(r^{i}_{\text{overnight},m} - r^{i}_{\text{intraday},m}\right) \tag{9}
$$

**Retail demand proxy.** Three proxies are constructed: (i) the maximum daily close-to-close return over the past month (MAX), capturing lottery-seeking behavior; (ii) retail investor ownership (the proportion of ETF shares held by retail investors, estimated as total shares minus institutional 13F holdings, then standardized); and (iii) net fund flow from retail investors (quarterly change in retail holdings divided by total shares). A composite retail demand index averages ranks across available proxies.

**Arbitrage constraint proxy.** Four proxies are combined: idiosyncratic volatility (standard deviation of CAPM residuals over 12 months), bid-ask spread (average closing bid-ask spread), Amihud illiquidity ratio, and AP concentration (inverse of number of authorized participants). A composite arbitrage constraint index averages ranks across available proxies.

**Retail order imbalance (equation 10, p. 11).** Using the Boehmer et al. (2021) sub-penny price improvement algorithm to identify retail orders in TAQ:

$$
\text{retail order imbalance}_{it} = \frac{\text{buy volume}_{it} - \text{sell volume}_{it}}{\text{buy volume}_{it} + \text{sell volume}_{it}} \tag{10}
$$

This is computed separately for the first 5 minutes after the open and the last 5 minutes before the close.

## Empirical specifications

**Portfolio sorts (R2-R4).** At the beginning of each month $$t$$, ETFs are sorted into three groups (bottom 30%, middle 40%, top 30%) based on each lagged proxy. Equal-weighted portfolios are held throughout month $$t$$. The reported monthly overnight-intraday return differentials are averaged over the 2004-2021 sample. Standard errors follow Newey and West (1987) with three lags.

**Fama-MacBeth cross-sectional regressions (R5, Table 5).** Each month, the cross-sectional regression:

$$
\text{NDdiff}_{i,t} = \alpha_t + \boldsymbol{\beta}' \mathbf{X}_{i,t-1} + \varepsilon_{i,t} \tag{FM}
$$

is run with $$\mathbf{X}_{i,t-1}$$ including Day Risk, Night Risk (or Day Beta, Night Beta), Retail Ownership, Beta, Log Cap, Turnover, and Momentum. Time-series averages of $$\hat{\beta}_t$$ are reported with Newey-West standard errors (3 lags). Sample: 200,118 monthly observations, January 2004 to December 2021.

**Panel regression with benchmark-time fixed effects (R5 robustness, Table 7).** To control for underlying asset fundamentals and test H3 jointly:

$$
\text{NDdiff}_{i,t} = \alpha + \beta_1 \text{Retail\_demand}_{i,t-1} + \beta_2 \text{Arbi\_constraint}_{i,t-1} + \boldsymbol{\gamma}' \mathbf{Z}_{i,t-1} + \mu_{j \times t} + \varepsilon_{i,t} \tag{PR}
$$

where $$\mu_{j \times t}$$ is a benchmark $$j$$ by time $$t$$ fixed effect (one cell per benchmark-month pair), $$\mathbf{Z}_{i,t-1}$$ includes Beta, Log Cap, Turnover, and Momentum. Standard errors are clustered at the benchmark level. Sample: 42,700 monthly observations, ETFs with at least two peers sharing the same benchmark.

**EIP causal identification regression (R6, Table 8).** Restricting to January 2020 to December 2021 and comparing EIP months to non-EIP months in the same pandemic window:

$$
\text{NDdiff}_{i,t} = \alpha + \beta_1 \text{Retail\_demand}_{i,t-1} + \beta_2 \text{EIP}_t + \beta_3 (\text{Retail\_demand}_{i,t-1} \times \text{EIP}_t) + \beta_4 \text{Arbi\_constraint}_{i,t-1} + \boldsymbol{\gamma}' \mathbf{Z}_{i,t-1} + \varepsilon_{i,t} \tag{EIP}
$$

where $$\text{EIP}_t = 1$$ for months in which US households receive EIP payments (April and May 2020, December 2020 and January 2021, March and April 2021). Standard errors are clustered at the fund level. N = 45,492.

**Retail order imbalance regression (R7, Table 9).** Daily regressions of the overnight-intraday return difference on retail order imbalances near the open and close, with fund and benchmark-time fixed effects, sample January 2010 to December 2021 (excluding 2016-2018). The positive and significant coefficient on open-market imbalance and the insignificant coefficient on close-market imbalance confirm that the retail demand channel operates through open-price inflation, not close-price deflation.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CRSP Mutual Fund (CRSPMF) database | ETF identifier, fund metadata, NAV, total net assets, quarterly holdings, inception date, investment style codes | [CRSP Mutual Funds](/wiki/commercial/crsp-mutual-funds/) |
| CRSP Daily Stock (CRSPSTOCK) | Daily open, high, low, close prices; trading volume; shares outstanding; return adjustment factors | [WRDS](/wiki/commercial/wrds/) |
| TAQ database | NBBO mid-quotes (5-minute intervals at open and close); retail investor order imbalances via Boehmer et al. (2021) algorithm | [TAQ](/wiki/commercial/taq/) |
| Morningstar | ETF benchmark identifiers, authorized participant (AP) lists, benchmark-level performance | [Morningstar](/wiki/commercial/morningstar/) |
| Thomson Reuters s34 filings | Quarterly institutional holding shares; used to construct retail investor ownership as total minus institutional | [Thomson Reuters 13F](/wiki/commercial/thomson-13f/) |

Sample: 2,916 unique US ETFs, January 2004 to December 2021 (217 months). Final sample covers approximately 98% of net assets invested in the US ETF market at end-2021. TAQ analysis restricted to January 2010 to December 2021 (excluding 2016-2018); benchmark panel restricted to ETFs followed by at least two ETFs sharing the same benchmark.

## When to read the full paper

Read the [original](https://doi.org/10.1016/j.jbankfin.2025.107621) if you are: (i) building on the return decomposition methodology (equations 1-9, pp. 4-5) for ETF or fund research; (ii) studying the role of retail investors in ETF pricing or market microstructure; (iii) using the TAQ-based retail order imbalance measure following Boehmer et al. (2021) in an ETF context; (iv) working on the cost of ETF investing (the paper's Online Appendix contains daily-return robustness, value-weighted results, and equity-only subsamples at Tables OA1-OA4); or (v) designing a study that exploits COVID-19 EIPs as an instrument for retail demand shocks.

## Attribution and rights

Source: peer-reviewed, *Journal of Banking and Finance* vol. 185, article 107621 (2026). This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Liu, Xin, Tianyao (Terry) Zhang, and Yaodong Zhang.
> "A hidden cost of ETF investing: Retail demand shocks and limits to arbitrage."
> *Journal of Banking and Finance* 185 (2026): 107621.
> DOI: 10.1016/j.jbankfin.2025.107621. (c) 2026 The Authors. Published by Elsevier B.V.
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Housing Booms and Local Capital Misallocation: Liu, Zhao & Zhao (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jbf/2026/liu-housing-booms-capital-misallocation-2026/
# Distilled: Exploits China's 2010-11 housing purchase restriction (HPR) policy as a natural experiment to show housing booms crowd out bank credit to manufacturing firms via reduced household mortgage and local government loan demand, worsening capital misallocation; the HPR policy improvement raised China's aggregate industrial TFP by approximately 2-3%. Journal of Banking and Finance 2026, paywalled. Seven core results with source locators, datasets used, the DiD/event-study design, and the TFP aggregation equation.
# Tags: paper-summary, housing, banks, capital-misallocation, credit-supply
==============================================================================

**What this is.** The paper's core results, the economic mechanism (crowding-out of firm credit via housing booms and the LGFV government-loan channel), and the DiD/event-study design with the TFP aggregation equation: enough to know what it found and how, without reading all 16 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1016/j.jbankfin.2025.107584).

## TL;DR

Liu, Zhao, and Zhao (2026) exploit China's housing purchase restriction (HPR) policy, implemented city-by-city starting in 2010, as a natural experiment to study how housing booms crowd out credit to manufacturing firms. The HPR capped household home purchases (at most one new home), directly reducing housing demand and halting local house-price growth. Using a DiD design comparing 45 treated cities to 216 control cities over 2006-2013, the paper shows that the housing boom had significantly crowded out bank credit to manufacturing firms via two channels: (i) banks reduced mortgage loans to households (down 23-29%), and (ii) local governments' land-sales revenue fell, reducing their borrowing through local government financing vehicles (LGFVs, down 31-52%), freeing bank credit for firms. The credit relaxation raised firm leverage by about 3.8 percentage points and lowered firms' interest costs by about 1.6 pp. More financially constrained firms (smaller, higher initial MRPK) and cities with more constrained banks (higher bank leverage) benefited most. Increased credit access led to higher non-housing investment, industrial output, and value-added. The paper documents that within-city capital misallocation (MRPK dispersion) fell significantly in treated cities and estimates the HPR policy increased China's aggregate industrial TFP by approximately 2-3%, consistent with Chakraborty, Goldstein, and MacKinlay (2018) and extending that crowding-out literature to the Chinese manufacturing context. The comparison with Basco, Lopez-Rodriguez, Moral-Benito, and Moreno (2024) suggests the crowding-out channel dominates the collateral channel in China.

## Core results

Magnitudes and significance are as reported; \*\*/\*\*\* = 5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | HPR policy significantly raised firm leverage in treated cities | Table 1, col. 2, p. 6 | DiD coeff. 0.0378\*\*\* (se 0.0099); leverage = debt / total assets |
| R2 | HPR policy significantly lowered firm interest rate in treated cities | Table 1, col. 7, p. 6 | DiD coeff. -0.0162\*\*\* (se 0.0034); interest rate = total interest / total debt |
| R3 | Firms increased non-housing investment and industrial output after HPR | Table 2, cols. 1, 4-5, p. 7 | I(non-housing investment > 0): 0.0302\*\* (se 0.0119); log output: 0.1081\*\* (se 0.0529); log value-added: 0.1409\*\* (se 0.0679) |
| R4 | HPR policy significantly reduced household mortgage loans | Table 5, Panel A, col. 5, p. 8 | DiD coeff. on log total mortgage loans: -0.2893\*\*\* (se 0.0751) |
| R5 | HPR policy significantly reduced local government (LGFV) loans | Table 5, Panel B, col. 13, p. 8 | DiD coeff. on log total LGFV loans: -0.3152\*\* (se 0.1564) |
| R6 | HPR policy significantly reduced within-city cross-firm MRPK dispersion | Figure 5a, 5c, 5e, p. 13 | Event-study coefficients on Var(log-MRPK), 90th/10th, and 75th/25th MRPK ratios all significantly negative post-HPR |
| R7 | HPR policy increased China's aggregate industrial TFP by approximately 2-3% | Section 5.5, Eq. 5, p. 12-13 | Aggregation formula: 2.04%; reduced-form: treated-city TFP gain of ~6% x 51% national capital share = ~3.1% |

**Overall (paper's conclusion).** China's prolonged housing boom generated large unintended negative externalities on manufacturing firms by crowding out bank credit, particularly through the government-loan channel (LGFVs backed by land-sales collateral) and the household mortgage channel. The housing market decline induced by the HPR policy relieved these credit constraints, improved capital allocation across firms, and raised aggregate industrial productivity. The TFP improvement of 2-3% represents a substantial welfare gain attributable to cooling the housing boom through the credit channel.

## Theory / model

The paper presents no formal general-equilibrium model. The theoretical argument rests on two institutional features of China's credit market that create the conditions for housing-boom crowding-out.

**Geographic segmentation.** China's credit market is highly segmented geographically: about 89% of all bank loans are issued to borrowers in the issuing bank's own city (Gao et al., 2019, cited p. 3). When local housing demand rises, local banks increase mortgage lending to households and extend more credit to LGFVs (whose borrowing capacity is backed by rising land-sales collateral and land-conveyance revenue). Because banks face binding regulatory constraints (Basel III leverage ratio, loan-to-deposit ratio, and annual credit quotas from the People's Bank of China), credit to housing-related borrowers crowds out credit to local manufacturing firms.

**Identification.** The HPR policy was implemented in 46 cities based on whether housing prices were rising rapidly - a criterion orthogonal to manufacturing productivity trends. This quasi-random assignment provides a credible DiD design: manufacturing firms in treated and control cities showed no differential pre-existing trends in leverage, interest rate, investment, or output (Figure 2 and Figure A.4, pp. 9, 10). The exclusion restriction is that the HPR policy affected manufacturing firms only through the housing and credit market (no explicit government mandate to redirect lending to industry, confirmed in Section C of the online appendix).

**Tested hypotheses.** Three primary hypotheses follow:
1. Housing market declines raise firm credit access: leverage rises, interest rate falls.
2. The mechanism operates through reductions in household mortgage loans and LGFV loans, crowding in credit to manufacturing firms.
3. Effects are stronger in cities with more financially constrained banks (higher bank leverage, lower deposits) and in firms that are more ex-ante credit-constrained (smaller, lower MRPK).

Hsieh and Klenow (2009) provide the theoretical framework for how credit-access heterogeneity translates into capital misallocation: firms with high marginal revenue product of capital (MRPK) relative to the average are under-financed, and equalizing credit access across firms reduces the MRPK dispersion and raises aggregate TFP.

## Method

The headline estimator is a two-way fixed-effects DiD regression (p. 5):

$$
y_{i,c,t} = \beta \, \text{Treat}_c \times \text{Post}_t + \lambda X_{i,t} + \theta X_{c,t} + \gamma_i + \gamma_t + \epsilon_{i,c,t} \tag{1}
$$

where $$y_{i,c,t}$$ is leverage or interest rate of firm $$i$$ in city $$c$$ in year $$t$$; $$\text{Treat}_c$$ is a dummy equal to 1 for the 45 HPR-implementing cities; $$\text{Post}_t$$ is equal to 1 for years 2011 onward; $$\gamma_i$$ and $$\gamma_t$$ are firm and year fixed effects; $$X_{i,t}$$ includes firm age, age squared, log assets, and fixed-asset share; $$X_{c,t}$$ includes city-level contemporaneous real-estate investment share, total credit/GDP, log real GDP, and GDP growth rate. Standard errors are clustered at the city level.

Dynamic (event-study) effects use (p. 5):

$$
y_{i,c,t} = \sum_{\tau=-4}^{3} \beta_\tau \, \text{Treat}_c \times I_t^\tau + \lambda X_{i,t} + \theta X_{c,t} + \gamma_i + \gamma_t + \epsilon_{i,c,t} \tag{2}
$$

where $$I_t^\tau = 1$$ if year $$t$$ is $$\tau$$ years after the HPR shock (year 0 = 2011). Coefficients for $$\tau \in \{-4, \ldots, -1\}$$ test for pre-existing trends; none are found for any main outcome variable.

Heterogeneous effects by firm financial constraint are examined via triple-differences (p. 10):

$$
y_{i,c,t} = \beta_1 \, \text{Treat}_c \times \text{Post}_t \times \text{FirmFC}_i + \beta_2 \, \text{Treat}_c \times \text{Post}_t + \beta_3 \, \text{FirmFC}_i \times \text{Post}_t + \lambda X_{i,t} + \theta X_{c,t} + \gamma_i + \gamma_t + \epsilon_{i,c,t} \tag{3}
$$

where $$\text{FirmFC}_i$$ is a pre-reform financial constraint proxy (employment, total assets, property-holding status, or initial MRPK).

Capital misallocation is studied with an industry-city-year event study (p. 12):

$$
y_{c,j,t} = \sum_{\tau=-4}^{3} \alpha_\tau \, \text{Treat}_c \times I_t^\tau + \gamma_{c,j} + \delta_t + \epsilon_{c,j,t} \tag{4}
$$

where $$y_{c,j,t}$$ is one of three cross-firm MRPK dispersion measures in 2-digit industry $$j$$, city $$c$$, year $$t$$: variance of log-MRPK, log-ratio of 90th to 10th percentiles, and log-ratio of 75th to 25th percentiles. City-industry and year fixed effects are included.

Aggregate TFP is computed via the Sraer and Thesmar (2023) aggregation formula (p. 12):

$$
\Delta \log(\text{TFP}) \approx -\frac{a}{2}\!\left(1 + \frac{a\theta}{1-\theta}\right) K_s \Delta\Delta\sigma^2 - \frac{a}{2}\!\left(\frac{a\theta}{1-\theta}\right) \left[\Phi_s - K_s\right] \left(\Delta\Delta\bar{\mu} + \Delta\Delta\hat{\sigma}_{\text{MRPK},py}\right) + \frac{1}{2}\frac{a\theta}{1-\theta} \Delta\Delta\sigma^2 \tag{5}
$$

where $$a = 1/3$$ (capital share from Cobb-Douglas production), $$\theta = 0.83$$ (price elasticity of demand), $$K_s = 0.5114$$ (treated cities' share of national capital), $$\Phi_s = 0.4723$$ (treated cities' output share), and $$\Delta\Delta\sigma^2$$, $$\Delta\Delta\bar{\mu}$$, $$\Delta\Delta\hat{\sigma}_{\text{MRPK},py}$$ are the DiD estimates of changes in variance, mean, and covariance of log-MRPKs from Table 9 (p. 15).

## Empirical specifications

**Credit access (R1-R2).** Eq. (1) with firm leverage (debt/total assets) or interest rate (total interest/total debt) as the dependent variable, applied to approximately 1.85 million firm-year observations from the ASIF panel, 261 cities, 2006-2013. City controls, firm controls, and pre-trend interaction terms are added progressively (Table 1, cols. 1-5 for leverage, cols. 6-10 for interest rate). Results are replicated separately for private, state-owned, and foreign firms; private firms show the largest effects, consistent with private firms being most financially constrained.

**Real outcomes (R3).** Eq. (1) applied to real firm outcomes: an indicator of positive non-housing investment, log non-housing investment, log non-housing capital, log industrial output, and log value-added (Table 2). The 427,088 to 655,568 firm-year observations used here are the subset of firms reporting investment data.

**Crowding-out mechanism (R4-R5).** Eq. (1) applied at the province-treatment-year level with log total mortgage loans (China Real Estate Yearbooks, 31 provinces, 35 major cities) and at the city-year level with log total LGFV loans (WIND database, 261 cities). Table 5 Panel A shows that HPR reduced housing transaction volumes and prices (cols. 1-4) and total mortgage loans (cols. 5-6). Panel B shows that HPR reduced primary land-market prices and areas sold by local governments (cols. 7-10), local government land-sales revenue (cols. 11-12), and total LGFV loan volumes (cols. 13-14).

**Bank constraints heterogeneity.** Eq. (1) and a triple-differences variant separating cities by pre-reform bank leverage (above/below national median, Table 6, p. 10) and by bank deposits (Table 7). The HPR leverage effect is 8.62 pp for cities with high bank leverage vs. insignificant for low-leverage cities; the interest-rate effect is -2.91 pp vs. insignificant. Consistent with the crowding-out channel operating through bank balance-sheet constraints.

**Firm constraints heterogeneity.** Eq. (3) with four proxies for $$\text{FirmFC}_i$$ (Table 8, p. 14-15). Firms with higher initial MRPK (2009) show a 0.0084 pp larger leverage increase per unit of log-MRPK (t = 2.40) and a 0.0032 pp larger interest-rate decline, consistent with HPR differentially improving credit access for the most capital-constrained firms, reducing MRPK dispersion.

**Capital misallocation (R6).** Eq. (4) applied at the 2-digit industry by city by year level. All three MRPK dispersion measures (variance, 90/10 ratio, 75/25 ratio) show significantly negative event-study coefficients in the post-HPR period with no pre-trends, both for capital (MRPK) and labor (MRPL) misallocation (Figure 5, p. 13). Within-city capital misallocation contributes approximately 60% of overall MRPK dispersion in China.

**Aggregate TFP (R7).** Eq. (5) applied using parameter values from Table 9 (p. 15): $$\Delta\Delta\sigma^2 = -0.0998$$, $$\Delta\Delta\bar{\mu} = -0.0453$$, $$\Delta\Delta\hat{\sigma}_{\text{MRPK},py} = 0.0326$$. The formula yields a 2.04% TFP improvement. A reduced-form approach directly estimating TFP at the 2-digit industry by city by year level finds treated cities improved TFP by approximately 6% relative to control cities; scaling by the 51% national capital share of treated cities gives 3.1%.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Annual Survey of Industrial Firms (ASIF) | Main firm panel: leverage, interest rate, investment, output, MRPK; manufacturing firms with annual sales above 5 million RMB; approx. 1,845,072 firm-years | No page yet |
| WIND database (LGFV data) | Local government financing vehicle loan volumes for 261 cities, 2006-2013 | No page yet |
| China Land Transaction Monitoring System (landchina.com) | Primary land market transaction prices and areas sold by city-year, 2006-2013 | No page yet |
| CEIC China database | Secondary housing market transaction prices and sold housing area by city-year | No page yet |
| China Real Estate Yearbooks | Province-level and city-level home mortgage loan data, 31 provinces and 35 major cities | No page yet |
| National Tax Statistics Database (NTSD) | Firm-level property holding status and non-housing asset investments; approx. 700,000 firms annually | No page yet |
| China City Statistical Yearbooks | City-level covariates: GDP, credit/GDP ratio, real-estate investment share | No page yet |

Sample: 261 Chinese cities, 2006-2013 (2010 excluded because ASIF data for 2010 were never released). The 45 HPR-treated cities implemented the policy in 2010-2011; 216 cities serve as controls.

## When to read the full paper

Read the [original](https://doi.org/10.1016/j.jbankfin.2025.107584) to:
- Replicate the DiD estimates or extend to the post-2013 HPR relaxation period (after 2016).
- Examine the full battery of robustness checks: neighboring-city controls (Table 3), full ASIF sample (Table 4), balanced panel (Table A.5), alternative leverage and interest measures (Tables A.7-A.9), province-year fixed effects (Table A.10), and NTSD-supplemented estimates (Table A.11).
- Study the bank-constraints heterogeneity (Tables 6-7) and firm-level heterogeneity (Table 8, Figure 4) in detail.
- Apply the Sraer and Thesmar (2023) aggregation formula in other settings; parameters and derivation are in Table 9.
- Study the comparison with Basco et al. (2024) and Martín et al. (2021) on the relative importance of the crowding-out vs. collateral channel across countries.

## Attribution and rights

Source: peer-reviewed, *Journal of Banking and Finance* 182 (2026), article 107584. Copyright 2025 Elsevier B.V. All rights reserved. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. The source is paywalled; reproduction beyond brief extraction requires the publisher's permission.

> Liu, Yu, Peng Zhao, and Xiaoxue Zhao. "Housing booms and local capital misallocation." *Journal of Banking and Finance* 182 (2026) 107584. DOI: [10.1016/j.jbankfin.2025.107584](https://doi.org/10.1016/j.jbankfin.2025.107584).


==============================================================================
# Repurchasing Overpriced Shares: Oded (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jbf/2026/oded-firms-repurchase-shares-overpriced-2026/
# Distilled: Jacob Oded proposes an agency model in which firms repurchase shares even when overpriced because insiders' benefit from preventing free cash waste can outweigh the cost of overpaying. Journal of Banking and Finance vol. 182 (2026), paywalled. Five core results covering three equilibrium types and their governance determinants, with model equations and derivations.
# Tags: paper-summary, payout-policy, stock-repurchases, corporate-governance, agency-costs, peer-reviewed, unreplicated
==============================================================================

**What this is.** The model equations, equilibrium characterization, and core propositions of Oded (2026): enough to understand why an agency model generates overpriced repurchases, and what it predicts about governance and completion rates, without reading the full 19 pages plus appendices. To replicate or extend, read the original at [doi.org/10.1016/j.jbankfin.2025.107568](https://doi.org/10.1016/j.jbankfin.2025.107568).

## TL;DR

Firms are widely observed to repurchase shares even when those shares are overpriced, a puzzle for standard signaling models that predict repurchases only when shares are undervalued. This paper proposes an agency explanation rooted in the free cash cost framework of Jensen (1986). An insider-controlled firm holds uncertain value and generates free cash. If not disbursed, the free cash deteriorates (gets wasted), and insiders privately benefit from that waste. In a full repurchase equilibrium, insiders repurchase in both high-value and low-value states because the benefit from preventing free cash waste outweighs the cost of overpaying for shares. In a partial repurchase equilibrium they repurchase only when the stock is undervalued. The model predicts that program completion rates rise with governance quality (lower private benefits from waste) and insider ownership, and fall with value variance. Repurchases are a positive governance mechanism in this framework, not a sign of managerial abuse, as long as the private-benefit rate is low. The alternative model of Babenko et al. (2020) also generates overpriced repurchases via a current-shareholder-wealth channel rather than a waste-prevention channel; the two mechanisms are complementary. Oded (2005) is the prior framework by the same author on repurchase announcements, which this paper extends to the overvaluation setting.

## Core results

Propositions are derived analytically; proofs are in Appendix B of the source.

| # | Result | Locator | Key finding |
|---|---|---|---|
| R1 | Proposition 1: full repurchase equilibrium always involves repurchase of overvalued shares | Prop. 1, p. 6 | In state L (value = α − σ), the post-repurchase per-share value at t2 equals (1 − σ/α)(α + c) < p_f = α + c, so the repurchase price exceeds true per-share value; shares are overvalued at the repurchase price |
| R2 | Proposition 3: partial repurchase equilibrium never involves repurchase of overvalued shares | Prop. 3, p. 7 | In a partial repurchase equilibrium the firm repurchases only in state H (high value = α + σ); by Lemma 2 p_p < p_f, so the partial-repurchase price is lower, and shares are always undervalued at p_p |
| R3 | Proposition 4: full repurchase equilibrium existence condition | Prop. 4, p. 8 | Full repurchase equilibrium exists if and only if condition (5) holds: γ < β(1 − σ/(α(1 − δ))); this range is identical to the region where repurchase of overvalued shares occurs |
| R4 | Proposition 5: three equilibrium types as a function of private-benefit rate γ | Prop. 5, pp. 8-9 | γ < γ₁: full repurchase (firm always announces and repurchases regardless of value); γ₂ < γ < γ₃: partial repurchase in pure strategies (repurchase only when undervalued); γ > γ₃: no announcement; γ₁ < γ₂ < γ₃ by Lemma 4 |
| R5 | Section 4 empirical predictions: governance and completion rates | §4, p. 9 | Good governance (low γ), higher insider ownership (β), more free cash waste (1 − δ), and lower value variance (σ) each increase the probability of a full repurchase equilibrium and raise predicted program completion rates |

**Overall (paper's conclusion).** The model resolves the overpriced-repurchase puzzle without invoking managerial mistakes or irrationality. Repurchases prevent free cash waste; in firms with low private benefits from waste (good governance) this motive dominates the cost of overpaying for shares, generating repurchase regardless of mispricing. Repurchase programs are a rational and value-enhancing mechanism in such firms. The model also explains the observed underperformance of actual repurchase returns relative to benchmark strategies: when repurchases occur in overvalued states, realized long-run returns to repurchasers are lower than a pure undervaluation strategy would deliver.

## Theory / model

The model considers an all-equity firm that cannot raise external capital. There are three dates indexed by $$t_i$$, $$i \in \{0, 1, 2\}$$. The firm's only asset generates value $$X \in \{H, L\}$$ with equal probability at $$t_1$$, where $$H = \alpha + \sigma$$ and $$L = \alpha - \sigma$$, $$0 < \sigma < \alpha$$, so $$E[X] = \alpha$$. Free cash $$c$$ is also generated at $$t_1$$, with $$0 < c < 1$$. All information becomes public at $$t_2$$; the firm is liquidated and paid out at true value.

**Share structure and agency.** Insider shareholders control the firm with ownership fraction $$\beta < 1$$; outsiders hold $$1 - \beta$$. A subset $$q < 1 - \beta$$ of outsiders face a liquidity shock at $$t_1$$ and must sell their shares. Free cash retained at $$t_1$$ deteriorates between $$t_1$$ and $$t_2$$: retention rate $$\delta \in [0, 1]$$ means fraction $$1 - \delta$$ is wasted. Insiders earn private benefit $$\gamma \in [0, 1]$$ per dollar wasted (perks, empire building, and similar agency costs of free cash following Jensen (1986)). Parameter $$\gamma$$ therefore measures the quality of governance: low $$\gamma$$ (hard to waste) reflects good governance; high $$\gamma$$ (easy to waste with large private gain) reflects poor governance.

**Repurchase policy.** At $$t_0$$, insiders may announce an open-market repurchase program. The announcement does not commit them to repurchasing; execution is decided at $$t_1$$ after value $$X$$ is privately observed. A competitive market maker, observing neither $$X$$ nor the insider's $$t_1$$ decision, sets the $$t_1$$ price $$p$$ to earn zero expected profit from liquidity sellers.

**Insider repurchase condition.** Given market price $$p$$, insiders in state $$X \in \{H, L\}$$ repurchase if and only if their expected wealth with repurchase exceeds their expected wealth without. Rearranging (eq. 3, p. 5):

$$
\beta \delta + (1 - \delta)\gamma < \frac{\beta X}{p - c} \tag{3}
$$

The left side is the marginal benefit of keeping cash (insiders retain fraction $$\delta$$ of value and privately benefit at rate $$\gamma$$ from the waste fraction $$1 - \delta$$). The right side is the marginal share-value enhancement from repurchasing at price $$p$$. Repurchase dominates waste whenever the share-value gain exceeds the waste benefit.

**Full repurchase equilibrium.** Suppose insiders repurchase in both states H and L. The market maker earns zero profit from liquidity sellers in both states (eq. 1, p. 5):

$$
0 = \frac{1}{2}\!\left[\!\left(q - \frac{c}{p}\right)\!\left(\frac{H}{1 - \frac{c}{p}} - p\right) + \left(q - \frac{c}{p}\right)\!\left(\frac{L}{1 - \frac{c}{p}} - p\right)\right]
$$

which simplifies to (eq. 2, p. 5):

$$
p_f = \alpha + c \tag{2}
$$

Under $$p_f$$, the post-repurchase per-share value in state L at $$t_2$$ is $$(1 - \sigma/\alpha)(\alpha + c) < p_f$$, confirming that shares are overpriced at the repurchase price when value is low (Proposition 1, p. 6). The full repurchase equilibrium requires the repurchase condition (3) to hold in state L, the binding constraint. Substituting $$X = L = \alpha - \sigma$$ and $$p = p_f = \alpha + c$$ into (3) and rearranging (eqs. 4-5, p. 5):

$$
p_f < \frac{\beta(\alpha - \sigma)}{\beta\delta + (1 - \delta)\gamma} + c \quad\iff\quad \gamma < \beta\!\left(1 - \frac{\sigma}{\alpha(1 - \delta)}\right) \tag{5}
$$

Condition (5) is more easily satisfied when $$\gamma$$ (private benefit from waste) is low, $$\beta$$ (insider ownership) is high, $$\sigma$$ (value variance) is low, or $$\delta$$ (cash retention) is low.

**No-repurchase equilibrium.** If insiders never repurchase, the market maker prices at (eqs. 6-7, p. 6):

$$
p_n = \alpha + \delta c \tag{7}
$$

No repurchase dominates for insiders in both states when the benefit from waste exceeds the gain from repurchase even in the high-value state (H is binding). Rearranging the no-repurchase condition for $$\gamma$$ (eqs. 8-10, p. 6):

$$
\gamma > \beta\!\left(1 + \frac{\sigma c(1 - \delta)}{[\alpha - c(1 - \delta)](1 - \delta)}\right) \equiv \gamma_0 \tag{10}
$$

**Partial repurchase equilibrium.** In a partial repurchase equilibrium insiders repurchase only in state H. The market maker anticipates this, facing adverse selection from insiders who repurchase strategically. The zero-profit condition becomes (eq. 12, p. 6):

$$
0 = \frac{1}{2}\!\left[\!\left(q - \frac{c}{p}\right)\!\left(\frac{H}{1 - \frac{c}{p}} - p\right) + q(L + \delta c - p)\right]
$$

This yields a closed-form quadratic solution (eq. 13, p. 6):

$$
p_p = \frac{\Psi + \sqrt{\Psi^2 - 4\zeta}}{2} \tag{13}
$$

where $$\Psi \equiv \alpha + c + \tfrac{c}{2}\!\left(\delta + \tfrac{1}{q}\right)$$ and $$\zeta \equiv \left[\alpha + \sigma + c(1 + \delta q) + q(\alpha - \sigma)\right]\tfrac{c}{2q}$$. By Lemma 1 (p. 7), $$p_p < p_f = \alpha + c$$, since adverse selection from strategic repurchase lowers the expected gain from liquidity sellers. By Proposition 3 (p. 7), $$p_p \leq H/(1 - c/p_p)$$, so shares are always undervalued at the partial repurchase price.

The partial repurchase equilibrium holds when (eqs. 16-17, p. 7):

$$
\frac{\beta(\alpha - \sigma)}{\beta\delta + (1-\delta)\gamma} + c < p_p < \frac{\beta(\alpha + \sigma)}{\beta\delta + (1-\delta)\gamma} + c \tag{16}
$$

Rearranging in terms of $$\gamma$$:

$$
\frac{\beta}{1-\delta}\!\left(\frac{\alpha - \sigma}{p_p - c} - \delta\right) < \gamma < \frac{\beta}{1-\delta}\!\left(\frac{\alpha + \sigma}{p_p - c} - \delta\right) \tag{17}
$$

**Equilibrium existence (Proposition 5, p. 8-9).** Define three thresholds:

$$
\gamma_1 \equiv \frac{\beta}{1-\delta}\!\left(\frac{\alpha - \sigma}{\alpha} - \delta\right), \quad \gamma_2 \equiv \frac{\beta}{1-\delta}\!\left(\frac{\alpha - \sigma}{p_p - c} - \delta\right), \quad \gamma_3 \equiv \frac{\beta}{1-\delta}\!\left(\frac{\alpha + \sigma}{p_p - c} - \delta\right) \tag{23-25}
$$

with $$\gamma_1 < \gamma_2 < \gamma_3$$ (Lemma 4). Then:

- $$\gamma < \gamma_1$$: full repurchase equilibrium; firm always announces and repurchases regardless of value realization.
- $$\gamma_2 < \gamma < \gamma_3$$: partial repurchase equilibrium in pure strategies; firm announces and repurchases only when undervalued.
- $$\gamma > \gamma_3$$: no announcement; firm never announces a program.
- $$\gamma_1 < \gamma < \gamma_2$$: partial repurchase equilibrium in mixed strategies (Appendix A, Proposition 9).

Corollary 1 (p. 9) summarizes: a firm announces a repurchase program if and only if a full or partial repurchase equilibrium can hold.

## Method

The model is solved by backward induction using subgame perfection as the solution concept (Definition 1, p. 5). The equilibrium consists of (1) a repurchase policy (announce or not at $$t_0$$), (2) a market maker price $$p$$ given the policy, and (3) an execution strategy by insiders at $$t_1$$ given $$p$$ and realized value. The three steps are:

**Step 1: market maker pricing.** For each candidate insider strategy, the market maker sets $$p$$ to earn zero expected profit from liquidity sellers. Three candidate prices arise: $$p_f = \alpha + c$$ (full repurchase, eq. 2), $$p_n = \alpha + \delta c$$ (no repurchase, eq. 7), and $$p_p$$ from the quadratic eq. (13) (partial repurchase).

**Step 2: insider execution condition.** Given $$p$$, insiders in each state compare wealth with and without repurchase via condition (3). Full repurchase requires (3) to hold in state L (the binding constraint); partial repurchase requires (3) to hold in H but fail in L; no repurchase requires (3) to fail in both.

**Step 3: announcement decision.** Insiders compare expected wealth under each equilibrium before the value realization. Expected insider wealth under each type (eqs. 19-22, pp. 7-8):

$$
W_f = \beta(\alpha + c), \tag{19}
$$

$$
W_n = \beta(\alpha + \delta c) + (1 - \delta)c\gamma, \tag{20}
$$

$$
W_p = \beta\alpha + \frac{c}{2}\!\left(\beta\delta + (1-\delta)\gamma + \beta\frac{\alpha + \sigma}{p_p - c}\right). \tag{22}
$$

Lemma 3 shows $$W_f > W_n$$ whenever condition (5) holds. Lemma 5 shows $$W_p > W_n$$ whenever the partial repurchase equilibrium can hold. These comparisons determine which equilibrium prevails for each range of $$\gamma$$.

**Extensions.** Section 5 adds three extensions:

(a) *Outsiders in control* (Proposition 6): outsiders have no benefit from waste, so they prefer a full repurchase equilibrium whenever insiders do; for partial repurchase they announce only when $$\gamma_3 < 1$$ (condition 26, p. 10), which holds less often. Firms with outsider control have higher completion rates but fewer strategic repurchases.

(b) *Dividends* (Propositions 7-8, pp. 9-11): with dividend tax rate $$T_D > 0$$, insiders prefer repurchases over dividends whenever a full repurchase equilibrium holds. A partial repurchase equilibrium always dominates dividends for insiders (since $$T_{D1}(\gamma) < 0 < T_D$$), while outsiders prefer dividends below a threshold tax rate.

(c) *Debt* (Section 5.2.3, p. 12): debt removes free cash mechanically and prevents waste, an advantage over repurchases, but introduces bankruptcy costs and eliminates flexibility. Repurchases dominate debt as a waste-prevention mechanism in the model.

## Empirical specifications

This is a pure theory paper. The data availability statement in the source reads: "No data was used for the research described in this article" (p. 19).

Section 4 (pp. 9-10) derives testable comparative statics from Proposition 5. The main predictions, in terms of model parameters:

- **Governance quality** (lower $$\gamma$$): reduces the lower bound $$\gamma_1$$ and expands the full repurchase region, predicting more repurchase program announcements and higher completion rates. Empirical proxies: GIM index (Gompers and Metrick (2003)), E-index, Tobin's Q. Guthrie (2020) provides supporting evidence that governance quality governs whether buybacks harm or benefit shareholders.

- **Insider ownership** (higher $$\beta$$): increases all three thresholds $$\gamma_1$$, $$\gamma_2$$, $$\gamma_3$$ proportionally, expanding the full repurchase region and raising predicted completion rates.

- **Free cash waste** (higher $$1 - \delta$$): increases the benefit from repurchasing (more waste to prevent), raising the likelihood of the full repurchase equilibrium and predicted completion rates.

- **Value variance** (higher $$\sigma$$): lowers $$\gamma_1$$ and raises $$\gamma_3$$ (via $$p_p$$), making the full repurchase equilibrium harder to sustain; firms in riskier industries have lower predicted completion rates and more strategic (partial) repurchases.

The model also explains the post-announcement return underperformance documented by Bonaime et al. (2016): when full repurchase equilibria prevail, overpriced repurchases are executed, so realized long-run returns are lower than a pure undervaluation strategy, but this does not imply value destruction.

Appendix A provides a numerical example confirming all three equilibrium types at parameters $$\delta = 0.7$$, $$\gamma = 0.2$$, $$\beta = 0.3$$, $$\alpha = 3.5$$, $$\sigma = 0.5$$, $$c = 0.1$$, $$q = 0.2$$ (pp. 13-16). Figures 4-7 (pp. 14-16) trace equilibria as each of $$\gamma$$, $$\beta$$, $$\delta$$, $$\sigma$$ is varied holding others fixed.

## Datasets used

This is a pure theory paper. No empirical data are used.

| Dataset | Role in paper | Wiki page |
|---|---|---|
| None | All results are analytical; no estimation or calibration to data | N/A |

## When to read the full paper

Read the [original](https://doi.org/10.1016/j.jbankfin.2025.107568) if you are:
building an empirical test of the governance-repurchase completion-rate predictions (Section 4, pp. 9-10 give the exact comparative statics);
extending the model to board control, dividends, or debt (Section 5 covers all three, pp. 9-12);
working through proofs of Propositions 1-9 and Lemmas 1-6 (Appendix B, pp. 17-19);
or comparing the waste-prevention mechanism mechanically against the Babenko et al. (2020) current-shareholder-wealth channel.

## Attribution and rights

Source: peer-reviewed, *Journal of Banking and Finance* vol. 182 (2026), article 107568. JEL codes G14, G30, G35 are printed on the article. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. The paper is paywalled; only textual extracts are reproduced here under scholarly commentary norms (extract-only).

> Oded, Jacob. "Why do firms repurchase their shares when they are overpriced?" *Journal of Banking and Finance* 182 (2026): 107568. DOI: 10.1016/j.jbankfin.2025.107568. Copyright 2025 Elsevier B.V. All rights reserved.


==============================================================================
# Time-Varying Pollution Premium: Yin, Yu & Chen (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jbf/2026/yin-time-varying-pollution-premium-2026/
# Distilled: A long-short portfolio of high- versus low-emission US firms earns significant abnormal returns in constant factor models, but a semiparametric time-varying estimate shows the premium is significant only before 2005 and statistically indistinguishable from zero thereafter; risk aversion, macroeconomic uncertainty, natural disasters, and consumer sentiment are the most robust determinants of the time-varying pattern. Journal of Banking and Finance 187 (2026), paywalled. Seven core results with source locators, datasets used, the model, and the estimation equations.
# Tags: paper-summary, asset-pricing, factors, anomalies, esg, climate-finance
==============================================================================

**What this is.** The paper's core results, the semiparametric factor model framework, and the estimation procedure with key equations: enough to understand what was found and how, without reading all 15 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1016/j.jbankfin.2026.107693).

## TL;DR

Using EPA Toxic Release Inventory (TRI) data matched to CRSP and Compustat for US common stocks (October 1992 to September 2018), the paper sorts firms into quintiles by annual toxic emission intensity scaled four ways (total assets, PP&E, sales, market cap) and constructs a long-short portfolio. In constant factor models (CAPM through HXZ), the high-minus-low portfolio earns a statistically significant pollution premium of roughly 4 to 5 percent annually. A partially time-varying semiparametric model reveals, however, that the premium is concentrated before 2005: the nonparametric alpha estimate is statistically significant (90% confidence interval above zero) through approximately 2004-2005, then falls to a stable but indistinguishable-from-zero level. The paper identifies two channels driving this time variation: time-varying risk exposures on the supply side (especially the HML value factor) and time-varying investor preferences on the demand side (consumer sentiment, natural disaster severity, sustainable fund flows, and environmental litigation risk). Risk aversion and macroeconomic uncertainty exhibit negative associations with the premium. In a multivariate horse-race following Hsu et al. (2023), risk aversion, macroeconomic uncertainty, natural disasters, and consumer sentiment emerge as the most robust predictors.

## Core results

Magnitudes are as reported; `\*\*`/`\*\*\*` = 5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Long-short portfolio earns positive raw returns across all four emission scalings | Table 1, Panel A, p. 4 | H-L 4.57% annually (t=2.39, Sharpe=0.47, total-assets scaling); quintile raw returns increase almost monotonically from 6.82% (low) to 11.39% (high) (Q3=8.28% dips below Q2=10.60%) |
| R2 | Long-short portfolio earns significant alpha in all constant factor models | Table 2, p. 5 | FF3F alpha = 5.28%\*\*\* (t=3.56); CAPM 4.44%\*\*\* (t=3.10); FF4F 4.56%\*\*\* (t=3.29); FF5F 3.96%\*\*\* (t=2.70); HXZ 4.80%\*\*\* (t=3.44) |
| R3 | Time-varying semiparametric alpha is significant only before 2005; statistically indistinguishable from zero thereafter | Figure 2, p. 6; §4.2 | FF3F estimate approx. 0.40 monthly (4.8% annualized) in 1992, peaking near 1.0 around 2000; 90% CI above zero through approx. 2004-2005; falls to approx. 0.25 monthly and becomes insignificant post-2005 |
| R4 | HML (value factor) is the only risk factor consistently related to the pollution portfolio | Table 3, p. 7; §4.3 | FF5F OLS: HML = -0.288 (SE=0.079, t approx -3.65); FF3F OLS: HML = -0.191 (SE=0.097); MKT and SMB insignificant across all specifications |
| R5 | Consumer sentiment index is the strongest single economic-condition predictor of the time-varying alpha | Table 4, Panel C, p. 11 | beta(ICS) = 0.009\*\*\* (t=14.25, FF3F partially time-varying); Adj. R-sq = 0.394 |
| R6 | Natural disaster severity is a positive and robust predictor of the time-varying pollution premium | Table 6, Panel A, p. 11 | beta(Disasters) = 60.358\*\*\* (t=11.31, FF3F partially time-varying); Adj. R-sq = 0.554 |
| R7 | Multivariate horse-race: risk aversion, macroeconomic uncertainty, natural disasters, and consumer sentiment jointly explain most time variation | Table 9, Panel A, p. 12 | Adj. R-sq = 0.770 (CAPM) to 0.819 (FF5F), partially time-varying model (Panel B, fully time-varying: up to 0.864) |

**Overall (paper's conclusion).** The pollution premium is not a stable, time-invariant phenomenon: it reflected a priced systematic risk tied to environmental policy and investor preferences before 2005, but the abnormal returns became statistically indistinguishable from zero after that. The time variation is driven by both the HML value factor channel on the supply side and by macroeconomic conditions and investor environmental awareness on the demand side.

## Theory / model

The paper proposes no formal economic model. The economic object is the abnormal return (pollution premium) of a long-short portfolio sorted on toxic emission intensity, studied through two related semiparametric factor models adapted from Ang and Kristensen (2012) and Chen and Hong (2012).

**Hypotheses tested.** The null is that pollution exposure is not priced at all (alpha=0 in constant models), and the secondary null is that any pricing is time-invariant (alpha is constant over time). The paper rejects both nulls pre-2005 but fails to reject the second null post-2005.

**Identification.** The design is descriptive: no exogenous variation is exploited and no causal claim is made. The semiparametric model relaxes the constant-coefficient assumption, letting the data reveal the temporal shape of alpha without imposing a functional form (linear, structural-break, or otherwise). Driver regressions (Sections 5-6) establish associations between the estimated time-varying alpha and proxy variables; they do not identify causal effects.

## Method

### Partially time-varying factor model

The paper's baseline model has a time-varying intercept and constant factor loadings (Eq. 1, p. 2):

$$
r_t = \alpha_t + \sum_{j=1}^{J} \beta_j x_{jt} + \varepsilon_t, \qquad \varepsilon_t = \sigma_t e_t,
\quad t = 1, \ldots, T; \quad j = 1, \ldots, J \tag{1}
$$

where $$r_t$$ is the long-short portfolio excess return, $$x_{jt}$$ are observable common risk factors, $$\alpha_t$$ is the time-varying pollution premium, and $$\beta_j$$ is the constant factor loading. In matrix notation (Eq. 2):

$$
r_t = \alpha_t + Z_t^\top \beta + \varepsilon_t, \quad t = 1, \ldots, T \tag{2}
$$

### Semiparametric profile estimation

Following the `semiparametric-profile-estimation` approach of Fan and Huang (2005) and Chen and Hong (2012), $$\alpha_t$$ is estimated as a smooth function of rescaled time $$\tau_t = t/T \in [0,1]$$ by local linear kernel regression. Step 1 treats $$\beta$$ as known, computes the adjusted return $$\tilde{r}_t = r_t - Z_t^\top \beta$$, and minimizes the loss function (Eq. 5, p. 3):

$$
\sum_{t=1}^T \left[ \tilde{r}_t - \alpha(\tau) - \alpha^{(1)}(\tau)(\tau_t - \tau) \right]^2 K_h(\tau_t - \tau) \tag{5}
$$

where $$K_h(u) = K(u/h)/h$$ is the Epanechnikov kernel with bandwidth $$h$$ selected by leave-one-out cross-validation, and $$\alpha^{(1)}(\tau)$$ is the local linear slope. The infeasible local linear estimator of $$\alpha(\tau)$$ is (Eq. 6):

$$
\tilde{\alpha}(\tau) = [1,0] \left( D(\tau)^\top K_h(\tau) D(\tau) \right)^{-1} D(\tau)^\top K_h(\tau)(r - Z\beta) \tag{6}
$$

Step 2 substitutes $$\tilde{\alpha}$$ back into the original model, rearranges to isolate $$\beta$$, and estimates $$\hat{\beta}_{\text{ols}}$$ by OLS. The feasible semiparametric estimator is then (Eq. 10):

$$
\hat{\alpha}_{\text{semi}}(\tau) = S(\tau) \left( r - Z \hat{\beta}_{\text{ols}} \right) \tag{10}
$$

where $$S(\tau)$$ stacks the local-linear kernel smoothing matrices. Confidence bands are obtained by a wild bootstrap (5,000 replications; Appendix B), using the reflection method of Chen and Hong (2012) at boundary points to correct for boundary bias.

To correct for heteroscedasticity in the errors, the paper also constructs a weighted least-squares (WLS) estimator of $$\beta$$ (Appendix D), with weights $$\hat{\sigma}^2(\tau)$$ estimated from the kernel-local residuals.

### Fully time-varying factor model

An alternative model allows both $$\alpha_t$$ and factor loadings $$\beta_{jt}$$ to be time-varying (Eq. 11, p. 7):

$$
r_t = \alpha_t + \sum_{j=1}^{J} \beta_{jt} x_{jt} + \varepsilon_t, \quad t = 1, \ldots, T; \quad j = 1, \ldots, J \tag{11}
$$

Both functions are approximated locally by linear polynomials and estimated by minimizing the weighted sum of squared residuals (Eq. 13, Appendix A). This model allows the study of time variation in both the pollution premium level and the portfolio's exposure to common risk factors (Figures 3 and 4).

## Empirical specifications

### Portfolio construction

Following Hsu et al. (2023), the paper sorts firms into quintiles by annual toxic emission intensity, grouping firms relative to their Fama-French 49 industry peers each October (when updated TRI data become available). Emissions are scaled four ways: total assets (Panel A), property, plant and equipment (Panel B), sales (Panel C), and market capitalization (Panel D). Portfolios are value-weighted. The H-L long-short portfolio takes a long position in Quintile 5 (highest emission) and short in Quintile 1 (lowest emission).

Sample restrictions: common stocks (SHRCD = 10/11) on NYSE, AMEX, or NASDAQ; non-missing data in all three databases (CRSP, Compustat, EPA TRI); at least two years of Compustat history; financial firms (SIC 6000-6999) excluded.

### Constant factor model tests (Table 2)

Alphas for each quintile portfolio and the long-short portfolio are estimated under five factor models: CAPM (MKT only); FF3F (MKT, SMB, HML); FF4F (MKT, SMB, HML, UMD); FF5F (MKT, SMB, HML, RMW, CMA); HXZ q-factor model (MKT, SMB, I/A, ROE). Alphas are annualized (multiplied by 12); standard errors use Newey-West (1987) adjustments with 12 lags. The H-L alpha from the FF3F model is 5.28% annually (t=3.56, Table 2, Panel B), consistent across all factor models tested.

### Time-varying estimation (Figure 2; §4.2)

The partially time-varying model (Eq. 1) is estimated for each factor model. The semiparametric estimate $$\hat{\alpha}(\cdot)_{\text{semi}}$$ is plotted with 90% wild-bootstrap confidence bands (Figure 2). The FF3F-based estimate peaks near monthly alpha of 1.0 around 2000, then declines; the 90% CI lower bound is above zero from 1992 through approximately 2004-2005, after which the band crosses zero and the premium is statistically insignificant. The pattern holds across CAPM, FF4F, FF5F, and HXZ models.

Time-varying betas from the fully time-varying model (Figure 3 for CAPM; Figure 4 for FF3F) confirm that MKT beta is effectively zero throughout the sample, size beta is short-lived (significant only through the late 1990s), and HML beta is the key driver: significantly negative from 1992 to 1998 and from 2003 to 2013. Bansal et al. (2016) provide the baseline climate-risk-pricing motivation; Bekaert et al. (2022) provide the risk aversion and uncertainty indices used in the driver analysis.

### Driver regressions (Tables 4-9; §§5-6)

For each proxy variable $$Z_t$$, the paper regresses the estimated time-varying alpha:

$$
\hat{a}_t = \text{const} + \beta \times Z_t + \varepsilon_t
$$

Five categories of proxy variables are tested: (1) economic conditions (Shiller P/E ratio, luxury goods consumption, consumer sentiment index ICS; Table 4); (2) sustainable fund flows from Morningstar (percentage and dollar-amount; Table 5); (3) natural disaster severity from SHELDUS, defined as total fatalities and injuries relative to US population, scaled by 1,000 (Table 6); (4) environmental policy uncertainty via firm-level litigation risk, aggregated as total EPA civil cases filed against firms from year $$t+1$$ to $$t+5$$ (Table 7); (5) risk aversion and macroeconomic uncertainty indices from Bekaert et al. (2022), measured in monthly variance units and annual volatility percentage (Table 8).

The multivariate horse-race (Table 9) includes all proxies simultaneously. For the partially time-varying model (Panel A), risk aversion, macroeconomic uncertainty, natural disasters, and consumer sentiment are all significant at the 5% or 1% level across most factor models. Macroeconomic uncertainty exhibits the strongest negative effect (e.g., -6.094 for FF5F, t=-2.48) and natural disasters the strongest positive effect. The adjusted R-sq reaches 0.819 (FF5F) in Panel A and 0.864 (FF4F) in Panel B (fully time-varying model).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| EPA Toxic Release Inventory (TRI) | Annual firm-level toxic emission intensity (pounds released per year, 1991-2018); scaled four ways for emission-sorted quintile construction | [EPA TRI](/wiki/datasets/epa-tri/) |
| CRSP monthly returns and market data | Stock returns, market capitalization, exchange codes; monthly Oct 1992-Sep 2018 | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| Compustat annual fundamentals | Total assets, PP&E, sales for scaling; book values and firm characteristics for sample filters | [WRDS / Compustat](/wiki/commercial/wrds/) (licensed) |
| Fama-French factors (Kenneth French library) | FF3F, FF4F, FF5F factor returns; FF 49-industry classification for within-industry quintile sorts | [Ken French library](/wiki/datasets/ken-french/) |
| HXZ q-factors (Global-Q) | MKT, SMB, I/A, ROE factors for the HXZ model tests | No page yet (`data:global-q`) |
| Morningstar sustainable fund database | Quarterly total net assets, holdings, and returns for sustainable and ESG funds; 2007:Q4-2018:Q3 | No page yet |
| SHELDUS natural disaster database | County-level hazard event data (fatalities + injuries relative to US population), quarterly 12-quarter moving average; Arizona State University | No page yet |
| Bekaert et al. (2022) risk aversion and uncertainty indices | Monthly time-varying risk aversion and macroeconomic uncertainty indices (available at nancyxu.net) | No page yet |
| EPA ICIS enforcement and FRS data | Annual aggregate count of civil cases against firms from EPA Enforcement and Compliance History Online; matched to TRI via Facility Registry Service | No page yet |

Sample: October 1992 to September 2018 (312 months, monthly). TRI data are annual (from 1991), matched to stock data in October of each year.

## When to read the full paper

Use the [original](https://doi.org/10.1016/j.jbankfin.2026.107693) if you are: studying whether and how the pollution or ESG premium varies over time in the US; applying semiparametric partially or fully time-varying factor models to other long-short portfolios; investigating the supply-side (risk exposure) versus demand-side (investor preference) decomposition of time-varying premia; or seeking formal asymptotic derivations for the local-linear kernel estimator (Appendices A-D). The locators above point to exact tables and figures.

## Attribution and rights

Source: peer-reviewed, *Journal of Banking and Finance* vol. 187 (2026). This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. The article is paywalled; this page contains extracted text only and is not a redistribution of the work.

> Yin, Ximing, Deshui Yu, and Li Chen. "The time-varying pollution premium." *Journal of Banking and Finance* 187 (2026): 107693. DOI: [10.1016/j.jbankfin.2026.107693](https://doi.org/10.1016/j.jbankfin.2026.107693). © 2026 Elsevier B.V. All rights reserved.


==============================================================================
# Rookie Directors and Board Efficacy: Al Dah, Dah & Stathopoulos (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jcf/2026/aldah-rookie-directors-board-efficacy-2026/
# Distilled: Rookie board refreshment (not merely rookie presence) enhances CEO turnover-performance sensitivity, improves managerial incentives, and reduces discretionary accruals; seasoned refreshment improves investment efficiency and acquisition outcomes without hindering monitoring. Journal of Corporate Finance 96 (2026), CC BY 4.0. Eight core results with source locators, datasets used, the tested hypotheses, and estimating equations.
# Tags: paper-summary, corporate-governance, board-composition, dei, monitoring
==============================================================================

**What this is.** This is a distilled skeleton of Al Dah, Dah, and Stathopoulos (2026), "Rookie directors and board efficacy," Journal of Corporate Finance 96, article 102910. Read the [original paper](https://doi.org/10.1016/j.jcorpfin.2025.102910) to replicate or extend. Extraction is LLM-generated; it is not human-verified and has not been reproduced.

## TL;DR

Diversity, equity, and inclusion (DEI) pressures have increased the appointment of first-time ("rookie") directors on U.S. corporate boards. Critics argue that inexperienced directors harm governance by producing inefficient firm-director matches. The authors challenge this critique by separating two distinct concepts: the mere presence of rookie directors and rookie board refreshment, defined as the appointment of a rookie who brings new characteristics to the board relative to the director they replace. Using a Board Refreshment Index (BRI) across seven director characteristics for S&P 1500 firms from 2007 to 2022, they find that rookie refreshment (not mere presence) strengthens monitoring: it raises CEO turnover-performance sensitivity (R1), improves CEO pay-performance and pay-risk incentives (R2-R3), and reduces discretionary accruals (R4). Seasoned refreshment, by contrast, advances advising: it reduces overinvestment and underinvestment (R5), lowers acquisition frequency while improving deal quality (R6-R8). Neither refreshment type hinders the other's function. The results support H1 and H2 (rookie refreshment beats mere presence on monitoring and outperforms seasoned refreshment on monitoring) but only partly confirm H3 (seasoned refreshment does not dominate rookie refreshment on advising to a statistically significant degree for most outcomes).

## Core results

| # | Result | Locator | Magnitude as reported |
|---|--------|---------|----------------------|
| R1 | Rookie BRI raises CEO turnover-performance sensitivity | Table 2, p. 11 | Rookie BRI x ROA: -0.360\*\* (t=-2.667); Rookie BRI x BHAR: -0.149\*\*\* (t=-3.358); marginal effect approx. 22.5% (ROA) and 13.2% (BHAR) increase in forced CEO replacement probability per unit rise in rookie BRI |
| R2 | Rookie BRI improves CEO delta | Table 3 col 2, p. 12 | Rookie BRI: 0.024\*\*\* (t=3.374); 1-SD increase in rookie BRI -> +4.42% (+$37,992) relative to mean delta |
| R3 | Rookie BRI improves CEO vega | Table 3 col 5, p. 12 | Rookie BRI: 0.005\*\*\* (t=4.384); 1-SD increase -> +5.46% (+$7,915) relative to mean vega |
| R4 | Rookie BRI reduces discretionary accruals | Table 4 cols 2 and 5, p. 12-13 | Rookie BRI (modified Jones): -0.139\*\*\* (t=-4.217); ROA-adjusted Jones: -0.181\*\*\* (t=-2.942); 1-SD increase -> -3.74% SD in modified Jones model accruals |
| R5 | Seasoned BRI improves investment efficiency | Table 5 cols 3 and 6, pp. 13-15 | Seasoned BRI: 0.134\*\* (t=2.702, cash-leverage) / 0.110\*\*\* (t=3.863, residuals); 1-SD increase -> +4.05% and +3.32% investment for most underinvesting firms; -2.75% and -2.96% for most overinvesting firms |
| R6 | Seasoned BRI reduces acquisition count and likelihood | Table 6 cols 3 and 6, p. 15-16 | Seasoned BRI: count -0.008\*\* (t=-2.035) / likelihood -0.006\*\* (t=-1.972); 1-SD increase -> -2.57% acquisition count, -0.74% acquisition probability |
| R7 | Seasoned BRI raises announcement-period CAR | Table 7 col 6, p. 16-17 | Seasoned BRI: 0.001\*\* (t=2.042); 1-SD increase -> +5.25% SD in announcement-period CAR |
| R8 | Seasoned BRI raises post-acquisition BHAR | Table 8 cols 3 and 6, pp. 17-18 | Seasoned BRI: 0.009\*\*\* (t=2.946, 504-day) / 0.010\*\* (t=2.539, 756-day); 1-SD increase -> +5.2% SD (504-day) and +4.3% SD (756-day) daily BHAR |

**Overall (paper's conclusion).** Rookie refreshment enhances board monitoring without impairing advising; seasoned refreshment advances advising. The two refreshment types are complementary governance strategies. Neither the mere presence of rookies nor seasoned refreshment significantly affects monitoring, and neither rookies nor rookie refreshment significantly affects advising outcomes. Results survive entropy-balanced matching, firm fixed effects, 2SLS with two instruments (airport proximity and director supply from same-city firms), and multiple robustness tests including exclusion of gender from the BRI and restriction to pre-SB-826 periods (pp. 19-20).

## Theory / model

The paper has no formal mathematical model. Its theoretical framework draws on two bodies of prior work.

**Upper Echelons Theory (UET).** Hambrick (2007) argues that directors' observable traits significantly shape corporate outcomes. Applied here following Terbeck et al. (2022) and Dah et al. (2024): a rookie director's temporary "reputational window" during the first three years of service creates heightened career concerns (Fama and Jensen (1983); Jiang et al. (2016)) and cognitive independence from entrenched board coalitions (Hambrick and Fukutomi (1991)), motivating more diligent monitoring behavior. This career-concern effect fades as the director transitions to seasoned status, joining existing coalitions and socialization structures (Kang et al. (2016)).

**Agency and Resource Dependence Theory.** Agency theory predicts that stronger internal monitoring (CEO turnover-performance sensitivity, incentive alignment, accruals management) reduces managerial slack. Resource Dependence Theory (Pfeffer and Salancik 1978) predicts that experienced directors supply strategic advice by leveraging their established networks, long-horizon industry knowledge, and prior board experience (Kim et al. (2014); Adams et al. (2010)).

**Key distinction.** The paper argues that board refreshment and rookie status are distinct concepts. Rookie status is a temporary behavioral state driven by reputational incentives; refreshment records whether the incoming director changes the composition of the board along observable characteristics. A rookie non-refreshment (a rookie who replicates existing board characteristics) leaves dynamics unchanged; a rookie refreshment disrupts established norms. This motivates separating the Rookie Presence share (RookieP) from the Board Refreshment Index (BRI) in all empirical tests.

**Hypotheses** (p. 5-6):

- **H1**: Rookie refreshment has a more pronounced effect on board functionality than the mere presence of rookie directors.
- **H2**: Rookie refreshment provides more effective monitoring compared to seasoned refreshment.
- **H3**: Seasoned refreshment provides more effective advising compared to rookie refreshment.

## Method

The central measurement object is the **Board Refreshment Index (BRI)**, developed following Dah et al. (2024). The BRI is an additive index computed annually for each firm across seven director characteristics: gender, nationality, age, classification (insider/outsider), interlocks, education, and financial expertise (p. 6-7). For each characteristic, a refreshment score of +1 is assigned when the quality of the added director refreshes that dimension (e.g., a female added to a male-dominated board), and -1 when it de-refreshes. Net refreshment per characteristic is the difference between total refreshment and total de-refreshment. The BRI sums net refreshments across all seven characteristics. A **Rookie BRI** (or Seasoned BRI) is identified only when the board change involves a rookie (or seasoned) director.

The **OverInvest** measure (for the investment-efficiency test) is constructed as the average of two decile ranks of the firm's likelihood of overinvestment, following Biddle et al. (2009). Cash-leverage rank is based on cash level and leverage (multiplied by minus one); residuals rank uses the residual from eq. (3) (see below), estimated within each industry-year group. Each rank is rescaled to [0, 1] and averaged.

All regressions include industry dummies (Fama-French 48-industry) and year dummies. Standard errors are clustered at the firm level. CEO turnover specifications are probit; delta/vega and discretionary-accruals specifications are OLS; investment-efficiency specifications are OLS with interaction terms; acquisition count uses a zero-inflated Poisson (ZIP) regression; acquisition likelihood uses a probit; acquisition performance uses a Heckman two-stage estimator.

To address endogeneity of refreshment, Section 5.3 runs 2SLS using two instruments: (i) RDist, the reciprocal distance from the firm's headquarters to the nearest airport hub (proxying commute ease for directors, which facilitates board turnover without directly affecting governance outcomes); (ii) DirSupply, the average percentage of other companies headquartered in the same city (OCHSC) that refresh with rookies (or seasoned directors, matched by characteristic) from a category different from the board's dominant majority in the prior year. The Hansen J-test does not reject overidentification restrictions, and first-stage F-statistics exceed 10 in all specifications (p. 20).

## Empirical specifications

### Investment benchmark (equation 3, p. 8)

The residuals-based OverInvest measure requires first estimating the industry-year average investment level:

$$\text{Investment}_{j,t} = \beta_0 + \beta_1 \text{SalesGrowth}_{j,t-1} + \varepsilon_{j,t} \tag{3}$$

where $$j = 1, \ldots, 48$$ denotes Fama-French 48 industries. Investment is R&D + capex + acquisition expenditure minus cash from property/plant/equipment sales, all scaled by lagged total assets and multiplied by 100. Residuals are ranked into deciles within each industry-year group to form the residuals-based OverInvest.

### Discretionary accruals (equations 1-2, p. 7-8)

Two variants of the modified Jones model are used as monitoring proxies (Dechow et al. (1995); Kothari et al. (2005)):

$$\text{Accruals}_{i,t} = \beta_0 \!\left(\frac{1}{\text{ASSETS}_{i,t-1}}\right) + \beta_1 \frac{\Delta\text{REV}_{i,t} - \Delta\text{REC}_{i,t}}{\text{ASSETS}_{i,t-1}} + \beta_2 \!\left(\frac{\text{PPE}_{i,t}}{\text{ASSETS}_{i,t-1}}\right) + \varepsilon_{i,t} \tag{1}$$

$$\text{Accruals}_{i,t} = \beta_0 \!\left(\frac{1}{\text{ASSETS}_{i,t-1}}\right) + \beta_1 \!\left[\frac{\Delta\text{REV}_{i,t} - \Delta\text{REC}_{i,t}}{\text{ASSETS}_{i,t-1}}\right] + \beta_2 \!\left(\frac{\text{PPE}_{i,t}}{\text{ASSETS}_{i,t-1}}\right) + \beta_3 \text{ROA}_{i,t} + \varepsilon_{i,t} \tag{2}$$

Discretionary accruals are the residuals from these cross-sectional regressions run within each two-digit SIC code and year group. A decrease (increase) in discretionary accruals signals improvement (deterioration) in financial reporting oversight.

### CEO turnover-performance sensitivity (equation 4, p. 9; Table 2)

$$\text{Turnover}_{i,t+1} = \beta_0 + \beta_1 \frac{\text{RookieP}_{i,t}}{\text{BRI}_{i,t}} + \beta_2 \text{Performance}_{i,t} + \beta_3 \frac{\text{RookieP}_{i,t}}{\text{BRI}_{i,t}} \cdot \text{Performance}_{i,t} + \beta_4 \text{Controls}_{i,t} + \text{Year} + \text{Industry} + \varepsilon_{i,t} \tag{4}$$

Probit on forced CEO turnover. Performance is the industry-adjusted 3-year average ROA (columns 1-3) or the industry-adjusted 3-year average daily BHAR (columns 4-6). The variable of interest is $$\beta_3$$, the interaction of BRI with firm performance: a negative $$\beta_3$$ means higher BRI amplifies the negative impact of poor performance on CEO retention, i.e., the board replaces underperforming CEOs more aggressively. Controls include leverage, CAPX/Assets, Sales/Assets, Slack, ESG, Percent Outside, Percent Female, Board Size, E-Index, CEO Tenure, CEO Ownership, CEO Age, Duality, and CEO Gender. N = 17,463.

### CEO incentives (equation 5, p. 10; Table 3)

$$\frac{\text{Delta}_{i,t}}{\text{Vega}_{i,t}} = \beta_0 + \beta_1 \frac{\text{RookieP}_{i,t}}{\text{BRI}_{i,t}} + \beta_2 \text{Controls}_{i,t} + \text{Year} + \text{Industry} + \varepsilon_{i,t} \tag{5}$$

OLS on lagged delta (dollar change in CEO pay per 1% change in stock price) and vega (dollar change in CEO pay per 1% change in stock volatility). A higher delta aligns managerial pay with firm performance; a higher vega encourages value-enhancing risk-taking. Following Core and Guay (2002) and Coles et al. (2006). N = 14,045.

### Discretionary accruals (equation 6, p. 11; Table 4)

$$\text{DiscretionaryAccruals}_{i,t+1} = \beta_0 + \beta_1 \frac{\text{RookieP}_{i,t}}{\text{BRI}_{i,t}} + \beta_2 \text{Controls}_{i,t} + \text{Year} + \text{Industry} + \varepsilon_{i,t} \tag{6}$$

OLS on the next-period discretionary accruals residuals from eqs. (1) and (2). Following Faleye et al. (2011) and Kim et al. (2014). N = 13,170.

### Investment efficiency (equation 7, pp. 13-15; Table 5)

$$\text{Investment}_{i,t+1} = \beta_0 + \beta_1 \frac{\text{RookieP}_{i,t}}{\text{BRI}_{i,t}} + \beta_2 \text{OverInvest}_{i,t+1} + \beta_3 \frac{\text{RookieP}_{i,t}}{\text{BRI}_{i,t}} \cdot \text{OverInvest}_{i,t+1} + \beta_4 \text{GOV}_{i,t} + \beta_5 \text{GOV}_{i,t} \cdot \text{OverInvest}_{i,t+1} + \beta_6 \text{Controls}_{i,t+1} + \text{Industry} + \varepsilon_{i,t} \tag{7}$$

OLS on total investment (R&D + capex + acquisitions minus divestitures, scaled by lagged assets). GOV includes the entrenchment index, institutional ownership, and analyst following, as in Biddle et al. (2009). The key estimates are $$\beta_1$$ (the effect of BRI for underinvesting firms, where OverInvest ≈ 0; a positive $$\beta_1$$ means BRI increases investment, reducing underinvestment) and $$\beta_1 + \beta_3$$ (the combined effect for overinvesting firms; the PDF reports this sum is negative and significant, meaning BRI decreases investment and thereby reduces overinvestment propensity). N = 14,405.

### Acquisition count and likelihood (equations 8-9, p. 17; Table 6)

$$\text{AcquisitionCount}_{i,t+1/t+2} = \beta_0 + \beta_1 \frac{\text{RookieP}_{i,t}}{\text{BRI}_{i,t}} + \beta_2 \text{Controls}_{i,t} + \text{Year} + \text{Industry} + \varepsilon_{i,t} \tag{8}$$

$$\text{AcquisitionLikelihood}_{i,t+1/t+2} = \beta_0 + \beta_1 \frac{\text{RookieP}_{i,t}}{\text{BRI}_{i,t}} + \beta_2 \text{Controls}_{i,t} + \text{Year} + \text{Industry} + \varepsilon_{i,t} \tag{9}$$

Eq. (8) is a zero-inflated Poisson (ZIP) regression for acquisition count (number of completed deals in years t+1 and t+2, restricted to deal value >= $1 million and acquirer owns < 50% pre-deal). Eq. (9) is a probit for acquisition likelihood (binary, 1 if count > 0). Sample: U.S. acquirers from SDC Platinum, excluding simultaneous-announcement deals. N = 15,966. Following Faleye et al. (2011) and Kim et al. (2014).

### Acquisition performance (equation 10, pp. 17-18; Tables 7-8)

$$\text{AcquisitionPerformance} = \beta_0 + \beta_1 \frac{\text{RookieP}_{i,t}}{\text{BRI}_{i,t}} + \beta_2 \text{InverseMillsRatio} + \beta_3 \text{Controls}_{i,t} + \text{Year} + \text{Industry} + \varepsilon_{i,t} \tag{10}$$

Heckman second stage, where the first stage is the acquisition likelihood probit (eq. 9). Acquisition performance is either: (a) announcement-period CAR in the (-3, +3) window using the market model estimated on the 255 days before a 45-day gap period (Table 7, N = 4,070); or (b) long-run buy-and-hold daily abnormal returns over 504 days and 756 days post-acquisition effective date (Table 8, N = 3,302 and 3,276 respectively).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---------|--------------|-----------|
| Compustat (via WRDS) | Firm financial data: assets, leverage, ROA, investment, accruals, sales, capex, R&D (2007-2022) | [WRDS](/wiki/commercial/wrds/) |
| CRSP (via WRDS) | Daily stock returns for BHAR and industry-adjusted returns | [WRDS](/wiki/commercial/wrds/) |
| BoardEx | Director characteristics for all seven BRI dimensions: gender, nationality, age, classification, interlocks, education, financial expertise | no page yet |
| ExecuComp (via WRDS) | CEO compensation (cash, equity), tenure, ownership; delta and vega computations | [WRDS](/wiki/commercial/wrds/) |
| ISS (Institutional Shareholder Services) | Six anti-takeover provisions for the Bebchuk E-Index | no page yet |
| SDC Platinum | Acquisition transactions: deal value, completion date, target ownership share | [SDC Platinum](/wiki/commercial/sdc-platinum/) |
| I/B/E/S (via WRDS) | Analyst following count | [I/B/E/S](/wiki/commercial/ibes/) |
| Thomson Reuters 13F (via WRDS) | Institutional ownership (percentage held by institutions) | [Thomson 13F](/wiki/commercial/thomson-13f/) |
| MSCI | ESG disclosure scores | [KLD / MSCI ESG](/wiki/commercial/kld/) |

Sample scope: S&P 1500 firms, 2007-2022 (starting 2007 to account for FASB and SEC disclosure-standard shifts). Base sample 20,129 firm-years before attrition on missing variables; regression samples range from 13,170 (accruals models) to 17,463 (turnover models) and 15,966 (acquisition models). All continuous variables winsorized at 1%. Industries follow the Fama-French 48-industry classification.

## When to read the full paper

Read the full paper (doi:10.1016/j.jcorpfin.2025.102910) if you:

- Study DEI board mandates and need evidence that distinguishes refreshment effects from the mere addition of inexperienced directors (the BRI construct in Section 3.1 and Appendix A is the most portable contribution).
- Need baseline OLS/probit results on CEO turnover-performance sensitivity as a function of board composition (Table 2, p. 11).
- Are replicating the Faleye et al. (2011) or Kim et al. (2014) board-advising framework and want updated estimates for the 2007-2022 period using a finer board-change measure (Tables 5-8).
- Need the full robustness battery: entropy-balanced matching, firm fixed effects, 2SLS with airport-proximity and director-supply instruments, alternative monitoring proxies, board diversity mandate exclusions (Section 5 and Online Appendices S1-S5).
- Are modeling how Chen and Keefe (2020) or Kang et al. (2016) results generalize to U.S. S&P 1500 boards and need comparisons across rookie-presence vs. rookie-refreshment specifications side-by-side (all main tables).

## Attribution and rights

This article is published open access under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

> Bilal Al Dah, Mustafa A. Dah, Konstantinos Stathopoulos, "Rookie directors and board efficacy," *Journal of Corporate Finance* 96 (2026) 102910. https://doi.org/10.1016/j.jcorpfin.2025.102910. Copyright 2025 The Authors. Published by Elsevier B.V. under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/).

This page is an LLM-distilled extract (paper-distiller, claude-sonnet-4-6, 2026-06-26). It is not human-verified and does not reproduce figures, tables, or extended text beyond brief excerpts needed for locators. The full article is freely accessible via the DOI above.


==============================================================================
# Local Peer Effects and Corporate Investment: Bao & Goetz (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jcf/2026/bao-local-peer-effects-corporate-investment-2026/
# Distilled: Using staggered U.S. state corporate income tax changes as an instrument within cross-state Economic Areas, Bao and Goetz identify a positive causal effect of local peer firms' investment on a firm's own investment, confirmed separately for physical and intangible capital, with learning from same-type peers as the primary mechanism. Journal of Corporate Finance vol. 97 (2026), paywalled. Seven core results with source locators, datasets used, and empirical specifications.
# Tags: paper-summary, corporate-finance, corporate-investment, peer-effects
==============================================================================

**What this is.** A distilled skeleton of Bao and Goetz (2026). Read the
original at https://doi.org/10.1016/j.jcorpfin.2025.102935 to replicate or
extend.

## TL;DR

Bao and Goetz study how a firm's investment is shaped by the investment of
neighboring peer firms within the same local Economic Area (EA) and
Fama-French industry. Using a large panel of U.S. public firms from 1989
to 2014, OLS results confirm a positive correlation between a firm's
investment and local peer firms' average investment, consistent with
Dougal et al. (2015) and Bustamante and Fresard (2021). To establish a
causal link, the paper exploits
staggered increases in U.S. state corporate income tax rates. Because EAs
span multiple states, a tax increase in one state depresses investment in
the taxed state without directly affecting investment conditions for peer
firms in other states of the same EA. The resulting variation in peer
investment is used as an instrument in a 2SLS framework. 2SLS results
confirm a positive causal peer effect: a one-standard-deviation increase in
instrumented peer investment raises a firm's total investment by roughly
1.57 percentage points (about 6.9% of average total investment). Separating
physical and intangible capital, the paper finds that peer effects in
physical investment do not spill over to a firm's intangible investment, and
vice versa. This type-specificity is consistent with managers learning from
peers who invest in the same type of capital. Further, peer effects in
physical investment are stronger among firms with weaker information
precision (higher earnings or equity volatility relative to local peers),
and peer effects in intangible investment are stronger in knowledge-intensive
local industries, consistent with a learning mechanism. Strategic competition
(product-market substitution) does not explain the results.

## Core results

| # | Result | Locator | Magnitude as reported |
|---|---|---|---|
| R1 | OLS: local peer total investment on firm total investment | Table 2 col 2, p. 6 | 0.088\*\*\* (SE 0.016); 0.407 pp per 1-SD peer investment change |
| R2 | 2SLS causal peer effect on total investment (fraction-of-peers IV) | Table 5 col 1, p. 12 | 0.772\*\*\* (SE 0.285); 1.57-pp increase per 1-SD instrumented peer investment (~6.9% of mean total investment) |
| R3 | First-difference OLS: state corporate income tax rise on firm total investment | Table 3 col 1, p. 9 | -0.629\*\* (SE 0.258), coefficients x100; approximately -63 bp drop |
| R4 | 2SLS: local peer physical investment on firm physical investment | Table 10 col 2, p. 17 | 0.628\*\*\* (SE 0.235) |
| R5 | 2SLS: local peer intangible investment on firm intangible investment | Table 10 col 4, p. 17 | 0.869\*\* (SE 0.425) |
| R6 | No cross-type peer effect: intangible peer investment on firm physical investment | Table 11 Panel A col 1, p. 18 | -0.115 (SE 0.197), not significant |
| R7 | Signal precision moderates physical peer effect (equity-vol interaction) | Table 12 Panel A col 1, p. 20 | Interaction: 0.914\*\*\* (SE 0.298); base peer effect: 0.497\*\*\* (SE 0.147) |

**Overall (paper's conclusion).** Local peer firms exert a positive causal
influence on a firm's investment behavior. This result is robust to
alternative IV constructions, exclusion of indirect tax-spillover channels
(customer-supplier links, subsidiaries in taxed states), local expansion
opportunity concerns, local demand shocks, and alternative clustering of
standard errors. The type-specificity of peer effects (R4, R5, R6) and the
learning-incentive heterogeneity (R7 and Table 12 Panel B) are consistent
with managers learning from peers who invest in the same type of capital,
particularly when information about future investment conditions is scarce.

## Theory / model

The paper has no formal model. It derives sign predictions from two
competing theoretical mechanisms and tests which one dominates.

**Learning / information sharing (positive peer effects).** Research on
social learning argues that managers can infer information about future
conditions by tracking the investment behavior of neighboring peers
(Scharfstein and Stein (1990); Bikhchandani et al. (1992)). When a
manager's own signal about the future is noisy and informational
asymmetries are significant, observing peers' investment reduces uncertainty
and induces correlated investment behavior (strategic complements). This
force predicts same-sign peer effects that are stronger when the learning
incentive is high (weaker own signal precision, i.e., higher earnings or
equity volatility) and when knowledge can plausibly diffuse locally (higher
R&D intensity of the peer set).

**Strategic product-market competition (negative peer effects).** An
increase in local investment may raise the price of shared local inputs and
intensify product-market competition, inducing neighboring firms to reduce
investment (Dixit (1980); Gal-Or (1987)). This force predicts negative peer
effects.

**Type-specificity prediction.** Drawing on learning theories, the paper
hypothesizes that peer effects in physical investment influence a firm's
physical investment but not its intangible investment, because observing a
neighbor's factory-building decision provides a clearer signal about
physical investment conditions than about R&D conditions, and vice versa.
This prediction is supported by Table 11 (R6 above).

The identification assumption is that a neighboring state's decision to
raise corporate income taxes is exogenous to the investment of firms in
other states of the same EA. Table 4 (p. 10) shows that state-level
aggregate investment and neighboring states' aggregate investment do not
predict state tax increases, and that neighboring states' tax policies are
not correlated with a home state's decision to raise taxes, supporting
exogeneity of the instrument.

## Method

The paper applies 2SLS within a first-difference panel framework. The
first-difference transformation eliminates time-invariant firm-level
unobservables; EA, industry, and year fixed effects absorb remaining
common variation.

**Investment measures** (PDF p. 4, Appendix A):

Physical investment rate (eq. 1):

$$I^{phy}_{i,t} = \frac{capx_{i,t}}{K^{total}_{i,t-1}} \tag{1}$$

Intangible investment rate (eq. 2):

$$I^{int}_{i,t} = \frac{\text{R\&D} + (0.3 \times \text{SG\&A})}{K^{total}_{i,t-1}} \tag{2}$$

Total investment rate (eq. 3):

$$I^{total}_{i,t} = I^{phy}_{i,t} + I^{int}_{i,t} \tag{3}$$

where $$K^{total}$$ is the replacement cost of physical capital (Compustat
item ppegt) plus intangible capital, both estimated following
Peters and Taylor (2017).

**Instrumental variables.** Three instruments capture the exogenous component
of the average peer investment change induced by state corporate income tax
changes. The first-stage instruments exploit variation across states within
the same cross-state EA:

1. *Fraction of local peers affected* ($$\%\text{LocalPeersAffected}$$): the
   fraction of firm $$i$$'s local peers located in a state that raises
   corporate income taxes in year $$t$$. A higher fraction produces a larger
   negative shock to average peer investment.

2. *Predicted state-specific* $$\Delta \bar{I}$$: the coefficient on the tax
   increase dummy from equation (5) is estimated state by state to recover the
   state-specific investment effect of a tax rise; the average predicted peer
   investment change across other local peers is then computed.

3. *Predicted state-industry-specific* $$\Delta \bar{I}$$: the same
   procedure at the state-industry level, capturing heterogeneity in how
   the tax shock transmits across industries.

All three instruments are highly significant in first-stage regressions
(KP Wald F-statistics: 130.4, 211.7, and 283.3 for the three 2SLS
specifications in Table 5, p. 12), satisfying instrument relevance.
The first-stage coefficients on the fraction-of-peers instrument (-0.017***)
and the predicted investment changes (+0.842***; +0.681***) have the
expected signs (Panel B, Table 5).

The cross-type peer effect analysis in Section 6.2 uses separate instruments
for physical and intangible peer investment. The signal-precision
heterogeneity analysis in Section 6.3 interacts the instrumented peer
investment with above-median dummy variables for equity volatility, ROA
volatility, and local-industry R&D intensity.

## Empirical specifications

**Eq. (4): Benchmark first-difference OLS** (PDF p. 5)

$$\Delta I_{i,t} = \beta \Delta\bar{I}_{-i,a,j,t} + \Delta X'_{i,t} \rho + \delta_{a/j/t} + \varepsilon_{i,t} \tag{4}$$

where $$\Delta I_{i,t}$$ is the annual change in firm $$i$$'s total
investment rate; $$\Delta\bar{I}_{-i,a,j,t}$$ is the change in average investment rate of
firm $$i$$'s local peers in the same EA $$a$$ and Fama-French 12 industry
$$j$$, excluding firm $$i$$; $$X'_{i,t}$$ includes two additional controls
for the general industry investment trend (firms in the same industry outside
the EA) and the local area investment trend (firms in the same EA but
different industries); $$\delta_{a/j/t}$$ are EA, industry, and year fixed
effects. Standard errors are clustered at the firm level. Estimated on
75,858 firm-years (level model) and 64,675 firm-years (first-difference
model, Table 2).

OLS peer effect (Table 2 col 2, p. 6): $$\hat{\beta} = 0.088^{***}$$ (SE
0.016), implying a 0.407 pp increase in total investment per one-SD increase
in local peer investment. OLS does not allow causal interpretation due to
common local latent factors.

**Eq. (5): Tax effect first-difference OLS** (PDF p. 8)

$$\Delta I_{i,t} = \beta_1 \text{TaxInc}_{s,t-1} + \beta_2 \text{TaxCut}_{s,t-1} + \gamma \Delta X_{i,t} + \delta_{a/j/t} + \varepsilon_{i,t} \tag{5}$$

where $$\text{TaxInc}_{s,t-1}$$ ($$\text{TaxCut}_{s,t-1}$$) equals 1 if
state $$s$$ increases (decreases) its corporate income tax rate in year
$$t-1$$, and 0 otherwise; $$X_{i,t}$$ includes firm-level controls (Tobin's
Q, cashflow, log assets) and macroeconomic state-level controls (GSP growth,
unemployment, union penetration, population growth, per capita income
growth). EA and industry fixed effects are included. Standard errors are
clustered at the firm level. Coefficients multiplied by 100; sample 67,319
to 59,291 firm-years (Table 3). A tax increase reduces total investment by
approximately 63 basis points ($$\hat{\beta}_1 = -0.629^{**}$$, SE 0.258,
Table 3 col 1). This is consistent with Mukherjee et al. (2017), who find
state tax increases reduce innovative investment, and motivates using the tax
shock as an instrument for peer investment.

**Eq. (6): Pre/post event dynamics** (PDF p. 9)

$$I_{i,t} = \sum_{k=-4}^{4} \beta_k \text{TaxInc}_{s,t+k} + \delta_i + \delta_t + \varepsilon_{i,t} \tag{6}$$

This regression traces the investment path four years before and after a
state corporate income tax increase, with the tax increase year as the
reference. Figure 2 (p. 8) shows flat pre-trends (no anticipatory effects)
and declining investment in post-event years, supporting the parallel-trends
assumption and the validity of the instrument.

**2SLS second stage** (Table 5, p. 12)

$$\Delta I_{i,t} = \alpha \Delta \hat{\bar{I}}_{-i,a,j,t} + \Delta X'_{i,t} \rho + \delta_{a/j/t} + \varepsilon_{i,t}$$

where $$\Delta \hat{\bar{I}}_{-i,a,j,t}$$ is the instrumented change in
peer investment from one of the three first-stage IV constructions. The
sample is restricted to firms in EAs spanning more than one state (3,871
firms, 28,066 firm-years). Year, industry, and EA fixed effects are
included. All three IV specifications yield positive and significant peer
effect estimates (0.772***, 0.829***, 1.082*** across the three columns of
Table 5 Panel A). Economic magnitude: 1.57 pp increase in total investment
per one-SD of instrumented peer investment (Table 5 col 1 discussion, p. 12),
equal to about 6.9% of average total investment.

**Signal precision / learning heterogeneity** (PDF p. 19)

$$\Delta I_{i,t} = \beta_1 \Delta \bar{I}_{-i,a,j,t} \times \text{Above}_{i,t} + \beta_2 \Delta \bar{I}_{-i,a,j,t} + \Delta X'_{i,t} \rho + \delta_{a/j/t} + \varepsilon_{i,t}$$

where $$\text{Above}_{i,t}$$ is a dummy equal to 1 if the firm's equity
volatility (or ROA volatility, or local-industry R&D ratio) exceeds the
sample median of its local peers. A significant positive $$\beta_1$$
indicates stronger peer effects for firms with weaker own information
precision. Physical investment: $$\hat{\beta}_1 = 0.914^{***}$$ (SE 0.298)
for equity volatility; intangible investment: $$\hat{\beta}_1 = 1.244^{***}$$
(SE 0.387) for local-industry R&D stock ratio (Table 12, p. 20). The peer
investment variable is instrumented using the state-industry predicted IV
throughout.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CRSP/Compustat Merged | Firm-level investment (capx, R&D, SG&A), total capital (ppegt), assets, Tobin's Q, cashflow, stock returns and equity volatility; NYSE, AMEX, NASDAQ; 1989-2014 | [WRDS](/wiki/commercial/wrds/) |
| BEA Economic Areas | Geographic definition of local peer groups as regional markets; 2004 BEA boundaries; cross-state EAs identify the IV subsample (Fig. 1, p. 5) | no page yet |
| Fama-French 12 industries | Industry classification for peer group construction and industry fixed effects | [Ken French library](/wiki/datasets/ken-french/) |
| State corporate income tax rates | Exogenous investment shock; Heider and Ljungqvist (2015) panel of 121 U.S. state tax changes 1989-2011, extended to 2014 using Tax Foundation data | no page yet |
| State macroeconomic controls | GSP growth (BEA), unemployment rate (BLS), union penetration (Hirsch and Macpherson 2003), population growth and per capita income growth (Census) | no page yet |

**Sample.** OLS sample: 9,099 publicly listed U.S. firms on NYSE, AMEX, or
NASDAQ with non-missing total investment data, fiscal years 1989-2014
(75,858 firm-years). Firms with fewer than five local peers in a given year
are excluded. Average firm assets: $3.0 billion; average total investment
rate: 22.7% of total capital (one third physical, two thirds intangible).
Average number of local peer firms per EA: 42. Approximately 47.9% of
sample firms are headquartered in cross-state EAs. 2SLS subsample: 3,871
firms, 28,066 firm-years (cross-state EAs only). All variables winsorized
at the 0.5 percentile in each tail.

## When to read the full paper

Read the full paper if you need:
- A causal IV design for local peer effects in corporate investment using
  state corporate income tax shocks (Tables 5-9), including robustness for
  indirect tax-spillover channels (Table 6), local expansion opportunities
  (Table 7), local demand shocks (Table 8), and fixed-effects alternatives
  (Table 9).
- Evidence on the type-specificity of peer effects: physical capital peers
  affect physical investment but not intangible investment, and vice versa
  (Table 11), with heterogeneity by firm operational strategy (Panel B).
- Tests of the learning-from-peers mechanism via signal-precision and
  knowledge-spillover proxies (Table 12).
- The cross-state EA identification strategy, which can be adapted to other
  firm-level outcomes affected by local conditions.

## Attribution and rights

Paywalled. Access at https://doi.org/10.1016/j.jcorpfin.2025.102935.
No open-access or CC license found in Crossref metadata (checked 2026-06-26;
Elsevier TDM and STM-ASF licenses only). Rights held by Elsevier B.V.

Citation: Bao, Y. and Goetz, M. R. (2026). Local peer effects and corporate
investment. *Journal of Corporate Finance*, 97, 102935.
https://doi.org/10.1016/j.jcorpfin.2025.102935

This page is LLM-distilled, not human-verified, and not a reproduction of
the paper. All quantitative results are extracted from the source PDF with
source locators.


==============================================================================
# Pay Restrictions and Labor Investment: Cao, Hasan, Huang & Zhao (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jcf/2026/cao-pay-restrictions-labor-investment-2026/
# Distilled: Exploiting China's 2014 SOE executive compensation reform as a quasi-natural experiment, this paper shows pay restrictions reduce abnormal labor investment in state-owned enterprises by 3.91 to 4.82 percent, operating through strengthened internal governance and reduced social comparison between executives and rank-and-file employees. Journal of Corporate Finance 2026, paywalled. Eight core results with source locators, datasets used, and the empirical specifications.
# Tags: paper-summary, corporate-governance, executive-compensation, labor-investment
==============================================================================

**What this is.** A distilled skeleton of Cao, Hasan, Huang & Zhao (2026), "Pay restrictions and labor investment," *Journal of Corporate Finance* 99, 102990. Read [the original](https://doi.org/10.1016/j.jcorpfin.2026.102990) to replicate or extend. This summary is LLM-distilled, not human-verified, and not reproduced.

## TL;DR

Exploiting China's 2014 SOE executive compensation reform as a quasi-natural experiment, this paper shows that pay restrictions causally reduce abnormal labor investment (ALI) in state-owned enterprises by 3.91 to 4.82 percent relative to non-SOEs, with the reduction driven by over-investment (not under-investment). The reform strengthens internal governance (narrowing the pay gap between CEOs and subordinate executives) and reduces social comparison (narrowing the pay gap between executives and rank-and-file employees), both of which in turn reduce ALI. Additional results show that the reform cuts over-hiring specifically, improves labor quality, and increases employee well-being (stock grants, work safety, vocational training, director communication channels).

## Core results

| # | Result | Locator | Magnitude as reported |
|---|---|---|---|
| R1 | Pay restrictions reduce SOE abnormal labor investment | Table 4 cols 1-4, p. 15 | SOE x Restriction = -0.039\*\*\* to -0.062\*\*\*; 3.91% to 4.82% reduction in ALI |
| R2 | Reduction concentrated in over-investment, not under-investment | Table 4 cols 5-6, p. 15 | SOE x Restriction = -0.075\*\*\* (over-invest); 0.003 n.s. (under-invest); F-test p < 0.001 |
| R3 | Reform cuts average SOE executive compensation | Table 3 Panel A col 1, p. 14 | SOE x Restriction = -0.096\*\*\*; 10.08% reduction (~49,000 CNY/year) |
| R4 | Reform increases internal governance effectiveness by 14.05% | Table 5 col 1, p. 16 | SOE x Restriction = 0.141\*\* on internal governance score |
| R5 | Higher internal governance firms see 5.57% reduction in ALI | Table 5 cols 2-3, p. 16 | SOE x Restriction = -0.056\*\* (high IG), -0.032 n.s. (low IG); F-test p = 0.076 |
| R6 | Reform reduces CEO-to-rank pay disparity in SOEs | Table 6 Panel A col 1, p. 17 | SOE x Restriction = -1.180\*\*\*; 4.68% ALI reduction for low-disparity firms |
| R7 | Reform reduces over-hiring by 8.73% | Table 8 col 1, p. 23 | SOE x Restriction = -0.087\*\*\*; under-hiring, over-firing, under-firing insignificant |
| R8 | Reform improves labor quality and employee well-being | Table 9, p. 23 | Labor quality +1.33% (col 1); stock ownership +7.63%, work safety +6.22%, training +3.41%, communication +3.73% |

**Overall.** China's 2014 SOE executive compensation reform significantly reduces abnormal labor investment, primarily over-investment (over-hiring), through two channels: enhanced internal governance (aligning interests between CEOs and subordinate executives) and reduced social comparison (narrowing pay disparities between executives and rank-and-file employees). Additional effects include improved labor quality and employee well-being, suggesting pay restrictions promote more efficient labor resource allocation in weak institutional environments.

## Theory / model

The paper has no formal mathematical model. Instead, it develops two competing hypotheses grounded in agency theory and institutional economics, then adjudicates between them empirically:

**H1a (agency / governance hypothesis):** Executive pay restrictions are associated with lower SOE abnormal labor investment. Excessive executive pay misaligns managerial incentives (Bebchuk and Fried (2003)), enabling executives to over-hire for political or status reasons beyond shareholder value. Prior work documents that CEO-director ties exacerbate labor investment inefficiency (Khedmati, Sualihu & Yawson (2020)); this paper examines whether pay restrictions address the underlying agency problem. Pay restrictions curb this self-serving over-investment.

**H1b (efficiency-wage / shirking hypothesis):** Pay restrictions are associated with higher SOE abnormal labor investment. Caps on compensation weaken incentives, reduce executive effort (Holmstrom (1979)), and may lead to greater labor investment inefficiency through reduced monitoring and risk-averse decision-making.

**Mechanism H2 (internal governance channel):** Pay restrictions enhance internal governance by narrowing the compensation gap between CEOs and key subordinate executives, improving monitoring incentives and the effectiveness of bottom-up oversight, which reduces ALI (Cheng, Lee & Shevlin (2016)).

**Mechanism H3 (social comparison channel):** Pay restrictions narrow pay disparity between executives and rank-and-file employees, reducing perceptions of inequity, increasing employee loyalty and organizational commitment, and thereby reducing ALI (Qian et al. (2024); Cao and Rees (2020)).

The identification argument relies on the 2014 reform being issued as a strict directive by the Political Bureau of the Central Committee, exogenous to individual firm decisions. Chinese firms cannot influence decisions of the Political Bureau, making the reform timing plausibly exogenous to firm-level ALI.

## Method

**Abnormal labor investment measure.** Following Jung, Lee & Weber (2014) and Cao and Rees (2020), the paper first estimates expected (normal) net hiring by running a cross-sectional regression each year within each industry (equation 1, p. 7):

$$\begin{aligned}
\text{Net Hire}_{it} &= \alpha + \beta_1 \text{SalesGrowth}_{i,t-1} + \beta_2 \text{SalesGrowth}_{it} + \beta_3 \Delta\text{ROA}_{it} + \beta_4 \text{ROA}_{it} + \beta_5 \text{ROA}_{i,t-1} + \beta_6 \text{SizeRank}_{i,t-1} \\
&\quad + \beta_7 \Delta\text{QR}_{i,t-1} + \beta_8 \Delta\text{QR}_{it} + \beta_9 \text{QR}_{i,t-1} + \beta_{10} \text{Leverage}_{i,t-1} + \textstyle\sum_{X=1}^{5} \beta_{10+X} \text{LossBin}X_{i,t-1} + \varepsilon_{it}
\end{aligned} \tag{1}$$

where Net Hire is the percentage change in employees, Sales Growth is the percentage change in sales revenue, ROA is net profit to total assets, Size Rank is the percentile rank of log market value, Quick Ratio is (cash + short-term investments + receivables) / current liabilities, and LossBinX (X = 1, ..., 5) are loss-bin dummies for intervals of ROA profitability 0.005 wide from -0.025 to 0. Abnormal Labor Investment (ALI) is the absolute value of the residuals $$|\varepsilon_{it}|$$; a positive residual indicates over-investment, a negative residual indicates under-investment.

**Internal governance measures** (equations 2 and 3, p. 9):

$$\text{Executive Horizon}_{it} = 65 - \text{Average Age of Key Subordinate Executives}_{it} \tag{2}$$

$$\text{Executive Ability}_{it} = \frac{\text{Average Annual Compensation of Key Subordinate Executives}_{it}}{\text{Annual Compensation of CEO}_{it}} \tag{3}$$

Both variables are standardized and summed to construct a composite Internal Governance score.

**Excess executive compensation model** (equation 4, p. 9):

$$\text{Executive Compensation}_{it} = \alpha + \beta_1 \text{FirmSize}_{it} + \beta_2 \text{Leverage}_{it} + \beta_3 \text{ROA}_{it} + \beta_4 \text{ROA}_{i,t-1} + \beta_5 \text{SalesGrowth}_{it} + \beta_6 \text{Dual}_{it} + \beta_7 \text{SOE}_{it} + \beta_8 \text{Independent}_{it} + \text{Industry FE} + \text{Year FE} + \varepsilon_{it} \tag{4}$$

Residuals from equation (4) are Excess Compensation, used in robustness tests.

The technique builds on `difference-in-differences` applied via `panel-regression` with firm and year fixed effects, building from Jung et al. (2014) and Cao and Rees (2020) on the ALI measure, Cheng et al. (2016) on internal governance, and Kuang et al. (2024) and Qian et al. (2024) on pay-disparity measurement.

## Empirical specifications

**Baseline DiD for abnormal labor investment** (equation 9, p. 11): the main identifying specification. The key variable of interest is the interaction $$\text{SOE} \times \text{Restriction}_{it}$$, which equals one for SOE firm-year observations in the post-reform period (2015-2019) and zero otherwise:

$$\text{Abnormal Labor Investment}_{it} = \alpha + \beta_1 (\text{SOE} \times \text{Restriction})_{it} + \beta_2 \text{SOE}_{it} + \beta_3 \text{Restriction}_{it} + \gamma \text{Controls}_{it} + \text{Fixed Effects} + \varepsilon_{it} \tag{9}$$

Controls: Leverage, Book-to-Market, ROA, Quick Ratio, Firm Age, Firm Size, Fixed Assets, Herfindahl Index, Institution Share, SA Index, CEO Gender, CEO Age, CEO Education. Firm and year fixed effects (columns 3-4 of Table 4). Standard errors clustered at the firm level. The DiD coefficient $$\beta_1$$ ranges from -0.039*** to -0.062*** across specifications (Table 4, p. 15), representing a 3.91% to 4.82% reduction in ALI.

**Internal governance channel** (equations 10-12, p. 11): tests H2. Equation (10) uses Internal Governance as the dependent variable to show the reform strengthens governance ($$\beta_1 = 0.141^{**}$$, Table 5 col 1). Equations (11) and (12) re-run the baseline DiD separately for above-median (high) and below-median (low) internal governance effectiveness subsamples. The reduction in ALI is 5.57% for high-IG firms ($$\beta_1 = -0.056^{**}$$) vs. insignificant for low-IG firms, with an F-test across subsamples significant at p = 0.076 (Table 5 cols 2-3).

**Social comparison channel** (equations 13-15, p. 11): tests H3. Equation (13) uses three pay-disparity measures as the dependent variable (CEO-to-rank ratio, Top-3-to-rank ratio, All-executives-to-rank ratio). All three are significantly reduced post-reform (SOE x Restriction = -1.180\*\*\*, -1.004\*\*\*, -0.335\*\*\* respectively; Table 6 col 1 of Panels A, B, C). Equations (14) and (15) split the sample at the sample median of pay disparity: the reduction in ALI is significant only for low-disparity firms (4.68-6.24% reduction), with F-tests across subsamples significant (p = 0.048, 0.042, 0.011 across the three disparity measures).

**Executive compensation and misconduct** (equations 5-8, p. 10):

$$\text{Compensation Variables}_{it} = \alpha + \beta_1 (\text{SOE} \times \text{Restriction})_{it} + \beta_2 \text{SOE}_{it} + \gamma \text{Controls}_{i,t-1} + \text{Fixed Effects} + \varepsilon_{it} \tag{5}$$

$$\text{Executive Misconduct}_{it} = \alpha + \beta_1 (\text{SOE} \times \text{Restriction})_{it} + \beta_2 \text{SOE}_{it} + \beta_3 \text{Restriction}_{it} + \gamma \text{Controls}_{it} + \text{Fixed Effects} + \varepsilon_{it} \tag{7}$$

Column (1) of Table 3 Panel A shows a 10.08% reduction in executive compensation ($$\beta_1 = -0.096^{***}$$, p. 14). The reform also reduces perks by 12.08% ($$\beta_1 = -0.049^{**}$$, col 5) and cuts the probability of executive misconduct by 3.10% ($$\beta_1 = -0.031^{***}$$, Panel B col 1).

**Robustness:** parallel trend tests (pre-reform coefficients jointly insignificant, post-reform coefficients jointly negative and significant; Figure 1, p. 21); entropy-balanced sample replication (Table 7 Panel B); 1,000-iteration placebo tests with randomly assigned treatment events (Figure 2, p. 22); alternative ALI measures (industry FE and industry+year FE predictions in Table 7 Panel C); Oster (2019) bound-estimate approach for omitted-variable bias (Panel E: $$\delta = 3.432$$ for the full sample, above the threshold of 1, indicating results unlikely to be driven by omitted variables).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CSMAR (China Stock Market & Accounting Research) | Financial variables, stock market data, firm characteristics, executive compensation, abnormal labor investment construction, executive misconduct data | [/wiki/commercial/csmar/](/wiki/commercial/csmar/) |
| CNRDS (Chinese Research Data Services) | Firm ownership data, CEO characteristics, employee well-being data (stock grants, work safety, training, communication channels), labor quality | no page yet |
| Procuratorial Yearbook of China | Provincial-level embezzlement and bribery cases (political corruption measure) | no page yet |
| China Statistical Yearbook | Provincial-level civil servant counts (denominator for political corruption measure) | no page yet |

Sample: 14,988 firm-year observations representing 2,889 unique A-share listed firms in China across 71 CSRC 2012 industries, spanning 2009-2019 (event year 2014 excluded; financial firms and ST-designated firms removed). The main sample is annual.

## When to read the full paper

Read Cao et al. (2026) if you are studying: (1) the labor investment consequences of executive compensation regulation in emerging markets; (2) the corporate governance of state-owned enterprises, particularly how internal pay structures affect investment efficiency; (3) causal identification of executive compensation effects using quasi-natural experiments; or (4) the mechanisms through which pay inequality between executives and workers affects organizational outcomes. Key tables are Table 4 (main DiD), Table 5 (internal governance channel), Table 6 (social comparison channel), Table 8 (over/under-hiring decomposition), and Table 9 (labor quality and employee well-being).

## Attribution and rights

Cao, J., Hasan, I., Huang, Z., & Zhao, J. (2026). Pay restrictions and labor investment. *Journal of Corporate Finance*, 99, 102990. https://doi.org/10.1016/j.jcorpfin.2026.102990

Copyright 2026 Elsevier B.V. All rights reserved. This page reproduces no substantial text from the article; results, magnitudes, and table references are extracted for scientific commentary under fair-use principles (extract-only). The article is paywalled; access requires a subscription to the Journal of Corporate Finance or institutional Elsevier access. This summary is LLM-distilled by paper-distiller (claude-sonnet-4-6) and has not been human-verified or independently reproduced.


==============================================================================
# Mandatory CSR Spending and Firm Risk: Chauhan, Ghosh & Jadiyappa (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jcf/2026/chauhan-mandatory-csr-firm-risk-india-2026/
# Distilled: Exploiting India's 2013 mandatory CSR regulation as a quasi-natural experiment, this paper finds that firms subject to mandatory CSR spending exhibit higher systematic risk (equity beta) than non-subject firms, with operating leverage as the primary transmission channel. Journal of Corporate Finance vol 98 (2026) 102965, paywalled (Elsevier). Eight core results with source locators, datasets used, the identification strategy, and the estimating equations.
# Tags: paper-summary, csr, esg, systematic-risk, india, corporate-finance
==============================================================================

**What this is.** A distilled skeleton of Chauhan, Ghosh and Jadiyappa (2026), extracted by an LLM from the published PDF. Read the [original](https://doi.org/10.1016/j.jcorpfin.2026.102965) to replicate or extend.

## TL;DR

India's Companies Act 2013 (Section 135) created a mandatory CSR regulation, requiring firms with net profit above INR 50 million, net worth above INR 5 billion, or revenue above INR 10 billion to spend at least 2% of their three-year average net profit on specified CSR activities. The paper exploits this as a quasi-natural experiment. The treatment group comprises 662 Indian non-financial listed firms that did not engage in CSR before the regulation but began doing so afterward; the control group comprises 268 firms that remained below all statutory thresholds and never engaged in CSR. Using a difference-in-differences (DiD) design over 2010-2019, the paper finds that mandatory CSR spending raises firms' systematic risk (equity beta) by approximately 8-13% of the sample mean beta. Three complementary robustness strategies (propensity-score matched DiD, a multivariate regression discontinuity design, and tail-risk measures MES and delta CoVaR) all confirm the direction. The paper identifies operating leverage as the primary transmission channel: mandatory CSR outlays act as quasi-fixed costs, raising firms' degree of operating leverage (DOL) and making earnings more sensitive to macroeconomic fluctuations.

## Core results

| # | Result | Locator | Magnitude as reported |
|---|---|---|---|
| R1 | Univariate DiD: treatment firms exhibit higher beta post-regulation | Table 2, p. 8 | DiD = 0.112\*\*\* (p<1%); ~12.7% above sample mean beta of 0.88 |
| R2 | Multivariate DiD (OLS + FE): baseline panel regression | Table 3 cols. 1-6, pp. 8-9 | CSR Dummy x Treatment Firms: 0.117\*\*\* (t=4.49) no controls; 0.098\*\*\* (t=3.754) with controls; 0.061\*\* (t=2.325) with firm FE |
| R3 | PSM-matched DiD corroborates baseline | Table 5 cols. 1-2, p. 10 | 0.078\*\* (t=2.571) without firm FE; 0.100\*\*\* (t=4.155) with firm FE |
| R4 | MRDD: Beta is higher at the CSR eligibility cutoff | Table 6, p. 12 | Discontinuity at M=0: 0.241-0.292 (BW ±0.50, all significant); 0.318-0.427 (BW ±0.10, all significant) |
| R5 | Treatment firms exhibit higher marginal expected shortfall | Table 7 cols. 1-2, pp. 14-15 | DiD = 0.344\*\* (t=2.305) col. 1; 0.392\*\*\* (t=3.222) col. 2 |
| R6 | Treatment firms exhibit higher delta CoVaR | Table 7 cols. 3-4, pp. 14-15 | DiD = 1.886\*\*\* (t=4.65) col. 3; 0.841\*\* (t=2.243) col. 4 |
| R7 | Operating leverage mediates the CSR-to-beta channel | Table 9 cols. 1-2, p. 17 | Triple interaction (CSR Dummy x Treatment Firms x ΔDOL): 0.013\*\*\* (t=3.556) col. 1; 0.013\*\*\* (t=2.632) col. 2 |
| R8 | Treated firms' profits are more sensitive to GDP growth post-regulation | Table 10 cols. 1-2, p. 19 | Triple interaction (CSR Dummy x Treatment Firms x GDP growth): 0.007\*\* (t=2.296) col. 1; 0.006\*\* (t=2.213) col. 2 |

**Overall (paper's conclusion).** Across all specifications and outcome measures, firms subject to India's mandatory CSR spending regulation exhibit significantly higher systematic risk than non-subject firms in the post-regulation period. The evidence supports the mechanism that mandatory CSR introduces quasi-fixed costs, raises firms' operating leverage, and amplifies their sensitivity to aggregate economic fluctuations. The findings imply that universal CSR mandates can impose costs on firms at the expense of shareholders, in contrast to the risk-reducing role of voluntary CSR documented in prior literature.

## Theory / model

The paper has no formal model. It tests the theoretical prediction of Albuquerque et al. (2019) in a mandatory CSR setting. In their framework, CSR investment enables firms to cultivate consumer loyalty and reduce demand elasticity, thereby stabilizing revenues and lowering systematic risk. A key implication of their model is that these risk-mitigating benefits depend on CSR remaining a relatively selective differentiating practice: as adoption becomes widespread, the strategic distinctiveness erodes and the insurance mechanism dissipates. Their model therefore predicts a reversal in the CSR-risk relationship once CSR adoption becomes ubiquitous.

The paper argues that India's 2013 mandate creates precisely this condition. Before the regulation, roughly 35-38% of Indian listed firms engaged in CSR voluntarily, using it as a differentiation signal. After the regulation, participation rose to 48-56%. This shift from selective adoption to universal compliance simultaneously (i) dilutes CSR's differentiation value, eroding the demand-stabilization channel, and (ii) imposes new quasi-fixed costs on formerly non-CSR firms, raising their degree of operating leverage.

The operating leverage channel follows Harjoto (2017): when firms cannot pass CSR costs onto customers or offset them with higher contribution margins, CSR expenditures function as fixed costs. Higher fixed costs raise the operating leverage ratio (DOL), amplifying the sensitivity of operating profit to sales fluctuations. Because sales fluctuations co-move with aggregate demand, higher DOL translates into higher systematic risk (equity beta).

The tested hypothesis (Section 3.2, p. 4): firms subject to mandatory CSR regulation would experience higher levels of systematic risk than firms not subject to the mandatory CSR regulation in the post-regulation period.

## Method

The primary estimator is a panel DiD comparing treatment firms (those legally compelled to begin CSR spending after 2014) against control firms (those consistently below all statutory thresholds). Standard errors are clustered at the firm level.

**Parallel trends validation.** An event-study specification (eq. 2 below) tests whether pre-regulation beta differences between treatment and control firms are statistically zero. Figure 1 (p. 8) confirms no significant pre-trend differences for 2010-2013, with positive and significant post-treatment coefficients from 2014 onward.

**PSM-matched DiD.** To address observable pre-period differences, each CSR-exposed firm is matched (nearest-neighbor with replacement) to a non-exposed control on Leverage, Tangibility, Sales Growth, Market-to-Book ratio, and Firm Age during the pre-shock period (Section 5.5, p. 7). The matched sample comprises 5,258 firm-year observations.

**MRDD.** Following Manchiraju and Rajgopal (2017), a multivariate regression discontinuity design (MRDD) constructs a composite binding score M = min(R1, R2, R3), where R1 = (Profit - 50)/50, R2 = (Book value - 5,000)/5,000, and R3 = (Sales - 10,000)/10,000 (all thresholds in INR millions). Firms with M > 0 are treated; M < 0 are controls. The rdrobust command (Calonico et al., 2014) provides bias-corrected and robust inference (Section 5.6, pp. 8-9).

**Tail-risk measures.** MES and delta CoVaR (Section 7.3, pp. 14-15) are used as alternative outcome measures. MES captures a firm's average return on days when the market falls in the bottom 5% of its distribution; delta CoVaR captures how much the system's downside VaR worsens when a firm moves from its median to its distressed state. Both are estimated over the same DiD framework as eq. 1.

**DOL mediation.** A two-stage approach identifies the operating leverage channel (Section 8.2, pp. 17-18). Stage 1 estimates each firm's ΔDOL from a regression of log EBIT on log Sales interacted with the CSR Dummy. Stage 2 tests whether post-mandate changes in DOL account for the increase in beta, via a triple interaction (CSR Dummy x Treatment Firms x ΔDOL) in the beta regression.

## Empirical specifications

**Main DiD regression** (eq. 1, p. 6), producing R1-R2:

$$\text{Beta}_{it} = \alpha + \beta_1\text{CSR Dummy}_t + \beta_2\text{Treatment Firms}_i + \beta_3(\text{CSR Dummy}_t \times \text{Treatment Firms}_i) + \beta_4\text{Size}_{it} + \beta_5\text{ROA}_{it} + \beta_6\text{Tangibility}_{it} + \beta_7\text{MB}_{it} + \beta_8\text{Leverage}_{it} + \beta_9\text{Firm Age}_{it} + \varepsilon_{it} \tag{1}$$

where $$\text{Beta}_{it}$$ is equity beta estimated from daily returns against the NSE Nifty 50 index (minimum 100 trading days per year); $$\text{CSR Dummy}_t = 1$$ for 2015-2019 (post-regulation) and 0 for 2010-2014; $$\text{Treatment Firms}_i = 1$$ for the 662 firms legally compelled to begin CSR spending. The coefficient of interest is $$\beta_3$$. All specifications include year and industry (firm x year) fixed effects; errors clustered at the firm level (Table 3).

**Parallel trends event study** (eq. 2, p. 7):

$$\text{Beta}_{it} = \sum_{p \neq 0} \delta_p (D_p \times \text{Treatment Firms}_i) + X_{it}\gamma + \mu_i + \lambda_t + \varepsilon_{it} \tag{2}$$

where $$D_p$$ is a year dummy (omitted base year: 2014), $$\mu_i$$ are firm fixed effects, and $$\lambda_t$$ are year fixed effects. Pre-regulation coefficients $$\delta_{-4}, \delta_{-3}, \delta_{-2}, \delta_{-1}$$ should be indistinguishable from zero; post-regulation coefficients $$\delta_1$$ to $$\delta_5$$ are expected to be positive (Figure 1, p. 8).

**MES definition** (eq. 3, p. 14), producing R5:

$$\text{MES}_{j,t} = \frac{1}{N_t} \sum_{d \in D_t} R_{j,d} \tag{3}$$

where $$D_t$$ is the set of trading days on which the market falls in the bottom 5% of its return distribution, $$N_t$$ is the number of such days in year $$t$$, and $$R_{j,d}$$ is firm $$j$$'s return on day $$d$$. MES is multiplied by $$-1$$ so higher values indicate greater systemic vulnerability.

**Delta CoVaR** (eqs. 4-6, p. 14), producing R6. The tail-event VaR for firm $$j$$ is defined by:

$$\Pr\!\left(r_j \leq \text{VaR}_{j,5\%}\right) = 5\% \tag{4}$$

The system's CoVaR conditional on firm $$j$$ being in distress:

$$\Pr\!\left(r_{\{j\}} \leq \text{CoVaR}_{\text{system}|j},\; r_j = \text{VaR}_{j,5\%}\right) = 5\% \tag{5}$$

$$\Delta\text{CoVaR}(r_{\{j\}}, 1\%) = \text{CoVaR}(r_{\{j\}} \mid j, 1\%) - \text{CoVaR}(r_{\{j\}} \mid j, 50\%) \tag{6}$$

Delta CoVaR is estimated via quantile regression conditioning on firm $$j$$'s median vs distressed state, with lagged market return, volatility (VIX), Treasury bill yield, and term premium as state variables.

**DOL estimation** (eq. 7, p. 17), providing the baseline DOL measure:

$$\ln(\text{EBIT}_{it}) = \alpha + \beta \ln(\text{Sales}_{it}) + \varepsilon_{it} \tag{7}$$

where $$\beta$$ is the firm's degree of operating leverage.

**Extended DOL with CSR mandate** (eq. 8, p. 17), producing R7 first stage (Table 8):

$$\begin{aligned} \ln(\text{EBIT}_{it}) = \alpha &+ \beta_1\text{CSR Dummy}_t + \beta_2\text{Treatment Firms}_i + \beta_3(\text{CSR Dummy}_t \times \text{Treatment Firms}_i) \\ &+ \beta_4(\text{Treatment Firms}_i \times \ln\text{Sales}_{it}) + \beta_5(\text{CSR Dummy}_t \times \text{Treatment Firms}_i \times \ln\text{Sales}_{it}) + X_{it}\gamma + \varepsilon_{it} \end{aligned} \tag{8}$$

$$\beta_5$$ captures whether mandated firms become more sensitive to sales in the post-regulation period. The estimate 0.023\*\* (t=2.270, Table 8 col. 2) implies a 1% change in sales corresponds to a ~2.3% larger change in operating profit for treatment firms after the mandate.

**First-stage DOL mediation** (eq. 9, p. 18), estimated per firm to recover ΔDOL:

$$\ln(\text{EBIT}_{it}) = \alpha + \beta_1\text{CSR Dummy}_t + \beta_2(\text{CSR Dummy}_t \times \ln\text{Sales}_{it}) + \beta_3\ln\text{Sales}_{it} + X_{it}\gamma + \varepsilon_t \tag{9}$$

Here $$\beta_2$$ measures the firm's change in DOL (ΔDOL) in the post-regulation period. The coefficient on $$\text{CSR Dummy}_t \times \text{Treatment Firms}_i \times \ln\text{Sales}_{it}$$ in Table 8 confirms ΔDOL is positive and significant for mandated firms.

**Second-stage DOL mediation** (eq. 10, p. 18), producing R7 (Table 9):

$$\text{Beta}_{it} = \alpha + \beta_1\text{CSR Dummy}_t + \beta_2(\text{CSR Dummy}_t \times \text{Treatment Firms}_i) + \beta_3(\text{CSR Dummy}_t \times \text{Treatment Firms}_i \times \Delta\text{DOL}_i) + \beta_4\text{Treatment Firms}_i + X_{it}\gamma + \varepsilon_{it} \tag{10}$$

$$\beta_3$$ tests whether firms with larger post-mandate increases in DOL exhibit correspondingly higher systematic risk. The estimate 0.013\*\*\* (t=3.556, Table 9 col. 1) is economically significant: with SD(ΔDOL) = 2.4, a one-standard-deviation rise in ΔDOL is associated with a beta increase of approximately 0.031 for mandated firms.

**Profit cyclicality** (eq. 11, p. 19), producing R8 (Table 10):

$$\Delta\text{ROA}_{it} = \alpha + \beta_1\text{CSR Dummy}_t + \beta_2(\text{CSR Dummy}_t \times \text{Treatment Firms}_i) + \beta_3(\text{CSR Dummy}_t \times \text{Treatment Firms}_i \times \text{GDP growth}_t) + \text{Controls} + \varepsilon_{it} \tag{11}$$

$$\beta_3$$ captures whether mandated firms' operating profits become more sensitive to GDP growth after the regulation. The estimate 0.007\*\* (Table 10 col. 1) confirms that CSR-exposed firms are more vulnerable to aggregate economic shocks, consistent with higher systematic risk via the operating leverage channel.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CMIE Prowess | Primary source: firm financials (profit/loss, balance sheet, financial ratios, share prices, stock returns), annual 2010-2019 for all Indian non-financial listed firms | no page yet |

Sample: 8,671 firm-year observations from 930 unique firms (2010-2019), of which 662 treatment firms (6,332 firm-years) and 268 control firms (2,339 firm-years). Beta estimated from daily returns against the NSE Nifty 50 index; firm-years with fewer than 100 trading days excluded. All variables winsorized at the 2% tails.

Rajgopal and Tantri (2023) noted that firms that voluntarily spent more than 2% of average profits on CSR prior to the mandate subsequently reduced their CSR expenditures; the paper excludes these voluntary pre-spenders from the treatment group to keep it clean.

## When to read the full paper

Read the original when studying: (1) the causal effect of mandatory ESG or CSR regulations on corporate risk, using India's Section 135 as a quasi-natural experiment with a clean statutory eligibility rule; (2) the two-stage DOL mediation design (eqs. 7-10) for tracing a compliance-cost channel through to systematic risk; or (3) MRDD methodology applied to multi-threshold eligibility rules following the composite score approach (Table 6). The appendix (Table A1, p. 21) defines all variables; Tables A2-A10 provide the full robustness suite (pre-mandate CSR orientation, voluntary spenders, direct CSR-spending regressions, industry concentration, alternative systematic risk measures, confounding governance reforms, and advertising substitution).

## Attribution and rights

Chauhan, Y., Ghosh, C., and Jadiyappa, N. (2026). Mandatory CSR spending and firm risk: New evidence from regulatory intervention in India. *Journal of Corporate Finance*, 98, 102965. https://doi.org/10.1016/j.jcorpfin.2026.102965

Copyright 2026 Elsevier B.V. All rights reserved, including those for text and data mining, AI training, and similar technologies. This page is an LLM-distilled extract (not human-verified, not reproduced). Access the original at [doi.org/10.1016/j.jcorpfin.2026.102965](https://doi.org/10.1016/j.jcorpfin.2026.102965).


==============================================================================
# Lenders Pricing Cybersecurity Risk: Choi, Degryse & Smedts (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jcf/2026/choi-lenders-price-cybersecurity-risk-2026/
# Distilled: Using syndicated loan data for U.S. non-financial firms (2012-2018), lenders charge 4 to 13 basis points higher loan spreads for firms with rising ex-ante cybersecurity risk, with commercial banks pricing more conservatively than non-bank lenders and pricing concentrated among lenders who are themselves aware of cybersecurity risk. Cybersecurity insurance does not mitigate the higher spreads. Journal of Corporate Finance vol. 98, 2026, paywalled; eight core results with source locators, the regression specifications, and datasets used.
# Tags: paper-summary, cybersecurity, credit-risk, syndicated-loans, banking, non-bank-lenders, panel-regression, peer-reviewed, unreplicated, data:wrds, data:edgar
==============================================================================

**What this is.** Core results, tested hypotheses, and regression specifications from this paper, distilled from the published PDF. To replicate or extend the analysis, read the full source at the [original](https://doi.org/10.1016/j.jcorpfin.2026.102958).

## TL;DR

Using 5,957 syndicated loan facilities for U.S. non-financial firms from 2012 to 2018, the paper asks whether lenders price firms' ex-ante cybersecurity risk in loan spreads. The main measure of cybersecurity risk, from Florackis et al. (2023), captures the textual similarity of a firm's 10-K disclosures to those of firms that experienced data breaches. The paper's key findings are: (1) a one standard deviation increase in cybersecurity risk raises the All-in-Spread-Drawn (AISD) by about 2%, equivalent to roughly 4 basis points; (2) first-time exposure to non-zero cybersecurity risk adds about 13 basis points; (3) commercial banks price the risk more strictly than non-bank lenders; (4) pricing depends on the lender's own awareness: only lenders who discuss cybersecurity risk in conference calls price it in spreads; (5) cybersecurity insurance does not reduce the premium; and (6) the credit risk channel, identified via distance-to-default, mediates the pricing. Prior studies such as Huang and Wang (2021) and Sheneman (2017) documented ex-post pricing after data breaches; this paper is the first to examine ex-ante pricing from the lenders' perspective.

## Core results

Magnitudes and significance are as reported; \*/\*\*/\*\*\* = 10%/5%/1%. All main regressions use the log AISD as the outcome and include firm fixed effects and industry-year fixed effects. Standard errors are clustered at the firm level. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Lenders price cybersecurity risk at the **intensive margin**: within-firm increases in the cybersecurity risk score raise the AISD | Table 2, col. (3), p. 6 | Coefficient on Cybersecurity risk = 0.021\*\*; ~2.1% increase in AISD per 1 SD, ~4.15 bps (mean AISD = 195 bps) |
| R2 | At the **extensive margin**, first exposure to non-zero cybersecurity risk carries a large premium | Table 2, col. (4), p. 6 | AboveZero coefficient = 0.063\*\*\*; firms with any cybersecurity exposure face ~12.71 bps higher AISD on average |
| R3 | **Commercial banks price cybersecurity risk more strictly** than non-bank lenders; non-bank participation nearly offsets the premium | Table 3, cols. (1)-(4), p. 7 | CBank-only: Cybersecurity risk = 0.029\*\* (intensive), AboveZero = 0.087\*\*\* (extensive); NonBank interaction: -0.033\*\*\*, AboveZero x NonBank = -0.092\*\*\* (offsets base) |
| R4 | Commercial banks **attach more financial covenants** as cybersecurity risk rises | Table 4, col. (1), p. 8 | Cybersecurity risk coefficient = 0.065\*; ~0.065 additional covenants per 1 SD, ~6% of mean covenant count; effect not significant in full sample with non-bank lenders |
| R5 | **Cybersecurity insurance does not mitigate** the higher loan spreads | Table 5, cols. (3)-(4), p. 8 | Insurance coefficient = 0.003-0.004 (ns); Cybersecurity risk x Insurance = 0.001-0.007 (ns); both small and insignificant |
| R6 | Pricing is **driven by lender awareness**: the premium appears only when lead arrangers discuss cybersecurity risk in their own conference calls | Table 6, cols. (1)-(2), p. 9 | No-mention subsample: coefficient = -0.014 (ns); Mention subsample: 0.028\*\*\*; Insured lenders: 0.042\*\*\* |
| R7 | Aware lenders **reduce their loan-share exposure** to riskier borrowers after recognizing their own cybersecurity risk | Table 8, col. (1), p. 11 | Cybersecurity risk x Insured: -1.310\*\* on lead arranger share (%); paper reports a 2.51% (1.31%) decrease per 1 SD when lead arranger discussed cybersecurity risk (insurance policy) |
| R8 | Cybersecurity risk is priced **via the credit risk channel**: it lowers borrower distance-to-default | Table 9, p. 12 | Cybersecurity risk coefficient on distance-to-default: -0.209\* (loan-level sample), -0.181\* (firm-year sample); ~3% average decrease per 1 SD |

**Overall (paper's conclusion).** Lenders do price ex-ante cybersecurity risk, but only through within-firm changes in risk scores (not cross-sectional differences), and only when they are themselves engaged with cybersecurity risk management. Commercial banks are more conservative than non-bank lenders. The credit risk mechanism, proxied by distance-to-default, mediates the pricing. Cybersecurity insurance neither reduces breach probability nor provides comprehensive loss coverage, so it does not lower the credit premium.

## Theory / model

The paper has no formal structural model. It tests five empirical hypotheses using within-firm variation in cybersecurity risk scores over time:

- **H1 (Pricing)**: Lenders charge higher loan spreads for firms with higher ex-ante cybersecurity risk, based on within-firm changes in risk exposure. The null of no ex-ante pricing is motivated by the observation that cross-sectional comparisons (Table A.4, p. 20) find no significant effect, suggesting the risk is idiosyncratic and firm-specific rather than industry-wide.
- **H2 (Lender heterogeneity)**: Commercial banks, subject to tighter regulation and lower risk tolerance (Aldasoro et al. (2022)), price cybersecurity risk more strictly than non-bank lenders (hedge funds, private equity funds, mutual funds, insurance companies, and finance companies).
- **H3 (Lender awareness)**: Pricing of borrower cybersecurity risk depends on the lender's own engagement with cybersecurity risk, measured via conference call discussions of cybersecurity and insurance policies (Jamilov et al. (2021)). Lenders who discuss cybersecurity risk are hypothesized to price it; those who do not, to ignore it.
- **H4 (Insurance)**: Cybersecurity insurance may signal adverse selection (riskier firms are more likely to buy it) and may induce moral hazard (weakening incentives to improve security), so it need not lower loan spreads.
- **H5 (Credit risk channel)**: Cybersecurity risk is correlated with default probability. To test this channel, the Merton (1974) distance-to-default is used as the dependent variable in place of loan spreads (see Empirical specifications for the formula).

**Identification.** The paper identifies effects from within-firm variation in cybersecurity risk scores over time, controlling for firm fixed effects (absorbing time-invariant firm characteristics) and industry-year fixed effects (absorbing sector-wide and time shocks). The cybersecurity risk score varies substantially within firms over the 2012-2018 window as cyber disclosures became more detailed following the SEC's 2011 guidance. Cross-sectional regressions without firm fixed effects (Table A.4) find no significant effect, consistent with the risk being idiosyncratic. Standard errors are clustered at the firm level.

## Method

**Cybersecurity risk measure.** The primary measure follows Florackis et al. (2023), who apply textual analysis to the "Item 1 A. Risk Factors" section of 10-K filings for U.S. non-financial firms from 2007 to 2018. The measure captures the similarity between a firm's current cybersecurity disclosure and the pre-breach disclosures of firms that subsequently experienced significant data breaches. A higher score indicates higher ex-ante cybersecurity risk, reflecting both the quantitative and qualitative intensity of cybersecurity disclosures. The paper standardizes the measure within the sample (subtracting the mean, dividing by the standard deviation). An indicator variable AboveZero captures firms with a positive risk score.

**Lender awareness.** Conference call data from Jamilov et al. (2021) identify whether a firm's quarterly earnings calls mention cybersecurity-related terms within 50 words of insurance topics. Lenders that discuss cybersecurity risk in their own calls (Discussed) or also mention insurance (Insured) are classified as aware. The Intensity variable counts cybersecurity-related keywords per call.

**Estimation.** All main regressions are OLS on a panel of loan facilities, building on the `panel-regression` technique with high-dimensional fixed effects. Following Lattanzio and Ma (2023), the paper employs either year + industry + firm fixed effects (Tables 2, cols. 1-2) or industry-year + firm fixed effects (Tables 2, cols. 3-4, and all subsequent tables). Industry-year FE absorb sector-time shocks and allow focus on within-firm variation. Loan-level controls include log loan amount, maturity, secured and covenant indicators, number of lead arrangers, non-bank participation indicator, and a relationship-lending indicator. Borrower controls (lagged one year) include log total assets, leverage, ROA, interest coverage, fixed-asset ratio, R&D intensity, log patents, book-to-market, a technology director indicator, and the share of AI-knowledge employees.

## Empirical specifications

**Main regression (Eq. 1, p. 5).**

$$
\log AISD_{i,j,t} = \beta_1 \text{CybersecurityRisk}_{i,t-1} + \gamma X_{i,j,t-1} + \text{FE} + \varepsilon_{i,j,t} \tag{1}
$$

where $$\log AISD_{i,j,t}$$ is the log of the All-in-Spread-Drawn (bps over LIBOR plus facility fee) for loan facility $$j$$ granted to firm $$i$$ in year $$t$$; $$\text{CybersecurityRisk}_{i,t-1}$$ is the standardized Florackis et al. (2023) score; $$X_{i,j,t-1}$$ is the vector of loan and borrower controls; FE are fixed effects (industry-year and firm in the preferred specification). A positive $$\beta_1$$ indicates lenders price cybersecurity risk in spreads. Results: Table 2, p. 5; full controls in Table A.3, p. 19.

**Lender composition interaction (Eq. 2, p. 6).** To test whether commercial banks differ from non-bank lenders:

$$
\log AISD_{i,j,t} = \beta_1 \text{CybersecurityRisk}_{i,t-1} + \beta_2 \text{LenderComposition}_j + \beta_3 \text{CybersecurityRisk}_{i,t-1} \times \text{LenderComposition}_j + \gamma X_{i,j,t-1} + \text{FE} + \varepsilon_{i,j,t} \tag{2}
$$

where LenderComposition is a categorical variable: the base group is commercial-bank-only loans; InvestmentBank includes investment bank participation; NonBank includes non-bank lender participation. Results: Table 3, p. 6. Commercial bank-only regressions (Table 3, cols. 1-2) exclude all non-commercial-bank loans and estimate Eq. (1) directly. The interaction $$\beta_3$$ tests whether non-bank or investment-bank participation dilutes the cybersecurity premium.

**Financial covenants.** Equation (1) is re-estimated replacing $$\log AISD_{i,j,t}$$ with the number of financial covenants attached to the loan facility, testing whether lenders adjust monitoring intensity. Results: Table 4, p. 7.

**Insurance effect.** A logit model on a firm-year panel tests which firms hold cybersecurity insurance. Equation (1) is then augmented with Insurance and a Cybersecurity risk x Insurance interaction term, testing whether insurance mitigates the spread premium. Results: Table 5, p. 8.

**Lender awareness subsamples.** Equation (1) is estimated separately on subsamples split by whether the lead arranger mentioned cybersecurity risk (Mention vs. NoMention) and whether they discussed insurance (Insurance vs. NoInsurance). Single-lead-loan regressions (Table 7, p. 10) add lender fixed effects and firm-lender fixed effects, with the interaction Cybersecurity risk x Insured testing whether lenders who have adopted cybersecurity insurance policies price borrower risk more. Results: Table 6, p. 9; Table 7, p. 10.

**Lender share (exposure reduction).** The outcome is each lender's percentage share in the syndicated loan. The interaction of Cybersecurity risk with Discussed and Insured tests whether aware lenders reduce exposure to riskier borrowers. Results: Table 8, p. 11.

**Credit risk mechanism: Merton distance-to-default.** To test the credit channel, the dependent variable is replaced with the Merton (1974) distance-to-default (DD), calculated following the Appendix C procedure (p. 18). Firm value $$V$$ and asset volatility $$\sigma_V$$ are solved simultaneously from the Black-Scholes-Merton equity pricing equation (Eq. 4, p. 18) and the relationship between equity volatility and asset volatility (Eq. 7, p. 18). The distance-to-default (Eq. 8, p. 18) is:

$$
DD = \frac{\ln\!\left(\dfrac{V}{F}\right) + \left(\mu - 0.5\sigma_V^2\right) T}{\sigma_V \sqrt{T}} \tag{8}
$$

where $$F$$ is the face value of debt, $$\mu$$ is the estimated annual return on firm assets (risk-free rate plus 0.06 as equity premium proxy), and $$T = 1$$ year. A negative coefficient on Cybersecurity risk in Eq. (1) re-estimated with DD as the outcome would confirm the credit channel (higher risk reduces distance-to-default). Results: Table 9, p. 12. Both Kamiya et al. (2021) and prior credit rating agency analyses motivate this mechanism test.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Florackis et al. (2023) cybersecurity risk scores | Main ex-ante cybersecurity risk measure derived from 10-K filings (2007-2018) | No page yet (based on SEC EDGAR 10-K filings: [EDGAR](/wiki/datasets/edgar/)) |
| Jamilov et al. (2021) conference call data | Borrower insurance coverage and lender awareness of cybersecurity risk from earnings call transcripts | No page yet |
| Thomson Reuters LPC DealScan | Syndicated loan facilities (spreads, covenants, lender identities, amounts, maturity, collateral) 1988-2019; sample 2012-2018 | [WRDS](/wiki/commercial/wrds/) (licensed) |
| Compustat | Borrower financial characteristics (lagged one year): total assets, leverage, ROA, R&D, book-to-market | [WRDS](/wiki/commercial/wrds/) (licensed) |
| Kogan et al. (2017) patent data | Patent count control for firm technological intensity | No page yet |
| Babina et al. (2024) AI employee data | Share of employees with AI-related knowledge | No page yet |
| WRDS Audit Analytics Cybersecurity | Data breach events; used to exclude post-breach observations | [WRDS](/wiki/commercial/wrds/) (licensed) |

Sample: 5,957 loan facilities from 1,714 unique U.S. non-financial borrowers, originated 2012-2018 (one year after the SEC's October 2011 cybersecurity disclosure guidance). Facility-level (loan as the unit of observation). Borrower variables are annual, lagged one year.

## When to read the full paper

Read the [original](https://doi.org/10.1016/j.jcorpfin.2026.102958) if you are: studying how non-standard risks (outside financial statements) enter credit pricing; building a model of lender heterogeneity in risk assessment; designing policies to raise bank awareness of operational or cyber risks (the stress-test policy implication is spelled out in the conclusion); or extending the analysis to international markets, different loan types, or other unconventional risk measures. Tables A.3 and A.4 (pp. 19-20) provide the full coefficient vectors and the cross-sectional robustness check.

## Attribution and rights

Source: peer-reviewed, *Journal of Corporate Finance* vol. 98, 2026, article 102958. DOI: [10.1016/j.jcorpfin.2026.102958](https://doi.org/10.1016/j.jcorpfin.2026.102958). This distillation was extracted by an LLM (claude-sonnet-4-6) on 2026-06-26 and is **not human-verified or independently reproduced**. The paper is paywalled (Elsevier; no CC license). Only text excerpts and numeric results appear here under extract-only use; the verbatim PDF is not hosted or redistributed.

Choi, Bok Min, Hans Degryse, and Kristien Smedts. "Do lenders price firms' cybersecurity risk?" *Journal of Corporate Finance* 98 (2026): 102958. DOI: 10.1016/j.jcorpfin.2026.102958.


==============================================================================
# M&As and Innovation: Farida, Fidrmuc & Zhang (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jcf/2026/farida-mas-innovation-private-targets-2026/
# Distilled: Acquiring private rather than public targets raises acquirer patent quantity, quality, and economic value by 8 to 15 percent more, with larger innovation synergies and inventor-network growth, in a matched US sample 1990-2020. Journal of Corporate Finance 96 (2026) 102905, CC BY 4.0. Seven core results with source locators, datasets used, the hypothesis framework, and the PPML difference-in-differences specification.
# Tags: paper-summary, mergers-acquisitions, innovation, patents, private-firms
==============================================================================

**What this is.** The paper's core results, the hypothesis it tests, and the PPML difference-in-differences specification it uses: enough to understand what it found and how, without reading all 20 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1016/j.jcorpfin.2025.102905).

## TL;DR

Using a 1:1 propensity-score-matched sample of US private versus public target acquisitions by publicly listed US firms (1990-2020) and a PPML difference-in-differences design, the paper documents that patent quantity, forward citation quality, and patent economic value all increase significantly more at acquirers after private-target deals than after public-target deals. The magnitudes are 8 to 15 percent across the three headline patent outcomes and 20 to 25 percent for combined-entity synergy measures and new inventor collaborations. The gap is strongest when acquirers have prior private-target M&A experience or employ complementary financial advisors, is concentrated in breakthrough-technology sectors, and appears regardless of whether the target held granted patents at acquisition. Announcement abnormal returns (CAR) are 1.2 pp higher for private-target acquirers, and this return premium is partially explained by the expected post-acquisition innovation improvements.

## Core results

Magnitudes and significance as reported; `\*\*` = 5%, `\*\*\*` = 1%. All Panel A results are from matched-pair PPML with matched-pair and calendar-year fixed effects (23,219 firm-event-year observations). IRR = exp(β) - 1 conversion as stated in the paper (p.6).

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Private-target acquisitions yield **higher post-acquisition patent count** at the acquirer | Table 4, Panel A, col. 1, pp. 5-6 | β = 0.142\*\*\* (s.e. 0.049); IRR 15.3% higher than public-target acquisitions |
| R2 | Private-target acquisitions yield **higher patent quality** (forward citations) | Table 4, Panel A, col. 2, pp. 5-6 | β = 0.080\*\* (s.e. 0.038); IRR 8.3% more forward citations |
| R3 | Private-target acquisitions yield **higher patent economic value** | Table 4, Panel A, col. 3, pp. 5-6 | β = 0.121\*\*\* (s.e. 0.034); IRR 12.9% higher patent value |
| R4 | **Combined entity patents rise more** for private-target deals, reflecting integration synergies | Table 4, Panel A, col. 4, pp. 5-7 | β = 0.225\*\*\* (s.e. 0.054); IRR 25.2% higher combined patent count; combined forward cites: β = 0.175\*\*\*, IRR 19.1% higher |
| R5 | **Inventor network grows more** after private-target acquisitions | Table 4, Panel A, col. 6, pp. 5-7 | β = 0.137\*\*\* (s.e. 0.043); IRR 14.7% more total inventors |
| R6 | **New cross-firm inventor collaborations** are significantly larger for private targets | Table 4, Panel A, col. 7, pp. 5-7 | β = 0.183\*\*\* (s.e. 0.051); IRR 20.1% more new inventors collaborating with acquirer incumbents |
| R7 | **Acquirer 5-day CAR is 1.2 pp higher** for private-target deals; deals with larger expected innovation gains earn even more | Table 10, col. 1 (baseline) and Cols. 2-8 (innovation quartile interactions), p. 16 | Private dummy: 0.012\*\*\* (s.e. 0.004); Private × ΔInn Q2 and Q3 interactions: 0.021-0.040\*\*\*; effect is not present for public-target acquirers |

**Overall (paper's conclusion).** The results support the hypothesis that acquisitions of private targets by public acquirers are associated with larger post-acquisition innovation gains than acquisitions of public targets. The mechanism runs through complementary capabilities: private targets embed tacit, exploratory knowledge that combines with the acquirer's commercialization assets in ways that are harder to replicate at arm's length. The private-public innovation gap predicts acquirer announcement returns, linking innovation complementarities to value creation and underscoring M&A as a boundary-of-the-firm mechanism through which public companies access and scale early-stage innovation from private firms.

## Theory / model

The paper has no formal model. It develops and tests the following hypothesis (p. 3):

> Acquisitions of private targets by public acquirers are associated with a larger post-acquisition change in innovation quantity, quality, value, and innovation synergies than acquisitions of public targets by public acquirers.

The economic rationale follows Teece (1986): private targets hold tacit, exploratory, earlier-stage knowledge that is less codified and more complementary to a public acquirer's downstream commercialization assets (manufacturing, distribution, regulatory expertise, and access to capital markets). In contrast, public targets operate under incentive structures and capital-market pressures closer to the acquirer's own, limiting marginal complementarities (Holmstrom 1989; Ferreira et al. 2014). Prior work by Sevilir and Tian (2012) established that M&A is positively associated with innovation outcomes, while Phillips and Zhdanov (2013) examine how acquisition prospects affect innovation incentives across large and small firms; this paper extends the analysis to the private-versus-public target dimension with post-acquisition patent outcomes.

The identification strategy exploits propensity-score-matched comparison of acquirers of private versus public targets, conditioning on acquirer year, FF30 industry, size, and book-to-market to control for observed firm-level differences. The DiD design holds the acquisition event fixed and lets target status (private vs public) be the only remaining systematic difference. Three mechanism channels are tested empirically:

1. **Acquirer expertise** (AE): acquirers with prior experience acquiring private targets, and those using two financial advisors (full-service bank plus boutique), show larger private-public gaps (Table 7), consistent with information-friction reduction around private firms' tacit intangibles.
2. **Breakthrough-technology sectors**: the innovation effects are concentrated in industries where breakthrough patents account for a larger share of activity (Table 8), where the advantages of private-firm exploration (tolerance for failure, long horizons) matter most.
3. **Target patent status**: both targets with granted patents (WP) and those without (WoP) contribute to the post-acquisition innovation gains (Table 9), consistent with value residing in unpatented know-how (Teece 1986).

## Method

The core estimator is Poisson pseudo-maximum likelihood (PPML), applied as a conditional-mean DiD. PPML imposes no restriction on the domain of the outcome beyond non-negativity, and Cohn et al. (2022) show it is valid for continuous non-negative outcomes as well as count data. The estimating equation (equation 1, p. 5) is:

$$
\mathbb{E}[\text{Inn}_{i,t} \mid X_{i,t}] = \exp\!\left(\alpha_1 \text{Private}_i + \alpha_2 \text{Post}_t + \beta(\text{Private}_i \times \text{Post}_t) + \lambda' \mathbf{X}_{i,t-1} + \delta_j + \theta_y\right) \tag{1}
$$

where $$i$$ indexes deals (private-target and matched public-target), $$j$$ indexes matched pairs, $$t \in \{-5,\ldots,5\}$$ is event time, and $$\text{Inn}_{i,t}$$ is one of seven innovation outcomes measured at the acquiring firm in year $$t$$. $$\text{Private}_i = 1$$ for private-target deals and 0 for matched public-target deals. $$\text{Post}_t = 1$$ for $$t \in \{0,\ldots,5\}$$ and 0 otherwise. The key DiD parameter is $$\beta$$; since the conditional-mean is exponential, the incidence-rate ratio is $$\exp(\beta)$$ and the percentage change is $$\exp(\beta) - 1$$. $$\delta_j$$ are matched-pair fixed effects and $$\theta_y$$ are calendar-year fixed effects. The control vector $$\mathbf{X}_{i,t-1}$$ (lagged one year) contains total sales, R&D, leverage, net income, and industry concentration. Standard errors are clustered by matched pair.

The mechanism test for acquirer expertise (equation 2, p. 12) triples the interaction:

$$
\mathbb{E}[\text{Inn}_{i,t} \mid X_{i,t}] = \exp\!\Bigl(\alpha_0 \text{AE}_i + \alpha_1^{\text{AE}} \text{Private}_i \times \text{AE}_i + \alpha_1^{\text{NAE}} \text{Private}_i \times \text{NAE}_i + \alpha_2^{\text{AE}} \text{Post}_t \times \text{AE}_i + \alpha_2^{\text{NAE}} \text{Post}_t \times \text{NAE}_i
$$

$$
+ \gamma^{\text{AE}} \text{Private}_i \times \text{Post}_t \times \text{AE}_i + \gamma^{\text{NAE}} \text{Private}_i \times \text{Post}_t \times \text{NAE}_i + \lambda' \mathbf{X}_{i,t-1} + \delta_j + \theta_y\Bigr) \tag{2}
$$

where $$\text{AE}_i = 1$$ for acquirers with prior private-target M&A experience and $$\text{NAE}_i = 1 - \text{AE}_i$$. The triple-interaction coefficients $$\gamma^{\text{AE}}$$ and $$\gamma^{\text{NAE}}$$ measure the DiD effect for experienced versus inexperienced acquirers. Analogous triples test breakthrough-technology sectors (Table 8) and target patent status (Table 9).

The announcement return test (Table 10, p. 16) extends the prior finding by Faccio et al. (2006) that private-target acquirers earn higher CARs, testing whether the expected post-acquisition innovation improvement drives this premium. The regression uses OLS:

$$
\text{CAR}(-2,2)_i = \beta_0 + \beta_1 \text{Private}_i + \sum_{k=2}^{4} \beta_k \Delta\text{Inn}_{Q_k,i} + \sum_{k=2}^{4} \gamma_k (\text{Private}_i \times \Delta\text{Inn}_{Q_k,i}) + \lambda' \text{Controls}_i + \delta_{\text{FF30}} + \theta_y + \varepsilon_i
$$

where $$\text{CAR}(-2,2)$$ is the acquirer 5-day abnormal return adjusted by the value-weighted market index, $$\Delta\text{Inn}_{Q_k}$$ are dummy variables for quartiles 2-4 of the change in each innovation outcome from pre- to post-acquisition (the lowest quartile $$Q_1$$ is the reference), and the interaction terms $$\gamma_k$$ capture the innovation-innovation return link separately for private versus public targets.

## Empirical specifications

**Sample construction (pp. 3-4).** The baseline sample covers publicly listed US acquirers of US stand-alone private or publicly listed targets from 1990 to 2020, drawn from SDC Platinum. Acquisitions must be completed equity deals not involving buyouts, spinoffs, or recapitalizations. Financial data require Compustat coverage; this restricts acquisitions to 1990 onwards. Patent data from KPSS (Kogan et al. 2017) end in 2015, so acquisitions are capped at 2015 to retain a 5-year post-deal patent window. "Both-type deals" (the same acquirer completes both a private and a public acquisition in the same calendar year) are excluded, yielding 13,448 deals with 2,161 public-target and 11,287 private-target observations.

**Matching (Table 1, Panel A, pp. 3-4).** Propensity scores predict the probability of acquiring a public target using total assets, book-to-market, FF30 industry fixed effects, and calendar year. Each public-target acquirer is matched 1:1 (without replacement) to the closest private-target acquirer in the same year and industry. After matching, 1,153 public-target and 1,153 private-target matched pairs are retained, with 23,219 firm-event-year observations spanning 5 years before and after each acquisition announcement. The matched sample satisfies balance on the matching covariates (Table 1, Panel B).

**Baseline results (Table 4, Panel A, R1-R6).** Equation (1) is estimated separately for seven outcome variables: patent count, forward cites, patent value (KPSS acquirer-level), combined patent count, combined forward cites (acquirer + target composite), number of all inventors, and number of new collaborating inventors (PatentsView). Matched-pair fixed effects ($$\delta_j$$) absorb any time-invariant deal-level heterogeneity; deal fixed effects are used in Panel B as a robustness check. Fig. 1 plots year-by-year incidence-rate ratios and shows flat pre-acquisition trends (supporting the parallel-trends assumption) and gradual post-acquisition build-up, peaking at $$t = +3$$.

**Full sample robustness (Table 5).** Equation (1) is re-estimated on the full unmatched sample of 10,942 deals with deal and calendar-year fixed effects; $$\beta$$ coefficients remain positive and significant across all outcomes, with magnitudes comparable to or slightly larger than Panel B of Table 4.

**Mechanism tests (Sections 5.1-5.3).** Equation (2) is applied in three variants:
- **Acquirer expertise (Table 7)**: $$\gamma^{\text{AE}} - \gamma^{\text{NAE}}$$ is positive and significant for patent count (0.368\*\*) and forward cites (0.358\*\*), confirming that experienced acquirers drive the effect. The two-advisor coefficient ($$\gamma^{2\text{FA}}$$) exceeds the one-advisor coefficient across most outcomes.
- **Breakthrough sectors (Table 8)**: $$\gamma^B$$ is positive and significant for all seven outcomes; $$\gamma^T$$ (traditional sectors) is significant only for patent value and combined counts, confirming breakthrough-sector concentration.
- **Target patent status (Table 9)**: both WP (with patent) and WoP (without patent) coefficients are positive, with WP stronger for patent count and WoP stronger for patent value and forward cites; the difference is significant only for patent value, combined patent count, and combined forward cites.

**Withdrawn deals (Table 6).** Successful private-target acquirers are compared to matched withdrawn private-target acquirers (following Seru 2014 and Bena and Li 2014). After trimming the top 1% of outcomes, forward cites, combined patent count, combined forward cites, and number of inventors show significant positive $$\beta$$ coefficients (Panel B), supporting the conclusion that the innovation gains are attributable to the acquisition rather than to acquirer innovation momentum. The pattern is reversed for public targets (Panels C-D), where the $$\beta$$ coefficients are not significant.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| KPSS patent database (Kogan et al. 2017) | Patent count, forward citations, patent economic value for acquirers and combined entities; sourced from GitHub | No page yet |
| KPST patent database (Kelly et al. 2021) | Technology classification; breakthrough patent identification; sourced from dimitris-papanikolaou.github.io | No page yet |
| SDC Platinum | M&A deal identification, deal type, transaction value, announcement and completion dates | [SDC Platinum](/wiki/commercial/sdc-platinum/) |
| Compustat (via WRDS) | Acquirer financial variables: total assets, R&D expenditure, leverage, net income, industry concentration | [WRDS](/wiki/commercial/wrds/) |
| CRSP (via WRDS) | Stock returns for acquirer 5-day CAR calculation around announcement | [WRDS](/wiki/commercial/wrds/) |
| PatentsView | Inventor tracking and new cross-firm collaboration links; matched to CRSP via KPSS patent numbers | No page yet |

Sample: acquisitions announced 1995-2015 (with 5-year patent windows yielding a data span of 1990-2020). 1,153 private-target and 1,153 public-target matched pairs; 23,219 firm-event-year observations. Innovation variables are measured annually at the acquirer level (patent count, forward cites, patent value, inventor counts) or at the combined acquirer-target level (combined patent count and forward cites). Announcement returns use a 5-day window centered on the deal announcement date.

## When to read the full paper

Use the [original](https://doi.org/10.1016/j.jcorpfin.2025.102905) if you are:
- replicating the matching procedure (Appendix A-D give the exact propensity-score model, patent matching, inventor assignment, and variable definitions);
- extending the analysis to non-US markets or longer post-acquisition horizons;
- studying mechanism channels in detail (Tables 7-9 cover acquirer expertise, breakthrough sectors, and target patent status, with full coefficient tables);
- examining the small-scale private-target case studies (Appendix E, 21 acquisitions with no granted patents) to understand how tacit innovation is identified qualitatively.

The locators above point to the exact tables and figures.

## Attribution and rights

Source: peer-reviewed, *Journal of Corporate Finance* 96 (2026) 102905. This distillation was extracted by an LLM on 2026-06-26 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Farida, Siti, Jana P. Fidrmuc, and Chendi Zhang.
> "M&As and Innovation: Evidence from Acquiring Private Firms."
> *Journal of Corporate Finance* 96 (2026): 102905.
> DOI: 10.1016/j.jcorpfin.2025.102905. © 2025 The Authors.
> Published by Elsevier B.V. under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Real Estate Collateral, Lender Screening, and M&A Performance: Gao, Luong & Qiu (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jcf/2026/gao-real-estate-collateral-ma-performance-2026/
# Distilled: Higher market value of corporate real estate (REMV) improves acquirer M&A deal quality measured by three-day announcement returns, operating through two channels: real estate collateral triggers tighter lender acquisition covenants (ex-ante screening), and REMV appreciation expands financial flexibility for constrained firms in high-growth industries. Journal of Corporate Finance 98, 2026, CC BY 4.0. Seven core results with source locators, the empirical specifications, and the REMV construction equations.
# Tags: paper-summary, mergers-acquisitions, real-estate, collateral, lender-screening, corporate-finance
==============================================================================

**What this is.** The paper's core results, the real estate market value measure it constructs (REMV), the identification strategy, and the two economic channels it tests: enough to know what it found and how, without reading the full 28 pages. To replicate or extend, read the original at [doi.org/10.1016/j.jcorpfin.2026.102962](https://doi.org/10.1016/j.jcorpfin.2026.102962).

## TL;DR

Gao, Luong & Qiu (2026) show that the market value of a firm's corporate real estate holdings (REMV) positively predicts the quality of its subsequent M&A decisions. Using 3,272 completed deals (2004-2020) and instrumental variables based on headquarters-state property taxes, crime rates, and natural-disaster exposure, the paper documents that a one-standard-deviation increase in REMV raises the acquirer's three-day announcement return (CAR3) by 0.455 percentage points, roughly 41% of the sample mean. Two complementary economic channels account for this effect. First, acquirers with high REMV are more likely to pledge real estate as collateral in M&A-related loans, and such loans embed substantially more restrictive acquisition covenants, consistent with lender ex-ante screening as in Rajan & Winton (1995). Second, REMV appreciation relaxes financing constraints for acquirers in industries with strong growth opportunities, enabling profitable acquisitions that would otherwise be unattainable. The paper extends the REMV construction method of Chaney, Sraer & Thesmar (2012) and contradicts Hossain et al. (2023), who report an insignificant CAR3 effect using a narrower CST-based measure and a shorter sample window.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%. Standard errors clustered at the acquirer level unless noted. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | REMV positively predicts acquirer CAR3 at the 1% level across all baseline specifications and sample restrictions | Table 3 Panel A Col. 1, p. 9 | Coefficient 0.025\*\*\* (SE 0.007); 1-SD increase in REMV (SD=0.182) raises CAR3 by 0.455 pp; sample mean CAR3 = 1.1%, so the effect = ~41% of mean |
| R2 | 2SLS IV estimates confirm the positive REMV-CAR3 relation; instruments pass overidentification test | Table 4 Cols. 2-3, p. 15 | Instrumented REMV: Col. 2 = 0.170\*\* (SE 0.075), Col. 3 = 0.169\*\* (SE 0.074); Cragg-Donald F = 13.321; Hansen J p-values = 0.376 and 0.395 |
| R3 | Real estate collateral is associated with significantly more restrictive M&A acquisition covenants in loan agreements | Table 6 Panels A-B, p. 17 | Univariate: 85.47% of RE-collateral loans have partial acquisition restrictions vs. 54.66% for non-RE-collateral loans (p < 0.001); multivariate: RE collateral coefficient = 0.147\*\*\* (SE 0.026) |
| R4 | Acquirers pledging real estate collateral in M&A-linked loans earn significantly higher announcement returns | Table 6 Panel A + Table 7 Col. 1, pp. 17-18 | Univariate: CAR3 = 2.26% (RE collateral, N=645) vs. 1.16% (all other loans, N=785), p = 0.0065; multivariate: RE collateral = 0.010\* (SE 0.005) on CAR3 |
| R5 | RE-collateral borrowers are more likely to withdraw M&A bids when deal announcement returns are negative | Table 8, pp. 19-20 | CAR3 × RE collateral coefficient = -0.431\*\* (SE 0.191) on bid-withdrawal probability; significant at 5% in both specifications (N=976) |
| R6 | REMV-CAR3 relation is concentrated in high industry-growth firms and financially constrained acquirers in high-growth sectors | Table 9 Panels A-B, p. 21 | High industry Q: REMV = 0.035\*\*\* (SE 0.010) vs. low: 0.017\* (SE 0.010); financially constrained + high-growth: 0.054\*\*\* (SE 0.020) vs. constrained + low-growth: -0.001 (n.s.) |
| R7 | Two-dimensional test: both lender screening and financial flexibility are active simultaneously; returns are highest when growth is high and covenants are tight | Table 10, p. 21 | High growth + M&A restriction: 0.092\*\*\* (SE 0.029); high growth + no restriction: 0.032\*\*\* (SE 0.009); low growth + restriction: 0.031 (n.s.); low growth + no restriction: 0.010 (n.s.) |

**Overall (paper's conclusion).** Real estate asset values shape M&A deal quality through two complementary channels: collateral-driven lender screening (restrictive covenants identifying high-quality projects ex ante) and enhanced financial flexibility (allowing financially constrained firms to pursue NPV-positive acquisitions in high-growth industries). Results are robust to IV identification, alternative REMV measures (MSA-level HPI, commercial property index, CST method), firm fixed effects, and exclusion of firms in real-estate and tradable-industry subsamples.

## Theory / model

The paper develops no formal mathematical model. It articulates two hypotheses grounded in existing theories of collateral and corporate investment.

**H1 (lender screening).** When a firm pledges real estate as collateral for M&A-related financing, lenders face stronger monitoring incentives because real estate is illiquid and difficult to redeploy relative to other collateral types such as receivables or inventory (Campello et al. (2022)). Rajan & Winton (1995) predict that lenders accepting riskier collateral will increase ex-ante screening intensity by imposing more restrictive covenants. In the M&A context this implies real-estate-secured loans embed more restrictive acquisition covenants, disciplining borrower deal-making and selecting for higher-quality acquisitions and better announcement returns. The paper tests whether pledging real estate as collateral rather than any asset activates this screening mechanism.

**H2 (financial flexibility).** When REMV appreciates, a firm's pledgeable collateral value increases, expanding its borrowing capacity and relaxing financing constraints. This enhanced access to capital is particularly valuable for financially constrained acquirers in industries with strong growth opportunities (high Tobin's Q, sales growth, or asset growth), where positive NPV acquisitions exist but the binding constraint is access to external finance. Jovanovic & Rousseau (2002) argue high-Q firms should acquire low-Q targets; REMV appreciation helps constrained high-Q firms act on this potential.

**H1 versus moral-hazard theories.** An alternative view (Stulz & Johnson (1985); Boot et al. (1991); Holmstrom & Tirole (1997)) predicts that collateral substitutes for bank monitoring rather than intensifying it, leading to looser screening and lower deal quality. The paper tests these opposing predictions empirically using hand-collected covenant data and deal-performance measures (R3, R4, R7 above).

## Method

The paper extends the REMV construction of Chaney, Sraer & Thesmar (2012) (CST) to cover firms entering the sample after 1993 and to account for real estate purchases and dispositions throughout the sample period. The book value of a firm's real estate assets at year $$t$$ is (p. 5):

$$BV_t = FATB_t + FATC_t + FATP_t \tag{BV}$$

where FATB = plant and equipment including buildings; FATC = construction in progress; FATP = land and improvements (all at historical cost). For firms continuously reporting accumulated depreciation of buildings (DPACB) after 1993, the average age of real estate assets is estimated as:

$$\text{Age}_{i,t} \; (\text{in years}) = 40 \times \frac{DPACB_t}{FATB_t}, \qquad \text{Year of Purchase}_t = t - \text{Age}_{i,t}$$

and the market value of real estate is inflated from the historical book value at purchase to the current year using a state-level Housing Price Index (HPI), substituting CPI where HPI is unavailable (p. 5):

$$MV_t = BV_{\text{Year of Purchase}} \times \frac{HPI_t}{HPI_{\text{Year of Purchase}}} \tag{MV}$$

For firms that cease reporting DPACB after 1993, the market value is updated recursively year-by-year (p. 6). A positive change in book value $$\Delta BV_{t+1} > 0$$ signals a new real estate purchase: the market value of new assets is added to the existing assets appreciated by HPI:

$$MV_{t+1} = MV_t \times \frac{HPI_{t+1}}{HPI_t} + \Delta BV_{t+1}, \qquad \text{if } \Delta BV_{t+1} > 0$$

A non-positive change signals a disposal: the existing portfolio is scaled down by the fraction of book value sold and appreciated by HPI:

$$MV_{t+1} = MV_t \times \frac{HPI_{t+1}}{HPI_t} \times \frac{BV_{t+1}}{BV_t}, \qquad \text{if } \Delta BV_{t+1} \leq 0$$

The normalized measure used throughout is:

$$REMV_t = \frac{MV_t}{\text{Total assets}_t}$$

Panel A of Appendix Table A2 (p. 24) confirms high correlation with the CST measure (0.856 for firms with non-missing values) while providing substantially broader sample coverage, especially post-2010. For the lender-screening tests, the paper hand-collects approximately 1,200 M&A-related loan contracts from SEC EDGAR 8-K filings, classifying each as RE-collateral, non-RE-collateral, or unsecured, and coding six categories of restrictive acquisition covenants (full restriction, expenditure limit, minimum profitability of target, no hostile acquisition, no diversifying acquisition, pro forma compliance, specific acquisition target; Table A6, appendix, pp. 26-27).

## Empirical specifications

**Baseline OLS (Eq. 1, p. 9).** The primary regression is a deal-level OLS with two-digit SIC industry and year fixed effects:

$$CAR3_i = \beta_0 + \beta_1 \, REMV_{i,t-1} + \beta \, X_{i,t-1} + \gamma_k + \lambda_t + \varepsilon_i \tag{1}$$

Controls $$X_{i,t-1}$$ include acquirer log market cap, M/B ratio, ROA, leverage, log cash, past stock return, top-5 institutional ownership, state-level real estate return, and deal characteristics: relative size, all-cash indicator, tender offer, diversifying indicator, private target, subsidiary target. $$\gamma_k$$ = 2-digit SIC industry FE; $$\lambda_t$$ = year FE; standard errors clustered at the acquirer level. $$CAR3$$ is the market-adjusted three-day cumulative abnormal return around deal announcement using the CRSP value-weighted index. The headline coefficient across all columns of Table 3 Panel A is 0.025\*\*\* to 0.026\*\*\* (SEs 0.007).

**Instrumental variable 2SLS (Table 4, p. 15).** To address omitted-variable bias and reverse causality, three instruments are used for REMV in the first stage:

1. *State property tax rate*: higher state taxes depress real estate values and REMV (Oates (1969); Hoyt et al. (2011)) without directly affecting M&A activity. Data from the Minnesota Department of Revenue, 2004-2019.
2. *State crime rate*: higher crime rates reduce property values and REMV (Gibbons (2004); Linden & Rockoff (2008)) without directly affecting acquisition quality. Data from the FBI Uniform Crime Reporting (UCR) program.
3. *Has disaster* (county-level binary): indicator for counties with severe natural-disaster exposure in the recent two years (SHELDUS database; top-decile property damage; hazards relevant to real estate: flooding, wildfire, severe storms, earthquake, hurricane, landslide, coastal, tsunami). Natural disasters reduce local real estate values without directly affecting M&A strategy.

First-stage coefficients (Table 4 Col. 1): state property tax = -2.796\*\*\* (SE 0.815), crime rate = -0.002\*\* (SE 0.001), Has disaster = -0.042\*\* (SE 0.021). Cragg-Donald F = 13.321. Second stage: Col. 2 instrumented REMV coefficient = 0.170\*\* (SE 0.075); Col. 3 = 0.169\*\* (SE 0.074). Hansen J overidentification test p-values = 0.376 and 0.395, supporting instrument validity.

**Lender-screening regressions (Tables 6-7, pp. 17-18).** Acquisition covenant restrictiveness is modeled as:

$$\text{M\&A restriction}_{i} = \alpha_0 + \alpha_1 \, \text{RE collateral}_i + \alpha \, X_{i,t-1} + \gamma_k + \lambda_t + \nu_i$$

with the same controls as Eq. (1). RE collateral = 0.147\*\*\* (SE 0.026) in the full-or-partial-restriction specification (Table 6 Panel B Col. 1); non-RE collateral = -0.210\*\*\* (Col. 2). A separate regression of CAR3 on RE collateral in the matched-loan subsample (Table 7 Col. 1, N=948) yields a coefficient of 0.010\* (SE 0.005). Controlling for restrictive acquisition covenants makes the RE collateral coefficient insignificant (Appendix Table A12), confirming that covenants mediate the collateral-CAR3 link.

**Bid-withdrawal test (Table 8, pp. 19-20).** Ex-ante screening is verified by testing whether negative announcement returns more strongly predict withdrawal for RE-collateral borrowers:

$$\text{Withdrawn}_i = \delta_0 + \delta_1 \, CAR3_i + \delta_2 \, \text{RE collateral}_i + \delta_3 \, (CAR3_i \times \text{RE collateral}_i) + \delta \, X_{i,t-1} + \gamma_k + \lambda_t + u_i$$

using a sample of completed and withdrawn deals (N=976). The interaction $$\delta_3 = -0.431^{**}$$ (SE 0.191): worse announcement returns predict withdrawal significantly more for RE-collateral borrowers, consistent with stricter ex-ante screening making borrowers more selective in completing deals that receive negative market signals.

**Growth-opportunity and covenant heterogeneity (Tables 9-10, p. 21).** The sample is split by industry-level growth opportunity (Tobin's Q, sales growth, or asset growth relative to the annual cross-industry median) and by whether the matched loan includes M&A acquisition restrictions. The four cells of the high/low growth times restriction/no-restriction partition produce REMV coefficients on CAR3 ranging from 0.092\*\*\* (high growth, tight covenants) to 0.010 (n.s., low growth, loose covenants), confirming that lender screening and financial flexibility are simultaneously active and complementary.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Compustat annual (via WRDS) | Acquirer firm characteristics; DPACB, FATB, FATC, FATP for REMV construction; total assets, M/B, ROA, leverage, cash | [WRDS / Compustat](/wiki/commercial/wrds/) |
| CRSP (via WRDS) | CAR3 construction using value-weighted index; acquirer past stock return; market capitalization | [WRDS / CRSP](/wiki/commercial/wrds/) |
| Thomson One Banker SDC | M&A deal sample: 3,272 completed deals 2004-2020; deal value, method of payment, target type | no page yet |
| DealScan (via WRDS) | Loan collateral type identification; cross-check for RE-collateral classification | [WRDS / DealScan](/wiki/commercial/wrds/) |
| SEC EDGAR 8-K filings | Hand-collected loan contracts (~1,200 agreements) for acquisition covenant data and RE collateral classification | [SEC EDGAR](/wiki/datasets/edgar/) |
| State-level HPI / CPI | Inflating historical real estate book values to current market values in the REMV construction | no page yet |
| FBI UCR crime data | Instrument 2: state-level crime rates 2004-2019 | no page yet |
| SHELDUS natural disasters | Instrument 3: county-level severe natural-disaster exposure, 2-year rolling window | no page yet |
| Minnesota Dept. of Revenue | Instrument 1: state property tax rates scaled by state personal income, 2004-2019 | no page yet |

Sample: 3,272 M&A deals announced 2004-2020 (deal level); acquirer characteristics lagged one year. Lender-screening subsample: 1,430 deals with matched loan filings (8-K + DealScan); RE-collateral loan subsample: 645 (8-K filings) and 948 deals (including DealScan matches).

## When to read the full paper

Use the [original](https://doi.org/10.1016/j.jcorpfin.2026.102962) if you are: studying how asset collateral shapes deal quality beyond the announcement-return evidence; replicating the extended REMV construction (Section 3 and Appendix Table A2); examining how acquisition covenants in credit agreements reflect lender screening (Tables 6-8); or testing the interaction of real estate wealth with financial constraints and industry growth opportunities (Tables 9-10). Tables 3-4 contain the main result and IV estimates; Tables 6-7 the covenant and collateral evidence.

## Attribution and rights

Source: peer-reviewed, *Journal of Corporate Finance* 98 (2026) 102962. This distillation was extracted by an LLM on 2026-06-26 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Gao, Mingze, Thanh Son Luong, and Buhui Qiu.
> "Real estate collateral, lender screening, and M&A performance."
> *Journal of Corporate Finance* 98 (2026): 102962.
> DOI: 10.1016/j.jcorpfin.2026.102962. © 2026 The Authors.
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Generalist CEO and Managerial Challenge: Gelman, Fralich, Bitektine & Zahraei (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jcf/2026/gelman-generalist-ceo-managerial-challenge-2026/
# Distilled: CEO generalist experience raises announcement CARs only when the hiring firm faces a managerial challenge (high complexity or prior poor performance); the pooled average effect is zero. CEO compensation carries a stable experience premium regardless of firm challenge. Journal of Corporate Finance vol. 97 (2026), CC BY 4.0. Nine core results with source locators, datasets used, the CEO job-market toy model, and the event-study and OLS interaction specifications.
# Tags: paper-summary, corporate-governance, executive-compensation, ceo-turnover, managerial-ability
==============================================================================

**What this is.** This is a machine-distilled skeleton of the paper. Read the
[original article](https://doi.org/10.1016/j.jcorpfin.2025.102917) to replicate or extend the results.

## TL;DR

Gelman, Fralich, Bitektine and Zahraei study whether investors react positively to generalist CEO
experience at new CEO announcements, and why prior studies found no such reaction. Using 1095 CEO
turnovers in S&P 1500 firms from 2000 to 2015 and the General Ability Index (GAI) of Custodio,
Ferreira and Matos (2013), they find no significant average CAR response to CEO experience. However,
when the firm is complex (large in scale and scope of operations) or has performed poorly prior to the
CEO change, investors react positively to higher GAI. One standard deviation of experience raises
market capitalization by about 0.85% for a complex firm and 0.82% for a poorly performing firm; in
firms facing at least one challenge (roughly two-thirds of the sample), the effect is 0.68%. CEO
compensation, by contrast, carries a positive and stable experience premium (7.3% per SD) regardless
of firm challenge, consistent with a CEO job-market model where the outside option is priced against
average-firm challenge rather than the specific hiring firm.

## Core results

| # | Result | Locator | Magnitude as reported |
|---|---|---|---|
| R1 | Base effect of CEO generalist experience (GAI) on 5-day announcement CARs: no significant average effect | Table 4, col 1, p. 11 | GAI coeff = 0.00252 (se = 0.00248), not significant |
| R2 | Firm complexity positively moderates the GAI-CAR relation (linear interaction) | Table 4, col 2, p. 11 | GAI × Complexity index = 0.00405\*\* (se = 0.00183) |
| R3 | High-complexity firms show positive investor reaction to CEO experience; low-complexity firms do not | Table 4, col 3, p. 11 | GAI × High complexity = 0.00900\*\* (se = 0.00372); F-test split-sample p = 0.0433 |
| R4 | Prior firm performance negatively moderates the GAI-CAR effect (higher performance, smaller GAI benefit) | Table 4, col 4, p. 11 | GAI × Performance index = −0.00654\*\*\* (se = 0.00241); F-test split-sample p = 0.0028 |
| R5 | Low-performing firms: CEO experience raises CARs by 0.82%; high-performing firms: no significant effect | Table 4, col 5, p. 11 | GAI × Low performance = 0.00874\*\*\* (se = 0.00329); GAI × High performance = −0.00439 (not sig) |
| R6 | Any-challenge firms: one SD in GAI raises CARs by 0.68%; Double Challenge shows the largest effect | Table 4, cols 7-8, p. 11 | GAI × Any challenge = 0.00727\*\*\* (se = 0.00281); GAI × Double challenge = 0.0162\*\*\* (se = 0.00545); vs No Challenge p = 0.0031 |
| R7 | CEO compensation premium for generalist experience is positive and unmoderated by firm challenge | Table 5, col 1, p. 13 | GAI = 0.0728\*\* (se = 0.0307); 7.3% per SD; no significant moderation by complexity or performance |
| R8 | GAI Surprise robustness: challenge firms show positive CARs; no-challenge firms show negative CARs | Table 6, col 8, p. 15 | GAI\_SURP × Any challenge = 0.00743\*\* (se = 0.00289); GAI\_SURP × No challenge = −0.00738\* (se = 0.00442) |
| R9 | Long-term ROA: complexity positively moderates the effect of residual CEO experience on operational performance | Table 7, col 2, p. 16 | GAI\_SURP × Complexity index = 0.00592\*\* (se = 0.00252) |

**Overall.** Generalist CEO experience benefits investors when the firm faces at least one dimension of
managerial challenge (complexity or prior poor performance), and appears to destroy value for firms
without such challenge. CEO compensation does not share this conditionality: boards pay a consistent
premium for experience irrespective of firm challenge, consistent with the theoretical prediction that
CEO outside options are tied to market-average challenge, not the specific firm. Betzer, van den
Bongard and Limbach (2020)'s absence of a broad investor reaction is explained by pooling challenge
and non-challenge firms where the effects partially cancel.

## Theory / model

The paper develops a two-round CEO job-market model (Appendix A, pp. 21-22) grounded in the
competitive assignment literature of Gabaix and Landier (2008), Pan (2017), and Tervio (2008).
The model has limited market participation: only one CEO candidate enters per period and faces one
firm in round one, with probability $$\pi$$ the candidate also meets a second firm in round two.

**Total surplus.** The production function assumes complementarity between firm challenge $$c_j$$ and
CEO generalist experience $$m_i$$, so total surplus from a match is:

$$T(c_j, m_i) = c_j \cdot m_i \tag{A1}$$

Investor (residual) surplus is the difference between total surplus and CEO compensation $$p$$:

$$v(c_j, m_i) = c_j \cdot m_i - p \tag{A2}$$

**CEO outside option and round-1 compensation.** If round two occurs, the CEO faces firm $$k$$ with
challenge $$c_k \sim U[C_{\min}, C_{\max}]$$ and receives pay equal to total surplus. The expected
round-2 compensation, conditional on entering round two, is:

$$E[p_2 \mid \text{Round two}] = E[c_k] \cdot m_i = \frac{C_{\min} + C_{\max}}{2} \cdot m_i \tag{A4}$$

The CEO's outside option in round one is the probability-weighted average of round-two pay and
current-position pay $$p_0$$:

$$\text{outside\_option} = \pi \cdot \frac{C_{\min} + C_{\max}}{2} \cdot m_i + (1-\pi) \cdot p_0 \tag{A5}$$

Because firms match the outside option exactly (they have stronger bargaining power in round one),
first-round CEO compensation is:

$$p_1 = \pi \cdot \frac{C_{\min} + C_{\max}}{2} \cdot m_i + (1-\pi) \cdot p_0 \tag{A6}$$

Taking the derivative with respect to experience shows that CEO compensation is independent of the
hiring firm's specific challenge level $$c_j$$:

$$\frac{\partial p_1}{\partial m_i} = \pi \cdot \frac{C_{\min} + C_{\max}}{2} \tag{A7}$$

**Investor surplus and the threshold.** Substituting (A6) into (A2) yields investor surplus:

$$v(c_j, m_i) = c_j \cdot m_i - \left\{ \pi \cdot \frac{C_{\min} + C_{\max}}{2} \cdot m_i + (1-\pi) \cdot p_0 \right\} \tag{A8}$$

subject to the firm participation constraint (investor surplus is non-negative):

$$c_j \cdot m_i \geq \pi \cdot \frac{C_{\min} + C_{\max}}{2} \cdot m_i + (1-\pi) \cdot p_0 \tag{A9}$$

The marginal effect of CEO experience on investor surplus is:

$$\frac{\partial v(c_j, m_i)}{\partial m_i} = c_j - \pi \cdot \frac{C_{\max} + C_{\min}}{2} \tag{A10}$$

This is positive if and only if $$c_j > \pi(C_{\max} + C_{\min})/2$$: investor surplus is increasing
in CEO experience only when the firm's challenge level exceeds the threshold. For firms below the
threshold, the CEO captures the full increment in total surplus as compensation, leaving investors no
better off with a more experienced hire.

**Hypotheses (pp. 4-5).** Drawing on the managerial challenge concept of Hambrick et al. (2005)
and the firm complexity literature, the model yields four empirical predictions:
- H1: Firm complexity positively moderates the GAI effect on investor reaction to a new CEO.
- H2: CEO experience has a positive effect on investor reaction when the firm is sufficiently complex.
- H3: Prior firm performance negatively moderates the GAI effect on investor reaction.
- H4: CEO experience has a positive effect on investor reaction when prior performance is sufficiently poor.

## Method

**General Ability Index.** Following Custodio, Ferreira and Matos (2013), CEO generalist experience
is measured by GAI, the first principal component of five standardized career-breadth variables
(p. 6):

$$\text{GAI}_{it} = 0.494 \cdot \#\_\text{positions} + 0.585 \cdot \#\_\text{firms} + 0.508 \cdot \#\_\text{industries} + 0.316 \cdot \text{CEO\_Exp\_Dummy} + 0.238 \cdot \text{Cong\_Exp\_Dummy} \tag{1}$$

where all inputs are standardized. Number of positions, firms, and industries come from BoardEx merged
with Execucomp; conglomerate experience from Worldscope. The final GAI is also standardized.

**Complexity Index.** Firm complexity captures the scale and scope dimensions of Chandler (1994)'s
analysis: the average of standardized log(number of employees) (from Compustat/Execucomp) and
number of product segments (from Worldscope, counting non-zero-sales segments in the CEO appointment
year). High complexity = upper tercile; low complexity = lower two terciles (p. 6-7).

**Performance Index.** Firm performance is isolated from industry and firm-characteristics effects
using two components (p. 7): (1) CAPM Alpha (Jensen, 1969) estimated on monthly CRSP returns
over the three years before the CEO announcement; (2) Residual Firm Efficiency from
Demerjian, Lev and McVay (2012), the residual of a regression of total factor productivity on firm
characteristics, averaged over years t-3 to t-1. The Performance Index is the average of these two
standardized measures. Low performance = below-median index in the year of turnover.

**GAI Surprise.** To partially address endogeneity from non-random CEO-firm matching, the paper
constructs GAI Surprise as the OLS residual from regressing incoming CEO GAI on Complexity,
Performance, firm controls, and departing CEO compensation (Table 3, col 4, pp. 8-10). This
isolates the portion of incoming CEO experience not predicted by observable firm characteristics
and the departing CEO's pay, which may reflect unobserved challenge.

## Empirical specifications

**Announcement CAR regression (primary, Table 4, p. 11).** The dependent variable is the
Carhart (1997) 4-factor cumulative abnormal return over a 5-day window [-2, 2] centered on the
new CEO announcement date, estimated using a [-255, -46] pre-announcement window. The baseline
regression and interaction specifications are:

$$\text{CAR}_{22,it} = \alpha + \beta_1 \text{GAI}_{it} + \beta_2 \text{GAI}_{it} \times \text{Challenge}_{it} + \beta_3 \text{Challenge}_{it} + \gamma' X_{it} + \delta_t + \delta_s + \varepsilon_{it}$$

where $$\text{Challenge}_{it}$$ is either the Complexity Index (col 2, linear) or the Performance Index
(col 4, linear), or piecewise high/low dummies (cols 3, 5), or intersection dummies for "No
Challenge" (high performance × low complexity), "One Challenge" (low performance × low complexity
or high performance × high complexity), "Double Challenge" (low performance × high complexity), and
"Any Challenge" (One + Double, cols 7-8). Controls $$X_{it}$$ include prior 3-year sales growth, log
assets, prior 3-year firm efficiency, complexity, performance, previous-year stock return and
volatility, Fasttrack, MaleY0, Insider, Forced, Unclassified, and CEO age dummies. Year and
2-digit SIC industry fixed effects are included throughout; standard errors are clustered by firm.

**Compensation regression (Table 5, p. 13).** Log total CEO compensation (TDC1 from Execucomp,
first full calendar year as CEO) is regressed on GAI and the same challenge interactions with
identical controls. This tests whether firm challenge moderates the compensation-experience relation,
as it does for CARs; the prediction is that it does not.

**GAI Surprise robustness (Table 6, p. 15).** GAI Surprise replaces GAI as the main explanatory
variable, replicating all challenge-interaction specifications with the same control set. The
purpose is to show that results hold when the experience measure isolates the component not
predicted by observable firm and departing-CEO characteristics.

**Long-term performance (Tables 7-8, p. 16-17).** Industry-adjusted ROA ($$\bar{\Omega}\_{\text{IND\_ADJ\_ROA}}_{t+1:t+3}$$)
and Tobin's Q, averaged over years t+1 to t+3 after appointment, are the outcome variables.
Lagged industry-adjusted performance ($$\Omega\_{\text{IND\_ADJ\_ROA}}_{t-1}$$) controls for mean
reversion. GAI Surprise is used as the explanatory variable to address selection; the same
challenge-interaction structure as Table 4 applies. Year and sector fixed effects; firm-clustered
standard errors. A propensity-score-matched sample (matched on complexity, performance, firm
controls, year and industry) is used as a further robustness check (Internet Appendix 7).

**CEO tenure (Cox hazard model, Table 9, p. 19).** The hazard of CEO succession is modeled as:

$$h(t) = h_0(t) \exp\!\left(\beta_1 \text{GAI}_{it} + \beta_2 \text{GAI}_{it} \times \text{Challenge}_{it} + \gamma' X_{it}\right)$$

using CEO departure dates from Execucomp as of June 30, 2024. Year and 2-digit SIC dummies; same
controls as Table 4. A higher (lower) coefficient means shorter (longer) expected CEO tenure.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CRSP (via WRDS) | Stock prices for 4-factor Carhart CAR estimation; stock return and volatility controls | [WRDS](/wiki/commercial/wrds/) |
| Compustat (via WRDS) | Firm financial characteristics (assets, sales growth, firm size) | [WRDS](/wiki/commercial/wrds/) |
| Execucomp (via WRDS) | CEO identification; compensation (TDC1); GAI components (positions, prior CEO role) | [WRDS](/wiki/commercial/wrds/) |
| Worldscope | Number of product segments (scope dimension of Complexity Index); conglomerate experience for GAI | [Worldscope](/wiki/commercial/worldscope/) |
| BoardEx | CEO career history (positions, firms, industries) for GAI construction; merged with Execucomp and Worldscope | no page yet |
| Factiva | News articles used to classify CEO departure reason as exogenous, forced, or unclassified per Eisfeldt and Kuhnen (2013) | no page yet |
| Demerjian, Lev and McVay (2012) data | Residual firm efficiency scores (managerial ability measure); from public data supplement at faculty.washington.edu/pdemeri | no page yet |

Sample: 1095 CEO turnovers in S&P 1500-listed firms, January 2000-December 2015 (after excluding
interim CEO appointments, financial firms, and turnovers without media announcement data). Long-term
ROA/Tobin's Q analyses use 890-891 observations; CEO tenure analysis uses 1086 observations.

## When to read the full paper

Read the original when: studying CEO succession and whether CEO human capital creates value
differentially by firm type; examining whether competitive CEO assignment models with market
frictions can explain the compensation-investor-reaction disconnect; interested in the GAI
measure of Custodio, Ferreira and Matos (2013) and its interaction with firm characteristics. The
main CAR results are in Table 4 (pp. 11-12); compensation results in Table 5 (pp. 13-14); the
identification robustness using GAI Surprise in Table 6 (p. 15); and long-term performance and
CEO tenure in Tables 7-9 (pp. 16-19). The toy model and threshold derivation are in Appendix A
(pp. 21-22).

## Attribution and rights

This article is open access under a Creative Commons Attribution (CC BY 4.0) license.
Published by Elsevier B.V.

> Gelman, S., Fralich, R., Bitektine, A., & Zahraei, S. (2026). When does a generalist CEO create
> shareholder value? The effect of managerial challenge. *Journal of Corporate Finance*, 97, 102917.
> https://doi.org/10.1016/j.jcorpfin.2025.102917

Machine-distilled by paper-distiller (claude-sonnet-4-6), 2026-06-26. Not human-verified; not
reproduced. Extraction role: extracted only.


==============================================================================
# Competition and the Value of Innovation: Hu & Ma (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jcf/2026/hu-competition-value-of-innovation-2026/
# Distilled: Using a stock-market-based patent value measure, Hu and Ma (2026) document a negative relationship between product-market competition intensity and the economic value of newly granted patents among US public firms 1986-2020; a quasi-experimental design exploiting horizontal M&A events confirms causality, with non-merging peers' patents gaining an average 2.8% in value after such deals. Journal of Corporate Finance vol. 96 (2026) 102909, CC BY 4.0. Six core results with source locators, datasets used, the hypotheses, and the estimating equations.
# Tags: paper-summary, innovation, competition, patents, mergers-acquisitions
==============================================================================

**What this is.** This is a distilled skeleton of the published article. Read the original at [https://doi.org/10.1016/j.jcorpfin.2025.102909](https://doi.org/10.1016/j.jcorpfin.2025.102909) to replicate or extend. All results below carry exact table/figure/page locators from the PDF.

## TL;DR

Hu and Ma (2026) ask how product-market competition shapes the economic returns firms capture from successful innovations. Using the Kogan et al. (2017) (KPSS) stock-market-based patent value measure for all patents granted to US public firms from 1986 to 2020, they document a negative association between a markup-based competition index and the economic value of newly granted patents: a 1% higher competition intensity is associated with 0.3 to 1.9% lower patent value depending on the specification. To address endogeneity, they exploit horizontal M&A announcements as quasi-natural experiments that plausibly reduce competition for non-merging industry peers. Patents issued to those peers just after such announcements are worth on average 2.8% more than patents issued just before, consistent with lower competition raising the expected monopoly rents from a new patent. Effects are larger for stealth mergers (likely-anti-competitive deals just below HSR antitrust thresholds) and for deals in concentrated industries with high product similarity, while non-horizontal M&A announcements produce no effect. Cross-sectional analysis further shows the negative competition-patent value relationship is especially pronounced for technology leaders in low-gap industries, economically valuable patents, and pioneering patents.

## Core results

| # | Result | Locator | Magnitude as reported |
|---|---|---|---|
| R1 | Baseline OLS: higher competition (SIC) associated with lower patent value | Table 2, Panel A, Col 4, p.6 | log(Competition) = -0.717\*\*\* (SE 0.199) |
| R2 | Quasi-experiment: horizontal M&A (4-digit SIC) raises patent value for non-merging peers | Table 4, Panel A, Col 1, pp.10-11 | Post-merger = +0.028\*\*\* (SE 0.010), approx 2.8% |
| R3 | Stealth mergers near upper HSR antitrust threshold: larger patent value gain | Table 5, Panel A, Col 3, p.12 | Post-merger = +0.137\* (SE 0.075), approx 13.7% |
| R4 | Non-horizontal M&A placebo: no patent value change for non-merging peers | Table 4, Panel C, Col 1, p.11 | Post-merger = 0.009 (SE 0.019), insig |
| R5 | High-value patents: amplified negative competition effect | Table 8, Col 5, p.16 | High Value x Competition = -0.864\*\*\* (SE 0.131) |
| R6 | Pioneer patents: amplified negative competition effect | Table 8, Col 7, p.16 | Pioneer x Competition = -1.839\*\*\* (SE 0.431) |

**Overall (paper's conclusion).** The evidence consistently supports the hypothesis that higher competition reduces the economic value of patents. Because firms base R&D investment decisions on the expected economic gains from innovations, lower patent value under intense competition implies weaker incentives to innovate. Effects are heterogeneous: technology leaders in low-gap industries, economically and scientifically important patents, and pioneering patents all face a more negative competition-patent value relationship. Science-based patents are relatively insulated. The paper also documents that firms operating in industries affected by likely-anti-competitive M&A subsequently increase their patent filings, consistent with higher post-merger innovation rents raising the reward to patenting (Table 7, p.14-15).

## Theory / model

The paper has no formal economic model. It develops two competing sets of hypotheses from the theoretical literature (Section 2, pp.3-4):

**Escape-competition channel** (Arrow 1962; Aghion et al. (2005)): In highly competitive markets, firms operate with thinner pre-innovation profit margins and face stronger survival pressure. A patent that grants a temporary monopoly or strong product differentiation allows the patentholder to escape competition and capture a large leap in profitability. The marginal gain from a patent is therefore larger when pre-innovation profits are lower, predicting a positive competition-patent value relationship:

> **Hypothesis 1a:** If the benefits of competition outweigh its negative effects, higher competition intensity will be associated with greater patent value.

**Competitive erosion channels** (Suetens 2005; Engel and Kleine 2015; Igami 2017; Teece 1986): Intense competition accelerates imitation of innovations, encourages rapid follow-up innovations by rivals, and shortens the period of exclusive benefits from a patent, all of which reduce the long-run value of the monopoly rents the patent secures. This predicts a negative relationship:

> **Hypothesis 1b:** If the negative effects of competition outweigh its benefits, higher competition intensity will be associated with lower patent value.

To identify the causal direction, the paper proposes a quasi-experimental design and translates the above into testable hypotheses about the direction of patent value change following an exogenous reduction in competition:

> **Hypothesis 2a:** All else equal, an increase in competition intensity leads to an increase in the value of a patent.
> **Hypothesis 2b:** All else equal, an increase in competition intensity leads to a decrease in the value of a patent.

The paper's results favor Hypothesis 1b and 2b. It finds no evidence of a non-monotonic (inverted-U) relationship at the patent level, in contrast to the firm/industry-level inverted-U documented by Aghion et al. (2005).

## Method

**Competition measure (p.5).** Industry-level competition intensity is measured as one minus the sales-weighted average markup among firms in the industry:

$$\text{Competition}_{st} = 1 - \sum_{i \in s} \frac{\text{Sales}_{it}}{\text{Sales}_{st}} \text{Markup}_{it} \tag{defn}$$

Markup is estimated following Aghion et al. (2005) and Nickell (1996):

$$\text{Markup}_{it} = \frac{\text{Operating Profit}_{it} - \text{Financial Cost}_{it}}{\text{Sales}_{it}}$$

where Operating Profit = Sales - COGS - SG&A - Depreciation, and Financial Cost is the product of a capital cost of 0.085 and the capital stock measured by the perpetual inventory method. A value of Competition close to one indicates near-perfect competition (zero markups), while lower values indicate greater monopoly power. Negative markups are truncated at zero. The primary measure uses the 4-digit SIC industry (Compustat full sample); an alternative uses the TNIC-based industry (Hoberg and Phillips (2016)).

**Patent value measure (p.4; following Kogan et al. (2017)).** Patent value is the KPSS stock-market-based measure: the present value of future cash flows associated with a newly granted patent, estimated as the 3-day cumulative abnormal return (CAR) around the patent grant date, adjusted for estimation noise and scaled by the firm's market capitalization. For firm-days with multiple patents, each patent's value is the total estimated dollar amount divided by the number of patents granted that day. Values are expressed in millions of 1996 US dollars and are winsorized at the 1st and 99th percentiles.

**Identification design (pp.7-9; Figure 2).** The quasi-experiment compares patents granted to non-merging, same-industry peers within a narrow window around horizontal M&A announcements. The three-step sample construction is: (1) retain all patents in the event industry issued within the [-35, +35] day window around the announcement; (2) exclude patents granted to the acquirer or target; (3) apply a [-7, +7] day exclusion window to minimize stock-return contamination. Pre-event patents are those in [-35, -8] and post-event patents are those in [+8, +35]. The identifying assumption is that USPTO patent grant dates are randomly distributed across the two windows (plausible because grant timing follows a 2-3 year examination process driven by examiners, not by firms). Balance tests (Table 3) confirm no systematic differences in patent or firm characteristics between the two groups.

## Empirical specifications

**Baseline OLS regression (Eq. 1, p.6).** The primary estimating equation for the competition-patent value relationship:

$$\log(\text{Patent Value}_{ijst}) = \alpha_i + \beta \log(\text{Competition}_{st}) + \gamma X_{ijst} + \delta_{jt} + \kappa_m + \varepsilon_{ijst} \tag{1}$$

Indices: $$i$$ = firm, $$j$$ = 3-digit CPC patent class, $$s$$ = 4-digit SIC industry, $$t$$ = patent-granting year, $$m$$ = patent-granting month. Fixed effects: firm ($$\alpha_i$$), patent class x year ($$\delta_{jt}$$), year x month ($$\kappa_m$$). Controls $$X_{ijst}$$: log total assets, leverage, ROA, market-to-book, R&D intensity, institutional ownership, patent examination time, quadratics of adjusted citations (backward and forward), quadratics of same-day patent count, annual change in industry firm count, annual change in industry employment, firm aggregate patent economic value and scientific value in the prior year. Standard errors are clustered at industry and year levels. The preferred specification (Table 2, Panel A, Col 4) includes all three sets of fixed effects and yields $$\hat{\beta} = -0.717$$ (SE 0.199, p < 0.01) for single-patent grants.

**Quasi-experimental regression (Eq. 2, pp.8-9).** To identify the causal impact of competition on patent value using horizontal M&A events:

$$\log(\text{Patent Value}_{ijst}) = \alpha_i + \beta \text{Post-merger}_{jst} + \gamma X_{ijst} + \delta_{jt} + \varepsilon_{ijst} \tag{2}$$

$$\text{Post-merger}_{jst} = 1$$ if patent $$j$$ is granted to a non-merging peer during the [+8, +35] event window, $$= 0$$ if granted in the [-35, -8] pre-event window. Fixed effects: firm ($$\alpha_i$$), patent class x year ($$\delta_{jt}$$). Same control vector $$X$$ as Eq. (1). Standard errors clustered at industry and year levels. The baseline result (Table 4, Panel A, Col 1) for horizontal M&A defined by 4-digit SIC industry: $$\hat{\beta} = 0.028$$ (SE 0.010, p < 0.01), approximately 2.8% higher patent value post-merger.

**Heterogeneity regression (Eq. 4, p.14).** Cross-sectional variation in the competition-patent value relationship:

$$\log(\text{Patent Value}_{ijst}) = \alpha_i + \beta_1 \text{Competition}_{st} + \beta_2 \text{Characteristic}_{ijt} + \beta_3 \text{Competition}_{st} \times \text{Characteristic}_{ijt} + \gamma X_{ijst} + \delta_{jt} + \kappa_m + \varepsilon_{ijst} \tag{4}$$

where Characteristic is a firm-level variable (e.g., indicator for technology leader, above-median markup, above-median sales share) or patent-level variable (e.g., indicator for high economic value, high scientific citation count, pioneer patent, science-based patent). The interaction $$\beta_3$$ captures differential competition sensitivity. Key results (Table 8, p.16): technology leaders in low-gap industries face a significantly more negative competition effect ($$\beta_3 = -0.887$$, SE 0.290); high-value patents have $$\beta_3 = -0.864$$ (SE 0.131); pioneer patents have $$\beta_3 = -1.839$$ (SE 0.431).

**Stealth merger and concentrated-industry sub-samples (pp.11-13; Tables 5-6).** To sharpen identification, the paper examines two subsets of horizontal mergers that are more likely to be anti-competitive. Following Kepler et al. (2021), stealth mergers are deals with values just below the HSR antitrust notification thresholds, defined as falling within 5% below the lower or upper threshold. Consistent with the market-power explanation, stealth mergers produce significantly larger increases in peer patent values than non-stealth mergers, while non-stealth mergers produce no significant effect (Table 5). For industry concentration and product similarity, the paper adopts the IPS text-based product-similarity measure developed by Fathollahi et al. (2022), which ranges from zero to one; deals in concentrated industries with high IPS scores show significantly larger patent value increases than deals in non-concentrated or low-IPS industries (Table 6).

**DiD markup analysis (Eq. 3, p.14).** To confirm that likely-anti-competitive M&A raises rivals' market power:

$$\text{Markup}_{ist} = \delta \text{post}_{it} + \gamma_1(\text{treated}_i \times \text{post}_{it}) + \gamma_2(\text{treated}_i \times \text{post}_{it}) \times H_{it} + \Gamma X_{ist} + \alpha_i + \sigma_{jt} + \varepsilon_{ist} \tag{3}$$

where $$H_{it}$$ is an indicator for stealth mergers or mergers in concentrated high-IPS industries, $$\text{treated}_i$$ indicates propensity-score-matched non-merging peers of merging firms, $$\alpha_i$$ are firm FEs, and $$\sigma_{jt}$$ are sector-year FEs. Using 1-to-1 nearest-neighbor matching by same-industry year and sector (2-digit SIC), stealth mergers near the upper HSR threshold produce 2.4% larger peer markup increases than matched control firms (Table 7, p.14-15), consistent with the anti-competitive mechanism driving the patent value results.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CRSP | Daily stock returns for KPSS patent value estimation (3-day CAR); firm market cap, beta, idiosyncratic volatility as controls | [WRDS](/wiki/commercial/wrds/) |
| Compustat | Firm financials: assets, leverage, ROA, market-to-book, R&D intensity, markup (operating profit, SG&A, depreciation, capital stock); 4-digit SIC segment data | [WRDS](/wiki/commercial/wrds/) |
| USPTO patent grants | Filing dates, grant dates, 3-digit CPC patent classes, citation counts (backward/forward); firm match via KPSS procedure | no page yet |
| SDC Platinum | M&A deal data: deal type, deal value, acquirer/target industries (SIC), combined market share | [SDC Platinum](/wiki/commercial/sdc-platinum/) |
| Hoberg-Phillips TNIC | Text-based Network Industry Classification (10-K product descriptions); IPS industry product similarity scores; alternative industry proxy | [TNIC](/wiki/datasets/tnic/) |
| Thomson/Refinitiv 13F Holdings | Institutional ownership (quarterly 13F filings) | [WRDS](/wiki/commercial/wrds/) |
| ISS / Execucomp | Independent board indicator; CEO ownership (governance controls in robustness, Table 9) | [WRDS](/wiki/commercial/wrds/) |
| US Census BDS | Annual changes in firm count and employees by industry; proxies for industry business cycle dynamics | no page yet |
| Marx and Fuegi (2020, 2022) | Science-based patent identification: patents in top-3 quartiles of non-patent literature citations within their class-year | no page yet |

Sample: Patents granted 1986-2020 to US public firms with Compustat financials and CRSP returns. Excludes financial (SIC 6000-6999), utility (SIC 4900-4999), and miscellaneous industries (SIC codes ending in 9). All continuous variables winsorized at 1st and 99th percentiles annually. Dollar amounts deflated to 1996 USD.

## When to read the full paper

Read the original if you are working on:

- **Innovation incentives**: studying how product-market competition shapes firms' economic returns from R&D and patenting, or calibrating innovation models that require a competition-returns elasticity.
- **M&A natural experiments**: using horizontal acquisitions as plausibly-anti-competitive shocks to market structure; the three-step event window construction (Figure 2, p.8) is directly reusable.
- **Patent valuation**: applying or extending the Kogan et al. (2017) KPSS measure; Table 1 (p.6) reports the measure's summary statistics and the paper discusses its relationship to citation-based scientific value measures.
- **Capital budgeting in R&D-intensive firms**: understanding how competition intensity affects the expected economic value of innovations as an input to investment and licensing decisions (managerial implications, p.19).
- **Cross-sectional heterogeneity in innovation**: examining which types of patents or firms are more sensitive to competitive pressure (Table 8, p.16 covers technology leaders, markup quintiles, market-share quintiles, economic vs. scientific value, pioneer vs. follow-up, and science-based patents).

## Attribution and rights

Hu, M., & Ma, L. (2026). Competition and the value of innovation. *Journal of Corporate Finance*, 96, 102909. https://doi.org/10.1016/j.jcorpfin.2025.102909

© 2025 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).

This page is an LLM-distilled summary produced by claude-sonnet-4-6. It has not been human-verified and the results have not been reproduced. Refer to the original article for all citations, proofs, and full robustness checks.


==============================================================================
# Real Effects of Tick-Size Adjustments: Lin, Yao & Zou (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jcf/2026/lin-real-effects-tick-size-ma-2026/
# Distilled: Using the SEC's 2016 Tick Size Pilot as an exogenous shock to stock liquidity, this paper shows that pilot firms required to quote and trade at a larger minimum price increment significantly reduce M&A investment intensity, shift toward smaller private targets, cut stock payment, and retain only deals with better announcement returns during the two-year pilot; the effect reverses partially after the pilot ends. Journal of Corporate Finance 96 (2026), paywalled (Elsevier). Nine core results with source locators, the DID specification, and channel evidence on information asymmetry and valuation. LLM-distilled, not human-verified.
# Tags: paper-summary, mergers-acquisitions, stock-liquidity, tick-size, information-asymmetry, panel-regression, natural-experiment, peer-reviewed, unreplicated, data:wrds, data:sdc-platinum
==============================================================================

**What this is.** The core results and the estimating specification of the paper: enough to know what it found, how, and which datasets it requires. To replicate or extend, read the full source at the [original](https://doi.org/10.1016/j.jcorpfin.2025.102890).

## TL;DR

Using the SEC's 2016 Tick Size Pilot as an exogenous shock to stock liquidity, this paper shows that firms whose stocks are assigned to larger minimum price increments (test groups 2 and 3, required to quote and trade at the $0.05 increment) significantly reduce the intensity of their M&A activity relative to control firms during the two-year pilot period. Firms in test groups 2 and 3 also cut stock payments in M&A deals, avoid large and public-target acquisitions, exhibit higher deal completion rates, and concentrate the reduction in horizontal and diversifying (rather than vertical) mergers. The retained deals earn better announcement returns. The evidence points to information asymmetry and valuation costs of lower stock liquidity as the dominant driver, not just the higher trading costs hypothesized by Huang et al. (2024): the reduction in M&A intensity persists even after excluding stock-financed deals, ruling out the trading-cost-only explanation. A partial reversal after the pilot ends corroborates the causal interpretation. Edmans et al. (2012) provide the motivating framework for studying real effects of stock markets on corporate investment.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%. Locators point to the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Firms in test groups 2&3 significantly reduce M&A intensity during the pilot | Table 2, col (2), p. 9 | TestGroup2&3\*Post = -0.020\*\* (t = -2.72); ~10.9% decline relative to SD of 0.183; TG1 insignificant (0.001, t = 0.16) |
| R2 | Reduction persists for non-stock-financed deals, supporting information asymmetry over pure trading-cost explanation | Table 2, col (6), p. 9 | -0.015\* (t = -2.10) after excluding all stock-financed deals from the sample |
| R3 | Channel (information asymmetry): lower ex-ante information asymmetry amplifies the reduction | Table 3, Panel A, col (1), p. 12 | Triple-DID b1 = -0.028\*\* (t = -2.54) with high-analyst-coverage proxy; combined b1+b2 = -0.026 (p = 0.02) |
| R4 | Channel (valuation): higher pre-pilot stock valuation or larger price drop amplifies the reduction | Table 3, Panel B, col (1), p. 12 | Triple-DID b1 = -0.033\*\* (t = -2.85) for high ex-ante stock valuation; combined b1+b2 = -0.034 (p = 0.00) |
| R5 | Pilot firms reduce the share of M&A deal value paid by stock | Table 4, col (1), p. 13 | -13.7 pp\*\*\* (t = -3.11); TG1 insignificant (-3.3 pp, t = -0.64) |
| R6 | Conditional on undertaking M&A, pilot firms avoid public targets and large deals, and complete more deals | Table 5, cols (2), (3), (5), p. 14 | Public target: -0.074\* (t = -2.01); deal size: -0.273\*\*\* (t = -3.09); completed deal: +0.265\*\* (t = 2.27) |
| R7 | The decline in M&A is concentrated in horizontal and diversifying mergers, not vertical | Table 6, cols (1)-(2), p. 16 | Multinomial logit relative to vertical: Diversifying = -1.887\*\* (t = -2.08); Horizontal = -1.482\* (t = -1.71) |
| R8 | Deal announcement returns improve for retained M&A deals of pilot firms | Table 7, col (1), p. 17 | CAR(-2,2): +0.023\* (t = 2.04); pre-pilot TG2&3 coefficient -0.015\*\* (t = -2.45), consistent with retained deals being higher quality |
| R9 | M&A intensity partially recovers after the pilot ends | Table 8, p. 18 | TestGroup2&3\*Post2 = +0.012\*\*\* (t = 3.62); ~28% of SD of M&A intensity; TG1 insignificant (0.011, t = 1.51) |

**Overall (paper's conclusion).** The tick-size increase causes firms to reduce M&A investment by cutting lower-quality acquisitions (horizontal, diversifying) and structuring retained deals more carefully (smaller size, private targets, less stock payment). The disciplinary effect of reduced stock liquidity raises deal announcement returns for pilot firms. M&A activity shows a partial reversal after the pilot ends, consistent with a causal liquidity shock rather than a pre-existing trend.

## Theory / model

The paper has no formal economic model. It tests the *information asymmetry and valuation* hypothesis: a mandatory increase in the minimum price increment reduces stock liquidity (wider bid-ask spreads, lower trading activity), which harms M&A activity through two channels.

**Channel 1 (information asymmetry).** Lower liquidity increases adverse selection by informed traders and raises uncertainty about acquirer value. Target shareholders find it harder to assess the value of acquirer stock offered as deal consideration, raising the effective cost of stock-financed acquisitions. More broadly, lower liquidity increases the cost of equity financing via higher required returns (Albuquerque et al. (2020)), and can raise debt financing costs by reducing stock price informativeness and weakening monitoring (Brogaard et al. (2017), Brogaard et al. (2021)). These forces make the acquirer a less attractive counterparty, particularly for public targets that scrutinize valuation carefully.

**Channel 2 (valuation).** Pilot stocks experience actual price declines due to higher required returns (Albuquerque et al. (2020)). Firms with a higher pre-pilot stock valuation (high Tobin's Q) or those that actually experience a large price drop face a more severe collateral-and-financing shock. The paper predicts these firms cut M&A activity more, which is confirmed by R3-R4.

The competing *trading cost-based hypothesis* (Huang et al. (2024)) predicts that lower acquirer stock liquidity primarily discourages stock-financed deals by making post-merger cash-out trading costly for target shareholders. This hypothesis predicts no reduction in the intensity of non-stock-financed deals, which is the main discriminating prediction tested in R2 (Table 2, col 6).

**Identification.** The SEC's 2016 Tick Size Pilot randomly assigned 1199 control stocks and 400 stocks per test group from among FINRA-listed small and mid-cap common stocks (market cap at most $3 billion, closing price at least $2, average daily volume at most one million shares) via a stratified random sampling process disclosed publicly. The randomization is confirmed by Table 1, Panel C: no statistically significant pre-pilot differences in M&A intensity or acquirer characteristics between test and control groups (t-tests and Kolmogorov-Smirnov tests, p. 9). Test group 1 stocks are required to quote at $0.05 but can continue to trade at the pre-pilot $0.01 increment; test groups 2 and 3 are required to both quote and trade at $0.05 (with test group 3 additionally subject to a trade-at rule). The paper pools test groups 2 and 3 because they face the same effective tick size in trading, while test group 1 is used as an additional control.

## Method

The main estimating equation (Eq. 1 in the paper, p. 7) is a firm-by-year-quarter OLS panel regression estimated on a symmetric window of seven pre-pilot quarters and seven post-pilot quarters, with the event quarter (2016Q4) excluded:

$$
\text{M\&A intensity}_{it} = \eta_i + \psi_{jt} + \beta_1 \, \text{TG1}_i \times \text{Post}_t + \beta_2 \, \text{TG2\&3}_i \times \text{Post}_t + \beta_4 X_{it} + \varepsilon_{it} \tag{1}
$$

where $$\eta_i$$ are firm fixed effects, $$\psi_{jt}$$ are industry (2-digit SIC)-by-year-quarter fixed effects, $$\text{TG1}_i$$ is a dummy for assignment to test group 1, $$\text{TG2\&3}_i$$ is a dummy for assignment to test groups 2 or 3, $$\text{Post}_t$$ equals one for the seven pilot quarters (2017Q1-2018Q3) and zero for the seven pre-pilot quarters (2015Q1-2016Q3), and $$X_{it}$$ is a vector of controls (log total assets, market-to-book ratio, leverage, cash flow, institutional ownership). The coefficient $$\beta_2$$ is the average treatment effect of the tick-size increase on M&A intensity for firms in test groups 2 and 3 relative to the control group. Standard errors are clustered at both the firm and year-quarter level throughout.

This specification builds on `difference-in-differences` identification and `panel-regression` with two-way fixed effects.

**Channel analysis (triple-DID).** To identify the information asymmetry channel, the specification adds a third interaction layer (Section 4.4, p. 11):

$$
\text{M\&A intensity}_{it} = \eta_i + \psi_{jt} + b_1 \, \text{TG2\&3}_i \times \text{Post}_t \times \text{LowAsymm}_i + b_2 \, \text{TG2\&3}_i \times \text{Post}_t + \cdots + \beta_4 X_{it} + \varepsilon_{it}
$$

Three proxies for low ex-ante information asymmetry (*LowAsymm*) are used: high analyst coverage (top tercile of coverage over the seven pre-pilot quarters), low earnings forecast dispersion (bottom tercile of EPS forecast standard deviation over absolute mean EPS), and small bid-ask spread (average daily quoted spread at most 3 cents in the pre-pilot period, following Albuquerque et al. (2020) and Ran and Ye (2024)). The prediction is $$b_1 < 0$$ and the combined $$b_1 + b_2 < 0$$ with $$|b_1 + b_2| > |b_2|$$: firms that were more transparent pre-pilot experience a larger decline in M&A intensity when the tick size increases, because the cost of the information asymmetry shock is proportionally larger for them. An analogous triple-DID tests the valuation channel, replacing *LowAsymm* with *HighValuation* (pre-pilot Tobin's Q in top tercile) or *LargePriceDrop* (actual stock price decline from pre-pilot to pilot period in top tercile).

**Deal-level outcomes and deal type.** For deal-level binary outcomes (public target, completed deal) and the continuous stock payment fraction and deal size, the estimator is OLS DID at the deal level with industry (2-digit SIC) and time fixed effects only (insufficient within-firm variation for firm FE given the infrequent deal sample). For deal-type classification (diversifying, horizontal, or vertical), the paper uses a multinomial logit DID with the same treatment structure and vertical mergers as the reference category. Announcement returns (CAR) use the `event-study` approach: abnormal return on day $$d$$ is actual return minus the normal return predicted by a value-weighted CRSP index market model estimated over (-210, -11), with at least 120 non-missing daily returns required. The five-day window (-2, +2) around the announcement date is then summed to form CAR(-2,2).

## Empirical specifications

**Main DID (Tables 2, 4, 5, 7, 8; results R1, R2, R5, R6, R8, R9).** Eq. (1) above, estimated with firm FE + industry-time FE at the firm-quarter level, or with industry (SIC2) FE + time FE at the deal level. The seven-quarter pre-period is 2015Q1-2016Q3 and the seven-quarter pilot period is 2017Q1-2018Q3 (2016Q4 excluded as the event quarter). For the post-pilot reversal test (Table 8), the window shifts to compare the five pilot quarters 2017Q3-2018Q3 against the five post-pilot quarters 2018Q4-2019Q4, stopping at 2019Q4 to avoid COVID-19 contamination. Controls ($$X_{it}$$): log total assets (Firm size), leverage (sum of long-term and short-term debt over total assets), market-to-book (M/B), cash flow from operations over total assets, institutional ownership fraction. All continuous variables winsorized at the 1st and 99th percentiles.

**Parallel trends check (Table 2, col 5; Fig. 1, p. 10).** A dynamic lead-lag DID defines two-quarter indicator dummies in the pre-pilot period backward from the reference quarter (-1 = 2016Q3) and one- or two-quarter indicators in the pilot period. No significant pre-trend divergence between test groups 2-3 and the control group is found in M&A intensity before the pilot (Fig. 1); the statistically significant drop emerges in periods +3 and +4 after pilot initiation.

**Triple-DID channel tests (Table 3, p. 12; R3 and R4).** Same two-way fixed effects structure as Eq. (1), adding $$\text{TG2\&3}_i \times \text{Post}_t \times \text{Channel}_i$$ as the key coefficient (b1). Observations are restricted to test groups 2-3 and the control group; test group 1 interactions enter as separate controls. Both the asymmetry proxies (Panel A: three proxies) and the valuation proxies (Panel B: two proxies) yield significant b1 at the 5% level, with the combined b1+b2 significant at the 5% level or better and negative, supporting both channels of the information asymmetry and valuation hypothesis.

**Stock payment DID (Table 4, p. 13; R5).** Dependent variable is the percentage of deal value paid by acquirer stock. For deals with missing stock payment percentage but inferrable zero values (remaining payment components sum to 100%), the stock fraction is set to zero. Only industry and time fixed effects used at the deal level. Massa and Xu (2013) provides the benchmark for interpreting these results.

**Deal characteristics DID (Table 5, p. 14; R6).** Three separate OLS DIDs with Public target, Deal size (deal value over acquirer total assets), and Completed deal as dependent variables, each including industry FE and time FE. The increase in deal completion (+0.265**, t = 2.27, col 5) implies that conditional on undertaking M&A, pilot firms in test groups 2-3 are approximately 27 percentage points more likely to complete a deal than control firms.

**Multinomial logit deal types (Table 6, p. 16; R7).** Vertical merger (vertical relatedness coefficient between acquirer and target SIC-4 industries exceeding 1%, from the 2012 US BEA Input-Output tables) is the reference category. Coefficients on $$\text{TG2\&3}_i \times \text{Post}_t$$ are negative and significant for both diversifying (-1.887**, t = -2.08 in col 1; -2.014**, t = -2.17 with controls in col 3) and horizontal (-1.482*, t = -1.71 in col 2; -1.605*, t = -1.80 with controls in col 4) relative to vertical.

**Announcement return DID (Table 7, p. 17; R8).** The five-day CAR(-2,2) is regressed on the same DID structure as Eq. (1) at the deal level, controlling for deal size, all-cash indicator, diversifying deal indicator, and public target. The key coefficient $$\text{TG2\&3}_i \times \text{Post}_t = +0.023$$ (t = 2.04) represents a 2.3-percentage-point improvement in announcement returns for retained deals; this holds for non-stock-financed deals too (col 2, +0.030*, t = 2.04), consistent with the information asymmetry and valuation hypothesis rather than the trading cost-based hypothesis.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CRSP (via WRDS) | Daily stock returns, prices, market capitalization, bid-ask spreads; pilot stock identification and market model estimation | [WRDS / CRSP](/wiki/commercial/wrds/) |
| Compustat (via WRDS) | Total assets, leverage, market-to-book, cash flow, book equity; acquirer financial characteristics | [WRDS / Compustat](/wiki/commercial/wrds/) |
| Thomson Reuters Institutional Holdings (via WRDS) | Institutional ownership fraction (13F-based) | [WRDS](/wiki/commercial/wrds/) |
| I/B/E/S (via WRDS) | Analyst coverage count; earnings forecast dispersion (proxies for ex-ante information asymmetry in channel tests) | [WRDS](/wiki/commercial/wrds/) |
| SDC Platinum Mergers and Acquisitions Database | M&A deal identification, announcement dates, deal size, payment method (stock fraction), target type (public/private), deal completion status, deal type | no page yet |
| FINRA tick size pilot stock list | Assignment of CRSP stocks to control group and test groups 1, 2, 3 via stratified random sampling | no page yet |
| BEA 2012 US Input-Output Table | Vertical relatedness coefficients between any two 4-digit SIC industries; used to classify acquisitions as vertical, horizontal, or diversifying | no page yet |

Sample: seven pre-pilot quarters (2015Q1-2016Q3) and seven post-pilot quarters (2017Q1-2018Q3); 211 unique acquirer firms with at least one M&A deal in the pre-pilot period (39 in test group 1, 34 in test group 2, 36 in test group 3, 102 in control); 2929 firm-quarter observations and 451 M&A deals (deal-level sample).

## When to read the full paper

Read the [original](https://doi.org/10.1016/j.jcorpfin.2025.102890) if you are: examining the three information asymmetry proxies and two valuation proxies in the triple-DID channel tests (Table 3); studying the full set of deal structure and type shifts (Tables 5-6); or tracing the parallel trends plots quarter by quarter (Fig. 1). Table 8 documents the partial post-pilot reversal and is useful for understanding recovery dynamics after a liquidity shock.

Unlike Ye et al. (2023), who study how the same 2016 pilot affected the sensitivity of internal investment (Tobin's Q) to stock prices through managerial learning, this paper focuses on external M&A investment; the authors note that, taken together with Ye et al. (2023), the evidence speaks to a potential bright side of a larger tick size.

## Attribution and rights

Source: peer-reviewed, *Journal of Corporate Finance* 96 (2026), article 102890. Published by Elsevier B.V. This distillation was extracted by an LLM on 2026-06-26 and is **not human-verified or independently reproduced**. The article is paywalled; no CC licence is present. Extract-only redistribution.

> Lin, Chen, Wentao Yao, and Hong Zou. "The real effects of tick-size adjustments: Evidence from the 2016 tick-size pilot." *Journal of Corporate Finance* 96 (2026): 102890. DOI: 10.1016/j.jcorpfin.2025.102890. Published by Elsevier B.V. All rights reserved.


==============================================================================
# Deposit Insurance and LLP Discretion: Pugachev, Robin, Wang & Yang (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jcf/2026/pugachev-deposit-insurance-loan-loss-provisioning-2026/
# Distilled: The 2008 EESA expansion of US deposit insurance from $100,000 to $250,000 caused affected banks to provision more conservatively, increasing discretionary loan loss provision by approximately 3.4 basis points of lagged loans (38% of the mean LLP level), with effects concentrated at banks that increased risk most and faced the most regulatory scrutiny. Journal of Corporate Finance vol. 99, 2026, paywalled. Seven core results with source locators, the LLP prediction model, and the DiD specifications. LLM-distilled; not human-verified.
# Tags: paper-summary, banking, deposit-insurance, accounting-conservatism
==============================================================================

**What this is.** The paper's core results, the hypotheses it tests, the LLP prediction model used to measure accounting discretion, and the difference-in-differences specifications: enough to know what it found and how, without reading all 24 pages. To replicate or extend it, read the full source at [doi.org/10.1016/j.jcorpfin.2026.102995](https://doi.org/10.1016/j.jcorpfin.2026.102995).

## TL;DR

The paper studies whether the Emergency Economic Stabilization Act (EESA) of 2008, which increased the FDIC deposit insurance ceiling from $100,000 to $250,000, changed bank accounting behavior. It uses a difference-in-differences design, exploiting the fact that 126 Massachusetts state-chartered savings banks and cooperatives were already fully insured by private state deposit insurance (DIF/SIF) and thus experienced no change in coverage. Relative to these controls, banks exposed to the EESA shock shift toward more conservative (income- and capital-reducing) discretionary loan loss provisioning. The effect is approximately 3.4 basis points of lagged loans (38% of the mean LLP level). Effects are strongest for banks that increased risk most and for those subject to the greatest regulatory scrutiny, consistent with two channels: DI-induced risk-taking prompting demand for conservative reporting, and heightened regulator incentives once the insurer's exposure rises.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%. All regressions include bank and quarter fixed effects with standard errors clustered by bank. DLLP is the discretionary component of LLP scaled to units where each unit = 10 basis points of lagged loan portfolio (LLP multiplied by 1,000).

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | DI coverage fraction (INSDEP) positively predicts conservative LLP in broad 75-quarter panel | Table 4, Col. 1, p. 10 | INSDEP = 1.880\*\*\* (t = 7.857); n = 123,310 bank-quarters; result holds excluding crisis years and in post-EESA subsample |
| R2 | Treated banks shift to more conservative DLLP relative to controls after EESA (main DiD result) | Table 5, Col. 1, p. 11 | TREATPOST = 0.336\*\*\* (t = 2.769); equivalent to ~3.4 bps of lagged loans; 38% of the mean LLP level of 8.88 bps |
| R3 | Banks with highest fraction of newly insured deposits shift most toward conservatism (intensive margin) | Table 7, Col. 1, p. 13; Fig. 4, p. 13 | Q5NIDEPPOST = 0.577\*\* (t = 2.323); DLLP change is monotone across DI-exposure quintiles |
| R4 | Banks that increase nonperforming loans most post-EESA shift most toward conservatism (risk channel) | Table 7, Col. 3, p. 13 | Q5NPLPOST = 2.442\*\*\* (t = 6.878); approximately seven times the full-sample baseline coefficient |
| R5 | Banks whose z-score falls most post-EESA shift most toward conservatism (risk channel) | Table 7, Col. 2, p. 13 | Q1ZSCOREPOST = 1.232\*\*\* (t = 4.400); approximately four times the full-sample baseline coefficient |
| R6 | Least-capitalized banks shift most toward conservatism (regulatory scrutiny channel) | Table 7, Col. 4, p. 13 | Q1T1CAPITALPOST = 0.635\*\* (t = 2.028); consistent with regulators focusing on banks closest to the default boundary |
| R7 | Among treated banks, only those provisioning opportunistically pre-EESA shift toward conservatism; pre-conservative banks do not shift or shift back | Table 6, Cols. 1 and 4, p. 12 | TREATPOSTDLLP<0 = 0.440\*\*\* (t = 2.728); POSTHPRE\_DLLP = -0.545\*\* (t = -2.413); consistent with strategic constraints on banks already at or past their conservatism target |

**Overall (paper's conclusion).** DI expansion increases conservative LLP discretion through two non-mutually exclusive channels: the risk-taking channel (DI-induced moral hazard leads banks to take more risk, prompting creditors and regulators to demand conservative accounting) and the regulatory scrutiny channel (the insurer's larger exposure heightens diligence). The finding contrasts with Huang (2021), who shows that deregulation reduced accounting conservatism among public bank borrowers; here the regulatory tightening from expanded DI moves bank provisioning in the opposite direction. From a policy perspective, this suggests US bank regulators provide sufficient monitoring to limit the moral hazard problem associated with DI, because the risk-taking behavior induced by DI is mitigated through the accounting channel.

## Theory / model

The paper has no formal structural model. It develops two hypotheses and an identification strategy.

**H1** (p. 4): Banks that experience greater DI coverage adopt more conservative LLP discretion relative to other banks. The prediction follows from two mechanisms that DI activates simultaneously:

1. *Risk channel*: DI reduces depositor monitoring incentives, allowing banks to take more risk, as Calomiris and Jaremski (2019) document across US banking history. Riskier banks face greater demand from creditors and equity-holders for conservative accounting (loss recognition that lowers reported income and erodes capital), as established by Beatty and Liao (2014), Kim et al. (2013), and Balakhrishnan et al. (2016).

2. *Regulatory scrutiny channel*: DI shifts the monitoring role from depositors to bank regulators (the FDIC). Regulators are known to prefer conservative accounting (Qiang (2007), Lobo and Zhou (2006)), and DI expansion increases their incentive to enforce conservative practices as their exposure grows.

**Identification strategy.** The paper exploits the EESA of 2008, which increased the FDIC deposit insurance ceiling from $100,000 to $250,000 per account. Diamond and Dybvig (1983) established that DI stabilizes banking systems by converting deposits to a risk-free asset; the EESA shock re-prices the coverage for roughly all US banks. The key identification assumption is that Massachusetts state-chartered savings banks and cooperatives are unaffected by EESA: their deposits were already fully insured by private state schemes (DIF/SIF) initiated in the 1930s. Assignment to treated vs. control status is thus pre-determined by a decision to incorporate in Massachusetts as a state-chartered savings or cooperative institution, typically 200 years before EESA (p. 2), satisfying the exogeneity requirement for DiD.

To improve comparability, the paper applies propensity score matching (PSM) on 22 pre-shock bank characteristics, selecting treated banks that resemble controls on observable traits. Pre-shock parallel trends in DLLP are documented in Figure 1 (p. 8), supporting DiD validity. A battery of placebo tests and three alternative control samples (non-Massachusetts state-chartered savings banks; Massachusetts federally chartered savings banks; Massachusetts state-chartered commercial banks) confirm robustness.

## Method

The method has two steps: constructing the discretionary LLP measure, then running the DiD.

**Step 1: LLP discretion (Eq. 1, p. 4).** Following Nicoletti (2018) with bank fixed effects added, the paper estimates a predicted LLP for each bank-quarter from its loan portfolio fundamentals:

$$
\text{LLP}_{b,t} = \alpha_1 \text{DNPL}_{b,t+1} + \alpha_2 \text{DNPL}_{b,t} + \alpha_3 \text{DNPL}_{b,t-1} + \alpha_4 \text{DNPL}_{b,t-2} + \alpha_5 \text{EBLLP}_{b,t}
$$

$$
+ \alpha_6 \text{TIER1}_{b,t-1} + \alpha_7 \text{LSIZE}_{b,t-1} + \alpha_8 \text{DLOAN}_{b,t} + \mu_b + \tau_t + \varepsilon_{b,t} \tag{1}
$$

where LLP is scaled to 1,000 bps of lagged loans; subscripts $$b$$ and $$t$$ index bank and quarter; $$\text{DNPL}$$ captures changes in nonperforming loans; $$\text{EBLLP}$$ is earnings before LLP and taxes; $$\text{TIER1}$$ is the Tier 1 capital ratio; $$\text{LSIZE}$$ is log assets; $$\text{DLOAN}$$ is loan growth; $$\mu_b$$ are bank fixed effects; $$\tau_t$$ are quarter fixed effects. Appendix B (p. 22) reports coefficient estimates. The residuals from Eq. (1), denoted $$\text{DLLP}_{b,t}$$, measure discretionary LLP: positive (negative) values indicate more conservative (opportunistic) provisioning than fundamentals predict.

**Step 2: Broad-sample panel regression (Eq. 2, p. 8).** To establish the motivating association before the DiD, the paper estimates:

$$
\text{DLLP}_{b,t} = \beta_1 \text{INSDEP}_{b,t} + \gamma' \text{Controls} + \mu_b + \tau_t + \varepsilon_{b,t} \tag{2}
$$

where $$\text{INSDEP}_{b,t}$$ is the fraction of bank $$b$$'s deposits below the FDIC insurance limit. $$\beta_1$$ measures the contemporaneous relationship between DI coverage and LLP conservatism over the full 75-quarter broad panel (n = 123,310 bank-quarters). Because $$\text{DLLP}$$ is a residual used as a dependent variable, the paper follows Chen et al. (2018) and includes all first-stage controls in the second-stage regression to mitigate bias from using residuals as dependent variables.

## Empirical specifications

**Main DiD specification.** In Eq. (2), $$\text{INSDEP}$$ is replaced by the treatment indicator:

$$
\text{DLLP}_{b,t} = \beta_1 \text{TREATPOST}_{b,t} + \gamma' \text{Controls} + \mu_b + \tau_t + \varepsilon_{b,t} \tag{3}
$$

where $$\text{TREATPOST}_{b,t} = 1$$ for all banks in the PSM-matched treated sample in every quarter after 3Q2008 (the EESA passage date). Controls include past, present, and future changes in nonperforming loans ($$\text{DNPL}$$), earnings before LLP ($$\text{EBLLP}$$), Tier 1 capital ($$\text{TIER1}$$), log assets ($$\text{LSIZE}$$), and loan growth ($$\text{DLOAN}$$). All continuous variables are winsorized at 1% tails. Standard errors are clustered by bank. Bank and quarter fixed effects are included in all specifications.

Sample period: 4Q2005-4Q2011 (12 quarters before and 12 quarters after EESA). Final PSM sample: 9,568 bank-quarters (314 treated, 74 control banks). Observations require both periods to exist (no survivorship-biased exits), total assets above $25 million, no quarterly asset growth above 10% (to exclude acquisition-driven changes), and at most 25% missing values.

**Intensive-margin specification.** To test whether effects scale with DI exposure, $$\text{TREATPOST}$$ is replaced by $$\text{Q5NIDEPPOST}$$, an indicator for banks in the highest quintile of the fraction of 3Q2008 deposits in the $100K-$250K range (the newly insured band). Following Lambert et al. (2017), the paper estimates newly insured deposits as the value of 3Q2009 deposits above the $250K ceiling minus the value of 3Q2008 deposits above the $100K ceiling, scaled by 3Q2008 total deposits. Table 7 Column 1 (p. 13) reports results from the treated-bank sample only; control banks are replaced by lower-exposure treated banks.

**Channel specifications.** To decompose mechanisms, $$\text{TREATPOST}$$ is interacted with quintile indicators for four cross-sectional variables measured pre-to-post EESA: change in z-score ($$\text{Q1ZSCOREPOST}$$), change in NPL ($$\text{Q5NPLPOST}$$), Tier 1 capital pre-EESA ($$\text{Q1T1CAPITALPOST}$$), and bank size pre-EESA ($$\text{Q5SIZEPOST}$$). These indicators equal one for banks in the highest risk-change or regulatory-scrutiny quintile. Table 7 Columns 1-5 (p. 13) report each interaction separately; Column 6 includes all interactions simultaneously.

**Pre-shock heterogeneity.** Table 6 (p. 12) adds $$\text{TREATPOSTDLLP<0}$$, which interacts $$\text{TREATPOST}$$ with an indicator for banks with negative average pre-EESA DLLP (opportunistic provisioners). This tests whether the shift toward conservatism is concentrated in banks that had room to become more conservative.

**Robustness.** Three alternative DLLP constructions (Kanagaretnam et al. 2010; Bushman and Williams 2012; Basu et al. 2020) and three alternative control samples yield similar $$\beta_1$$ estimates (Table 8, p. 15-16). Placebo tests using randomly assigned treatment or placebo event windows yield insignificant coefficients. The crisis period (2007-2010) is also excluded in one robustness column to address the concern raised by Huizinga and Laeven (2012) that LLP discretion systematically rises during financial crises; the TREATPOST coefficient remains similar in the non-crisis subsample.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| FDIC Statistics on Depository Institutions (SDI) | Main bank financial data: LLP, nonperforming loans, earnings, Tier 1 capital, total assets, loan balances, deposit composition (core, large, demand deposits), securities, write-offs; also TAG participation | no page yet |
| Census Bureau | County-level unemployment rate (UNEMP) for bank's main office county; matching covariate | no page yet |
| Federal Housing Finance Agency (FHFA) | County-level housing price index (LHPI) for bank's main office county; matching covariate | [FHFA House Price Index](/wiki/datasets/fhfa-hpi/) |
| U.S. Treasury | TARP Capital Purchase Program participation indicator; matching covariate | no page yet |

Sample: 4Q2005-4Q2011 (PSM baseline). Broad panel for motivating regressions: 1999-2018 (75 quarters, 123,310 observations). Data availability statement: "Data are available from the public sources cited in the text" (p. 22).

## When to read the full paper

Use the [original](https://doi.org/10.1016/j.jcorpfin.2026.102995) if you are: studying how banking regulation affects accounting discretion; designing DiD studies using EESA as a natural experiment for DI-related hypotheses; comparing the Nicoletti (2018) and Beatty and Liao (2014) LLP discretion constructions; or investigating whether regulatory intervention can attenuate moral hazard from deposit insurance. The exact PSM procedure (22 matching variables, Table 2, p. 7) and the placebo designs (Table 8, p. 15-16) are important for replication.

## Attribution and rights

Source: peer-reviewed, *Journal of Corporate Finance* 99, 2026, article 102995. This distillation was extracted by an LLM on 2026-06-26 and is **not human-verified or independently reproduced**. The paper is paywalled (© 2026 Elsevier B.V. All rights reserved); only textual extract is provided here.

> Pugachev, Leo, Ashok Robin, Dilin Wang, and Rong Yang. "Deposit insurance and
> discretion in loan loss provisioning." *Journal of Corporate Finance* 99 (2026): 102995.
> DOI: [10.1016/j.jcorpfin.2026.102995](https://doi.org/10.1016/j.jcorpfin.2026.102995).
> © 2026 Elsevier B.V. All rights reserved. Extract-only; not reproduced.


==============================================================================
# Insider Trading with Options: Vacca (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jcf/2026/vacca-insider-trading-options-employees-2026/
# Distilled: Using Finnish securities registry data (1995-2014), Vacca (2026) documents that rank-and-file employees' open-market purchases of own-company call options predict weekly excess stock returns of approximately 60 basis points, peaking before earnings announcements and extending to supply-chain partners. Journal of Corporate Finance 98 (2026) 102963, CC BY 4.0. Seven core results with source locators, datasets used, and the identification strategy.
# Tags: paper-summary, insider-trading, options-trading, information-asymmetry
==============================================================================

**What this is.** The paper's core results, the datasets, and the identification
strategy: enough to assess the scope of informed employee option trading and its
channels, without reading all 21 pages. To replicate or extend, read the full
source at the [original](https://doi.org/10.1016/j.jcorpfin.2026.102963).

## TL;DR

Using daily securities registry data from Euroclear Finland (January 1995 to
December 2014), Vacca (2026) documents that open-market purchases of own-company
call options by employees predict positive subsequent stock returns on short
horizons. The average market-adjusted weekly (five-day) return after an employee
buys own-company options is 64 basis points, compared with approximately zero for
purchases of options written on unrelated firms. Rank-and-file employees (below
the level of manager or primary insider) account for the vast majority of
own-company option purchases and their trades are the most informative, with a
spread of 81 basis points versus unrelated firm purchases. Consistent with the
argument of Black (1975) that informed investors prefer options for their embedded
leverage, employees trade own-company options at five to eight times the rate they
trade own-company stocks. The predictability concentrates before earnings
announcements (approximately 150 basis points in the preceding week), extends to
Nokia supply-chain partner stocks, and disappears entirely for former employees
(a key falsification). The paper also uncovers a tipping channel: anonymous
retail accounts that co-trade with employees earn similarly informed returns.

## Core results

All returns are market-adjusted weekly (five-day) returns, reported multiplied by
100 so that 0.64 = 64 basis points. Standard errors are clustered at the
stock-trade date level. Locators point to the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Own-company call option purchases by employees predict positive weekly stock returns; purchases of options on unrelated firms do not | Table 2, Panel A, p. 7 | Own-company avg = 0.64% (N=4,091); unrelated firm avg = 0.04% (N=3,250); difference = 0.59\*\*\*, p=0.000 |
| R2 | Rank-and-file employees drive the predictability; their own-company trades are more informative than those of higher-ranked employees | Table 2, Panel B, p. 7; Fig. 2, p. 7 | Rank-and-file own-company avg = 0.71% (N=3,413); unrelated = -0.10% (N=2,587); difference = 0.81\*\*\*, p=0.000 |
| R3 | Own-company option purchases in the week before earnings announcements are more informative than purchases at other times | Fig. 4, p. 9; text p. 8 | ~150 bps (annualized ~120%) for buys in window [-5,-1] before announcement; near zero in windows far from announcement date; 281 of 332 pre-announcement trades are by rank-and-file employees |
| R4 | Former employees' option purchases do not predict positive returns (falsification of the information-advantage story) | Table 3, Panel A, p. 13 | Former employees avg = -0.11% (N=2,275); current employees avg = 0.64% (N=4,091); difference = 0.76\*\*\*, p=0.000 |
| R5 | Employees at Nokia customer and supplier firms earn informed returns by trading Nokia options (supply-chain channel) | Table 4, Panel A, p. 14 | Nokia cluster employees avg = 0.58% (N=1,260); employees at other firms = 0.15% (N=1,767); difference = 0.43\*\*, p=0.019 |
| R6 | Tipping identified: anonymous retail accounts that co-trade with employees are also informed | Table 5, Panels B-C, p. 15 | Correlated (employee + tippee same day) avg = 0.83\*\*\* (N=3,187); tippee-only avg = 0.59\*\*\* (N=6,421); 783 anonymous informed accounts detected |
| R7 | Employee own-company option activity is associated with contemporaneous increases in retail option buying | Table 6, col. (1)-(2), p. 17 | Employee activity indicator coefficient = 10.88\*\* (t=2.26) on daily retail buy count (stock + day FE); 10.01\*\* (t=2.16) with option + day FE; positive also on log retail buy volume |

**Overall (paper's conclusion).** Between 3% and 9% of all retail demand in the
Finnish market for single-name equity derivatives can be attributed to employees
who likely have an information advantage; accounting for tipping raises this to
over 15% of retail investors and 10% of open-market option purchases. Rank-and-file
employees, not primary insiders or managers, are the primary source of informed
retail option trading. The evidence points to a disclosure gap: most informed
trading by employees goes undetected because rank-and-file employees face no
mandatory disclosure requirements.

## Theory / model

The paper has no formal theoretical model. It tests an information-advantage
hypothesis using the following chain of predictions.

**Theoretical foundation.** Black (1975) argues that informed investors prefer
options over stocks because of embedded leverage, which amplifies the return to
private information for a given capital outlay. If rank-and-file employees hold
private, price-relevant information about their employer, they should trade
own-company options at a higher rate relative to stocks than uninformed investors
do. The data confirm this: employees' share of Finnish retail option demand is
5 to 8 times their share of retail stock demand (Fig. 7, p. 12).

**Employee information.** Green, Huang, Wen and Zhou (2019) show that crowdsourced
employee reviews contain price-relevant information about firms, establishing that
employees hold private knowledge that reaches beyond their immediate job function.
Augustin, Brenner and Subrahmanyam (2019) document that informed options trading
occurs before corporate events such as takeovers. This paper extends both lines of
evidence by examining whether rank-and-file employees also trade on private
information through option purchases, not only around extraordinary events but
also as a recurring pattern around earnings announcements.

**Identification hypotheses.** The paper tests the information-advantage story
against the following predictions:

1. Own-company option purchases by current employees should predict positive
   short-horizon stock returns; purchases on unrelated stocks should not (R1, R2).
2. Predictability should peak immediately before earnings announcements, when
   private information about upcoming news is most valuable (R3).
3. Information advantage derives from active employment; former employees should
   lose access to current firm-specific information, so their trades should not
   predict returns (R4, falsification).
4. Economic linkages can transmit inside information to supply-chain partners.
   Prior studies find that primary insiders trade derivatives of supply-chain
   partners to circumvent own-company restrictions (Deuskar, Khatri and
   Subrahmanyam 2025). Rank-and-file employees at Nokia's customer and supplier
   firms should also exploit this channel (R5).
5. Employees may share private information with family members or acquaintances
   (tipping); correlated trading between employee and anonymous accounts should
   also be informed (R6).

Following Pan and Poteshman (2006), who link signed option demand to subsequent
stock price movements, this paper uses short-horizon market-adjusted returns as
the primary measure of option-trade informativeness.

## Method

The paper uses three main empirical tools.

**Return measurement.** For each employee trade on stock $j$ at trade date $t$,
the raw return over event window $\tau$ (expressed in trading days) is (p. 5):

$$
\text{Return}_{j,t} = \frac{P_{j,t+\tau} - P_{j,t}}{P_{j,t}} \tag{1}
$$

Market-adjusted returns subtract the contemporaneous market return, following the
approach of Brown and Warner (1985). Standard errors are clustered at the
stock-trade date level throughout; main results are robust to clustering by
stock-month or two-way clustering by stock and trade date.

**Tipping identification.** The procedure (Appendix C) identifies pairs $$(a, i)$$
where $$a$$ is an anonymous retail account and $$i$$ is a current employee at firm
$$j$$, such that $$a$$ buys call options on $$j$$ on the same day as $$i$$, with this
co-occurrence repeated at least $$k = 2$$ times and constituting a significant
fraction ($$p = 0.1$$) of $$a$$'s total own-company option activity during $$i$$'s
employment tenure. Very active traders are excluded to avoid false positives.

**Logit for option purchase decision.** The determinants of own-company option
purchasing are estimated by a logit model (Table 7, p. 18) with binary outcome
equal to 1 if employee $$i$$ at firm $$j$$ purchases at least one own-company call
option in month $$m$$. Regressors cover four domains: risk preferences (Female,
recent large portfolio losses/gains), probability of detection (Primary insider
status, which requires mandatory disclosure), habit (number of own-company option
buys in the prior year), and financial market familiarity (prior stock and option
trading activity, portfolio size). Standard errors are two-way clustered at the
employee and firm-month level.

## Empirical specifications

**Main return comparisons (R1, R2, R4).** For each employee-level option purchase,
the paper computes the one-week market-adjusted return on the underlying stock.
The primary comparison is between own-company and unrelated firm option purchases
by the same pool of employees (Table 2), tested via a t-test with standard errors
clustered at the stock-trade date level. The comparison is repeated separately for
rank-and-file employees vs. higher-ranked employees (Fig. 2) and for current vs.
former employees (Table 3). All returns multiplied by 100.

**Earnings-window event study (R3).** Employee option purchases are sorted by
calendar distance to the next earnings announcement date. Average market-adjusted
weekly returns following purchases in each window ([-21,-6], [-5,-1], [0], [1,5],
[6,21] relative to the announcement) are compared between own-company purchases and
purchases on other firms (Fig. 4, p. 9). Purchases in the five-day pre-announcement
window are further stratified by the sign and magnitude of standardized unexpected
earnings (SUE = realized EPS minus EPS from four quarters ago, scaled by the
eight-quarter rolling standard deviation) to separate event-timing from
earnings-surprise components (Fig. 5, p. 10).

**Supply-chain analysis (R5).** For the Nokia cluster, the sample is restricted to
purchases of Nokia call options by non-Nokia employees at identified customer and
supplier firms (111 employees, 1,260 purchases). Average weekly returns are compared
against returns after Nokia option purchases by employees at non-Nokia cluster firms
(Table 4, p. 14). Appendix Table F6 adds firm-year fixed effects to control for the
amount of private information available at a given firm in a given year.

**Retail response regression (R7).** The contemporaneous link between employee
activity and broader retail demand is estimated by (Eq. 1, p. 16):

$$
Y_{o,t} = \alpha + \beta X_{o,t} + \gamma_s + \delta_t + \varepsilon_{o,t} \tag{2}
$$

where $$o$$ indexes an option written on underlying stock $$s$$ on trading day $$t$$;
$$Y_{o,t}$$ is the daily retail buy count (or log retail buy volume) excluding
own-company employee trades; $$X_{o,t}$$ is the Employee activity indicator, equal to
1 if at least one own-company option purchase occurs on day $$t$$; $$\gamma_s$$ is
an underlying stock fixed effect (or option fixed effect in column 2); and
$$\delta_t$$ is a day fixed effect. Standard errors are two-way clustered at the
underlying stock and month level (Table 6, N = 106,519 option-day observations).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Euroclear Finland (securities registry) | Daily records of all Finnish securities holdings and changes, Jan 1995-Dec 2014; source of all option and stock trades; granular trade-type identifier separates open-market purchases from other transaction types | No page yet |
| Alexander Incentives (executive compensation data) | Information on hundreds of employee and executive stock option plans issued by Finnish firms; provides employment-relationship identification for over 40,000 individuals; used to classify each individual as primary insider, manager, or rank-and-file employee | No page yet |

Sample: January 1995 to December 2014; 43 Finnish firms with employee option
trading observed; 890 current employees making 4,091 own-company call option
purchases. Nokia supply-chain sub-sample: 111 employees at 7 Nokia customer and
supplier firms making 1,260 Nokia option purchases.

## When to read the full paper

Use the [original](https://doi.org/10.1016/j.jcorpfin.2026.102963) if you are:
studying the legal and institutional setting for employee options trading in Finland
(Section 1.1 and Appendix A); examining robustness by derivative type (listed
options vs. warrants, Appendix D) or by employer firm (Appendix E); seeking the
detail of the tipping-identification algorithm (Appendix C); extending the analysis
to non-earnings information events (Section 2.3, Fig. 6); or reviewing the logit
analysis of determinants of option purchasing (Table 7, Section 6).

## Attribution and rights

Source: peer-reviewed, *Journal of Corporate Finance* 98 (2026) 102963. This
distillation was extracted by an LLM on 2026-06-26 and is **not human-verified
or independently reproduced**. The CC BY 4.0 licence permits mirroring; the
verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Vacca, Matteo.
> "Insider Trading with Options: Evidence from Rank-and-File Employees."
> *Journal of Corporate Finance* 98 (2026) 102963.
> DOI: 10.1016/j.jcorpfin.2026.102963. (C) 2026 The Author.
> Published by Elsevier B.V. Licensed under
> [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Deep Learning, Predictability, and Optimal Portfolio Returns: Babiak & Barunik (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jef/2026/babiak-deep-learning-optimal-portfolio-returns-2026/
# Distilled: Deep feedforward and LSTM recurrent neural networks deliver economically significant gains in certainty-equivalent returns and Sharpe ratios over linear predictive regressions for a two-asset optimal US equity portfolio. Journal of Empirical Finance 2026, paywalled. Six core results with source locators, datasets used, the investor model, and the neural network method with its defining equations.
# Tags: paper-summary, asset-pricing, return-predictability, machine-learning, neural-networks
==============================================================================

**What this is.** The paper's core results, the investor model, and the neural network forecasting method with its defining equations: enough to know what was found and how, without reading the full 22 pages. To replicate or extend, read the source at the [original](https://doi.org/10.1016/j.jempfin.2026.101705).

## TL;DR

The paper asks whether deep neural networks can generate economically meaningful portfolio gains for a long-horizon investor allocating between the US stock market (S&P 500) and a risk-free Treasury bill. Using 12 monthly macroeconomic and financial predictor variables from Welch and Goyal (2008), the paper trains feedforward networks (NN1, NN2, NN3) and an LSTM recurrent network and compares their out-of-sample portfolio performance to linear OLS benchmarks and the no-predictability expectations hypothesis (EH), rebalancing quarterly or annually over February 1955 to December 2018. All deep learning architectures achieve positive out-of-sample R-squared statistics where OLS models uniformly fail, and deliver certainty-equivalent returns and Sharpe ratios roughly double and triple those of the EH benchmark. The LSTM's gated memory cells, which capture long-range temporal dependence in predictor variables, provide additional gains during NBER recessions and with more frequent rebalancing. Earlier empirical papers by Feng et al. (2018) and Gu et al. (2020) established that deep learning improves statistical predictions for stock returns; this paper demonstrates the translation into economically significant portfolio performance gains.

## Core results

Magnitudes are as reported. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Deep NNs achieve positive OOS R-squared; all OLS models fail to beat the historical mean | Table 1, Panel A, p.8 | LSTM OOS-R2-q = 1.6% (p=0.002); NN1 = 7.1%, NN2 = 5.1%, NN3 = 5.6% (all p<0.01); OLS1 = -2.5%, OLS2 = -3.6%, OLS3 = -8.0%, OLS4 = -2.5% |
| R2 | LSTM roughly doubles the investor's CER and triples the Sharpe ratio vs no-predictability benchmark (quarterly rebalancing, whole sample) | Table 2, Panel A, p.10 | LSTM CER = 10.007% (p=0.000, SR = 0.175); EH CER = 4.737% (SR = 0.049); NN1 CER = 7.295%, NN2 CER = 6.984% |
| R3 | Deep NN gains are especially large during NBER recessions: LSTM CER nearly 9x higher than EH in downturns | Table 2, Panel A, p.10 | LSTM recession CER = 26.770% (p=0.000, SR = 0.358); EH recession CER = 3.311% (SR = -0.193); LSTM expansion CER = 7.998% |
| R4 | Annual rebalancing: NNs still outperform but gains are smaller with less frequent portfolio revision | Table 2, Panel B, p.10 | LSTM CER = 5.622% (p=0.012, SR = 0.118); EH CER = 4.542% (SR = 0.048); NN2 CER = 6.879% |
| R5 | Results robust to transaction costs: LSTM CER exceeds 7.9% even at high TC | Table 7, p.17 | Low TC (tau = 0.1%): LSTM CER = 9.592% (p=0.000, SR = 0.169); High TC (tau = 0.5%): LSTM CER = 7.910% (p=0.000, SR = 0.145) |
| R6 | Results robust to borrowing and short-selling constraints: NNs maintain large CER advantage | Table 8, Panel C, p.18 | LSTM with both constraints: CER = 7.775% (p=0.000, SR = 0.150) vs EH CER = 4.737% (SR = 0.049) |

**Overall (paper's conclusion).** Deep neural networks, particularly the LSTM recurrent architecture, deliver economically significant and statistically significant portfolio gains over the full 1955-2018 sample and across business-cycle subperiods. The gains survive transaction costs, borrowing and short-selling constraints, alternative rebalancing horizons, and a post-1969 20-year rolling window robustness check. The non-parametric capture of time-series structure in standard predictor variables from Welch and Goyal (2008) is the primary source of improvement over linear models, without any explicit modeling of conditional volatility.

## Theory / model

The paper formulates a standard power utility portfolio choice problem; it proposes no formal asset-pricing model and takes the 12 predictor variables as given. A representative investor with a T-period horizon maximizes expected utility by choosing an allocation $$\omega$$ between the S&P 500 index and a risk-free Treasury bill.

The power utility function is (p.6):

$$
U(r_{p,t+T}) = \frac{r_{p,t+T}^{1-\gamma}}{1-\gamma}, \qquad \gamma = 4 \tag{U}
$$

where $$r_{p,t+T}$$ is the cumulative portfolio return and $$\gamma$$ is the coefficient of relative risk aversion, calibrated to 4 following Johannes et al. (2014). The optimization problem is (Eq. 7, p.5):

$$
\max_\omega \mathbb{E}_t\!\left[U(r_{p,t+T})\right] \tag{7}
$$

subject to the cumulative portfolio return (Eq. 8, p.5):

$$
r_{p,t+T} = \prod_{\tau=1}^{T}\!\left[(1-\omega_{t+\tau-1})\exp\!\left(r^f_{t+\tau}\right) + \omega_{t+\tau-1}\exp\!\left(r^f_{t+\tau} + r_{t+\tau}\right)\right] \tag{8}
$$

where $$r^f_{t+\tau}$$ is the risk-free rate and $$r_{t+\tau}$$ is the excess equity log return. Two horizon settings: T = 6 months with quarterly rebalancing, and T = 24 months with annual rebalancing. Weights are bounded to $$-1 \leq \omega_{t,\tau} \leq 2$$.

Each predictor variable $$x^i_t$$ follows an AR(1) process (Eq. 12, p.6):

$$
x^i_{t+\tau} = \alpha^{x^i} + \beta^{x^i} x^i_{t+\tau-1} + \varepsilon^{x^i}_{t+\tau} \tag{12}
$$

The joint forecast errors $$\varepsilon_t = (\varepsilon^r_t, \varepsilon^x_t)$$ are drawn from a multivariate normal with estimated covariance $$\hat{\Sigma}_t$$, enabling simulation of future return paths for utility evaluation.

## Method

**Linear benchmark.** The standard predictive regression (Eq. 1, p.3):

$$
r_{t+1} = \alpha + \beta x_t + \varepsilon^r_{t+1} \tag{1}
$$

where $$x_t = (x^1_t, \ldots, x^n_t)$$ is the vector of 12 Welch and Goyal (2008) predictors. OLS1-OLS4 vary by predictor set (single dividend yield vs. kitchen-sink) and estimation window (expanding vs. rolling 10-year).

**Deep feedforward neural network.** The NN approximation replaces the linear link function (Eq. 2-3, p.4):

$$
r_{t+1} = \hat{f}_{W,b}(x_t) + \varepsilon^r_{t+1} \tag{2}
$$

$$
\hat{r}_{t+1} = f^{(L)}_{W^{(L)},b^{(L)}} \circ \cdots \circ f^{(1)}_{W^{(1)},b^{(1)}}(x_t) \tag{3}
$$

where each layer $$f^{(\ell)}$$ applies a sigmoid or ReLU activation to an affine transformation. With L = 2 and linear activations, Eq. (3) reduces to OLS. Three feedforward architectures: NN1 (16 neurons, 1 hidden layer), NN2 (32-16, 2 layers), NN3 (32-16-8, 3 layers).

**LSTM recurrent network.** The LSTM builds on `lstm-recurrent-network` by adding gated temporal state. A recurrent network introduces serial dependence into Eq. (3) via a lagged hidden state (Eq. 4, p.4):

$$
h_t = f(W_h h_{t-1} + W_x x_t + b_0) \tag{4}
$$

The LSTM variant extends this with gated memory cells. The hidden state (output gate) is (Eq. 5, p.5):

$$
h_t = \sigma\!\left(W_h^{(o)} h_{t-1} + W_x^{(o)} x_t + b_0^{(o)}\right) \circ \tanh(c_t) \tag{5}
$$

The memory cell combines a forget gate and an input gate (Eq. 6, p.5):

$$
c_t = \sigma\!\left(W_h^{(g)} h_{t-1} + W_x^{(g)} x_t + b_0^{(g)}\right) \circ c_{t-1} + \sigma\!\left(W_h^{(i)} h_{t-1} + W_x^{(i)} x_t + b_0^{(i)}\right) \circ \tanh(k_t) \tag{6}
$$

where $$\sigma(\cdot)$$ is the sigmoid function and $$k_t$$ is the new information flow. The forget gate term $$\sigma(\cdot) c_{t-1}$$ enables long-range dependence across many lags, while the input gate $$\sigma(\cdot)\tanh(k_t)$$ incorporates new predictor information. The LSTM uses three recurrent layers with 32-16-8 neurons and LSTM cells in the last layer, building on `time-series-forecasting` by learning nonlinear time-series patterns in the Welch and Goyal (2008) predictors.

Estimation uses the Adam optimizer with weight decay regularization (Kingma and Ba 2014) and dropout (0-60%). Hyperparameters (learning rate, weight decay, dropout, activation) are selected from 100 random combinations on a rolling validation sample each quarter using a combined statistical and economic criterion.

## Empirical specifications

**Data and rolling estimation.** Predictors and returns span January 1945 to December 2018. Portfolio evaluation starts February 1955 (after an initial 10-year training window). All models are re-estimated on a rolling 10-year window each quarter (annual: each year), matching the investor's rebalancing timing.

**Statistical accuracy: OOS R-squared.** Following Campbell and Thompson (2008) (Eq. 14-15, p.7), the OOS R-squared for quarterly rebalancing is:

$$
R^2_{q,oos} = 1 - \frac{\displaystyle\sum_{t \in \mathcal{T}_q}(r_t - \hat{r}_t^{\mathcal{M}_s})^2}{\displaystyle\sum_{t \in \mathcal{T}_q}(r_t - \bar{r}_t)^2} \tag{15}
$$

where $$\mathcal{T}_q$$ is the set of end-of-quarter months when the investor reoptimizes and $$\bar{r}_t$$ is the historical mean over the same sample. The Clark-West (2007) test evaluates $$H_0: R^2_{oos} \leq 0$$ (Table 1, p.8). All NNs achieve positive OOS R-squared; all OLS models are negative (R1).

**Portfolio performance: CER and Sharpe ratio.** Expected utility is computed by simulating 1,000,000 draws from $$(\varepsilon_{r,t+\tau}, \varepsilon^x_{t+\tau}) \sim N(0, \hat{\Sigma}_t)$$, iterating Eqs. (10)-(13) forward. The annualized certainty-equivalent return CER solves $$U(1 + \text{CER}/100) = \mathbb{E}_t[U(r_{p,t+T})]$$. Statistical significance for CER differences from EH is assessed by the one-sided Diebold and Mariano (2002) test (Table 2, p.10). Sharpe ratios are computed monthly.

**Robustness checks.** Transaction costs: $$\tilde{r}^{\mathcal{M}_s}_t = r^{\mathcal{M}_s}_t - \tau|\omega_t - \omega_{t-1} \hat{r}^{\mathcal{M}_s}_{t-1}|$$ with $$\tau \in \{0.1\%, 0.5\%\}$$ (Table 7, p.17). Borrowing constraint (non-negative allocation to risk-free) and short-selling constraint ($$0 \leq \omega_t \leq 1$$) imposed separately and jointly (Table 8, p.18). A 20-year rolling window provides out-of-sample evidence from February 1969 onward (Table 9, p.19). A timing strategy (Table 10, p.20) benchmarks monthly Sharpe ratios across specifications and compares to Rossi (2018).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Welch-Goyal monthly predictors | 12 monthly predictor variables (dividend yield, log earnings-price ratio, dividend payout ratio, book-to-market, net equity expansion, treasury bill rate, term spread, default yield spread, default return spread, cross-sectional premium, inflation growth, stock variance) for S&P 500 excess log returns, Jan 1945-Dec 2018; from Amit Goyal's website | No page yet |
| NBER business cycle reference dates | Define expansion and recession subsamples in Tables 2, 4, and Figure 2 | [NBER cycles](/wiki/datasets/nber-cycles/) |

Sample: monthly, January 1945 to December 2018 (888 monthly observations). Portfolio evaluation: February 1955 to December 2018. Estimation run on GPU cluster (2 servers with 2 NVIDIA GeForce RTX 2080 Ti GPUs; rolling window calibration takes approximately 2 days).

## When to read the full paper

Read the [original](https://doi.org/10.1016/j.jempfin.2026.101705) if you are:
building a two-asset dynamic portfolio strategy with aggregate return predictors and want the exact hyperparameter grid and GPU estimation setup; comparing LSTM vs. feedforward NN architectures (NN1-NN3) with a consistent utility-based criterion; benchmarking against Rossi (2018) boosted-tree portfolios or the Johannes et al. (2014) Bayesian predictive framework on the Welch and Goyal (2008) predictor set; or studying the subperiod and decade performance in Tables 4 and 9 (where the Gu et al. (2020) cross-section evidence is separately in the literature). Table 2 gives the headline CER and SR by NBER phase; Tables 7-9 contain the robustness evidence.

## Attribution and rights

Source: peer-reviewed, *Journal of Empirical Finance* 87 (2026) 101705. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. The paper is paywalled (Elsevier B.V., all rights reserved, including text and data mining and AI training); only short extracts are reproduced here.

> Babiak, Mykola, and Jozef Baruník. "Deep learning, predictability, and optimal portfolio returns." *Journal of Empirical Finance* 87 (2026) 101705. DOI: [10.1016/j.jempfin.2026.101705](https://doi.org/10.1016/j.jempfin.2026.101705). © 2026 Elsevier B.V. All rights reserved.


==============================================================================
# Salience Theory and Corporate Bond Returns: Chen, Wang, Wei, Wu & Zhang (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jef/2026/chen-salience-corporate-bond-returns-2026/
# Distilled: U.S. corporate bonds with high salience theory (ST) values underperform those with low ST values by 0.61% per month in decile sorts (annualized Sharpe ratio 2.52, more than double the comparable equity figure). The premium is primarily driven by the outperformance of bonds with salient downside rather than the underperformance of those with salient upside, reflecting the asymmetric payoff structure of corporate bonds. Journal of Empirical Finance 2026, paywalled. Seven core results with source locators, datasets used, the BGS salience model, and the estimation and testing methods.
# Tags: paper-summary, asset-pricing, cross-section, behavioral-finance, corporate-bonds
==============================================================================

**What this is.** The paper's core results, the salience model it adapts from Bordalo, Gennaioli, and Shleifer (2012, 2013), and the estimation and testing methods with the defining equations: enough to know what it found and how, without reading all 24 pages. To replicate or extend it, read the original at [doi.org/10.1016/j.jempfin.2026.101692](https://doi.org/10.1016/j.jempfin.2026.101692).

## TL;DR

This paper applies salience theory to the U.S. corporate bond market. Following Bordalo, Gennaioli, and Shleifer (2012) and Cosemans and Frehen (2021), the authors compute a bond-level salience theory value (ST) as the covariance between salience-distorted probability weights and daily bond returns over the prior two months, using daily transaction data from Enhanced TRACE. Bonds with the highest ST values earn lower returns in the subsequent month than bonds with the lowest ST values. The monthly quintile spread is -0.48% (t=-10.30) and the decile Sharpe ratio is 2.52 annualized, more than double the 0.89 observed for stocks over the same sample period.

The key asymmetric finding distinguishes bonds from stocks. In the equity market (Cosemans and Frehen 2021), the salience effect is mainly driven by salient upside: stocks with lottery-like positive payoffs are overpriced. In the corporate bond market, the premium is driven primarily by the outperformance of bonds with salient downside (portfolio 1), not by the underperformance of bonds with salient upside (portfolio 10). This reflects the limited-upside, substantial-downside payoff structure of corporate bonds: bond investors constrained by default risk pay more attention to extreme negative payoffs, making bonds with salient bad news overly attractive and thereby overpriced. The effect survives full Fama-MacBeth controls for FFL6 factor exposures, illiquidity, short-term reversal, lottery features (MAX, MIN), and bond characteristics. It is larger for junk bonds, long-maturity bonds, bonds with greater limits to arbitrage, and bonds with more retail investor demand, and strengthens during high economic uncertainty and bullish credit market sentiment.

## Core results

Magnitudes and significance are as reported in the source; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Quintile sort: high-ST bonds earn lower returns than low-ST bonds**, and the spread survives risk adjustment | Table 3, Panel A, p.10 | 5-1 return spread = -0.48%\*\*\* (t=-10.30); FF5 alpha = -0.44%\*\*\* (t=-7.01); FFL6 alpha = -0.44%\*\*\* (t=-7.22); characteristic-adjusted spread = -0.45%\*\*\* (t=-10.28) |
| R2 | **Decile sort: bond 10-1 spread of -0.61%/month delivers Sharpe ratio 2.52**, more than double the stock equivalent | Table 3, Panel B, p.10 | Bond 10-1 return spread = -0.61%\*\*\* (t=-10.96); annualized SR = 2.52; stock 10-1 spread = -0.78%\*\*\* (t=-3.74) with SR = 0.89 over the same July 2002-June 2021 period (the -1.22%/SR=1.34 figures in row 1 are for the longer 1931-2015 stock replication) |
| R3 | **Asymmetric pattern: bond salience premium mainly from salient-downside outperformance**, not salient-upside underperformance | Table 3, Panel B, pp.9-10 | Decile portfolio 1 (salient downside) excess return = 0.82%; portfolio 5 = 0.32%; portfolio 6 = 0.29%; portfolio 10 (salient upside) = 0.21%; downside minus upside asymmetry gap = 0.42%\*\*\* (t=2.90) |
| R4 | **FM regression: ST coefficient remains large with full controls** including ILLIQ, STR, MAX, MIN, MOM, LTR, COSKEW, SKEW, IVOL, ISKEW, and bond characteristics | Table 5, p.14 | Univariate: ST = -0.14\*\*\* (t=-10.46); with FFL6 betas only: -0.12\*\*\* (t=-13.30); with full controls: -0.07\*\*\* (t=-7.36); Adj-R2 rises from 0.015 to 0.313 |
| R5 | **Weekly frequency: negative ST-return relation persists at weekly horizon**, ruling out daily microstructure noise | Table 7, Panel A, p.17 | 5-1 weekly return spread = -0.19%\*\*\* (t=-10.61); FFL6 return alpha = -0.19%\*\*\* (t=-9.35) per week |
| R6 | **Rating and maturity heterogeneity: salience premium is larger for junk bonds and long-maturity bonds** | Table 8, Panels A1 and B1, pp.17-18 | Junk 5-1 spread = -0.48%\*\*\* (t=-6.89) vs AAA/AA = -0.23%\*\*\* (t=-4.53); long-maturity spread = -0.47%\*\*\* (t=-10.79) vs short-maturity = -0.22%\*\*\* (t=-5.81) |
| R7 | **Limits to arbitrage amplify the premium**: all five arbitrage-friction proxies produce Diff portfolios significant at 1% | Table 9, Panel A, pp.19-20 | IVOL Diff = -0.48%\*\*\* (t=-9.10); TVOL Diff = -0.43%\*\*\* (t=-9.10); ILLIQ Diff = -0.37%\*\*\* (t=-4.62); Size Diff = -0.20%\*\*\* (t=-3.47); Attention Diff = -0.22%\*\*\* (t=-3.41) |

**Overall (paper's conclusion).** The salience premium in corporate bonds cannot be attributed to conventional risk factors, past return patterns, or bond characteristics. The premium is pervasive across all bond rating and maturity segments, is larger where limits to arbitrage are more severe, is stronger for bonds with higher retail investor demand, and intensifies during periods of high economic uncertainty and bullish credit market sentiment. Bond investors, including institutional ones, are not immune to salience bias.

## Theory / model

The paper applies the Bordalo, Gennaioli, and Shleifer (2012) (BGS hereafter) salience model to the corporate bond market. The model relaxes the rational-expectations framework by having agents assign disproportionate weight to payoff states that stand out relative to the payoffs of comparable assets. Asset $i$ has payoffs in $S$ states ordered from most negative to most positive, with objective probabilities $\pi_s$ (equation 1, p.3):

$$
\left(R^i_{-m},\, \pi_{-m};\; \ldots;\; R^i_{-1},\, \pi_{-1};\; R^i_0,\, \pi_0;\; R^i_1,\, \pi_1;\; \ldots;\; R^i_{n-1},\, \pi_n \right)
$$

For each state $s$, the salience of asset $i$'s payoff relative to the average return $\bar{R}_s$ on similar assets in the same state is measured by the salience function (equation 2, p.4):

$$
\sigma(R^i_s,\, \bar{R}_s) = \frac{|R^i_s - \bar{R}_s|}{|R^i_s| + |\bar{R}_s| + \theta} \tag{2}
$$

where $\theta > 0$ is a smoothing parameter. The function has three properties: (i) ordering (salience depends on distance from the peer average), (ii) diminishing sensitivity (salience decreases as the magnitude of payoffs grows larger), and (iii) reflection (salience of a payoff depends on its distance from the mean, not its sign).

Payoffs are ranked by degree of salience and assigned salience rankings $k^i_s$ (rank 1 = most salient, rank $S$ = least salient). Agents with salience thinking replace objective probabilities with salience-distorted probabilities (equation 3, p.4):

$$
\bar{\pi}^{ST}_s = \pi_s \omega^{ST}_s, \qquad \omega^{ST}_s = \frac{\delta^{k^i_s}}{\sum_s \delta^{k^i_s} \pi_s} \tag{3}
$$

where $\delta \in (0, 1)$ governs the degree of distortion. When $\delta < 1$, states with more salient payoffs (lower rank $k^i_s$) receive greater overweighting relative to objective probabilities. When $\delta = 1$, objective probabilities are undistorted.

BGS (2013) show that for a salient thinker, the expected return on an asset negatively depends on the covariance between its salience weights and its returns (equation 4, p.4):

$$
E(R^i_s) = -\text{cov}(\omega^{ST}_s,\, R^i_s) \equiv -ST^i \tag{4}
$$

An asset has a positive $ST$ value when its positive returns are more salient than its negative returns. Such assets attract excess demand from salient thinkers, driving up prices and reducing future returns. This is the key asset-pricing implication: assets with higher $ST$ earn lower subsequent returns.

**Bond-specific asymmetry.** Corporate bonds have limited upside potential (bounded by par value) but substantial downside (default risk). This payoff asymmetry implies that salient downside states are more likely to attract attention from bond investors than salient upside states. The model therefore predicts that the bond salience premium is driven primarily by the outperformance of bonds with salient downside, which contrasts with the upside-dominated pattern found by Cosemans and Frehen (2021) in the stock market.

## Method

**ST value estimation.** For each trading day $d$ in evaluation period $\tau$, the salience degree of bond $i$'s daily return $R^M_{d,\tau}$ is measured relative to the equal-weighted average daily return across all bonds $\bar{R}^M_{d,\tau}$, using the salience function in equation (2) above with parameters $\theta = 0.1$ and $\delta = 0.7$ (BGS 2012 calibration, p.5). A maximum 7-day gap between consecutive trading days is imposed. The baseline evaluation window is the prior two months (requiring at least 10 daily observations), chosen to balance recency of information and estimation reliability.

Daily salience degrees are ranked in descending order to assign salience rankings $k^i_{d,\tau}$. Assuming equal objective probability $\pi_d = 1/N_t$ for each trading day, the salience weight is:

$$
\omega^{ST}_{d,\tau} = \frac{\delta^{k^i_{d,\tau}}}{\sum_d \delta^{k^i_{d,\tau}} \cdot (1/N_t)}
$$

where $N_t$ is the number of trading days in $\tau$. The salience theory value of bond $i$ at month $t$ is:

$$
ST^i_t = \text{cov}\!\left(\omega^{ST}_{d,\tau},\, R^i_{d,\tau}\right)
$$

A positive $ST^i_t$ means the bond's historical daily returns are positively correlated with the salience weights: days with more extreme (salient) returns tend to have positive returns, making the bond appear attractive to salient thinkers.

**Bond return construction.** The raw monthly return on bond $i$ in month $t$ is (equation 5, p.5):

$$
r^i_t = \frac{(P^i_t + AI^i_t) + C^i_t}{P^i_{t-1} + AI^i_{t-1}} - 1 \tag{5}
$$

where $P^i_t$ is the trade-size-weighted average intraday price, $AI^i_t$ is accrued interest, and $C^i_t$ is the coupon payment. If the last transaction in a month does not fall on the last trading day, an interpolated month-end price is used.

**Amihud illiquidity.** The Amihud (2002) illiquidity measure adapted to bonds (equation 6, p.5):

$$
ILLIQ^i_t = \frac{1}{days_{it}} \sum_{j=1}^{days_{it}} \frac{|r_{ij,t}|}{vol_{ij,t}} \tag{6}
$$

where $days_{it}$ is the number of trading days with daily returns, and $vol_{ij,t}$ is the dollar volume (in USD million) on day $j$.

**Factor exposure estimation.** Beta loadings for the six FFL6 factors (MKT, SMB, HML from Fama and French (1993), DEF, TERM, and the Lin et al. (2011) liquidity factor LIQ) are estimated each month $t$ by regressing monthly bond excess returns on factor excess returns over a rolling 60-month window.

## Empirical specifications

**Univariate portfolio sorts (R1, R2, R3).** At end of each month $t$, bonds are sorted into quintiles or deciles by $ST^i_t$ and held for one month. Equal-weighted excess returns are computed in month $t+1$. Factor-adjusted alphas are obtained by regressing portfolio monthly excess returns on FFL6 factors over the past five years. The high-minus-low (5-1 or 10-1) spread and its Newey-West (1987) adjusted $t$-statistic (12 lags) are the headline tests.

**Bivariate portfolio sorts (R1, control for characteristics).** At end of each month, bonds are sorted into 5x5 portfolios by one of 18 control variables and by $ST$ within each control quintile. The high-minus-low $ST$ return spread within each control quintile and averaged across quintiles confirms the salience effect is not driven by any individual characteristic. The absolute STR-controlled return spread is reduced by about one-half (from 0.48% to 0.22%) but remains highly significant.

**Fama-MacBeth cross-sectional regressions (R4).** The headline FM regression runs each month (equation 7, p.11):

$$
r^{i,e}_{t+1} = \gamma_0 + \gamma_1 \beta^i_{MKT,t} + \gamma_2 \beta^i_{SMB,t} + \gamma_3 \beta^i_{HML,t} + \gamma_4 \beta^i_{DEF,t} + \gamma_5 \beta^i_{TERM,t} + \gamma_6 \beta^i_{LIQ,t} + \gamma_7 ST^i_t + \delta' Z^i_t + \varepsilon^i_t \tag{7}
$$

where $r^{i,e}_{t+1}$ is bond $i$'s excess return over the one-month T-bill rate and $Z^i_t$ includes ILLIQ, STR, MAX (maximum daily return, a lottery-preference control following Bali, Cakici, and Whitelaw (2011)), MIN, MOM, LTR, COSKEW, SKEW, IVOL, ISKEW, Maturity, Age, Coupon, Size, and Rating. Each regressor is standardized by its cross-sectional standard deviation each month. Standard errors are Newey-West adjusted with 12 lags.

**Weekly frequency (R5).** The $ST$ value is recalculated using Wednesday-to-Wednesday weekly returns over the prior 24 weeks (at least 10 observations), and the FM regression is re-run at the weekly frequency to address daily microstructure noise concerns:

$$
r^{i,e}_{t+1,w} = \gamma_0 + \gamma_1 \beta^i_{MKT} + \gamma_2 \beta^i_{SMB} + \gamma_3 \beta^i_{HML} + \gamma_4 \beta^i_{DEF} + \gamma_5 \beta^i_{TERM} + \gamma_6 \beta^i_{LIQ} + \gamma_7 ST^i_{t,w} + \delta' Z^i_{t,w} + \varepsilon^i_{t,w}
$$

**By rating and maturity (R6).** Bonds are grouped into four rating portfolios (AAA/AA, A, BBB, Junk) and three maturity portfolios (Short: under 5 years; Medium: 5-10 years; Long: over 10 years). Within each group, bonds are further sorted by $ST$ and a high-minus-low (5-1) portfolio is formed. Separate FM regressions within each rating or maturity group confirm the pattern.

**Limits-to-arbitrage interaction (R7).** The FM regression adds an interaction $ST^i_t \times D^i_t$, where $D^i_t$ equals 1 when a limits-to-arbitrage proxy (issue size, TVOL, IVOL, ILLIQ, or Attention) exceeds its cross-sectional median. A significantly negative interaction coefficient indicates a larger salience premium for high-friction bonds (Table 9, Panel B).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Enhanced TRACE (FINRA) | Intraday corporate bond transaction data: prices, volumes, daily returns for ST computation and monthly return construction | [TRACE](/wiki/commercial/trace/) (licensed) |
| Mergent FISD | Bond characteristics: CUSIP, issuer info, maturity, coupon frequency, coupon rate, offering amount, and credit ratings | no page yet |
| CRSP / Compustat (via WRDS) | Equity characteristics for the matched-sample analysis; factor betas via stock returns | [WRDS](/wiki/commercial/wrds/) (licensed) |
| Kenneth French Data Library | FF5 factor returns (MKT, SMB, HML) and default spread (DEF) used as the risk benchmark | [Ken French library](/wiki/datasets/ken-french/) |
| eMAXX (Lipper) | Quarterly institutional holdings of corporate bonds (PCT_Ins) for the investor-type analysis in Section 5.3 | [eMAXX](/wiki/commercial/emaxx/) (licensed) |

Sample: July 2002 to June 2021 (19 years, monthly). Final sample: 438,724 bond-month observations for 15,383 bonds issued by 1,800 firms. Transaction data filtered following Bessembinder et al. (2008) to remove corrections, cancellations, reversals, and agency transactions.

## When to read the full paper

Read the [original](https://doi.org/10.1016/j.jempfin.2026.101692) when you need the full Internet Appendix (Tables IA1-IA26): alternative factor models (BBW4-D and FF-BBW7-D), alternative salience evaluation horizons (1 month to 5 years, Table IA12), alternative credit-rating and maturity groupings (Table IA17), callable-bond robustness (Table IA15), firm-level regressions (Table IA19), the yield-based ST measure STY (Tables IA16, Appendix B5), or the prospect theory (PT) value comparison (Table IA26). The detailed matched-sample analysis pairing bonds and equities issued by the same firm is in Table IA24. The subperiod analysis covering the 2007-2009 crisis period and three economic uncertainty indices is in Table 11 (p.22).

## Attribution and rights

Source: peer-reviewed, *Journal of Empirical Finance* 87 (2026) 101692. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. The article is paywalled (Elsevier subscription); no open-access license is present in the Crossref metadata. Extract-only; no PDF is hosted here.

> Chen, Xi, Junbo Wang, K.C.John Wei, Chunchi Wu, and Linti Zhang. "Salience theory and cross-sectional corporate bond returns." *Journal of Empirical Finance* 87 (2026): 101692. DOI: 10.1016/j.jempfin.2026.101692.


==============================================================================
# Factor Pricing Across Asset Classes: Dang, Hollstein & Prokopczuk (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jef/2026/dang-factor-pricing-across-asset-classes-2026/
# Distilled: Factor models specialized for one asset class have limited pricing power across others; markets are significantly but imperfectly integrated. An optimal eight-factor integrated model spanning five asset classes achieves a full-sample Sharpe ratio of 1.053, far exceeding the AMP global benchmark (0.171) and all single-class models. Journal of Empirical Finance 2026, CC BY-NC 4.0. Six core results with source locators, datasets used, the method, and the empirical specifications.
# Tags: paper-summary, asset-pricing, factors, factor-models, market-integration
==============================================================================

**What this is.** The paper's core results, the conceptual framework (degrees of market integration, SDF theory), the two-step factor selection method (PRS protocol and BS-CZZ model scan), and the empirical specifications: enough to know what was found and how, without reading all 19 pages. To replicate or extend the work, read the full source at the [original](https://doi.org/10.1016/j.jempfin.2026.101688).

## TL;DR

Dang, Hollstein, and Prokopczuk study whether the factors that price one asset class can also price others. Across 77 factor candidates drawn from seven major asset classes (U.S. equities, international equities, corporate bonds, commodities, currencies, equity indices, and government bonds), covering August 2006 to December 2019, they find that markets are significantly but imperfectly integrated: asset class-specific models fail to explain most factors from other classes, yet some notable cross-market linkages exist. Using the PRS factor-identification protocol of Pukthuanthong, Roll, and Subrahmanyam (2019) and the BS-CZZ Bayesian model selection of Barillas and Shanken (2018) and Chib, Zeng, and Zhao (2020), they identify an optimal integrated eight-factor model drawing on five asset classes. This model achieves a full-sample Sharpe ratio of 1.053 and out-of-sample Sharpe ratios at least 48% higher than any single-class model, and leaves only 111 of 15,968 mutual funds with a statistically significant positive alpha, making it a substantially stronger benchmark for fund evaluation.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*` = 5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **U.S. market factor explains some but not all asset classes.** R² is high for international equities and equity indices but near zero for government bonds; GRS rejects jointly zero alphas | Table 3, Panel A, p. 6 | Int. equities R²=71.2%, equity indices R²=78.6%, corporate bonds R²=24.1%, FX R²=36.1%, govt bonds R²=8.01%; GRS=3.72\*\* |
| R2 | **Optimal integrated model (winner1\_across) achieves full-sample SR=1.053**, far above AMP global model (SR=0.171) and all single-class models | Table 8, p. 13 | winner1\_across SR=1.053; winner2\_across SR=1.095; AMP\_across SR=0.171; best single-class model is winner\_cb SR=0.756 |
| R3 | **Integrated model significantly dominates all single-class models in pairwise squared SR tests**; all 16 pairwise differences are positive and statistically significant | Table 7, p. 12 | vs winner\_useq (best US equity model): +0.832\*\*; vs IRP (corp bonds): +0.501\*\*; vs AMP\_across: +0.980\*\* |
| R4 | **Out-of-sample Sharpe ratio of integrated model remains high**; at least 48% higher OOS SR than any single-class model | Table 8, p. 13 | T/2 PERF=1.059, PERFw=0.817; 2T/3 PERF=1.219, PERFw=0.750; vs AMP\_across PERFw=0.132 |
| R5 | **Integrated model explains most factors across all classes**; only 12 of 77 viable factors have a significant alpha; GRS fails to reject jointly zero alphas for 4 of 7 asset classes | Table 9, Section 5.3, pp. 12-17 | 12/77 (15.6%) factors with \|t\|≥1.96 vs integrated model; GRS p-values non-significant for U.S. equities, commodities, equity indices, government bonds |
| R6 | **Integrated model reduces spurious positive fund alphas**; average absolute fund alpha falls and the count of significantly positive alphas drops from thousands to 111 | Table 10, pp. 17-18 | Avg absolute alpha=0.16%/month; 111 sig. positive funds; vs FF5\_inteq: 0.23%/month avg, 2,903 sig. positive; vs winner\_useq: 0.23%/month avg, 235 sig. positive |

**Overall (paper's conclusion).** Factor models that specialize in one asset class typically fail to price factors from other classes. There is strong evidence of multiple underlying systematic risk drivers across markets, but also of interdependencies: markets are significantly but imperfectly integrated. The Fama and French (1993) equity size and value factors, for example, have limited reach across corporate bond and government bond classes. The AMP global three-factor model of Cooper, Mitrache, and Priestley (2022) achieves a full-sample Sharpe ratio of only 0.171 in this setting. The 48 value-and-momentum portfolios of Asness, Frazzini, and Pedersen (2013) are used as the main cross-asset test assets throughout. A unified eight-factor model (MKT\_useq, SMB\_inteq, MGMT\_inteq, QMJ\_inteq, Carry\_cb, MOMeq\_cb, MOM\_fxaqr, Carry\_eqi) drawn from five asset classes spans the majority of all prominent factors across the seven classes and provides a substantially better benchmark for multi-asset fund managers than any single-asset-class model.

## Theory / model

The paper has no new formal model. Its theoretical motivation draws on two classical results.

**SDF theory (Cochrane 2009).** Under no-arbitrage, a single stochastic discount factor $$M_t$$ prices all assets simultaneously. For any excess return $$R^e_{i,t+1}$$ (p. 1 of the paper):

$$
E_t\!\left[M_{t+1}\, R^e_{i,t+1}\right] = 0, \quad i = 1, \ldots, N \tag{1}
$$

This implies that, in principle, one set of risk factors should suffice to price all asset classes. A purely class-specific factor model is consistent with this only if there are no cross-market risk drivers, that is, if markets are completely disintegrated.

**Mean-variance efficiency and Sharpe ratios.** The Markowitz (1952) tangency portfolio of a candidate factor set $$\mathbf{F}$$ achieves the maximum Sharpe ratio attainable from that set:

$$
\text{SR}^2(\mathbf{F}) = \boldsymbol{\mu}_F'\, \boldsymbol{\Sigma}_F^{-1}\, \boldsymbol{\mu}_F \tag{2}
$$

where $$\boldsymbol{\mu}_F$$ and $$\boldsymbol{\Sigma}_F$$ are the mean vector and covariance matrix of factor excess returns. Barillas and Shanken (2018) show that comparing two candidate factor sets reduces to comparing their tangency Sharpe ratios: if $$\text{SR}^2(\mathbf{F}_j) > \text{SR}^2(\mathbf{F}_k)$$, model $$j$$ is preferred. This makes the Sharpe ratio the natural model-selection criterion.

**Three hypotheses tested.** The paper tests three mutually exclusive hypotheses (p. 2):

- *Perfect integration*: a common set of global factors prices all asset classes; no class-specific factors are needed.
- *Partial integration*: some factors are common across classes; others are class-specific; both are needed for full explanation.
- *Complete disintegration*: each asset class has its own independent set of risk factors; cross-class pricing power is zero.

The results reject both perfect integration (GRS=3.72\*\* in Table 3, and many significant cross-class alphas in Table 5) and complete disintegration (many significant cross-market factor loadings and an integrated model that prices most remaining factors).

## Method

The empirical strategy has two steps, applied first within each asset class and then across all classes jointly.

**Step 1: PRS factor identification (Pukthuanthong, Roll, and Subrahmanyam 2019, pp. 7-8).** For each asset class, extract the first $$K$$ principal components $$\mathbf{p}_{1:K,t}$$ from the full universe of test portfolios using the Connor-Korajczyk (1988) method. For each candidate factor $$f_{k,t}$$, compute canonical correlations between $$f_{k,t}$$ and $$\mathbf{p}_{1:K,t}$$ in two equal sub-periods. Factor $$f_k$$ is a viable risk factor if and only if:

$$
\overline{|\hat{t}_j|} > 1.96 \quad \text{and} \quad \overline{\hat{s}_j} > 0.25 \tag{3}
$$

where $$\overline{|\hat{t}_j|}$$ is the average absolute t-statistic of significant canonical correlations over both sub-periods and $$\overline{\hat{s}_j}$$ is the average fraction of significant canonical correlations out of $$K$$. Factors failing this test are discarded as non-viable.

**Step 2: BS-CZZ Bayesian model selection (Barillas and Shanken 2018; Chib, Zeng, and Zhao 2020, pp. 7-8 and Online Appendix OA4).** Among factors that pass Step 1, perform an exhaustive Bayesian model scan over all factor subsets. Each candidate model $$\mathcal{M}_j$$ receives a posterior model probability proportional to (via the Barillas-Shanken marginal likelihood):

$$
P(\mathcal{M}_j \mid \text{data}) \propto \left(1 + \hat{\text{SR}}^2(\mathbf{F}_j)\right)^{T/2} \cdot p(\mathcal{M}_j) \tag{4}
$$

where $$\hat{\text{SR}}^2(\mathbf{F}_j) = \hat{\boldsymbol{\mu}}_j' \hat{\boldsymbol{\Sigma}}_j^{-1} \hat{\boldsymbol{\mu}}_j$$ is the sample squared tangency Sharpe ratio of the factor set $$\mathbf{F}_j$$ over $$T$$ observations, and $$p(\mathcal{M}_j)$$ is the prior probability. The optimal model is

$$
\mathcal{M}^* = \operatorname*{arg\,max}_j\; P(\mathcal{M}_j \mid \text{data}) \tag{5}
$$

This is the `bs-czz-model-selection` procedure: it builds on `panel-regression` (spanning regressions to estimate means and covariances) and `portfolio-sort` (factor and test portfolio construction) to rank all candidate factor sets by their posterior model probability. High-collinearity factor pairs (pairwise correlation exceeding 0.8) are excluded before the scan.

## Empirical specifications

**Time-series spanning regression (Tables 3, 5, 6, 9).** For each factor or test-portfolio return $$r_{a,t}$$ from asset class $$A$$, regress on the candidate factor set $$\mathbf{F}_t$$ (from another class or the integrated model), pp. 5-6:

$$
r_{a,t} = \alpha_a + \boldsymbol{\beta}_a' \mathbf{F}_t + \varepsilon_{a,t}, \quad t = 1, \ldots, T \tag{6}
$$

The intercept $$\alpha_a$$ measures the portion of $$r_{a,t}$$ not spanned by $$\mathbf{F}_t$$. Standard errors use Newey and West (1987) with four lags. This specification ties to R1 (Table 3, market factor spanning) and R5 (Table 9, full-model spanning for all 77 viable factors).

**GRS test (Gibbons, Ross, and Shanken 1989).** To test whether all $$N$$ intercepts are jointly zero (Table 3 GRS statistic; Table 9 panel-level GRS):

$$
\text{GRS} = \frac{T - N - K}{N}\left(1 + \hat{\boldsymbol{\mu}}_F' \hat{\boldsymbol{\Sigma}}_F^{-1} \hat{\boldsymbol{\mu}}_F\right)^{-1} \hat{\boldsymbol{\alpha}}' \hat{\boldsymbol{\Sigma}}_\varepsilon^{-1} \hat{\boldsymbol{\alpha}} \;\sim\; F(N,\, T - N - K) \tag{7}
$$

where $$\hat{\boldsymbol{\alpha}}$$ is the vector of estimated intercepts and $$\hat{\boldsymbol{\Sigma}}_\varepsilon$$ is the residual covariance matrix. This specification ties to R1 (GRS=3.72\*\* for market factors, Table 3) and R5 (GRS fails to reject for 4 of 7 asset classes, Table 9).

**Pairwise equality of squared Sharpe ratios (Barillas, Kan, Robotti, and Shanken 2020, Table 7, p. 12).** For each pair of candidate factor models $$j$$ and $$k$$:

$$
H_0:\; \text{SR}_j^2 = \text{SR}_k^2 \quad \text{vs} \quad H_a:\; \text{SR}_j^2 \neq \text{SR}_k^2 \tag{8}
$$

The bias-adjusted test statistic uses the asymptotic distribution of the sample squared Sharpe ratio difference, correcting for estimation error in the factor means and covariances. All 16 pairwise tests of the integrated model against single-class models and existing integrated benchmarks are reported in Table 7; all differences favor the integrated model at the 1% level (R3).

**Mutual fund performance evaluation (Section 5.5, Table 10).** For each of 15,968 mixed-asset mutual funds with at least 100 monthly return observations in LSEG Datastream (U.S. dollars, euros, British pounds, or Japanese yen):

$$
R^f_{i,t} = \alpha_i + \boldsymbol{\beta}_i' \mathbf{F}_t + \varepsilon_{i,t} \tag{9}
$$

where $$R^f_{i,t}$$ is the excess fund return and $$\mathbf{F}_t$$ is the integrated factor model. Standard errors use Newey-West (1987) with four lags. The headline comparison is the number of significantly positive $$\hat{\alpha}_i$$ (t-stat $$\geq 1.96$$) under each benchmark model (R6, Table 10).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Kenneth French Data Library | U.S. equity factor candidates (MKT, SMB, HML, RMW, CMA, UMD, BAB, QMJ); international equity factors; characteristic-sorted test portfolios for U.S. and international equities | [Ken French library](/wiki/datasets/ken-french/) |
| Open Source Asset Pricing (Chen and Zimmermann 2022) | Supplementary U.S. equity factor candidates (IA, ROE, ME, etc.) | [Open Source Asset Pricing](/wiki/datasets/open-source-asset-pricing/) |
| Refinitiv Datastream | Equity index returns for 43 countries (Aug 2006-Dec 2019); government bond index returns for 22 countries; also global equity index factors | [Datastream](/wiki/commercial/datastream/) (commercial) |
| Commodity Research Bureau (CRB) | Nearest-to-maturity futures prices for 21 commodity contracts; rolled 2 months before expiration; basis for commodity factor construction | no page yet |
| AQR factor library (Asness-Frazzini, Frazzini-Pedersen, Asness et al., Ilmanen et al.) | U.S. equity BAB, QMJ; corporate bond factors; currency factors; time-series momentum factors | no page yet |
| Stambaugh and Yuan (2017); Daniel et al. (2020) | U.S. mispricing (MGMT, PERF, PEAD, FIN) and DHS behavioral factors; authors extend the series to the full sample | no page yet |
| Hanauer (2020); Jensen et al. (2023) | Global equity factor candidates including size, momentum, ROE, and international BAB/QMJ | no page yet |
| Kelly and Pruitt (2022); Lustig et al. (2011); Verdelhan (2018) | Corporate bond term-structure factors; currency carry and dollar factors | no page yet |

Sample: August 2006 to December 2019 (162 months). The common sample is determined by the availability of all 77 candidate factors. An extended sample (July 1990 to June 2022) is used for robustness in the Online Appendix.

## When to read the full paper

Use the [original](https://doi.org/10.1016/j.jempfin.2026.101688) if you are: building a multi-asset-class factor model and need the factor-selection algorithm (PRS + BS-CZZ) and the precise list of the eight integrated factors (Table 6); evaluating asset managers against a cross-asset benchmark and need the exact benchmark specification (Table 10 columns); comparing the explanatory reach of single-class factor models across markets (Tables 5 and 9 heat maps); or examining whether a candidate factor is priced globally (spanning regressions in Table 9, one panel per class). The Online Appendix contains full robustness results and an extended sample running to June 2022.

## Attribution and rights

Source: peer-reviewed, *Journal of Empirical Finance* 87 (2026) 101688. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. CC BY-NC 4.0: non-commercial reproduction with attribution is permitted; commercial use requires permission from Elsevier.

> Dang, Thuy Duong, Fabian Hollstein, and Marcel Prokopczuk.
> "Factor pricing across asset classes."
> *Journal of Empirical Finance* 87 (2026) 101688.
> DOI: 10.1016/j.jempfin.2026.101688. © 2026 The Author(s).
> Licensed under [CC BY-NC 4.0](http://creativecommons.org/licenses/by-nc/4.0/).
> This page is an extract by the Institute for Automated Research; **changes were made**.


==============================================================================
# The Decay of cay: Dauber & Lawrenz (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jef/2026/dauber-decay-of-cay-2026/
# Distilled: Documents a substantial decline over the last two decades in the predictive power of the consumption-wealth ratio (cay) for US stock market excess returns, attributing it to a structural shift in the cointegration relationship as asset wealth decouples from aggregate consumption and labor income. Proposes a top-10% household version of cay as the most stable remaining predictor. Journal of Empirical Finance 2026, CC BY 4.0. Six core results with source locators, datasets used, the model, and the method.
# Tags: paper-summary, asset-pricing, macro, return-predictability, predictive-regression
==============================================================================

**What this is.** The paper's core results, the model that motivates cay (the Campbell-Mankiw
intertemporal budget constraint approximation developed by Lettau and Ludvigson (2001)), and the
estimation and forecasting procedure with the defining equations: enough to understand what was found
and why, without reading all 20 pages. To replicate or extend, read the full source at
the [original](https://doi.org/10.1016/j.jempfin.2025.101668).

## TL;DR

The paper revisits the ability of the consumption-wealth ratio (cay) to forecast US stock market
excess returns and documents that its predictive power has declined substantially over the last two
decades. Using comprehensive in-sample, out-of-sample, and economic significance tests on quarterly
US data from 1952:1 to 2019:4, the authors show that aggregate cay has lost even its in-sample
predictive ability from the perspective of the most recent data. They trace this decay to a
structural shift in the underlying cointegration relationship between consumption, aggregate wealth,
and labor income, as asset wealth has become increasingly detached from aggregate consumption since
around 2000 due to rising wealth inequality. As a partial remedy, they propose a version of cay
constructed from the top 10% richest households (cay^PCE10), which remains the most stable and
significantly predictive alternative among those examined, though even this measure's predictive
advantage over a naive historical mean strategy has largely disappeared.

## Core results

Magnitudes and significance are as reported; `\*` = 10%, `\*\*` = 5%, `\*\*\*` = 1%.
Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Aggregate cay (NDS) has lost in-sample predictive power** for 1Q-ahead excess returns in the full sample | Table 3, p. 8 | beta = 0.241, t = 1.118, adj-R2 = 0.001; compare to Lettau and Ludvigson (2001) sample: beta = 2.165, t > 3, adj-R2 = 9% |
| R2 | **cay^PCE10 (top-10% PCE version) retains marginal in-sample significance** at the 1Q horizon | Table 3, p. 8 | beta = 0.629, t = 2.417\*\*, adj-R2 = 0.020; aggregate NDS coefficient is insignificant (t = 1.118) |
| R3 | **Cointegration between consumption, wealth, and income cannot be confirmed** for the full sample | Table 4 Panel B, p. 13 | Phillips and Ouliaris (1990) test p-values: NDS = 0.868, PCE = 0.577 (full sample 1952:1-2019:4); p-values below critical values for the Lettau and Ludvigson (2001) sample only |
| R4 | **Asset wealth coefficient in the cointegrating vector has declined steadily** since the late 1990s | Fig. 5, p. 13 | beta_a (NDS) falls from approximately 0.3 in the early 1990s toward 0.0 by 2019; becomes insignificant from the perspective of the full sample (Table 4 Panel A) |
| R5 | **Out-of-sample economic performance decays across all specifications** except cay^PCE10 | Fig. 4, pp. 11-12 | Risk-adjusted abnormal return theta starts at approximately 200bp for all specs in mid-1990s; all aggregate cay specs exhibit double-digit negative theta by 2019; cay^PCE10 ends near 10bp |
| R6 | **cay^PCE10 is the most stable and significant alternative** among four proposed improvements in the full sample | Table 6, p. 17 | Full sample 1Q t-statistics: cay^PCE10 = 2.417\*\*, cday = 1.693, cay^g = 1.732, cay^unfi = 1.299 |

**Overall (paper's conclusion).** The predictive ability of cay has fundamentally weakened over the
last roughly two decades. The decay is traceable to a structural shift in the cointegrating
relationship between consumption, aggregate wealth, and labor income: as asset wealth has become
increasingly detached from aggregate consumption (particularly since the global financial crisis),
the cointegrating parameters have drifted, undermining both the theoretical rationale and the
empirical performance of cay. Focusing on the richest 10% of households via cay^PCE10 mitigates but
does not eliminate the decay.

## Theory / model

The model builds on the representative agent's intertemporal budget constraint (p. 4, Eq. 1):

$$
W_{t+1} = (1 + R_{w,t+1})(W_t - C_t), \tag{1}
$$

where $$W_t$$ is aggregate wealth, $$C_t$$ is aggregate consumption, and $$R_{w,t+1}$$ is the return
on total wealth between periods $$t$$ and $$t+1$$. Following Campbell and Mankiw (1989) and
Lettau and Ludvigson (2001), log-linearizing around the steady state yields the log
consumption-wealth approximation (p. 4, Eq. 2):

$$
c_t - w_t \approx \mathbb{E}_t \sum_{i=1}^{\infty} \rho^i (r_{w,t+i} - \Delta c_{t+i}), \tag{2}
$$

so cay proxies for expected future returns and expected consumption growth. Because total wealth
$$W_t$$ includes unobservable human capital $$H_t$$, Lettau and Ludvigson (2001) approximate it
using log labor income $$y_t$$ (p. 4, Eq. 3-4):

$$
w_t \approx \alpha a_t + (1-\alpha) h_t, \quad r_{w,t} \approx \alpha r_{a,t} + (1-\alpha) r_{h,t}, \tag{3-4}
$$

where $$a_t$$ is log asset wealth, $$\alpha$$ is the average share of assets in total wealth, and
$$r_{a,t}$$ ($$r_{h,t}$$) is the return on assets (human capital). Together with a stationarity
argument, this gives the defining identity (p. 4, Eq. 6):

$$
\text{cay}_t = c_t - \alpha a_t - (1-\alpha) y_t. \tag{6}
$$

If $$c_t$$, $$a_t$$, and $$y_t$$ are cointegrated with the vector $$(1, -\alpha, -(1-\alpha))$$, then
$$\text{cay}_t$$ is stationary and forecasts future asset returns. The paper's central finding is that
this cointegrating relationship has become unstable: the coefficient $$\alpha$$ on asset wealth has
drifted steadily toward zero since the late 1990s, which it attributes to rising wealth inequality
making asset wealth increasingly detached from aggregate consumption.

**Top-10% version.** To better capture the representative investor, the paper constructs a
version of cay using consumption, wealth, and income of the top 10% richest households. Wealth share
$$AS^{\text{top10}}_t$$ and income share $$YS^{\text{top10}}_t$$ of this group are regressed on the
capital share $$KS_t$$ (p. 5, Eqs. 7-8):

$$
AS^{\text{top10}}_t = \alpha_A + \beta_A \, KS_t + \varepsilon_t, \quad YS^{\text{top10}}_t = \alpha_Y + \beta_Y \, KS_t + \varepsilon_t. \tag{7-8}
$$

The fitted values $$\widehat{AS}^{\text{top10}}_t$$ and $$\widehat{YS}^{\text{top10}}_t$$ are then
used to construct top-10% series for aggregate consumption, income, and wealth. The top-10% version
of cay is then (p. 5, Eq. 9):

$$
\text{cay}^{\text{top10}}_t = c^{\text{top10}}_t - \alpha^{\text{top10}} a^{\text{top10}}_t - (1-\alpha^{\text{top10}}) y^{\text{top10}}_t. \tag{9}
$$

The rationale, following Lettau et al. (2019), is that rich households own a disproportionate share
of stock market wealth and better approximate the marginal investor whose expectations drive
equity premia.

## Method

**Cointegration estimation (DLS).** The paper follows Lettau and Ludvigson (2001) in using the
dynamic-least-squares (DLS) technique of Stock and Watson (1993), which builds on `time-series-forecasting`
and the proposed `dynamic-least-squares` technique. To estimate the cointegrating parameters
$$\beta_a$$ and $$\beta_y$$, the following regression is run (p. 6, Eq. 10):

$$
c_t = \alpha + \beta_a a_t + \beta_y y_t + \sum_{i=-8}^{8} b_{a,i} \Delta a_{t-i} + \sum_{i=-8}^{8} b_{y,i} \Delta y_{t-i} + \varepsilon_t, \tag{10}
$$

where $$\Delta$$ denotes the first-difference operator. The 8-lead and 8-lag augmentation removes
regressor endogeneity that would otherwise cause OLS to be inconsistent. The estimated cay is then
the residual from the cointegrating equation (p. 6, Eq. 11):

$$
\widehat{\text{cay}}_t = c_t - \hat{\beta}_a a_t - \hat{\beta}_y y_t. \tag{11}
$$

**Stability analysis.** The paper also re-estimates the cointegration relationship in an expanding
window beginning in 1990, tracking the time-varying behavior of $$\hat{\beta}_a$$ and
$$\hat{\beta}_y$$. A Vector Error Correction Model (VECM) is estimated to assess the short-term
dynamics of the cointegration relationship for the two sample periods.

**OOS economic significance.** To measure economic performance, the paper follows the approach
of Della Corte et al. (2010), adapted for a short-selling constrained mean-variance investor who
allocates between the risk-free asset and the market portfolio. The optimal weight on the market
portfolio at time $$t$$ is (p. 11, Eq. 13):

$$
w_t = \frac{1}{\lambda} \frac{\mathbb{E}_t[r_{t+1} - r_{f,t+1}]}{\text{Var}_t[r_{t+1} - r_{f,t+1}]}, \tag{13}
$$

with $$\lambda = 3$$ as the coefficient of relative risk aversion. The ex-post performance of the
cay-timing strategy relative to a rolling historical mean strategy is evaluated by the risk-adjusted
abnormal return $$\theta$$ (Goetzmann et al. (2007)), measured in basis points.

## Empirical specifications

All results use quarterly data, 1952:1-2019:4. The main baseline sample excludes the Covid-19
episode (post-2019:4); a robustness check through 2022:4 is in Appendix H and does not affect
conclusions.

**In-sample predictive regression (R1, R2, R6).** The standard regression forecasting
excess returns at horizon $$H$$ is (p. 8, Eq. 12):

$$
\bar{r}_{t,H} = \alpha^k + \beta^k_{\text{cay}} \widehat{\text{cay}}^k_t + \varepsilon^k_{t,H}, \quad k \in \{\text{NDS, PCE, NDS10, PCE10}\}, \tag{12}
$$

where $$\bar{r}_{t,H} = r_{t+1} - r_{f,t+1} + \cdots + r_{t+H} - r_{f,t+H}$$ is the H-period
cumulative log excess return. Newey and West (1987) corrected t-statistics are reported. Table 3
(p. 8) covers horizons $$H \in \{1, 2, 4, 8, 12, 16, 20\}$$ quarters. A time-varying version
re-estimates the regression recursively in an expanding window from 1990 to track how the
coefficient $$\hat{\beta}_{\text{cay}}$$ has evolved (Fig. 3, pp. 9-10).

**Cointegration stability tests (R3, R4).** The paper re-estimates the cointegrating regression
in an expanding window from 1990, reporting the resulting $$\hat{\beta}_a$$ and $$\hat{\beta}_y$$
paths (Fig. 5, p. 13). Cointegration tests use the Phillips and Ouliaris (1990) test; results for
Engle and Granger (1987) and Johansen (1988, 1991) tests appear in Appendix C.2. Table 4
(p. 13) compares cointegrating parameters and test statistics across sample periods.

**Structural shift with time trend (R4 extended).** Section 4.2 augments Eq. (10) with a
deterministic time trend $$\hat{\pi} t$$ to assess whether the shift in $$\hat{\beta}_a$$ is driven
by an omitted trend. The cointegrating parameters are re-estimated with $$c_t - \hat{\pi} t$$ as
the dependent variable; the resulting time paths of $$\hat{\beta}_a$$, $$\hat{\beta}_y$$ and
$$\hat{\pi}$$ are tracked in an expanding window (Fig. 7, p. 16).

**OOS economic performance (R5).** The cay-timing strategy is initialized with a 43-year training
period (through 1994:4) and evaluated from 1995:1 onward. The risk-adjusted abnormal return
$$\theta_t$$ is plotted over time for all four cay specifications (Fig. 4, pp. 11-12).

**Comparison with alternatives (R6).** Table 6 (p. 17) runs Eq. (12) at the 1Q horizon for four
competing cay specifications (cay^PCE10, cday from Sousa (2010), cay^g from Guo (2006), and
cay^{unfi} from Kroencke (2017)) using PCE consumption, reporting coefficient estimates,
Newey-West t-statistics, and adjusted R2 statistics for both the Lettau and Ludvigson (2001)
sample period (through 1998:3) and the full sample (through 2019:4).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| BEA NIPA tables: nondurables and services (NDS) and PCE consumption; labor income | Consumption and income series for cay estimation | [FRED](/wiki/datasets/fred/) |
| Federal Reserve Financial Accounts (Flow of Funds) | Aggregate asset wealth series for cay estimation | [FRED](/wiki/datasets/fred/) |
| World Inequality Database (WID) / US Distributional National Accounts | Wealth and income shares of top 10% households (AS^top10, YS^top10) | no page yet |
| CRSP NYSE/NYSE MKT/NASDAQ/Arca Value-Weighted Market Index | Stock market excess return (dependent variable) | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| BLS nonfarm business labor share | Labor share for KS construction; interest rate proxy | [FRED](/wiki/datasets/fred/) |
| Federal Reserve H.15: 3-Month Treasury Bill rate | Risk-free rate proxy | [FRED](/wiki/datasets/fred/) |

Sample: quarterly, 1952:1-2019:4 (272 quarters). All nominal series deflated using the PCE deflator
from the BEA. Inequality shares from the World Inequality Database (Saez and Zucman (2016) series,
continued in the World Inequality Database (WID)).

## When to read the full paper

Read the [original](https://doi.org/10.1016/j.jempfin.2025.101668) if you are:
examining the structural stability of cay over time (Sections 4-4.2 and Fig. 5-7 give the fullest
treatment); comparing Welch and Goyal (2008) out-of-sample failure results against IS evidence in a
single paper; assessing whether adjusting cay for wealth inequality (Lettau et al. (2019) argument)
restores predictability; or using the Brennan and Xia (2005) look-ahead bias critique and want the
authors' detailed response. The Appendices (C-H) cover additional cointegration tests, the direct
multivariate regression, consumption predictability, OOS tests, alternative specifications, and
Covid-19 robustness.

## Attribution and rights

Source: peer-reviewed, *Journal of Empirical Finance* 85 (2026) 101668. This distillation was extracted
by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**.
The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Dauber, Moritz, and Jochen Lawrenz.
> "The Decay of cay."
> *Journal of Empirical Finance* 85 (2026): 101668.
> DOI: 10.1016/j.jempfin.2025.101668. © 2025 The Author(s).
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Mutual Fund Stars: Hounyo & Lin (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jef/2026/hounyo-mutual-fund-stars-pick-stocks-2026/
# Distilled: Hounyo and Lin identify a "duplicate observations" flaw in the Fama-French (2010) bootstrap for mutual fund performance tests and propose a wild bootstrap fix (CSDWB). Applied to U.S. equity mutual funds (1984-2019), CSDWB finds a measurable fraction outperform the market, concentrated before 2003. Journal of Empirical Finance 2026, paywalled. Six core results with source locators, datasets used, the regression framework, and the wild bootstrap method with its defining equations.
# Tags: paper-summary, mutual-funds, asset-pricing, performance-evaluation, bootstrap
==============================================================================

**What this is.** The paper's core results, the statistical model of fund performance, and the
cross-sectional dependent wild bootstrap (CSDWB) method with its defining equations: enough to
understand what was found and why existing bootstrap tests fail, without reading all 18 pages.
To replicate or extend, read the full source at
[DOI 10.1016/j.jempfin.2025.101673](https://doi.org/10.1016/j.jempfin.2025.101673).

## TL;DR

The Fama-French (2010) bootstrap test for mutual fund skill is severely undersized, primarily
because it jointly resamples factors and residuals, producing duplicate observation pairs in the
bootstrap sample. This distortion far exceeds the undersampling effect identified by Harvey and Liu
(2022) and persists even at large sample sizes. Hounyo and Lin propose the cross-sectional
dependent wild bootstrap (CSDWB), which multiplies cross-sectional residuals by random Rademacher
weights at each time point, preserving unbalanced panel structure and cross-sectional dependence
while eliminating duplicate observations. Simulations confirm CSDWB achieves near-optimal test
size under a range of cross-sectional dependence scenarios. Applied to 3,630 U.S. equity mutual
funds from January 1984 to May 2019, CSDWB finds that a measurable fraction of funds outperform
the market on the four-factor Carhart (1997) model, with outperformance concentrated before 2003
and absent afterwards.

## Core results

Magnitudes are as reported; `\*` = significant at the 5% level (likelihood exceeds 0.95).
Locators point to tables, figures, and sections of the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | FF joint resampling generates far more extreme bootstrap t-statistics than Fixed (no-duplicate) resampling, even at T=60, confirming duplicate observations (not undersampling) as the dominant distortion | Fig. 1, p. 4 | Joint & Complete bootstrap right-tail density is visibly larger than Fixed & Complete at both T=13 and T=60; the gap persists as T increases, showing duplicate-observation inflation does not vanish with sample size |
| R2 | CSDWB_I and CSDWB_II have near-optimal test size (~10%) at all percentiles across three cross-sectional dependence scenarios; FF is oversized at the max percentile and HL methods are unreliable for small T | Figs. 7-9, pp. 11-12 | CSDWB_I and CSDWB_II size ≈10% at max, 99.5, 99, 95, 90th percentiles under no, weak, and strong cross-sectional dependence; HL_II size diverges from 10% at small T |
| R3 | CSDWB detects genuine mutual fund outperformers across the performance distribution; the FF method misses them entirely | Table 2, p. 15 | CSDWB_I: max 0.915, 2nd 0.989\*, 5th 0.992\*, 99.9 0.988\*, 99.5 0.972\*, 99th 0.920; FF: max 0.154, 2nd 0.330, 5th 0.568; KTWW: max 0.424, 2nd 0.806, 5th 0.982\* |
| R4 | Outperformance is concentrated entirely in the pre-2003 period; no evidence of outperformance after 2003 | §4.1, p. 13-14; Table E.3 | Pre-2003 subperiod: strong evidence of outperformance using CSDWB; post-2003: no evidence; pattern consistent across bootstrap methods |
| R5 | Best-performing funds portfolio (238 funds, selected via MCS-based Algorithm 1 using 1984-1993 data) earns cumulative excess returns of 8.91% above the market over 4 years and 15.12% over 9 years | §4.2, p. 15; Fig. 11, p. 16 | Superior performance significant at 10% level for first 4 years; advantage fades over longer horizons; most notable in 1998, when the portfolio briefly underperforms the market |
| R6 | Time-varying (nonparametric) alpha for top-percentile funds is stable pre-2005 near the upper bootstrap confidence bound; post-2005 alpha becomes volatile and approaches the lower bound | §4.3, p. 16; Fig. 12, p. 17 | 99.5th percentile fund alpha ≈ 0.05-0.25% per month pre-2005; post-2005 widening confidence interval and alpha near lower bound suggest reduced and uncertain skill |

**Overall (paper's conclusion).** Duplicate observations in the Fama-French (2010) bootstrap
are the dominant source of its severe undersizing, not the undersampling problem previously
emphasized by Harvey and Liu (2022). The proposed CSDWB methods correct this flaw and confirm
that a small but measurable fraction of U.S. equity mutual funds outperform the market, with
outperformance concentrated in the pre-2003 period. Outperforming funds tend to have lower
expense ratios, higher total net assets, and higher turnover, consistent with active management
in a competitive environment before low-cost ETFs and product proliferation compressed alpha
opportunities post-2003.

## Theory / model

The paper has no structural economic model. The statistical framework is a four-factor panel
regression for fund excess returns with the null hypothesis of zero true alpha.

Huang et al. (2023) analyzed mutual fund performance using a bootstrap approach but used the FF method to generate simulated data, embedding the same duplicate-observations distortion into the DGP. The current paper shows that a valid bootstrap method applied to FF-generated data will detect excess extreme t-statistics, suggesting the presence of outperformers even under the null, and proposes CSDWB as the correct alternative.

The benchmark regression (equation 2.1, p. 2) for fund $$i$$ at time $$t$$ is:

$$
r_{it} = a_i + \sum_{k=1}^{K} \beta_{ik} F_{kt} + \varepsilon_{it}, \tag{2.1}
$$

where $$r_{it}$$ is fund $$i$$'s excess return over the risk-free rate, $$a_i$$ is fund $$i$$'s
true abnormal return (alpha), $$\beta_{ik}$$ is the factor loading on factor $$k$$, $$F_{kt}$$
is the return on factor $$k$$ at time $$t$$, and $$\varepsilon_{it}$$ is the residual. The
factors $$F_{kt}$$ include the market excess return ($$RMRF_t$$), the size factor ($$SMB_t$$)
and value-growth factor ($$HML_t$$) from Fama and French (1993), and the momentum factor
($$MOM_t$$) from Carhart (1997).

The null hypothesis of no skill is:

$$
H_0 : a_i = 0, \quad \text{for } i = 1, \ldots, N. \tag{2.2}
$$

A positive $$a_i$$ indicates fund $$i$$ outperforms the market after adjusting for factor
exposures. The paper tests this null by comparing the distribution of estimated
$$\hat{t}_{(i)}$$ statistics (ranked order statistics) against a bootstrap-approximated
distribution under $$H_0$$, focusing on extreme-right-tail percentiles (99th, 99.5th, max)
following Kosowski et al. (2006) and Fama and French (2010).

The key insight is that the Fama-French (2010) implementation distorts the bootstrap
null distribution by jointly resampling factor returns and fund residuals, so the same
$$(r_{it}^{*}, F_t^{*})$$ pair may appear multiple times in a bootstrap sample, inflating
extreme t-statistics and reducing the likelihood of rejection.

## Method

The paper proposes two variants of a cross-sectional dependent wild bootstrap (CSDWB).
Both variants build on `panel-regression` (the Carhart regression framework, equation 2.1)
and the `wild-bootstrap` primitive (Wu 1986, Liu 1988, Mammen 1993).

**Bootstrap pseudo-returns under the null** (equation 2.3, p. 3): factor returns and residuals
are combined to produce bootstrap pseudo-returns with zero true alpha:

$$
r_{it}^{*b} = \sum_{k=1}^{K} \hat{\beta}_{ik} F_{kt}^{*b} + \varepsilon_{it}^{*b}, \tag{2.3}
$$

where $$\hat{\beta}_{ik}$$ is the estimated factor loading, and $$F_{kt}^{*b}$$ and
$$\varepsilon_{it}^{*b}$$ are bootstrap factor returns and residuals. The CSDWB methods
differ from Kosowski et al. (2006) and Fama and French (2010) in how these quantities
are generated.

**CSDWB_I** (the preferred variant, p. 8): Let $$T_i$$ denote fund $$i$$'s observed period
and $$o_{it} = 1$$ if $$t \in T_i$$, 0 otherwise. Let $$\eta_t^{*b}$$ be a random weight
with mean 0 and variance 1, drawn independently at each period $$t$$ (Rademacher weights:
$$\eta_t^{*b} = \pm 1$$ with probability 0.5 each, following Cameron et al. 2008 and
Davidson and Flachaire 2008). The CSDWB_I bootstrap residuals and factor returns are:

$$
o_{it} \varepsilon_{it}^{*b} = o_{it} \hat{\varepsilon}_{it} \cdot \eta_t^{*b}, \quad
F_t^{*b} = F_t \cdot \eta_t^{*b}.
$$

The same scalar weight $$\eta_t^{*b}$$ is applied to the entire cross-section of residuals
$$(\hat{\varepsilon}_{1t}, \ldots, \hat{\varepsilon}_{N_t t})$$ at each period $$t$$, preserving
cross-sectional dependence. Missing observations remain intact. Because each observation uses
a different realization of $$\eta_t^{*b}$$, no exact duplicate $$(\varepsilon_{it}^{*b}, F_t^{*b})$$
pairs are generated, eliminating the FF duplicate-observations distortion.

**CSDWB_II** differs from CSDWB_I solely in that factor returns are not perturbed: $$F_t^{*b} = F_t$$.
This makes CSDWB_II closely related to both the KTWW and CSDB methods and is useful for
elucidating differences across approaches.

**Fund selection algorithm** (Algorithm 1, p. 9): A Model Confidence Set (Hansen et al. 2011)
sequential procedure identifies the best funds. At each iteration, CSDWB is used to derive
a 90% confidence level for the performance of the current best fund relative to all other
funds in the active set $$\mathcal{M}_0$$. The worst performer is eliminated if the null of
equal performance is rejected. The procedure iterates until no further rejection occurs,
yielding a final set $$\hat{\mathcal{M}}^*_{1-\lambda}$$ of best-performing funds.

## Empirical specifications

**Data sample**: CRSP Survivor-Bias-Free U.S. Mutual Fund Database, 3,630 open-end U.S.
domestic equity funds, January 1984 to May 2019, 425 months. Factor returns from the Kenneth
French Data Library (four-factor model: RMRF, SMB, HML, MOM). The final dataset is an
unbalanced panel with missing observations reflecting fund entry and exit.

**Bootstrap implementation**: For each fund $$i$$, OLS estimates of model (2.1) yield
$$(\hat{a}_i, \hat{\beta}_{ik})$$ and residuals $$\hat{\varepsilon}_{it}$$. Bootstrap
pseudo-returns are generated by subtracting $$\hat{a}_i$$ and applying CSDWB weights.
Each bootstrap iteration produces a ranking of the $$N$$ fund t-statistics; the procedure
is repeated $$B = 999$$ times to build the empirical null distribution. The test focuses
on extreme right-tail percentiles (max, 2nd, 5th, 99.9th, 99.5th, 99th, 98th, 97th, 95th, 90th)
following Kosowski et al. (2006).

**Performance likelihood** (Table 2, p. 15): For each percentile, the reported likelihood is
the proportion of 10,000 bootstrap replications in which the simulated t-statistics at that
percentile fall below the observed t-statistic. A likelihood exceeding 0.95 (marked with `\*`)
provides strong evidence of genuine outperformance at that percentile.

**Simulation design** (Section 3, pp. 9-12): Three parametric scenarios (no, weak, strong
cross-sectional dependence; $$\xi_i = 0, 0.1, 0.4$$) with $$N = 3630$$ funds over $$T = 425$$
periods, 5% of funds having positive alpha, $$\alpha^+ = 30\%$$. An enhanced empirical
simulation uses CSDWB_I to generate data under the null (Section 3.2.2), avoiding the FF
duplicate-observations distortion in the simulated DGP.

**Subperiod analysis** (Section 4.1, §4.2): Pre/post-2003 split to test persistence.
Fund selection period: January 1984 to December 1993 (10 years of training data, 238 funds
selected via Algorithm 1). Out-of-sample evaluation: January 1984 to December 2002.

**Time-varying alpha** (Section 4.3, equation 4.1, p. 16): A local least-squares estimator
estimates the time-varying alpha $$a_{it}$$ in:

$$
r_{it} = a_{it} + \sum_{k=1}^{K} \beta_{ikt} F_{kt} + \varepsilon_{it}, \tag{4.1}
$$

using a rolling-window approach following Cai (2007) and Cai et al. (2018). Bootstrap
confidence intervals are constructed with CSDWB_I fixing $$\hat{a}_i = 0$$.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CRSP Survivor-Bias-Free U.S. Mutual Fund Database | Monthly excess returns for 3,630 open-end U.S. domestic equity mutual funds, 1984-2019 | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| Kenneth French Data Library | Four-factor model returns (RMRF, SMB, HML, MOM) used as benchmark factors in regression (2.1) | [Ken French library](/wiki/datasets/ken-french/) |

Sample: January 1984 to May 2019, N = 3,630 funds, T = 425 months (unbalanced panel). Data access
provided via the University at Albany Center for Institutional Investment Management (CIIM).

## When to read the full paper

Use the [original](https://doi.org/10.1016/j.jempfin.2025.101673) if you are: implementing a
bootstrap procedure for mutual fund or hedge fund performance in an unbalanced panel (the CSDWB
algorithm is fully described in Section 2.3 and Internet Appendices); diagnosing why the
Fama-French (2010) method undersizes (Sections 2.1-2.2 with the dominant-effect analysis);
comparing CSDWB against HL, KTWW, CSDB, and FF across simulation designs (Section 3);
or applying the MCS-based fund-selection Algorithm 1 to build a best-funds portfolio.

## Attribution and rights

Source: peer-reviewed, *Journal of Empirical Finance* 85 (2026) 101673. Paywalled; no
open-access or CC licence. This distillation was extracted by an LLM on 2026-06-25 and is
**not human-verified or independently reproduced**. Quotation of results is for scholarly
reference only (extract-only).

> Hounyo, Ulrich, and Jiahao Lin. "Can mutual fund 'stars' really pick stocks? New evidence
> from a wild bootstrap analysis." *Journal of Empirical Finance* 85 (2026): 101673.
> DOI: 10.1016/j.jempfin.2025.101673. © 2025 Elsevier B.V. All rights reserved.


==============================================================================
# Peer Effects in Financial Expectations: Thornton (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jef/2026/thornton-peer-effects-financial-expectations-2026/
# Distilled: Using the British Household Panel Survey and an instrumental variables strategy, Thornton (2026) provides causal evidence that neighborhood financial expectations positively influence individual financial expectations, with a one-standard-deviation peer effect equal to roughly 31% of the family effect in financial beliefs. Journal of Empirical Finance 2026, paywalled. Seven core results with source locators, datasets used, the identification strategy, and the empirical specifications.
# Tags: paper-summary, household-finance, expectations, beliefs, peer-effects, social-finance, panel-regression, panel-data, peer-reviewed, unreplicated, data:bhps
==============================================================================

**What this is.** The paper's core results, the identification strategy, and the empirical specifications: enough to know what was found and how, without reading all 22 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1016/j.jempfin.2026.101712).

## TL;DR

Using 18 waves of the British Household Panel Survey (BHPS, 1991-2008) and an instrumental variables strategy adapted from Brown et al. (2008), Thornton (2026) provides causal evidence that neighborhood financial expectations positively influence individual financial expectations. The instrument is the average financial expectations of neighbors' nonlocal family members, which affects a neighbor's beliefs through family interaction but has no direct path to the focal individual. A one-standard-deviation increase in neighborhood financial expectations leads to a 2.8% increase in individual financial expectations (IV estimate), equal to roughly 31% of the corresponding family effect. Peer effects are larger for socially connected individuals, grow with time spent in a neighborhood (consistent with social interaction rather than sorting), and are informative only in neighborhoods with diversity in financial expectations and uniformity in income and voting behavior. These findings support the social transmission frameworks of Burnside et al. (2016) and Han et al. (2020). Individuals also act on their expectations: those expecting financial improvement are less likely to save.

## Core results

Magnitudes and significance as reported; `\*\*`/`\*\*\*` = 5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Baseline OLS: neighborhood financial expectations are positively correlated with individual expectations | Table 2, col 1, p. 7 | FINEXn coef = 0.555\*\*\* (t=30.01) |
| R2 | With individual FE, year FE, and time-varying controls: large, significant peer effect remains | Table 2, col 2, p. 7 | FINEXn coef = 0.272\*\*\* (t=16.88); 1-SD in FINEXn = 2.6% increase in FINEX; roughly 28% as large as the family effect (9.2%) |
| R3 | IV causal estimate using nonlocal-family expectations as instrument | Table 6, col 2, p. 12 | FINEXn coef = 0.505\*\*\* (t=2.60); restricted to different-region nonlocal family: 0.380\*\*\* (t=4.07, col 4) |
| R4 | Reverse causality test: previous-neighborhood expectations do not predict current expectations | Table 7, col 2, p. 13 | FINEXprev coef = 0.0484 (t=1.33), not significant |
| R5 | Peer effects grow with time in neighborhood; income and voting similarity do not converge | Fig. 1, p. 9; Figs. 2-3, p. 10 | Coef rises from 0.105 (insig) for 0 years to 0.264 (t=14.3) for 3+ years in neighborhood; income and voting coefficients flat or decreasing |
| R6 | Expecting financial improvement is associated with a lower probability of saving | Table 8, col 2, p. 16 | FINEX coef on Save = -0.0165\*\*\* (t=-7.37) |
| R7 | Peer effects are larger for socially connected individuals | Fig. 4, p. 14; Figs. 5-7, pp. 15-16 | Daily-talker subsample: coef = 1.01 (t=2.10); not significant for less-frequent interactors; same pattern for neighborhood-likers (Fig. 5) and organization members (Fig. 6) |

**Overall (paper's conclusion).** Financial expectations are causally transmitted among neighbors through social interaction, with a magnitude equal to roughly 31% of the family effect. The evidence is consistent across panel FE, IV, and IV robustness specifications; inconsistent with homophily (peer effects grow while income and political similarity do not); and supported by sociability heterogeneity (socially connected individuals show stronger transmission). Individuals also act on these expectations: optimistic individuals save less.

## Theory / model

The paper does not propose a formal model. The central hypothesis is that an individual's financial expectations (FINEX: whether the individual expects to be better off, about the same, or worse off financially in the coming year) are influenced by the financial expectations of her neighbors through social interaction.

The identification challenge is the reflection problem of Manski (1993): when neighbors have similar expectations, this correlation could arise from (1) endogenous social effects (social interaction), (2) contextual effects (shared local environment), or (3) correlated effects (similar individual characteristics). The paper tests three hypotheses:

- **H1 (social interaction):** Individuals take neighborhood expectations into account when forming their own; the coefficient $$\beta_1$$ in equation (1) captures a causal peer effect.
- **H2 (homophily):** Individuals sort into neighborhoods with like-minded residents; any observed correlation reflects selection rather than transmission.
- **H3 (contextual / correlated):** A shared local environment (e.g. the local labor market) drives correlated expectations; there is no individual-level transmission.

H2 is tested via the time-in-neighborhood design (Section 4.2): if sorting drives the result, peer effects should be strongest when individuals first move (closest to their selection decision) and weaken thereafter as neighbors diverge from the mover's baseline. The opposite pattern is found. H3 is addressed by the IV strategy and by the fact that income and political-preference similarity within neighborhoods do not grow over time (Figs. 2-3 on p. 10).

The paper also tests a joint hypothesis (Section 4.8): that survey expectations reflect actual beliefs and that individuals act on them. If this joint hypothesis holds, individuals expecting improvement should save less. This is confirmed in Table 8.

## Method

The paper applies two estimators: panel OLS with individual and year fixed effects, and two-stage least squares (2SLS) with the nonlocal-family instrument.

**Panel fixed effects.** The main estimating equation (p. 6, Eq. 1) is:

$$
\text{FINEX}_{it} = \gamma_i + \gamma_t + \beta_1 \, \text{FINEXn}_{it} + \lambda_i + \varepsilon_{it} \tag{1}
$$

where $$\gamma_i$$ are individual fixed effects absorbing time-invariant characteristics (race, religion, baseline sociability), $$\gamma_t$$ are year fixed effects absorbing sample-wide trends, $$\text{FINEXn}_{it}$$ is the average financial expectation of individual $$i$$'s neighbors in year $$t$$ (excluding $$i$$), $$\lambda_i$$ is a vector of time-varying controls (income, education, marital status, vote intention, job industry), and $$\varepsilon_{it}$$ is clustered at the interview-area (neighborhood) level throughout.

**IV strategy.** The instrument for $$\text{FINEXn}_{it}$$ is the average financial expectation of neighbors' nonlocal family members, denoted $$f\!amFINEXn_{it}$$. This instrument builds on the Brown et al. (2008) design: nonlocal family members are likely to influence their relative's expectations through family interaction, but are not subject to the same local environment as the focal individual. The first-stage regression is:

$$
\text{FINEXn}_{it} = \gamma_i + \gamma_t + \delta \cdot f\!amFINEXn_{it} + \lambda_i + \nu_{it}
$$

The instrument is constructed in two ways: (a) nonlocal family = family members living outside the focal neighborhood; (b) nonlocal family = family members living in a different UK region (19 regions), the more demanding robustness specification. The first-stage $$t$$-statistic in the full-controls specification is 3.61 (Table 4, col 2, p. 10), exceeding the Lee et al. (2022) tF critical value of 3.02 at the 5% level ($$F$$-statistic = 13.03).

**Sociability subsamples.** To provide additional evidence for social interaction as the mechanism, the IV specification from Table 6, col 2 is re-run on subsamples split by four sociability proxies: frequency of talking with neighbors (FRNA), opinion of neighborhood (Lknbr), local organization membership (Org), and desire to move (Lkmove). This approach, similar to Hong et al. (2004), uses sociability variation to test whether more connected individuals exhibit larger peer effects.

## Empirical specifications

**Main panel OLS (R1, R2).** Equation (1) is estimated with no controls (Table 2, col 1, $$N$$ = 218,149) and with individual FE, year FE, and time-varying controls (col 2, $$N$$ = 207,362). Standard errors clustered at the neighborhood level throughout. FINEX is coded 1 (better off), 0 (same), -1 (worse off). The focal individual is excluded from the neighborhood average.

**IV 2SLS (R3).** Table 6 (p. 12) reports four 2SLS specifications. Columns (1)-(2) define nonlocal family as living outside the focal neighborhood; columns (3)-(4) restrict to a different UK region. Columns (1) and (3) include only wealth as a time-varying control; columns (2) and (4) add the full set of controls plus individual and year FE. The 2SLS coefficient is stable across specifications: from 0.612\*\*\* (col 1) to 0.380\*\*\* (col 4, most demanding).

**Reverse causality test (R4).** Table 7 (p. 13) regresses individual FINEX on average financial expectations in the individual's *previous* neighborhood (FINEXprev), using the same controls as Table 2, col 2. A significant coefficient would indicate that the IV result is capturing ties to the previous neighborhood rather than the current one. The coefficient is 0.0484 (t = 1.33), not significant.

**Time-in-neighborhood trend (R5).** Using the specification from Table 2, col 2, four subsample regressions are run on individuals grouped by years-in-neighborhood (0, 1, 2, 3+). The peer-effect coefficient grows monotonically (Fig. 1, p. 9). The same subsampling is applied with income and vote intention as outcomes (Figs. 2-3, p. 10); those coefficients show no convergence, ruling out general assimilation to the local environment as the driver.

**Savings regression (R6).** Table 8 (p. 16) regresses a binary savings variable (Save = 1 if the individual saved over the past year) on individual FINEX, with individual FE, year FE, and time-varying controls. Standard errors clustered at the neighborhood level.

**Sociability subsamples (R7).** The IV specification from Table 6, col 2 is re-run separately for each level of each sociability proxy (Figs. 4-7, pp. 14-16). Peer effects are statistically significant only for the most socially connected subgroup in each proxy (daily talkers: coef = 1.01, t = 2.10; neighborhood-likers: coef = 0.583, t = 2.86; organization members: coef = 1.10, t = 2.38; non-movers: coef = 0.72, t = 2.79).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| British Household Panel Survey (BHPS), waves 1-18, 1991-2008; UK Data Service SN 5151-2 | All financial expectation measures (FINEX, FINEXn, famFINEXn), sociability proxies (FRNA, Lknbr, Org, Lkmove), savings dummy (Save), neighborhood identifiers (interview area, IVIA), and demographic controls | no page yet |

Sample: 18 annual waves, approximately 10,000 initial participants (later adding subsamples in 1997 and 1999), 250 interview areas averaging 41 residents each. Main panel-OLS samples: 207,362 (full controls, Table 2 col 2) to 218,149 (baseline, Table 2 col 1) person-year observations.

## When to read the full paper

Read the [original](https://doi.org/10.1016/j.jempfin.2026.101712) if you are: studying how beliefs (rather than behavior) spread through social networks; replicating or extending the nonlocal-family IV strategy of Brown et al. (2008); analyzing how sociability moderates peer effects (Figs. 4-7); testing whether peer learning is informative only in specific neighborhood types (Figs. 9-11 on expectation, voter, and income polarization); or connecting financial expectations to household saving behavior.

## Attribution and rights

Source: peer-reviewed, *Journal of Empirical Finance* 87 (2026) 101712. Paywalled; all rights reserved, © 2026 Elsevier B.V. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. Extract-only; the verbatim PDF is not hosted.

> Thornton, Joshua. "Peer effects in financial expectations." *Journal of Empirical Finance* 87 (2026) 101712. DOI: 10.1016/j.jempfin.2026.101712.


==============================================================================
# Dynamic Competition in Negotiated Price Markets: Allen & Li (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/allen-dynamic-competition-negotiated-price-2025/
# Distilled: Using Canadian mortgage contract data, Allen and Li document an "invest-and-harvest" pricing pattern and build a structural dynamic model of price negotiation with search and switching frictions to quantify market frictions and study counterfactual policies. J. Finance 2025, CC BY-NC 4.0. Eight core results with source locators, datasets used, the model, and the estimation method.
# Tags: paper-summary, household-finance, mortgages, search-frictions, switching-costs
==============================================================================

**What this is.** The paper's core results, the dynamic structural model of
price negotiation with search and switching frictions, and the maximum
likelihood estimation procedure: enough to know what it found and how, without
reading all 54 pages. To replicate or extend it, read the full source at the
[original](https://doi.org/10.1111/jofi.13408).

## TL;DR

Using anonymized Canadian mortgage contract data linked with credit bureau
records (January 2014 to July 2019), Allen and Li document an "invest-and-harvest"
pricing pattern: lenders charge loyal renewers 4 bps more than new
borrowers, while switchers pay 7.9 bps less than stayers. To explain these
patterns, they build a T-period dynamic game of price negotiation in which
borrowers face search and switching frictions and lenders are forward-looking,
competing aggressively ex ante to build a customer base (invest) and
extracting rents later (harvest). Structural estimation yields average
search costs of \$386, switching costs of \$462 (new) and \$829 (renewers), and
lender investment incentives of \$843 per borrower. Counterfactual experiments
show that dynamic competition attenuates the anticompetitive effects
of market frictions relative to static model predictions, and that the
mortgage stress test unintentionally distorts lender pricing strategies.

## Core results

Magnitudes and significance are as reported. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Invest-and-harvest pricing documented**: loyal renewers pay 4.0 bps more than new borrowers; switchers pay 7.9 bps less; only 9.55% of renewers switch | Table I, p. 569 | Loyalty premium = 8.53 bps; switching rate at renewal = 9.55% |
| R2 | **OLS confirms invest-and-harvest**: loyal renewal dummy +4.00 bps; switch renewal dummy -3.93 bps vs purchase baseline, controlling for borrower and contract characteristics | Table II col (1), p. 570 | R-squared = 0.72; bond rates, FSA house prices, lender FE, year and region FE included |
| R3 | **Structural estimates**: average per-borrower search cost \$386 (1.3% of interest cost); average switching cost \$462 (new), \$829 (renewers); average lender investment incentive \$843; lender annual discount factor 0.93 | Table III, p. 587; Table IV, p. 588 | Likelihood ratio test rejects static model (delta=0) at 0.1% level; LR statistic = 91.39 |
| R4 | **Dynamic model predicts smaller consumer gains from removing frictions than static model**: removing both frictions saves borrowers 1.4% at origination and 2.5% at renewal (dynamic); static model overpredicts at 3.0% and 4.6% | Table VIII Panel C, p. 597 | Joint removal saves 58% (30%) more than sum of individual effects at origination (renewal): interaction effect |
| R5 | **Removing switching costs alone hurts new borrowers** (+0.2% total cost) because lender investment incentives fall; renewers benefit (-0.6%); static model always predicts savings for both | Table VIII col (2), p. 597 | Investment incentive V drops to \$0.253 from \$0.806 at origination when switching costs removed |
| R6 | **Investment incentive is driven by both frictions**: average V = \$843; increasing in switching cost; effect of switching cost on V is stronger when search costs are high | Table IV, p. 588; Table V col (2), p. 590 | Interaction coefficient (next-period switching cost x next-period search cost) = 0.39 (SE 0.0038), R-squared = 0.96 |
| R7 | **25-year FRMs benefit all players in the dynamic model**: borrowers save \$344 in total cost; lenders earn \$423 more over 25 years; static model reverses sign, predicting lenders earn \$1,163 less | Table IX, p. 602 | Dynamic model: interest rate indifference spread = 1.79 bps; static = 10.14 bps; both far below observed 91 bps swap-adjusted spread |
| R8 | **Mortgage stress test at renewal unintentionally distorts pricing**: unqualified borrowers (10.2%) pay 22 bps more and incur 7.8% higher total costs; switching rate falls from 20.2% to 3.0%; lender investment incentive at origination quadruples to \$3,967 | Table XI, p. 606; Table XII, p. 607 | Anticipatory effect lowers origination rates 4.6% for affected borrowers but share of borrowers in financial distress (GDS>39%) rises from 4.69% to 5.81% |

**Overall (paper's conclusion).** Dynamic competition attenuates the
anticompetitive effects of search and switching costs: each friction is less
harmful when the other is also present (interaction effect), and lenders'
forward-looking investment incentives lower current prices more than a static
model predicts. A static framework therefore systematically overpredicts
consumer gains from removing frictions and misidentifies who benefits from
policy changes. The invest-and-harvest pattern is consistent with
Dube, Hitsch, and Rossi (2009) and Shcherbakov (2016), who find switching
costs can lower equilibrium prices in dynamic frameworks. The paper extends
Allen, Clark, and Houde (2019) (origination-only) to a full repeated-interaction
setting. Identification uses an English auction approximation following
Woodward and Hall (2012) and Allen, Clark, and Houde (2019). The framework for
jointly modelling search and switching follows Honka (2014) but adds
supply-side forward-looking responses and negotiated pricing. The stress-test
results are analogous to Agarwal et al. (2023a), who document that HARP
unintentionally strengthened incumbency advantages in U.S. mortgage refinancing.

## Theory / model

Consider a borrower seeking a mortgage with a fixed rate for $$m$$ years,
amortizing over $$T \times m$$ years. The game has $$T$$ periods. In each period
$$t$$, the home bank $$h^t$$ (the lender from the previous period, or a
premortgage-relationship lender at $$t=1$$) moves first. The period is divided
into two stages: (i) an initial quote stage, and (ii) a negotiation stage.

**Borrower preferences** (p. 574, eq. 1). The borrower chooses the lender
$$j$$ from choice set $$n^t$$ that maximizes expected present value:

$$
\max_{j \in n^t} \; v^t_j - p^t_j + \rho U^{t+1}_j, \tag{1}
$$

where $$v^t_j$$ is the borrower's valuation for lender $$j$$'s mortgage,
$$p^t_j$$ is the interest payment, $$\rho$$ is the borrower's discount factor,
and $$U^{t+1}_j$$ is the continuation value of being attached to lender $$j$$
next period. Products are homogeneous except for a switching disutility $$\lambda^t$$:

$$
v^t_j = \begin{cases} \bar{v}^t, & j = h^t \\ \bar{v}^t - \lambda^t, & \text{otherwise.} \end{cases}
$$

**Lending costs** (p. 575). Lender $$j$$'s cost in period $$t$$ is:

$$
c^t_j = \begin{cases} c^t, & j = h^t \text{ in initial quote stage} \\ c^t + \omega^t_j, & \text{otherwise,} \end{cases}
$$

where $$c^t \sim F(\cdot)$$ is the common funding cost (observable to lenders
but not the econometrician), and $$\omega^t_j \sim G(\cdot)$$ is a mean-zero
IID idiosyncratic match-value component drawn at the negotiation stage.

**Negotiation stage: English auction** (p. 577, eqs. 2-5). With $$n^t \geq 2$$
lenders in the choice set, lenders compete in a descending procurement auction.
Lender $$j$$ stays in the auction so long as the present value of winning exceeds
the present value of losing (p. 577, eq. 2):

$$
\bar{b}^t - (c^t + \omega^t_j) + \delta W^{t+1}_j \geq \delta L^{t+1}_j, \tag{2}
$$

where $$\delta$$ is the lender discount factor, $$W^{t+1}_j$$ is the continuation
value of winning, and $$L^{t+1}_j$$ is the continuation value of losing. The
weakly dominant strategy is to bid one's reservation value. Lender $$j$$'s
equilibrium drop-out bid (p. 577, eq. 3) is:

$$
b^t_j(c^t, \omega^t_j) = c^t + \omega^t_j - \delta(W^{t+1} - L^{t+1}). \tag{3}
$$

The term $$V^{t+1} \equiv W^{t+1} - L^{t+1}$$ is the net continuation value (investment
incentive): lenders bid below cost in period $$t$$ to secure the incumbency
advantage in period $$t+1$$. The equilibrium price given the state vector $$s^t$$
is (p. 578, eq. 5):

$$
p^{t*}(s^t) = \begin{cases}
c^t - \delta V^{t+1} + \omega^t_{(2)} + \lambda, & \omega_{h^t} - \lambda = \omega^t_{(1)} \\
c^t - \delta V^{t+1} + \omega^t_{(2)}, & \omega_{h^t} - \lambda \leq \omega^t_{(2)},
\end{cases} \tag{5}
$$

where $$\omega_{h^t}$$ is the home bank's idiosyncratic match value and
$$\omega^t_{(k)}$$ denotes the $$k$$th-order statistic among the rival banks'
adjusted costs. The home bank charges more when it ranks first in expected
utility (first line), capturing the switching cost $$\lambda$$ as a rent.

**Initial quote stage** (p. 578-580, eqs. 6-12). Given the home bank's initial
offer $$p^t_0$$, the borrower's expected gain from searching $$l$$ lenders
(relative to accepting) is:

$$
\Delta^t_l = \begin{cases}
0, & l = 1 \\
p^t_0 - \lambda - (c^t - \delta V^{t+1} + E[\omega^t_{(2)} | n^t = l]), & l = 2, 3, \ldots, N,
\end{cases} \tag{6}
$$

and the expected marginal benefit of adding lender $$l$$ to the choice set is
$$\kappa^t_l \equiv \Delta^t_l - \Delta^t_{l-1}$$ (eq. 7). The borrower chooses
choice set size to maximize net expected benefit of searching (eq. 8):

$$
n^t = \text{argmax}_l \; \Delta^t_l - (l-1)\kappa^t, \quad l = 1, 2, \ldots, N. \tag{8}
$$

The home bank anticipates search probabilities and sets the optimal initial
quote $$p^{t*}_0$$ to maximize expected profit (eq. 11-12):

$$
\underbrace{p^{t*}_0 - c^t + \delta W^{t+1}}_{\text{Profit from } p_0}
= \underbrace{E[\pi^{t*}_h | n^t = 2]}_{\text{Auction profit}}
+ \underbrace{\Delta^{t*}_2}_{\text{Rent from } H(\cdot)},
\tag{12}
$$

showing that the optimal initial price equals the expected auction profit plus
the rents the home bank can extract from search frictions.

**Continuation values** (p. 581-582, eqs. 13-15). The investment incentive
$$V^t = W^t - L^t$$ is determined by the search cost distribution $$H(\cdot)$$,
the idiosyncratic cost distribution $$G(\cdot)$$, the switching cost $$\lambda$$,
and the number of available lenders $$N$$:

$$
V^t = [1 - H(\Delta^{t*}_2)]\bigl(\Delta^{t*}_2 + E[\max\{\omega_{-h^t} - (\omega_{h^t} - \lambda), 0\} | n^t = 2]\bigr) + \sum_{l=2}^N \Pr(n^t = l) E[\max\{\omega_{-h^t} - (\omega_{h^t} - \lambda), 0\} | n^t = l] - \sum_{l=2}^N \Pr(n^t = l) \frac{l-1}{N-1} E[\max\{\omega^t_{-j} - \omega^t_j, 0\} | n^t = l]. \tag{15}
$$

Since $$V^t$$ depends only on the distributions $$H(\cdot)$$, $$G(\cdot)$$, $$\lambda$$,
and $$N$$, which are assumed time-invariant, $$V^t$$ does not depend on future
continuation values, greatly simplifying the solution.

**Equilibrium** is a Markov perfect equilibrium: (i) the home bank sets
$$p^{t*}_0$$ to maximize expected profit; (ii) the borrower sets $$n^t$$ to maximize
net search benefit; (iii) lenders in the choice set bid $$b^t_j(\cdot)$$ as in
eq. (3).

## Method

The model is estimated by maximum likelihood on a sample of 34,554 Canadian
mortgage contracts. The method builds on `blp-demand` (parametric demand)
and `search-bargaining-otc` (price as second-order statistic from an English
auction). The structural approach is required because search decisions are
unobserved: only the final contract rate, the home bank's identity, and the
switching decision are observed.

**Parametric specification** (p. 583-584). Per-unit common cost $$c^t_i$$ is
drawn from a Normal distribution $$N(\mathbf{x}_i^t \boldsymbol{\beta}, \sigma_c^2)$$,
where $$\mathbf{x}_i^t$$ includes credit score, outstanding balance, bond rate,
amortization, income, house price, and year/region fixed effects. The
idiosyncratic cost for lender $$j$$ in the negotiation stage is $$M_i^t \omega_{i,j}$$,
where $$\omega_{i,j} \sim \text{T1EV}(\gamma \sigma_\omega, \sigma_\omega)$$ (Type 1 Extreme Value).

Search costs follow an exponential distribution with a mean determined by
the borrower's age, credit score, and FSA-level income (p. 584):

$$
H_i(\kappa) = 1 - \exp\!\left(-\frac{\kappa}{\alpha_i}\right), \quad
\alpha_i = \exp(\alpha_0 + \alpha_{\text{credit}} \, \text{Credit}_i + \alpha_{\text{inc}} \, \text{Income}_i + \alpha_{\text{age}} \, \text{Age}_i).
$$

Switching costs are a linear function of borrower type, origination amount,
age, credit score, and income (p. 584):

$$
\lambda_i = \lambda_0 + \lambda_{\text{new}} + M_i^1 \times (\lambda_{\text{credit}} \, \text{Credit}_i + \lambda_{\text{inc}} \, \text{Income}_i + \lambda_{\text{age}} \, \text{Age}_i).
$$

The vector $$(\sigma_c, \sigma_\omega, \delta, \boldsymbol{\alpha}, \boldsymbol{\beta}, \boldsymbol{\lambda})$$
is estimated by maximizing the likelihood of observed switching decisions and
interest rates given the equilibrium of the model. The likelihood ratio test
rejects the static model ($$\delta = 0$$) at the 0.1% significance level
(LR statistic = 91.39; Table III, p. 587).

**Identification** (p. 584-585). Search and switching frictions are separately
identified because they have different effects on the correlation between the
number of lenders ($$N$$) and switching probability. Search costs reduce
pass-through from $$N$$ to the number of quotes $$n$$; switching costs do not.
Variation in $$N$$ across local markets (FSA-level) and variation in borrower
characteristics that differentially predict search versus switching costs
enable separation of the two frictions. The lenders' discount factor $$\delta$$
is identified by the relationship between amortization period and price
(p. 585-586): longer amortization implies larger outstanding balance at
renewal and hence stronger investment incentives.

## Empirical specifications

The paper combines descriptive reduced-form evidence with structural estimation.

**Descriptive evidence (R1, R2).** OLS regressions of mortgage rates and
switching decisions on borrower/contract characteristics (Table II, p. 570),
with bond rates, FSA house prices, transaction volume, lender fixed effects,
and year and region fixed effects as controls. The estimating equation for
rates is:

$$
\text{Rate}_{it} = \alpha + \mathbf{x}_{it} \boldsymbol{\gamma} + \text{BondRate}_t + \text{HousePrice}_{fsa} + \text{LenderFE} + \text{Year} + \text{RegionFE} + \varepsilon_{it},
$$

with standard errors clustered at the FSA level. For switching probability,
the linear probability model replaces rate as the outcome.

**Structural estimation and model fit (R3, R6).** Maximum likelihood
estimation over 34,554 observations. Parameters estimated: $$(\sigma_c, \sigma_\omega, \delta, \boldsymbol{\alpha}, \boldsymbol{\beta}, \boldsymbol{\lambda})$$. Model fit is assessed by simulating 1,000 samples of 34,554 borrowers from the benchmark and single-friction models and comparing the distribution of predicted switching probabilities and interest rates to the data (Figure 4, p. 595). The benchmark dual-friction model reproduces data patterns; single-friction models fail (Table VII, p. 593).

**Counterfactual exercises (R4, R5, R7, R8).** Simulate 100,000 borrowers
from the estimated dynamic model under alternative market structures (no
switching cost; no search cost; no frictions; 25-year FRM; mortgage stress test).
For each, solve the equilibrium home bank offer, borrower search decision, and
auction outcome; compute total financing cost (interest plus search/switching
costs incurred). Panel C of Table VIII (p. 597) sums origination and renewal
costs to obtain the lifetime comparison.

**Reduced-form validation of mechanisms (R6).** OLS of the investment
incentive on model-estimated search costs, switching costs, and their
interaction (Table V, p. 590), using the full final sample of 34,554 contracts.
Column (3) shows interest rate is decreasing in the estimated investment
incentive (-7.74 per unit, SE 1.09), consistent with the forward-looking
pricing mechanism.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| TransUnion credit bureau data (Canada) | Monthly credit bureau records for Canadian population: borrower characteristics (age, credit score, address at FSA level, nonmortgage debt), mortgage identity, switching activities, financial inquiries | no page yet |
| OSFI / federally regulated lender administrative data | Contract-level mortgage information: lender identity, loan size, funding date, monthly payment, outstanding balance, mortgage rate, amortization, LTV, total debt-servicing ratio | no page yet |
| 2016 FSA-level demographic data | Population and average household income at the forward sortation area (FSA) level | no page yet |
| Teranet quarterly FSA-level house price index | Local house price controls | no page yet |

Sample: January 2014 to July 2019 (cross-section of new borrowers and first-time renewers). Final estimation sample: 34,554 contracts (17,277 purchase, 15,627 loyal renewal, 1,650 switch renewal). Restricted to insured FRM-5Y contracts, excluding broker transactions, movers, and contracts without matching administrative data.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13408) if you are:
studying mortgage market competition, search and switching frictions in
negotiated-price markets, or the welfare implications of market frictions
in the presence of forward-looking firms; extending the framework to markets
with endogenous refinancing (U.S.) or broker intermediation; or designing
counterfactual macroprudential policy experiments using a structural dynamic
model. The locators above point to the exact tables and figures for
each result.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(1). This distillation
was extracted by an LLM on 2026-06-06 and is **not human-verified or
independently reproduced**. The CC BY-NC 4.0 licence permits noncommercial
reproduction; the verbatim PDF is not hosted in this batch.

> **Citation.** Allen, Jason, and Shaoteng Li.
> "Dynamic Competition in Negotiated Price Markets."
> *The Journal of Finance* 80, no. 1 (February 2025): 561-614.
> DOI: 10.1111/jofi.13408. © 2024 The Author(s).
> Licensed under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/).
> This page is an extract-only distillation by the Institute for Automated Research;
> it does not reproduce the full text and is for noncommercial research purposes only.


==============================================================================
# Superstar Returns: Amaral, Dohmen, Kohl & Schularick (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/amaral-superstar-returns-spatial-heterogeneity-2025/
# Distilled: Large metropolitan areas earn about 95 to 100 basis points less per year in total housing returns than the rest of the same country, driven by persistently lower rental yields that more than offset their well-known capital gain advantage. The return gap is rationalized as compensation for higher idiosyncratic and covariance risk in smaller, less liquid housing markets. J. Finance 2025, CC BY 4.0. Seven core results with source locators, datasets used, the theoretical framework (CAPM-style covariance pricing plus idiosyncratic risk), and the empirical specifications.
# Tags: paper-summary, housing, real-estate, asset-pricing, spatial-economics
==============================================================================

**What this is.** The paper's core results, the model it builds on (CAPM-style covariance pricing of housing risk), and the empirical design with the estimating specifications: enough to know what it found and how, without reading all 38 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13479).

## TL;DR

The paper introduces the first comprehensive city-level data set on housing returns for 15 OECD countries over up to 150 years (27 cities, balanced panel after 1950) and documents a new stylized fact: large metropolitan areas earn persistently lower total returns on housing than the rest of the country, roughly 95 to 100 basis points per year. Higher capital gains in large cities are more than offset by lower rental yields. This finding is confirmed for all 316 U.S. MSAs (1950-2018) and 42 West German cities (1975-2018). The paper rationalizes the pattern in a rational-expectations equilibrium: large cities are safer investments, with lower covariance between housing returns and local income growth and lower idiosyncratic price risk (because their housing markets are more liquid). Smaller, riskier locations must offer higher expected returns to attract capital.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Large cities earn persistently lower total housing returns** than the national portfolio and the rest of the country | Table III Panel A, p. 3072 | 27-city average difference vs. national: -0.95\*\*\* log points/year (SE 0.23); vs. rest-of-country: -1.04\*\*\* (SE 0.26); N = 1,767 |
| R2 | **Rental returns are the driver**: large cities have significantly lower rental yields that more than offset their capital gain advantage | Table III Panel A, p. 3072 | Rental return difference vs. national: -1.39\*\*\* (SE 0.04); vs. rest-of-country: -1.65\*\*\* (SE 0.05) |
| R3 | **Capital gains are higher in large cities**, but not enough to reverse the total return gap | Table III Panel A, p. 3072 | Capital gain difference vs. national: +0.43\* (SE 0.23); vs. rest-of-country: +0.61\*\* (SE 0.26) |
| R4 | **U.S. MSA evidence confirms the pattern**: large vs. small MSAs differ by -0.80 log points/year in total returns | Table V, p. 3077 | Large (top 5%) vs. rest: -0.52\*\*\* (SE 0.15); large vs. small (bottom 5%): -0.80\*\*\* (SE 0.20); N = 2,184 |
| R5 | **German city evidence confirms the pattern**: large vs. small German cities show the same negative return gap | Table VII, p. 3081 | Large (top 5%) vs. rest: -0.45\* (SE 0.25); large vs. small: -0.57\* (SE 0.35); N = 1,848 |
| R6 | **Covariance risk is lower in large cities**: housing return-income growth covariance is significantly smaller in large MSAs than in small ones | Table VIII row 1, p. 3086 | Large vs. rest: -0.55\*\* (SE 0.273, x10,000); large vs. small: -1.94\*\*\* (SE 0.573); N = 316 MSAs |
| R7 | **Idiosyncratic risk and illiquidity are higher in small cities**: sales-level idiosyncratic risk is roughly 25% higher in smallest MSAs; time on market is ~30 days longer | Figure 7 p. 3088; Table IX p. 3090 | Idiosyncratic risk: smallest MSAs 12.34%, largest 9.28% of sales price; time on market: large vs. small = -29.67\*\*\* days (SE 9.918) |

**Overall (paper's conclusion).** The conventional wisdom that large superstar cities are the best places to invest in residential real estate is overturned once rental yields are included. Capital gains favor large cities, but rental yields are persistently lower there, so total returns are lower by roughly 1 percentage point per year. This overturns the view in Demers and Eisfeldt (2022), who focused on local-market price volatility and found no significant return differences across U.S. MSAs (their sample excluded the 1970s period of high rental yields). The finding is also consistent with evidence on within-city return gradients, and extends the national housing return evidence of Jorda, Schularick & Taylor (2019) to the city level. The equilibrium framework follows the logic of Piazzesi and Schneider (2016): large cities are diversified, liquid, and less exposed to local income shocks, so investors require a lower risk premium. Riskier smaller locations must compensate with higher total expected returns.

## Theory / model

The paper has no new formal theory but applies the standard consumption CAPM pricing equation to housing markets. For a utility-maximizing household the excess expected log return on a housing asset equals a risk-premium proportional to the asset's covariance with consumption growth (p. 3085, equation 4):

$$
\ln \mathbb{E}[R_{t+1}] - \ln R_f = \gamma \, Cov\!\left[\ln\!\left(\frac{C_{t+1}}{C_t}\right),\, \ln R_{t+1} - \ln R_f\right] \tag{4}
$$

where $$R_{t+1}$$ is the total return, $$R_f$$ the risk-free rate, $$\gamma > 0$$ the coefficient of risk aversion, and $$C_{t+1}/C_t$$ is consumption growth (approximated by regional income growth due to data availability). An asset with higher co-movement with future income provides less insurance against income shocks and must earn a higher return. If large cities have lower covariance with local income, they are safer hedges and should earn lower returns.

A second channel is idiosyncratic risk. Because housing is large, indivisible, and illiquid, and most households own a single property, idiosyncratic price shocks are not diversified away. As Giacoletti (2021) and Sagi (2021) show, idiosyncratic volatility is priced in housing and is directly related to housing market illiquidity. The paper thus tests two predictions: (i) covariance of housing returns with income is lower in large cities, and (ii) idiosyncratic housing price risk is lower in large (more liquid) cities.

The equilibrium identity linking national, city, and rest-of-country returns through population weights (equation 3, p. 3069) is:

$$
\text{National return}_t = w_{t-1} \times \text{Large-city return}_t + (1 - w_{t-1}) \times \text{RoC return}_t \tag{3}
$$

where $$w$$ is the relative weight of the large city in the national housing series. This allows rest-of-country (RoC) returns to be backed out from national series and city-level data.

Total housing return decomposition, used throughout (equation 1, p. 3065):

$$
\text{Total return}_t = \underbrace{\frac{P_t - P_{t-1}}{P_{t-1}}}_{\text{Capital gain}} + \underbrace{\frac{R_t(1-c)}{P_{t-1}}}_{\text{Net rent return}} \tag{1}
$$

where $$P_t$$ is the house price at time $$t$$, $$R_t$$ the gross rent, and $$c$$ the share of operating costs in gross rent. Rental return series are constructed using the rent-price approach (equation 2, p. 3065):

$$
\frac{RI_{t+1}}{HPI_{t+1}} = \left(\frac{RI_{t+1}/RI_t}{HPI_{t+1}/HPI_t}\right) \frac{RI_t}{HPI_t} \tag{2}
$$

anchored to 2018 MSCI benchmark rent-price ratios and extrapolated backward using the ratio of the rent index to the house price index growth rates.

## Method

The paper's methodological contribution is data construction rather than a new estimator. The estimation strategy uses two variants:

**Paired t-tests (main international analysis, Tables III and IV).** For each of the 27 city-country pairs, the paper computes the average annual city-level log return and the average national return for the same country and period (post-1950 for the main analysis). Differences are paired by country. Standard errors are reported in parentheses. The null is equality of city and national means. This is not an OLS regression with fixed effects; it is a direct comparison of city-level to population-weighted national averages, with the rest-of-country (RoC) series derived by subtracting the large-city contribution from the national series.

**Random effects panel regression (U.S. MSA and German city sub-samples, Tables V-VII).** For the cross-section of cities within a country, the paper estimates:

$$
y_{s,t} = \alpha + \beta \cdot \mathbf{1}[\text{large city}]_s + \lambda_t + \varepsilon_{s,t}
$$

where $$y_{s,t}$$ is the dependent variable (log capital gain, log rental yield, or log total return) for city (MSA) $$s$$ in decade $$t$$, $$\mathbf{1}[\text{large city}]_s$$ is a dummy for being at or above the 95th percentile of the city population distribution, and $$\lambda_t$$ are year fixed effects. Standard errors are clustered at the city level. This is the `panel-regression` estimator from the registry.

**Idiosyncratic risk estimation (equation 5, p. 3087):**

$$
\Delta p_{l,i,t} = \Delta v_{l,t} + BX_i + \sigma_{l,\text{idiosyncratic}}\, \varepsilon_{i,t} \tag{5}
$$

where $$\Delta p_{l,i,t}$$ is the log capital gain on sale of property $$i$$ in MSA $$l$$ at time $$t$$, $$\Delta v_{l,t}$$ is the growth in local county house prices (FHFA/Zillow), $$BX_i$$ is a vector of house and transaction characteristics (including zip-code and time fixed effects), and $$\sigma_{l,\text{idiosyncratic}}$$ is the standard deviation of the sales-specific shock, estimated as the residual component. The estimator follows Giacoletti (2021).

## Empirical specifications

**Main specification (international panel, R1-R3).** The comparison is between the 27-city population-weighted average and the national housing portfolio return (Jorda, Schularick & Taylor (2019), extended to 2018), and separately the rest-of-country (RoC) portfolio. Paired t-tests on the time-series average of the annual city minus national (or city minus RoC) return difference. Sample: 1950-2018 (N=1,767 city-year obs). Standard errors computed from the paired differences. The result: -0.95\*\*\* log points per year (SE 0.23) total return gap vs. national; -1.04\*\*\* (SE 0.26) vs. RoC (Table III Panel A, p. 3072).

**U.S. MSA sub-sample (R4).** Gyourko, Mayer & Sinai (2013) extended to 2018 via American Community Survey data. 316 MSAs, decadal frequency 1950-2018. Random effects panel regression of log capital gain, log rental yield, and log total return on a large-MSA dummy (top 5% by 1950 population) and decade fixed effects; SE clustered at the MSA level. For total returns: large vs. rest = -0.52\*\*\* (SE 0.15), large vs. small = -0.80\*\*\* (SE 0.20); N=2,184 (Table V, p. 3077).

**German cities sub-sample (R5).** 42 West German cities 1975-2018 from Immobilienverband Deutschland (IVD) market reports. Random effects panel regression analogous to U.S. specification. Large cities = top 5% by 1975 population. Total return: large vs. rest = -0.45\* (SE 0.25), large vs. small = -0.57\* (SE 0.35); N=1,848 (Table VII, p. 3081).

**Covariance risk test (R6).** MSA-specific covariances computed as $$Cov_s = Cov(R_s - R_f, y_s)$$ over decadal periods 1950-2018, where $$R_s$$ is total log housing return and $$y_s$$ is log income growth from U.S. Census. Cross-sectional regression of covariance on a large-MSA dummy and robust SE. Result: large MSAs have significantly lower covariance with income growth (Table VIII, p. 3086). The identification is descriptive: no causal design.

**Idiosyncratic risk test (R7).** Corelogic repeat-sales transaction data for 248 U.S. MSAs 1990-2020. Residual idiosyncratic risk estimated following Giacoletti (2021) (equation 5). Liquidity proxied by median time on market and asking price discount from Zillow.com for 277 MSAs 2012-2020 (Table IX, p. 3090).

**Robustness.** Results hold (i) pre- and post-1990 sub-periods (Table IV Panel B), (ii) with alternative rental yield benchmarks from Numbeo and country-specific sources (Table IV Panel A), (iii) excluding rent-control periods and different tax regimes, (iv) using nominal series, (v) with outlier exclusion (top/bottom 5% annual returns), and (vi) using annual Agorastos et al. (2024) hedonic series for 21 U.S. cities 1920-2006 (Table VI).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Authors' new city-level housing return database (27 cities, 15 OECD countries, 1870-2018) | Main international analysis: house price indices, rent indices, total returns | No page yet (hand-collected from yearbooks, tax records, notaries, newspapers, MSCI) |
| Gyourko, Mayer & Sinai (2013) MSA database extended to 2018 via ACS | U.S. cross-section analysis: 316 MSAs, decadal returns 1950-2018 | [Superstar Cities (GMS)](/wiki/datasets/gyourko-mayer-sinai/) |
| German IVD / RDM market reports (42 cities, 1975-2018; 127 cities, 1992-2018) | German city cross-section: annual housing returns | No page yet |
| Jorda, Schularick & Taylor (2019) macro-history database | National housing return benchmarks for 15 OECD countries | No page yet |
| MSCI real estate investor yield data | 2018 rental yield benchmark for main data set | [no page yet](/wiki/datasets/) |
| Corelogic deeds / repeat-sales transaction data (248 MSAs, 1990-2020) | Idiosyncratic housing price risk estimation | [CoreLogic](/wiki/commercial/corelogic/) (licensed) |
| FHFA county house price indices + Zillow.com MSA data (277 MSAs, 2012-2020) | Liquidity proxies (time on market, asking price discount) and county price controls | [Zillow research](/wiki/datasets/zillow/) |
| Agorastos, Gray, Lyons & Shertzer (2024) hedonic series (21 U.S. cities, 1920-2006) | Long-run robustness for U.S. sub-sample | No page yet |

Sample: main international analysis 1950-2018 (balanced panel), with some series starting 1870. U.S. MSA analysis decadal 1950-2018; German cities annual 1975-2018.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13479) if you are: studying within-asset-class return heterogeneity across locations; building spatial housing pricing models; replicating the long-run city-level return series for a specific country (the Data Appendix details all sources); extending the U.S. MSA or German city analysis; or examining the role of housing risk in household portfolio choice. Detailed robustness tables (rent control, alternative benchmarks, outlier exclusion) are in the Internet Appendix.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(5). This distillation was extracted by an LLM on 2026-06-05 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Amaral, Francisco, Martin Dohmen, Sebastian Kohl, and Moritz Schularick.
> "Superstar Returns? Spatial Heterogeneity in Returns to Housing."
> *The Journal of Finance* 80, no. 5 (October 2025): 3057–3094.
> DOI: 10.1111/jofi.13479. © 2025 The Author(s).
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Arbitrage Capital of Global Banks: Anderson, Du & Schlusche (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/anderson-arbitrage-capital-global-banks-2025/
# Distilled: The 2016 U.S. money market fund reform cut banks' unsecured wholesale funding by about $600 billion; global banks responded by cutting liquid arbitrage positions (IOER and CIP arbitrage), not loan supply. J. Finance 2025, U.S. Government work (public domain in USA). Eight core results with source locators, datasets used, the model, and the identification strategy.
# Tags: paper-summary, banking, wholesale-funding, arbitrage, money-markets
==============================================================================

**What this is.** The paper's core results, the arbitrage capital framework it develops, and the Bartik IV identification strategy: enough to know what it found and how, without reading all 48 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13478).

## TL;DR

The 2016 U.S. money market fund (MMF) reform required institutional prime funds to adopt floating NAV, triggering a roughly $600 billion shift in unsecured wholesale funding away from global banks. Using granular daily supervisory transaction-level data on wholesale funding instruments (federal funds, Eurodollars, commercial paper, CDs, repos) for 58 top-tier global banks, the paper documents that banks hold about $1.5 trillion in potential "arbitrage capital": unsecured wholesale funding deployed in interest-on-excess-reserves (IOER) arbitrage (borrow short-term dollars below the IOER rate, park at the Fed) and covered interest parity (CIP) arbitrage (borrow dollars, lend in FX forward/swap markets). The primary response to the funding shock was a cutback in these liquid arbitrage positions, not a reduction in loan provision. This overturns the traditional bank lending channel prediction for this episode, and is attributed to post-GFC Basel III LCR regulations that effectively require unsecured wholesale funding to be invested in high-quality liquid assets.

## Core results

Magnitudes and significance are as reported; `\*\*\*`/`\*\*`/`\*` = 1%/5%/10%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | MMF reform significantly reduced **IOER arbitrage capital**: 1% decline in prime MMF unsecured funding -> 0.9% decline in potential IOER arbitrage capital | Table I Panel B col. 1, p. 2613 | IV coeff. 0.90 (SE 0.28)\*\*\* |
| R2 | MMF reform significantly reduced **CIP arbitrage capital**: 1% decline in prime MMF unsecured funding -> 0.91% decline in potential CIP arbitrage capital | Table I Panel B col. 2, p. 2613 | IV coeff. 0.91 (SE 0.28)\*\*\* |
| R3 | **IOER arbitrage position** declined; IOER arbitrageurs responded ~4x more than non-arbitrageurs | Table I Panel B cols. 3-5, p. 2613 | Full sample: 0.84 (SE 0.25)\*\*\*; arbitrageur interaction: 0.85 (SE 0.17)\*\*\* |
| R4 | Bank **total assets and cash** declined significantly; **loans** did not | Table IV cols. 1-3, p. 2620 | Total assets: 0.83 (SE 0.14)\*\*\*; cash: 0.76 (SE 0.14)\*\*\*; loans: 0.03 (SE 0.03), not significant |
| R5 | **No evidence of loan reduction** from MMF reform (contrast: European debt crisis 2011-2012 shows significant loan reduction) | Table IV col. 3, p. 2620; Internet Appendix Table IA.XIII | MMF reform: loans coeff. 0.03 (not sig.); European crisis: coeff. >0.20 and significant |
| R6 | **Funding costs rose** for longer-tenor unsecured instruments: ~5 bps increase for maturities >=6 months per 1% funding decline | Table VI col. 2, p. 2625 | IV coeff. -4.99 bps (SE 1.57)\*\*\*; <6M tenors insignificant |
| R7 | **LCR-constrained banks** (lowest HQLA/unsecured funding ratio) cut arbitrage positions most; high-HQLA banks show no significant response | Table VIII, p. 2633 | LCR2: 1.05 (SE 0.11)\*\*\*; LCR3: 1.61 (SE 0.52)\*\*\*; LCR1: 0.02 (not sig.) |
| R8 | **Quarter-end window-dressing** in unsecured funding declined markedly: share of unsecured quarter-end effects in total quarter-end effects fell from ~0.6 (2015Q4) to ~0.1 (2016Q4) | Figure 8, p. 2627 | Unsecured QE-end effects: ~$170B in 2015Q4 to ~$30B in 2016Q4 |

**Overall (paper's conclusion).** The 2016 MMF reform reduced the availability of potential arbitrage capital for global banks. Contrary to the traditional bank lending channel, banks did not reduce their credit supply, but instead cut down their liquid arbitrage positions: the IOER arbitrage and the CIP arbitrage. Banks that had tighter LCR constraints and a higher reliance on arbitrage activity (IOER arbitrageurs) responded most strongly. The use of unsecured wholesale funding as arbitrage capital makes large global banks more resilient to negative wholesale funding shocks, as they can swiftly reduce arbitrage positions in response to wholesale funding dry-ups. However, this business model shift also reduces the usefulness of short-term wholesale funding for maturity and liquidity transformation.

This result contrasts with Ivashina, Scharfstein, and Stein (2015), who find significant loan reduction following MMF funding shocks during the European debt crisis (2011-2012), before Basel III LCR regulations were in place. It also contrasts with Correa, Sapriza, and Zlate (2016), who document the bank lending channel during the European sovereign crisis. The IOER arbitrage framework builds on Bech and Klee (2011) and Keating and Macchiavelli (2017); the CIP arbitrage motivation comes from Du, Tepper, and Verdelhan (2018). The Bartik identification follows Goldsmith-Pinkham, Sorkin, and Swift (2020).

## Theory / model

The paper has no formal equilibrium model. It develops two measurement frameworks for arbitrage capital and positions, and uses them to study the transmission of a funding shock.

**IOER arbitrage.** A bank engages in IOER arbitrage by raising unsecured dollar funding at a rate below the IOER rate and parking the proceeds as reserves at the Federal Reserve. The potential IOER arbitrage capital for bank $$i$$ at time $$t$$ is defined as the total outstanding unsecured wholesale funding borrowed at a rate below the IOER rate (equation 1, p. 2602):

$$
Y_{i,t}^{\text{IOER}} = \sum_{n,k} y_{i,n,k,t} \left[ y_{i,n,k,t} \, | \, r_{i,n,k,t-n} < r_{t-n}^{\text{IOER}} \right], \tag{1}
$$

where $$y_{i,n,k,t}$$ denotes the outstanding amount at time $$t$$ for instrument $$k$$ with remaining maturity $$n$$ issued by bank $$i$$, and $$r_{i,n,k,t-n}$$ denotes the issuing rate on the issuance date $$t-n$$. A proxy for the IOER arbitrage position is then:

$$
Q_{i,t}^{\text{IOER}} = \min(\text{ExcessReserves}_{i,t},\, Y_{i,t}^{\text{IOER}}),
$$

where $$\text{ExcessReserves}_{i,t}$$ denotes excess reserve balances held at the Federal Reserve by bank $$i$$ at time $$t$$ (p. 2603).

**CIP arbitrage.** A bank engages in CIP arbitrage by borrowing unsecured dollars at a rate below the swapped yen rate and lending the dollars in the FX forward/swap markets. The swapped yen rate in dollars is (p. 2603):

$$
r_{n,t}^{\yen \to \$} = r_{n,t}^{\yen} - \rho_{n,t}^{\yen \to \$},
$$

where $$r_{n,t}^{\yen}$$ is the yen OIS rate with tenor $$n$$ and $$\rho_{n,t}^{\yen \to \$}$$ is the FX forward premium to swap yen into dollars. Analogously to equation (1), potential CIP arbitrage capital is (equation 2, p. 2603):

$$
Y_{i,t}^{\text{CIP}} = \sum_{n,k} y_{i,n,k,t} \left[ y_{i,n,k,t} \, | \, r_{i,n,k,t-n} < r_{n,t-n}^{\yen \to \$} \right]. \tag{2}
$$

**Identification logic.** The key challenge is that changes in equilibrium quantities could reflect both MMF funding supply shocks and banks' own demand for funding. The paper isolates the funding supply shock via a Bartik-style shift-share instrument. The exclusion restriction is that banks' preform exposure shares to different prime fund complexes are uncorrelated with unobserved bank-level demand shocks. The paper follows the Goldsmith-Pinkham, Sorkin, and Swift (2020) diagnostic protocol and shows pre-trend tests pass (Figure 7, p. 2618).

## Method

**Baseline OLS regression.** The benchmark OLS specification regresses changes in bank $$i$$'s outcome variable $$\Delta Y_{i,t}$$ on changes in prime funds' holdings of bank $$i$$'s unsecured debt, normalized by 2014 total assets $$\text{Asset}_{i,0}$$ (p. 2611):

$$
\Delta Y_{i,t} / \text{Asset}_{i,0} = \alpha + \beta \,\Delta \text{hold}_{i,t}^{\text{Unsec}} / \text{Asset}_{i,0} + \epsilon_{i,t}.
$$

Changes are quarterly over the reform implementation period (October 2015 to October 2016), using four non-quarter-end quarterly windows to avoid window-dressing contamination. All specifications include time fixed effects; standard errors are clustered at the bank level.

**Bartik IV.** To isolate the supply shock from banks' funding demand, the Bartik-style shift-share instrument for $$\Delta \text{hold}_{i,t}^{\text{Unsec}} / \text{Asset}_{i,0}$$ is (equation 3, p. 2611):

$$
\hat{B}_{i,t} = \sum_j \text{Share}_{i,j,0} \cdot \text{Shift}_{j,t} = \sum_j (s_{i,j,0} / \text{Asset}_{i,0}) \times \Delta\text{aum}_{j,t}, \tag{3}
$$

where $$s_{i,j,0}$$ is the lagged (May 2014) share of bank $$i$$ in fund complex $$j$$'s portfolio, and $$\Delta\text{aum}_{j,t}$$ is the change in AUM of all prime funds within complex $$j$$. The instrument exploits the pooled-exposure design: banks with higher preform exposure to funds that suffered larger AUM losses received larger funding shocks, independently of their own demand. The "share" component captures preform exposure; the "shift" component (aggregate AUM changes at the fund complex level) is not attributable to individual bank funding demand. Banks borrow from 84 fund complexes, so the Bartik is a weighted average of 84 individual instruments.

**Leave-one-out variant.** To address concerns that a bank's own demand shifts the aggregate AUM of a fund complex, the paper also implements a leave-one-out estimator (equation 4, p. 2619):

$$
\tilde{B}_{i,t} = \sum_j (s_{i,j,0} / \text{Asset}_{i,0}) \times \Delta(\text{aum}_{j,t} - \text{aum}_{i,j,t}). \tag{4}
$$

Results are very similar to the baseline Bartik.

**Arbitrage profit regressions.** The volume-weighted IOER arbitrage profit is (p. 2628):

$$
\pi_t^{\text{IOER}} = \sum_{i,n,k} (y_{i,n,k} / Y_{i,t}^{\text{IOER}})(r_{t-n}^{\text{IOER}} - r_{i,n,k,t-n}). \tag{5 (unnumbered in paper)}
$$

Changes in daily average arbitrage profits are then regressed on daily changes in potential arbitrage capital, with a post-reform indicator and an interaction, for both IOER and CIP arbitrage:

$$
\Delta \pi_t^{\text{IOER}} = \alpha + \beta \,\Delta Y_t^{\text{IOER}} + \gamma \,\text{Post}_t + \delta \,\text{Post}_t \times \Delta Y_t^{\text{IOER}} + \epsilon_t. \tag{5}
$$

$$
\Delta \pi_{n,t}^{\text{CIP}} = \alpha + \beta \,\Delta Y_t^{\text{CIP}} + \gamma \,\text{Post}_t + \delta \,\text{Post}_t \times \Delta Y_t^{\text{CIP}} + \epsilon_t. \tag{6}
$$

## Empirical specifications

**Sample.** 58 global banks that frequently trade with U.S. prime MMFs (accounting for ~90% of total prime fund holdings of bank securities). Quarterly observations from October 2015 to October 2016 (four quarters), avoiding quarter-ends. Main regression samples: N = 232 (all banks x four quarters). Balance sheet regressions use 200 observations (50 banks with FR 2644 data).

**Main specifications (Tables I, IV, V, VI, VIII).**

- **Arbitrage capital and positions (Table I).** Dependent variable: quarterly change in potential IOER capital (col. 1), CIP capital (col. 2), or IOER arbitrage position proxy (cols. 3-5) as share of 2014 total assets. Regressor: quarterly change in prime funds' unsecured holdings of bank debt as share of 2014 assets. Time fixed effects; SEs clustered at bank level. Reported for both OLS (Panel A) and Bartik IV (Panel B).

- **Balance sheet adjustments (Table IV).** IV regressions of quarterly changes in balance sheet items (total assets, cash, loans, securities, FF and repo, other assets on asset side; total liabilities, borrowing, deposits, trading liabilities, other liabilities, net-due-to on liability side) on the Bartik-instrumented change in prime-fund unsecured holdings. Sample: U.S.-based entities (FR 2644), N = 200. Time FE; SEs clustered at bank.

- **Intraoffice positions (Table V).** IV regressions decomposing the change in the net-due-to (NDT) position into the Eurodollar (ED) component and the residual, and comparing NDT, ED, and IOER arbitrage position responses. IOER arbitrageur interaction shows ED market is the key source for IOER arbitrage funding.

- **Funding costs (Table VI).** IV cross-sectional regressions of changes in funding rates (for each instrument and benchmark tenor from overnight to 12M) on changes in prime-fund unsecured holdings. Instrument-type and tenor fixed effects; SEs clustered at bank. Separate estimates by instrument type (ED, FF, CD, CP).

- **LCR heterogeneity (Table VIII).** Banks sorted into three LCR-proxy terciles; baseline IOER arbitrage position regression repeated separately for each tercile and interacted. Banks with lowest HQLA/unsecured funding ratio show strongest response.

- **Arbitrage profits (Table VII).** Daily time-series regressions of changes in volume-weighted arbitrage profits on daily changes in aggregate arbitrage capital, with Post dummy and interaction. Estimated separately for non-month-ends, month-ends, quarter-ends (QE), and non-QE month-ends for IOER; and by CIP tenor (1W, 1M, 3M) and QE/non-QE for CIP.

All regressions include time fixed effects. Standard errors are clustered at the bank level for cross-sectional regressions and are robust for time-series regressions.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| DTCC Solutions LLC (CP transactions) | Commercial paper transaction-level data (issuer, volume, rate, maturity) for unsecured funding and arbitrage capital measurement | No page yet |
| FR 2420 Report of Selected Money Market Rates | Daily transaction-level data on federal funds (FF), Eurodollars (ED), and negotiable certificates of deposit (CD) for U.S. banks and FBOs; backbone of IOER and CIP arbitrage capital measures | No page yet |
| FRBNY tri-party repo data | Position-level tri-party repo data to measure secured funding from MMFs | No page yet |
| FR 2644 Weekly Report (Selected Assets and Liabilities) | Weekly bank balance sheet data for U.S.-based entities (domestic banks and FBO branches); key asset and liability categories | No page yet |
| Federal Reserve Board reserve balance data | Daily excess reserve balances by bank; used to construct IOER arbitrage position proxy | No page yet |
| N-MFP (SEC Monthly Schedule of Portfolio Holdings) | Month-end MMF portfolio holdings at security level; AUM per fund; used to construct Bartik shares and shifts | [Form N-MFP](/wiki/datasets/n-mfp/) |
| Dealscan (Refinitiv) | Dollar-denominated syndicated loan origination by sample banks (lead arranger credit); used to test loan supply response | [DealScan](/wiki/commercial/dealscan/) (licensed) |
| SNL Financial | Bank holding company total assets in 2014 for normalization; also credit ratings and CET1 ratios | No page yet |
| Bloomberg (JPY OIS rates) | Dollar-yen OIS rates at granular maturities to construct swapped yen rate and CIP arbitrage profit | No page yet |

Sample: October 2015 to June 2017 (daily); main regression period October 2015 to October 2016. 58 global banks, comprising ~90% of total prime fund holdings of bank securities. Banks had at least 100 transactions with all U.S. MMFs and total assets of at least $100 billion in 2014.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13478) if you are: studying the post-GFC role of unsecured wholesale funding and arbitrage in banks' business models; evaluating credit channel transmission of funding shocks in a post-Basel III regulatory environment; applying Bartik shift-share designs in banking contexts (the paper follows the Goldsmith-Pinkham, Sorkin, and Swift (2020) protocol closely); analyzing IOER or CIP arbitrage dynamics and their relationship to unconventional monetary policy and excess reserves; or assessing the likely effects of the 2023 SEC MMF reform on global bank funding. The locators above point to the exact tables and figures.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(5), pp. 2591-2638. The PDF states "This article is a U.S. Government work and is in the public domain in the USA." Crossref DOI metadata records the Wiley termsAndConditions#vor licence with no CC designation; a rights signal conflict exists. This distillation was extracted by an LLM on 2026-06-05 and is **not human-verified or independently reproduced**. Treated as extract-only pending licence clarification.

> Anderson, Alyssa, Wenxin Du, and Bernd Schlusche. "Arbitrage Capital of Global Banks."
> *The Journal of Finance* 80, no. 5 (October 2025): 2591–2638. DOI: 10.1111/jofi.13478.


==============================================================================
# Are CEOs Rewarded for Luck: Andreani, Ellahie & Shivakumar (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/andreani-ceos-rewarded-luck-evidence-2025/
# Distilled: Using the 2017 Tax Cuts and Jobs Act as a quasi-natural experiment, the paper shows that weakly scrutinized CEOs are compensated for one-off windfall tax gains (deferred tax liability remeasurement) but not penalized for corresponding tax losses, consistent with rent extraction rather than optimal contracting. J. Finance 2025, CC BY-NC-ND 4.0. Six core results with source locators, datasets used, the empirical design, and the estimating equations.
# Tags: paper-summary, executive-compensation, corporate-governance, pay-for-luck
==============================================================================

**What this is.** The paper's core results, the quasi-natural experiment it exploits (TCJA 2017 deferred-tax remeasurement), the triple-difference estimating equation, and the evidence for asymmetric pay-for-luck: enough to know what it found and how, without reading all 48 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1111/jofi.13448).

## TL;DR

The 2017 Tax Cuts and Jobs Act (TCJA) lowered the U.S. corporate tax rate from 35% to 21%, generating one-off book gains for firms with large deferred tax liabilities and book losses for firms with large deferred tax assets. Because these gains and losses arise from past firm transactions, are exogenous to current managerial effort, and vary heterogeneously across firms even within industries, they are a clean "luck" shock to earnings. Using a triple-difference (DDD) panel regression on 9,225 firm-year observations (2,081 U.S. firms, 2013-2019), the paper finds that CEOs facing weak external pay scrutiny receive significantly higher total compensation in the TCJA transition period when their firms have larger deferred tax liabilities (windfall gains), but are not penalized when their firms have larger deferred tax assets (windfall losses). CEOs facing high pay scrutiny show no such association. The asymmetry is concentrated in discretionary (variable) pay, is observed in CEO and CFO pay but not in other named executive officers, and is not explained by investment incentives, political activities, talent retention, or catch-up for underpaid CEOs. The results support the rent-extraction view of executive compensation (Bertrand and Mullainathan (2001)) and challenge optimal contracting theories. The paper also provides cleaner evidence for the asymmetric pay-for-luck pattern predicted by Garvey and Milbourn (2006) and restores credibility to findings challenged by Daniel, Li, and Naveen (2020). It extends the earlier analysis of cash windfalls and CEO pay in Blanchard, Lopez-de-Silanes, and Shleifer (1994) to a large-sample setting with a TCJA-specific identification strategy.

## Core results

Magnitudes and significance as reported; `\*\*` / `\*\*\*` = 5% / 1%. All regressions include firm and year fixed effects and standard errors clustered by firm. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Weakly monitored CEOs earn significantly more in the TCJA period when their firm has larger net deferred tax liabilities (windfall gains); effect absent for high-scrutiny CEOs | Table III col (2)-(3), p. 2275 | Tax Shock x NDTL = 2.17 (t=2.40)\*\*; Tax Shock x NDTL x Pay Scrutiny = -3.81 (t=-2.68)\*\*\*; effect absent at Pay Scrutiny = 1 |
| R2 | Separating DTA from DTL: CEO pay is positively associated with deferred tax liabilities (windfall gains) in the TCJA period, concentrated in low-scrutiny firms; a firm in the 3rd DTL quartile pays CEO 19.7% more than a firm in the 1st quartile | Table IV col (2), p. 2276-2278 | Tax Shock x DTL = 3.311 (t=3.20)\*\*\*; Tax Shock x DTL x Pay Scrutiny = -5.814 (t=-3.62)\*\*\* |
| R3 | CEO pay is not reduced for windfall tax losses (DTA remeasurement); the gain-loss asymmetry is statistically significant (F-test p=0.000) | Table IV col (2)-(4), p. 2276, 2278 | Tax Shock x DTA coefficients all statistically insignificant; F-test rejects equality of Tax Shock x DTA and Tax Shock x DTL magnitudes |
| R4 | The windfall pay-for-luck is concentrated in discretionary (variable) pay; the fixed salary effect is substantially smaller in magnitude | Table IV col (5)-(6), p. 2276-2279 | Tax Shock x DTL (Disc. Comp) = 6.61 (t=3.69)\*\*\*; Tax Shock x DTL (Fixed Comp) = 0.802 (t=3.23)\*\* and substantially smaller in magnitude; Tax Shock x DTL x Pay Scrutiny (Disc. Comp) = -7.175 (t=-1.97)\* |
| R5 | The reward is specific to CEOs and CFOs, not shared broadly with other named executive officers | Table VIII col (1)-(2), p. 2287 | CFO: Tax Shock x DTL = 2.32 (t=2.21)\*\*; Other NEOs: Tax Shock x DTL and all three-way interactions insignificant |
| R6 | Within-calendar-year identification exploiting staggered fiscal years confirms results; no pre-trend in DTA/DTL compensation prior to TCJA | Table IV col (4), p. 2278-2279 | Results robust; interaction of calendar-year FE with DTA and DTL show no differential pre-trends |

**Overall (paper's conclusion).** The evidence supports the rent-extraction view: weakly scrutinized CEOs are paid for good luck but not penalized for bad luck, the pattern is concentrated in discretionary pay and senior executives, it is unrelated to future investments or shareholder payouts, and it cannot be explained by talent retention, political activity, or formulaic compensation plans. Shareholder-value-maximizing theories alone are insufficient to explain the observed pay patterns.

## Theory / model

The paper has no formal structural model. It tests two competing theoretical frameworks using the TCJA shock as a natural experiment:

**Optimal contracting theory** (also called efficient contracting or shareholder-value-based theory): managers should be compensated only for performance attributable to their own effort, not for gains from factors beyond their control. Under this view, CEO pay should not respond to one-off tax windfalls that are unrelated to managerial actions.

**Rent-extraction model** (Bertrand and Mullainathan (2001)): managers with limited pay scrutiny aim to extract maximum compensation from their firms. A key prediction is that pay for luck should be asymmetric: CEOs seek rewards for favorable luck but not penalties for adverse luck. Three empirical conditions distinguish this view: (i) pay for luck factors genuinely beyond CEO control, (ii) asymmetric response (good luck rewarded, bad luck not penalized), and (iii) effect concentrated in firms with weak pay scrutiny.

The identification logic exploits three features of TCJA-driven deferred tax remeasurements (pp. 2257-2258):
1. The trigger (government decision to lower the tax rate) is exogenous to managers' decisions.
2. The TCJA was largely unexpected until Trump's surprise election victory in November 2016 and passed in less than three months, making anticipation or influence by firms implausible.
3. A firm's TCJA gain or loss was determined solely by its pre-existing stock of deferred tax assets and liabilities, which are related to past transactions, not current effort.

These features make TCJA tax effects a clean measure of luck: compensation for these effects cannot reflect rewards for CEO effort or performance in the transition period.

## Method

The paper proposes a novel composite pay-scrutiny measure as a methodological contribution, arguing that post-Sarbanes-Oxley regulatory changes have homogenized internal governance variables (board independence, committee composition) so they no longer differentiate well-governed from poorly-governed firms. The composite proxy focuses on external monitoring by investors.

**Pay Scrutiny construction** (p. 2268): The first principal component is extracted from five firm-level proxies measured at the beginning of each fiscal year. This component explains 78% of the variation in the five proxies:

$$
\text{Pay Scrutiny}_j = \text{PC}_1(\text{Market Value}_j, \text{Liquidity}_j, \text{Trading Volume}_j, \text{Nonzero Return Days}_j, \text{Analyst Coverage}_j)
$$

Each proxy is normalized to the [0, 1] interval by subtracting the minimum and dividing by the range. The component is then re-normalized to [0, 1].

The `factor-analysis-pca` technique (five external monitoring proxies collapsed into one index) builds on the `panel-regression` approach for the causal estimates. The method generalizes the internal-governance proxies of Bertrand and Mullainathan (2001) toward market-based external scrutiny, motivated by post-2006 regulatory changes (Say-on-Pay voting, 2011 SEC mandate) that increased the salience of investor and media monitoring.

## Empirical specifications

The main estimating equation is a triple-difference (DDD) pooled regression for CEO compensation in fiscal year $$t$$ (p. 2266, equation 1):

$$
\text{Total Comp}_t = \beta_0 + \beta_1 \text{Tax Shock}_t + \beta_2 \text{DTA}_{t-1} + \beta_3 \text{Tax Shock}_t \times \text{DTA}_{t-1}
$$

$$
+ \beta_4 \text{DTL}_{t-1} + \beta_5 \text{Tax Shock}_t \times \text{DTL}_{t-1} + \beta_6 \text{Pay Scrutiny}_{t-1}
$$

$$
+ \beta_7 \text{Tax Shock}_t \times \text{Pay Scrutiny}_{t-1} + \beta_8 \text{DTA}_{t-1} \times \text{Pay Scrutiny}_{t-1}
$$

$$
+ \beta_9 \text{Tax Shock}_t \times \text{DTA}_{t-1} \times \text{Pay Scrutiny}_{t-1} + \beta_{10} \text{DTL}_{t-1} \times \text{Pay Scrutiny}_{t-1}
$$

$$
+ \beta_{11} \text{Tax Shock}_t \times \text{DTL}_{t-1} \times \text{Pay Scrutiny}_{t-1}
$$

$$
+ \gamma' X_{t-1} + \theta' \lambda_j + \pi' \tau_t + \phi' \tau_t \times \text{DTA}_{t-1} + \psi' \tau_t \times \text{DTL}_{t-1} + \varepsilon_t \tag{1}
$$

where:
- $$\text{Tax Shock}_t = 1$$ for fiscal years ending between December 31, 2017 and March 31, 2019 (the TCJA transition period); 0 otherwise.
- $$\text{DTA}_{t-1}$$ = beginning-of-year deferred tax assets scaled by book total assets (Compustat `txndba` / `at`).
- $$\text{DTL}_{t-1}$$ = beginning-of-year deferred tax liabilities scaled by book total assets (Compustat `txndbl` / `at`).
- $$X_{t-1}$$ = control vector: log total assets (Size), return on assets (Profitability), size- and industry-adjusted 12-month past returns (Past Returns), book-to-price ratio, idiosyncratic volatility, financial leverage, log CEO age, log CEO tenure.
- $$\lambda_j$$ = firm fixed effects; $$\tau_t$$ = calendar-year fixed effects.
- Standard errors clustered by firm.

**Primary variables of interest:** the double interactions $$\text{Tax Shock} \times \text{DTA}$$ (captures pay for windfall tax losses) and $$\text{Tax Shock} \times \text{DTL}$$ (captures pay for windfall tax gains), and the triple interactions $$\text{Tax Shock} \times \text{DTA} \times \text{Pay Scrutiny}$$ and $$\text{Tax Shock} \times \text{DTL} \times \text{Pay Scrutiny}$$ (capture whether these relationships differ by scrutiny level).

**Within-calendar-year identification (column 4, Table IV, p. 2278-2279):** Because the TCJA transition period is staggered across fiscal years (firms with different year-ends recognize TCJA tax effects in different calendar years), the authors add calendar-year-by-DTA and calendar-year-by-DTL fixed effects. Identification within a calendar year compares CEO compensation across firms whose financial statements are more vs. less likely to reflect TCJA tax effects, holding the calendar year constant. This provides a sharper test and the results strengthen in magnitude and significance.

**Alternative explanations tested (Tables V-X):** The windfall-related pay is decomposed via a two-stage regression to test whether it relates to future CAPX, ROA, share repurchases, or total payout (Table V). Political leanings of CEOs (Table VI), CEO tenure (Table VII), sharing with other executives (Table VIII), underpaid CEOs (Table IX), and talent transferability (Table X) are each tested by extending equation (1) with the relevant interaction terms.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| ExecuComp, S&P Capital IQ People Intelligence, ISS Incentive Lab | CEO and CFO total compensation, fixed salary, discretionary pay; CEO age and tenure | [WRDS / Compustat](/wiki/commercial/wrds/) (licensed) |
| Compustat North America (annual) | Deferred tax assets (txndba), deferred tax liabilities (txndbl), total assets (at), book equity, ROA, CAPX, dividends, share repurchases, leverage, sales | [WRDS / Compustat](/wiki/commercial/wrds/) (licensed) |
| CRSP monthly and daily | Market value of equity, stock returns, trading volume, daily returns for illiquidity and nonzero return days | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| I/B/E/S (Institutional Brokers' Estimate System) | Analyst coverage (number of equity research analysts with EPS forecasts) | [WRDS / I/B/E/S](/wiki/commercial/wrds/) (licensed) |
| Federal Election Commission (FEC) campaign finance database | CEO personal political donations (Republican vs. Democratic leanings) | No page yet |
| Boardex | Board structure (independence of directors, CEO-Chairman separation) | No page yet |
| FactSet | Share ownership data (institutional shareholders, blockholder presence) | No page yet |
| RavenPack | Media scrutiny (firm-specific full-size articles) | No page yet |

Sample: January 2013 to December 2019 (annual frequency). 9,225 firm-year observations from 2,081 unique U.S. firms after excluding three years around CEO turnover events and requiring at least one observation before and during the TCJA transition period.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13448) if you are:
- Evaluating the empirical validity of pay-for-luck models or the rent-extraction view of executive compensation.
- Using the TCJA as a natural experiment for corporate-level or compensation research (institutional details in Section I; variable definitions in Appendix B).
- Building a composite external-monitoring index from market-based proxies (Pay Scrutiny construction, Section III and Appendix B).
- Studying asymmetries in executive pay response to gains vs. losses.
- Extending the analysis to other executive types (CFO, named executive officers) or other windfall settings.

Tables III-IV (main results), Table V (investment/payout tests), and Tables VI-X (alternative explanations) contain the full specification details.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(4), August 2025, pp. 2255-2302. This distillation was extracted by an LLM on 2026-06-05 and is **not human-verified or independently reproduced**. Licensed CC BY-NC-ND 4.0: use and distribution in any medium permitted provided the original work is properly cited, the use is non-commercial, and no modifications or adaptations are made.

> Andreani, Martina, Atif Ellahie, and Lakshmanan Shivakumar.
> "Are CEOs Rewarded for Luck? Evidence from Corporate Tax Windfalls."
> *The Journal of Finance* 80, no. 4 (August 2025): 2255-2302.
> DOI: 10.1111/jofi.13448. © 2025 The Author(s).
> Licensed under [CC BY-NC-ND 4.0](https://creativecommons.org/licenses/by-nc-nd/4.0/).
> This page extracts core results only; no modifications to the original were made in the excerpts cited above.


==============================================================================
# Crisis Interventions in Corporate Insolvency: Antill & Clayton (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/antill-crisis-interventions-corporate-insolvency-2025/
# Distilled: A general-equilibrium model shows that optimal insolvency interventions can favor either liquidation or reorganization depending on which externality dominates: a fire-sale externality (fewer liquidations optimal) or a collateral externality (more liquidations optimal). J. Finance 2025, paywalled. Six core results with source locators, the model, and the propositions with their equations.
# Tags: paper-summary, corporate-finance, insolvency, bankruptcy, financial-intermediation
==============================================================================

**What this is.** The paper's core propositions, the GE model it builds, and the comparative statics behind its policy conclusions: enough to know what it found and how, without reading all 36 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13421).

## TL;DR

Antill and Clayton (2025) build a two-period general-equilibrium model with collateral-constrained banks to study the optimal resolution of insolvent firms. The model features two externalities that work in opposite directions: (i) a fire-sale externality following Shleifer and Vishny (1992) (more liquidations depress asset prices, reducing all banks' recovery) and (ii) a collateral externality (reorganizations tie up bank balance sheets, congesting lending capacity and raising borrowing rates). Banks have a monitoring advantage over households in firm lending, as in Diamond (1984). The paper shows that socially optimal policy can encourage either more or fewer liquidations relative to the private equilibrium depending on which externality dominates. A simple uniform tax or subsidy on liquidations decentralizes the optimum without requiring the planner to know individual firms' long-run values. The framework extends and complements earlier GE insolvency models of Corbae and D'Erasmo (2021a) and Corbae and D'Erasmo (2021b), and the mechanism-design approach of Philippon (2021). Applied to aggregate statistics for Japan's nonperforming loan crisis (studied empirically by Caballero, Hoshi, and Kashyap (2008)) and the U.S. COVID crisis, the model predicts that optimal policy would have subsidized liquidation in Japan and subsidized reorganization in the United States, consistent with observed policy responses.

## Core results

Magnitudes and significance are as reported. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | The socially optimal liquidation rule is a threshold rule; its threshold differs from the privately optimal threshold in two ways: (i) a fire-sale externality pushes the social threshold below the private threshold (fewer liquidations) and (ii) a collateral externality pushes it above (more liquidations). | Prop. 2, eq. 17, p. 890 | $$V_* = \delta_*(\gamma + c) - (\delta_* - 1)\xi_\gamma\gamma$$; relative to $$V_P = \delta_P(\gamma + c)$$ (Prop. 1) |
| R2 | A uniform liquidation tax $$\tau$$ decentralizes the social optimum; optimal intervention favors liquidation subsidies ($$\tau < 0$$) when corporate distress is high (large $$c$$, $$d_0$$), bank monitoring advantage large (high $$t$$), or fire-sale prices are low (low $$\gamma$$). Lower fire-sale prices can sometimes call for more, not fewer, liquidations. | Prop. 3-4, eqs. 19-20, pp. 893-894 | $$\tau = (\gamma + c)\delta_P(1-M) + (M\delta_P - 1)\xi_\gamma\gamma$$; $$\tau < 0$$ if $$\tilde\xi_\gamma < (1 + c/\gamma)\phi t/(p\bar g + \bar u)(1 - \phi(1+t))pd_0$$ |
| R3 | Relative to the U.S. COVID crisis, Japan's nonperforming-loan crisis featured higher corporate leverage (126.8% vs 80.9% of GDP), lower firm profitability (EBIT/Rev = 5.5% vs 14%), higher bank lending share (83.5% vs 38%), and lower GDP growth (0.1% vs 1.8%); all four comparative-static factors imply liquidation subsidies were optimal for Japan and reorganization subsidies for the United States. | Table II, pp. 904-906; §VI | Model qualitative comparative statics applied to Table II statistics |
| R4 | Ex ante macroprudential regulation (a debt tax on banks of $$\tau_0^b = (M-1)\delta_P$$) complements ex post liquidation subsidies: both target the collateral externality. Macroprudential regulation does NOT directly target the fire-sale externality (Envelope Theorem). | Prop. 5, p. 897 | Socially optimal equity satisfies $$\Psi'(A_0) = \delta_*$$; debt tax $$\tau_0^b = (M-1)\delta_P > 0$$ |
| R5 | Bailouts to banks (weakly) dominate bailouts to solvent firms in welfare terms because banks obtain an additional collateral multiplier benefit; bailouts to distressed firms equal bailouts to banks conditional on an insolvency rule. | Prop. 6, pp. 899-900 | Marginal welfare: $$\delta_*$$ for banks or distressed firms; $$(1 - \phi t/(1-\phi))\delta_*$$ for solvent firms |
| R6 | With heterogeneous banks (varying collateral haircuts $$\phi_b$$), the optimal seniority structure bifurcates banks: high-$$\phi$$ banks become secured creditors (liquidate and lend) while low-$$\phi$$ banks become distressed lenders (reorganize); bailouts should go to secured creditors. | §IV.D, eq. 23, p. 901 | $$V_* = \delta_*^1 c + \delta_*^2\gamma - (\delta_*^2 - 1)\xi_\gamma\gamma$$ with $$\delta_*^2 > \delta_*^1$$ |

**Overall (paper's conclusion).** Optimal crisis interventions in insolvency need not favor reorganization. When bank lending capacity is tight (the collateral externality dominates), subsidizing more liquidations improves welfare by freeing balance-sheet capacity for new lending to solvent firms. The optimal Pigouvian tax or subsidy is simple, uniform, and does not require firm-specific information. Macro-prudential regulation and bailouts are natural complements to insolvency interventions, but they target only the collateral externality, not the fire-sale externality; insolvency policy must handle both.

## Theory / model

The model has two dates (date one and date two) and four types of agents: firms, banks, arbitrageurs, and households (pp. 881-886).

**Firms.** A fraction $$p$$ of firms are solvent; each has a project worth $$v_S$$ at date two and an investment opportunity $$I_S$$. Solvent firms maximize date-two cash flows net of debt repayment (eq. 1, p. 883):

$$
\max_{I_S} \; g_S(I_S) - \frac{I_S}{Q_S} + v_S - d_0 \tag{1}
$$

with first-order condition $$g'_S(I_S) = Q_S^{-1}$$ (eq. 2, p. 883). A fraction $$1-p$$ of firms are insolvent with idiosyncratic long-run payoff $$v \in [\underline{v}, \bar{v}]$$ and a date-one operating loss $$c \geq 0$$ that must be paid to avoid liquidation.

**Banks.** Banks choose new household borrowing $$B_1$$, loans to solvent firms $$D_1$$, and an insolvency resolution rule $$\rho(v) \in [0,1]$$ (the probability that a firm with viability $$v$$ is liquidated). The bank budget constraint (eq. 3, p. 884) is:

$$
pD_1 + (1-p)\!\int (1-\rho(v))\,c\,f(v)\,dv \;\leq\; B_1 - b_0 + (1-p)\!\int \rho(v)\,\gamma\,f(v)\,dv \tag{3}
$$

An agency friction limits bank borrowing from households via a collateral constraint (eq. 4, p. 884):

$$
B_1 \leq \phi\,Q_B\;p\!\left(d_0 + \frac{D_1}{Q_S}\right) \tag{4}
$$

Banks maximize equity value (eq. 5, p. 885):

$$
\max_{B_1,D_1,\rho} \; p\!\left(\frac{D_1}{Q_S} + d_0\right) + (1-p)\!\int (1-\rho(v))\,v\,f(v)\,dv - \frac{B_1}{Q_B} \tag{5}
$$

**Households.** Households maximize utility over date-one consumption (eq. 7, p. 886):

$$
\max_{B_H, D_H} \; u\!\left(e - B_H - p(1+t)D_H\right) + \frac{B_H}{Q_B} + p\,\frac{D_H}{Q_S} \tag{7}
$$

with first-order conditions (eqs. 8-9, p. 887):

$$
u'\!\left(e - B_H - p(1+t)D_H\right) = \frac{1}{Q_B}, \qquad u'\!\left(\cdot\right) = \frac{1}{(1+t)Q_S} \tag{8-9}
$$

Market clearing requires $$D_1 + D_H = I_S$$, $$(1-p)\int \rho(v)f(v)dv = L$$, and $$B_H = B_1$$ (eqs. 10-12, p. 887).

**Social planner objective.** The planner maximizes aggregate utilitarian welfare (p. 890):

$$
p\bigl(g_S(I_S) + v_S\bigr) + (1-p)\!\int(1-\rho(v))\,v\,f(v)\,dv + a(L) - \gamma L + u\!\left(e - B_1 - p(1+t)D_H\right)
$$

## Method

**Private equilibrium: Proposition 1** (p. 888). In any competitive equilibrium, the bank's privately optimal liquidation rule is a threshold rule $$\rho(v) = \mathbf{1}(v \leq V_P)$$ where

$$
V_P = \delta_P\,(\gamma + c), \qquad \delta_P = \underbrace{\frac{1}{Q_S}}_{\text{Direct}} + \underbrace{\frac{\phi\frac{Q_B}{Q_S}}{1 - \phi\frac{Q_B}{Q_S}}}_{\text{Collateral Multiplier}} \times \underbrace{\left(\frac{1}{Q_S} - \frac{1}{Q_B}\right)}_{\text{Excess Return}} \tag{13-14}
$$

The effective return $$\delta_P$$ is the bank's marginal value of an additional date-one dollar, combining the direct return from lending and the shadow value of collateral (which lets banks borrow more from households).

**Social optimum: Proposition 2** (p. 890). The socially optimal liquidation rule is also a threshold rule $$\rho(v) = \mathbf{1}(v \leq V_*)$$ where

$$
V_* = \underbrace{\delta_*(\gamma + c)}_{\text{Weakly Bigger than } V_P} - \underbrace{(\delta_* - 1)\xi_\gamma\gamma}_{\text{Fire-Sale Externality} \geq 0} \tag{17}
$$

$$
M = \left[1 - \frac{\sigma_S\sigma_H\phi t}{\bigl(p(1+t)\sigma_H + \sigma_S\bigr)\bigl(1-\phi(1+t)\bigr)}\,pQ_B d_0\right]^{-1} \geq 1 \tag{18}
$$

The social effective return is $$\delta_* = M\delta_P$$, where $$M \geq 1$$ captures the collateral multiplier effect that banks do not internalize. The elasticity $$\xi_\gamma \equiv -(L/\gamma)\partial\gamma/\partial L \geq 0$$ measures how much each additional liquidation depresses the fire-sale price.

**Decentralization: Proposition 3** (p. 893). The liquidation tax or subsidy that decentralizes the social optimum is

$$
\tau = \underbrace{(\gamma + c)\delta_P(1-M)}_{\text{Subsidy component}} + \underbrace{(M\delta_P - 1)\xi_\gamma\gamma}_{\text{Tax component}} \tag{19}
$$

The subsidy component (negative) reflects the collateral externality; the tax component (positive) reflects the fire-sale externality. Under Assumption 1 (log production, log utility, iso-elastic arbitrageur demand), a sufficient condition for $$\tau < 0$$ (subsidize liquidations) is (eq. 20, p. 894):

$$
\tilde\xi_\gamma < \left(1 + \frac{c}{\gamma}\right)\frac{\phi t}{(p\bar g + \bar u)(1-\phi(1+t))}\,pd_0 \tag{20}
$$

## Empirical specifications

The paper does not run regressions. The empirical section (§VI, pp. 903-907) applies the model's comparative statics to aggregate statistics for two historical crises. The approach is as follows:

1. **Data.** Country-year-quarter panel from the IMF for corporate debt/GDP; country-quarter panel from FRED for bank lending and bond market shares; country-quarter GDP growth rates; firm-year Compustat North America and Compustat Global data for EBIT/Revenue ratios. Sample: Japan 1990-2005 and US 2020-2023 (Table II, p. 905).

2. **Mapping parameters to observables.** Each model parameter ($$d_0$$, $$c$$, $$t$$, $$\bar g$$) is mapped to an observable aggregate statistic. Proposition 4's sufficient condition for $$\tau < 0$$ (eq. 20) then yields a qualitative prediction about the direction of optimal intervention in each crisis.

3. **Four comparative-static dimensions.** (i) Corporate leverage $$d_0$$: Japan 126.8% of GDP vs. US 80.9%, so higher $$d_0$$ favors liquidation subsidies. (ii) Firm profitability (proxy for $$c$$): EBIT/Revenue Japan 5.5% vs. US 14%, so lower profitability (higher $$c$$) favors liquidation subsidies. (iii) Bank lending share (proxy for $$t$$): Japan 83.5% vs. US 38%, so higher $$t$$ favors liquidation subsidies. (iv) GDP growth (proxy for $$\bar g$$): Japan 0.1% vs. US 1.8%, so lower growth (more permanent shock) favors liquidation subsidies.

All four factors point in the same direction: the model predicts liquidation subsidies were optimal for Japan and reorganization subsidies for the United States, consistent with Japan's Takenaka Plan (which promoted liquidations of nonperforming loans) and U.S. COVID-era policies (which promoted reorganization).

The two testable empirical implications offered for future work are: (a) bank capital requirements bind more tightly when corporate interest rates are high; (b) banks with higher pledgeability $$\phi_b$$ are more likely to liquidate a given distressed firm (pp. 906-907).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| IMF International Financial Statistics | Country-year corporate debt/GDP for Japan (1990-2005) and US (2020-2023) | No page yet |
| FRED (Federal Reserve Bank of St. Louis) | Country-quarter bank lending and bond market shares; quarterly GDP growth rates | [FRED](/wiki/datasets/fred/) |
| Compustat North America | Firm-year EBIT and Revenue for large US firms (2020-2023) | [WRDS / Compustat](/wiki/commercial/wrds/) (licensed) |
| Compustat Global | Firm-year EBIT and Revenue for large Japanese firms (1990-2005) | [WRDS / Compustat](/wiki/commercial/wrds/) (licensed) |

Sample: Japan crisis defined as 1990-2005 (Caballero, Hoshi, and Kashyap (2008)); US COVID crisis defined as March 2020 to May 2023.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13421) if you are: designing insolvency interventions and need the full set of propositions and Internet Appendix extensions (acquisitions by arbitrageurs or solvent firms, endogenous $$p$$, distinct liquidation deadweight losses); studying the interaction between macroprudential regulation and insolvency rules; or extending the framework to dynamic models with temporary versus permanent shocks. The locators above point to the exact propositions and equations.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(2), April 2025, pp. 875-910. DOI: [10.1111/jofi.13421](https://doi.org/10.1111/jofi.13421). Published by Wiley on behalf of the American Finance Association; paywalled (Wiley VoR terms; not CC). This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. Extract-only; the verbatim PDF is not hosted here.

> Antill, Samuel, and Christopher Clayton. "Crisis Interventions in Corporate Insolvency." *The Journal of Finance* 80, no. 2 (April 2025): 875-910. DOI: 10.1111/jofi.13421. © 2025 the American Finance Association.


==============================================================================
# Feedback Effects and Systematic Risk Exposures: Banerjee, Breon-Drish & Smith (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/banerjee-feedback-effects-systematic-risk-2025/
# Distilled: Models feedback effects when managers learn discount rates (not just cash flows) from stock prices, applied to climate-exposed investment. Shows cash-flow and price maximization both fail to maximize welfare because neither internalizes hedging and risk-sharing benefits of investment. J. Finance 2025, paywalled. Seven core results with source locators, the model equations, and the equilibrium investment rules under each objective.
# Tags: paper-summary, asset-pricing, feedback-effects, climate-risk, esg, corporate-investment
==============================================================================

**What this is.** The paper's core propositions, the equilibrium investment rules, and the welfare results: enough to know what it found and how, without reading all 48 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1111/jofi.13427).

## TL;DR

The paper builds a rational-expectations feedback model in which a manager learns both cash-flow news and discount-rate news from the stock price when deciding whether to invest in a project exposed to a systematic risk factor (climate risk). This extends the feedback effects literature surveyed by Bond, Edmans, and Goldstein (2012) and Goldstein (2023), and the risk-averse feedback models of Dow and Rahi (2003), to settings where the manager also learns about the factor risk premium. A cash-flow-maximizing manager treats discount-rate variation as noise; a price-maximizing manager internalizes it. This difference reverses how project "greenness" affects investment: higher climate exposure reduces investment under cash-flow maximization (noisier price signal) but can raise it under price maximization (more volatile NPV makes ex-ante unattractive projects more likely to become desirable). Neither objective maximizes investor welfare, because neither manager internalizes the hedging and risk-sharing benefits of investment in climate-exposed assets, a mechanism related to Pastor, Stambaugh, and Taylor (2021) on green asset pricing. The paper derives welfare-maximizing rules and shows when feedback reduces rather than improves welfare. The welfare gap between price maximization and welfare maximization is analogous to the quality-choice distortion in Spence (1975): the price reflects marginal disutility of the last share, while welfare depends on average disutility across all shares.

## Core results

Magnitudes and significance are as reported. Results are analytical propositions from the theoretical model; no empirical estimation is involved. Locators cite PDF pages.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Under cash-flow maximization, higher project climate exposure (higher \|alpha\|) makes the price a noisier signal, reducing investment for ex-ante profitable projects and increasing it for ex-ante unprofitable ones | Prop. 3, p. 997; Fig. 2, p. 999 | Prob(invest) decreases with tau\_theta and \|alpha\| when mu\_theta > c; increases with tau\_theta and \|alpha\| when mu\_theta < c (eq. 22) |
| R2 | Under price maximization, climate exposure has an additional "variance of NPV" channel that can increase investment for ex-ante unprofitable projects and decrease it for ex-ante profitable ones, creating the opposite pattern relative to cash-flow maximization | Prop. 4, p. 998; Fig. 2, p. 999 | Prob(invest) decreases with greenness alpha iff (mu\_theta - c - (gamma n / tau\_eta) + (alpha gamma tau\_Z mu\_Z) / tau\_eta) sgn(alpha) > 0 (eq. 23, part v) |
| R3 | Cash-flow maximization leads to more investment than price maximization if and only if the discount-rate premium exceeds the ex-ante profitability advantage: (gamma / tau\_eta)(n - alpha mu\_Z) > -(tau\_theta / tau\_p)(mu\_theta - c) | Corollary 1, p. 999 | Threshold condition eq. 24: s\_P > s\_C iff (gamma / tau\_eta)(n - alpha mu\_Z) > -(tau\_theta / tau\_p)(mu\_theta - c) |
| R4 | With homogeneous investor climate exposures, price maximization always leads to underinvestment relative to welfare maximization; cash-flow maximization leads to underinvestment iff the condition in eq. 34 holds, and overinvestment otherwise | Prop. 5, p. 1003; eqs. 34-36, pp. 1003-1005 | Underinvestment gap: s\_P - s\_W = (1/2)(gamma / tau\_eta)n > 0 always; for cash-flow: s\_C - s\_W > 0 iff (gamma / tau\_eta) alpha mu\_Z - (tau\_theta (mu\_theta - c)) / tau\_p - (1/2)(gamma / tau\_eta)n > 0 |
| R5 | The welfare-maximizing rule can be implemented by a manager maximizing a weighted average of expected price and expected cash flows, with weight delta on price; delta is between 0 and 1 iff the ex-ante profitability condition holds | Prop. 6, eqs. 39-40, p. 1006 | delta = [(tau\_p/(tau\_theta + tau\_p))(mu\_theta - c) - (tau\_p/(tau\_theta + tau\_p))(gamma / tau\_eta)(alpha mu\_Z - n/2)] / [(tau\_theta/(tau\_theta + tau\_p))(mu\_theta - c) - ... + (1/2)(gamma / tau\_eta)n] (eq. 40) |
| R6 | With heterogeneous investor exposures and zero share endowment (n = 0), welfare maximization always requires investment; both cash-flow and price maximization lead to underinvestment relative to welfare maximization | Prop. 7, eq. 42, p. 1008 | arg max W(k; s\_p) = 1 for all s\_p when n = 0 and heterogeneous exposures (1/tau\_zeta > 0) |
| R7 | Feedback reduces welfare when firm size n is small or when gains from risk-sharing are large (tau\_zeta small), even though feedback always improves the manager's objective (expected cash flows or price) | Prop. 8, p. 1008; Fig. 3, p. 1009 | Welfare is lower with feedback than without when n -> 0 or tau\_zeta -> 0; the risk-sharing channel is unaffected by n but is the dominant welfare effect in this limit |

**Overall (paper's conclusion).** Neither cash-flow maximization nor price maximization aligns with welfare maximization because the stock price reflects the marginal disutility of the last outstanding share, not the average disutility across all investors. Investing in a climate-exposed project improves investors' ability to hedge systematic risk (risk-sharing channel) and reduces information uncertainty about climate exposures (value of information channel), but neither channel is fully captured by standard managerial objectives. As a result, both objectives can lead to underinvestment in green projects and, sometimes, overinvestment in brown projects. Incentivizing managers via climate scores can improve welfare even when it reduces stock prices and future profitability.

## Theory / model

The model has four dates (t = 1, 2, 3, 4) and two securities (risk-free and risky). A continuum of CARA investors indexed by i in [0, 1] with risk aversion gamma each have initial endowment of n shares and idiosyncratic climate exposure z\_i = Z + zeta\_i (p. 986). The terminal cash flow per share, given investment choice k in {0, 1}, is (eq. 2, p. 987):

$$
V(k) = A + k\!\left(\theta + \alpha\eta_C + \sqrt{1-\alpha^2}\,\eta_I - c\right)
$$

where $$A \sim N(\mu_A, \tau_A^{-1})$$ are assets in place, $$\theta \sim N(\mu_\theta, \tau_\theta^{-1})$$ is the learnable cash-flow component, $$\eta_C \sim N(0, \tau_\eta^{-1})$$ are climate shocks, $$\eta_I \sim N(0, \tau_\eta^{-1})$$ are idiosyncratic shocks, $$\alpha \in [-1,1]$$ is the project's climate exposure ("greenness"), and c >= 0 is investment cost. Projects with $$\alpha > 0$$ are "green" (cash flows high when climate outcomes are good), projects with $$\alpha < 0$$ are "brown."

Investor i's terminal wealth is (eq. 1, p. 986):

$$
W_i = (n + X_{i1} + X_{i3})V - X_{i3}P_3 - X_{i1}P_1 - z_i\eta_C
$$

Investor i maximizes expected CARA utility (eq. 3, p. 988):

$$
\mathcal{W}_i \equiv \sup_{x \in \mathbb{R}} \mathbb{E}_{i1}\!\left[-e^{-\gamma W_i}\right]
$$

subject to market clearing at each date (eq. 4, p. 988):

$$
\int_i X_{it}\,di = 0
$$

**Two managerial objectives.** A cash-flow-maximizing manager solves (eq. 5, p. 988):

$$
k(P_1) = \arg\max_k \mathbb{E}[V|\mathcal{F}_m]
$$

A price-maximizing manager solves (eq. 6, p. 988):

$$
k(P_1) = \arg\max_k \mathbb{E}[P_3|\mathcal{F}_m]
$$

where $$\mathcal{F}_m = \sigma(P_1)$$ is the manager's information set at date 2.

**The equilibrium.** A threshold equilibrium (Definition 1, p. 991) features prices depending on the sufficient statistic $$s_p = \theta + \frac{1}{\beta}\alpha Z$$, with the price taking a piecewise-linear form: $$P_3 = P_1 = A_1 + B_1 s_p$$ when $$s_p > \bar{s}$$ and $$P_0$$ otherwise. The manager invests if and only if the price exceeds the no-investment price (i.e., $$s_p > \bar{s}$$). The key feature is that $$\beta = \frac{\tau_\eta}{\gamma\alpha}$$ (p. 993), so the price statistic $$s_p$$ mixes cash-flow news ($$\theta$$) and discount-rate news ($$\alpha Z$$), both of which are relevant to the project's NPV.

**The NPV rule under price maximization** (eq. 21, p. 995):

$$
\text{NPV} \equiv s_p - \bar{s}_P = \underbrace{\theta - c}_{\text{cash flows}} - \underbrace{\frac{\gamma}{\tau_\eta}(n - \alpha Z)}_{\text{discount rate}}
$$

The first term is the expected cash flows from the project net of investment costs. The second term is the discount rate: it is higher when the firm is larger (n is higher) and lower (higher) for green (brown) projects when Z > 0. Green projects carry lower discount rates because they reduce investors' aggregate climate exposure.

## Method

The paper uses a rational-expectations equilibrium approach in a CARA-Normal model with four dates and feedback. The solution method is backward induction (p. 1014, Appendix): equilibrium is conjectured in a piecewise-linear form (eq. A.1), then verified by working backwards from t = 4 to t = 1.

At t = 3, given investment decision k, investor i's optimal demand is (eq. 9, p. 992):

$$
X_{i3} = \frac{\mathbb{E}_{i3}[V(k)] + \gamma\mathbb{C}_{i3}(V(k),\eta_C)z_i - P_3}{\gamma\mathbb{V}_{i3}(V(k))} - (n + X_{i1})
$$

Market clearing at t = 3 implies the equilibrium price (eq. 10, p. 992):

$$
P_3 = \mu_A - \frac{\gamma}{\tau_A}n + k\!\left(\theta - c - \frac{\gamma}{\tau_\eta}(n - \alpha Z)\right)
$$

The welfare measure is the ex-ante expected utility of an arbitrary investor (eq. 25-26, pp. 1001-1002):

$$
\mathcal{W} \equiv \mathbb{E}\!\left[-e^{-\gamma W_i(k(s_p))}\right] = \Pr(k=1)\mathbb{E}\!\left[-e^{-\gamma W_i(1)}\big|k=1\right] + \Pr(k=0)\mathbb{E}\!\left[-e^{-\gamma W_i(0)}\big|k=0\right]
$$

The certainty equivalent (eq. 29, p. 1001) decomposes into four channels: a cash flow channel, a nonclimate risk channel, a climate risk channel, and a risk-sharing/value-of-information channel:

$$
CE(k) = \underbrace{\mathbb{E}[V(k)]n}_{\text{cash flow}} - \frac{\gamma}{2}\!\left(\frac{1}{\tau_A} + k^2\!\left(\frac{1}{\tau_\theta} + \frac{1-\alpha^2}{\tau_\eta}\right)\!\right)n^2 - \frac{\gamma}{2\tau_\eta}(\mu_Z - k\alpha n)^2(1+\Gamma) - \frac{1}{\gamma}\log(D(k))
$$

where $$\Gamma(k)$$ captures the amplification of climate risk disutility through exposure heterogeneity (eq. 31, p. 1002) and $$D(k)$$ captures the value of information (eq. 30, p. 1001).

## Empirical specifications

This is a pure theory paper. There are no regressions or empirical specifications. The paper's propositions are proved analytically (Appendix, pp. 1014-1028). The only "empirical" content is calibrated comparative statics in Figures 2 and 3, using parameter values (tau\_theta = tau\_eta = tau\_A = 1, tau\_Z = mu\_Z = 0.5, n = 0.1 for Figure 2; tau\_theta = 0.5, tau\_Z = 3, tau\_zeta = 2, mu\_A = 0, tau\_A = 5, mu\_theta = c = tau\_eta = gamma = mu\_Z = n = alpha = 1 for Figure 3).

**Figure 2 (p. 999).** Compares Prob(invest) vs climate exposure alpha under cash-flow and price maximization. For ex-ante profitable projects (mu\_theta > c, Panel A) and ex-ante unprofitable projects (mu\_theta < c, Panel B). Key finding: in Panel A, the cash-flow rule's probability of investment is U-shaped in alpha while the price rule shows an opposite pattern. The two rules produce qualitatively different comparative statics for greenness.

**Figure 3 (p. 1009).** Plots ex-ante welfare with and without feedback as functions of (i) precision of outside exposures tau\_zeta and (ii) asset supply n. Shows feedback reduces welfare when n is small (Panel B) or when risk-sharing gains are large (Panel A), even though the manager's objective always improves.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| No empirical data used | Pure theory paper; Figures 2-3 use analytical calibrations only | N/A |

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13427) if you are:
building or extending feedback-effects models that incorporate systematic risk factor loadings;
studying how managerial compensation tied to cash flows versus stock prices affects green investment;
interested in the welfare implications of feedback when investors have heterogeneous climate risk exposures;
looking for a tractable CARA-Normal framework linking production-based asset pricing models (Cochrane (1991)) to feedback effects.

The locators above point to the exact propositions. The proofs are in the Appendix (pp. 1014-1028) and Internet Appendix.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(2), April 2025, pp. 981-1028. DOI: [10.1111/jofi.13427](https://doi.org/10.1111/jofi.13427). Copyright 2025 the American Finance Association. Wiley VOR licence; no CC licence. This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. Extract only; the verbatim PDF is not reproduced here.

> Banerjee, Snehal, Bradyn M. Breon-Drish, and Kevin M. Smith. "Feedback Effects and Systematic Risk Exposures." *The Journal of Finance* 80, no. 2 (April 2025): 981-1028. DOI: 10.1111/jofi.13427.


==============================================================================
# Value without Employment: Barkai & Panageas (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/barkai-value-employment-2025/
# Distilled: Young firms have sharply reduced their contribution to aggregate employment since the early 1980s, yet their contribution to aggregate sales and market capitalization has remained stable, implying a rising average-to-marginal revenue product of labor (ARPL-to-MRPL ratio). A calibrated model of dynamic firm heterogeneity shows this single shift jointly explains the labor share decline, muted investment-share response, and declining job flows, while predicting only a modest (~8%) long-run drop in aggregate consumption. J. Finance 2025, CC BY 4.0. Eight core results with source locators, datasets used, the model (dynamic firm heterogeneity with monopolistic competition), and the method with equations.
# Tags: paper-summary, business-dynamism, labor-share, markups, firm-dynamics
==============================================================================

**What this is.** The paper's core results, the economic model connecting the
ARPL-to-MRPL ratio to aggregate trends, and the key equations: enough to
understand what it found and why, without reading all 46 pages. To replicate
or extend, read the full source at [doi.org/10.1111/jofi.13505](https://doi.org/10.1111/jofi.13505).

## TL;DR

Young firms' employment contribution has fallen sharply since the early 1980s,
but their contributions to aggregate sales and stock market value have not
fallen similarly. The ratio of market-value contribution to employment
contribution of young-firm IPO cohorts has more than doubled over 1985 to 2014
(Figure 1, p. 3734). Pitchbook exit-value data and NETS establishment-level
data corroborate this pattern for the broader universe of firms. The divergence
implies a rising average-to-marginal revenue product of labor (ARPL-to-MRPL
ratio) for recent young-firm cohorts: these firms earn similar or greater
revenues per dollar of market value while employing far fewer workers.

The paper introduces this feature into a standard model of dynamic firm
heterogeneity (monopolistic competition, heterogeneous productivity, endogenous
bankruptcy), shows it jointly explains a large set of empirical trends including
the labor share decline, and then proves analytically that a 40% drop in
young-firm employment contribution translates into only a 4.8% drop in
steady-state consumption (under the base-case assumption that the shift
reflects lower labor intensity). Even in the worst case where the shift
reflects a rise in economic rents, the long-run consumption decline is bounded
at 8.1%, roughly five times smaller than the employment decline.

## Core results

Magnitudes and significance are as reported; `**`/`***` = 5%/1%. Locators
point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Employment contribution of recent IPO cohorts fell, but market-value and sales contributions did not show a similar decline | Figure 1, Panels A-D, pp. 3733-3734 | Log employment contribution of 2010-2014 cohort bin is 0.71 lower than 1985-1989 cohort bin; log sales contribution is only 0.11 lower; log market-value contribution is 0.07 higher |
| R2 | The ratio of market-value contribution to employment contribution has more than doubled | Figure 1, Panel D, p. 3734 | Normalized log ratio (1985-1989 = 0) rises by about 0.8 log points for the 2010-2014 cohort bin relative to the earliest cohort |
| R3 | Recent cohort deflated exit values (Pitchbook) are at least as large as those of earlier cohorts, controlling for age | Figures 3-4, pp. 3737-3738 | With the exception of the 1995-1999 cohort (dot-com era outlier), each successive cohort bin has deflated exit values similar to or larger than its predecessors at the same cohort age |
| R4 | Young firms (post-2005) exhibit slower employment growth at acquired establishments relative to older acquirers, conditional on year, industry, state, and establishment age and size | Tables II-IV, pp. 3743-3745 | Young Acquirer coefficient: -0.024\*\* to -0.039\*\*\* in the first specification (Table II); Young Acquirer x Post-2005 interaction: -0.148\*\*\* to -0.161\*\*\* (Table III); coefficient is stable with establishment age and size controls (Table IV) |
| R5 | A single change in the ARPL-to-MRPL ratio (modeled as a rise in rent share, i.e., lower xi\*) simultaneously reproduces the labor share decline, stable investment share, falling young-firm employment share, declining job creation and destruction, and rising market-value-to-employment ratio | Figures 6-9, pp. 3755-3758 | Model labor share declines from 0.60 to 0.54 (data: 0.60 to 0.54); young-firm employment share falls from ~22% to ~8% (data: ~18% to ~10%); investment share remains roughly flat; job creation and destruction rates each decline by ~4-5 pp |
| R6 | The elasticity of steady-state consumption to a decline in young-firm output is approximately 12%, so a 40% drop in new-firm output implies only a 4.8% decline in consumption | pp. 3727, 3761 | Under base-case parameters (xi = 0.93, alpha = 0.62), the pass-through factor (1 - xi) / (alpha xi) is approximately 12%; a 40% drop in new-firm output implies a 4.8% consumption decline |
| R7 | In the worst case (rise in economic rents), the long-run consumption decline from the observed dynamics is bounded at 8.1% | pp. 3727-3728, 3764 | Markup-driven scenario (Barkai (2020) estimates): 8.1% drop in steady-state consumption between old and new steady states, which is roughly five times smaller than the 40-50% decline in young-firm employment |
| R8 | The multisector model quantifies the total steady-state consumption impact at -8.1%, with the main negative term being the rent-share channel (-14.2%), partially offset by positive factor-reallocation effects (+12.4%) | Table VI, p. 3767 | dC^SS/C^SS = -0.081; Term 1 (new-firm output) = -0.047; Term 2 (rents) = -0.142; Term 3 (factor intensity) = -0.016; Term 4 (factor reallocation) = +0.124 |

**Overall (paper's conclusion).** Young firms are not weaker than their
predecessors in terms of value creation, only in terms of employment creation.
This divergence implies a higher ARPL-to-MRPL ratio for recent cohorts.
Attributing this shift to a rise in rents (as in Decker et al. (2016a),
De Loecker, Eeckhout, and Mongey (2021)) provides a unified explanation for
the labor share decline, declining business dynamism, and declining job
flows. Akcigit and Ates (2023) document the same dynamism trends; this paper
argues those trends are not mirrored in market values and thus cannot be
interpreted as pure firm weakness. Decker et al. (2020) show that firms have
become less responsive to idiosyncratic shocks; the model captures this as
lower $$\xi^*$$ for new-type firms, implying a reduced standard deviation of
employment changes for new cohorts. De Loecker, Eeckhout, and Unger (2020)
provide the sector-level markup estimates used in the multisector model
calibration. Gutierrez and Philippon (2017) document investmentlessness;
consistent with the model, the investment share responds only modestly.
Haltiwanger et al. (2017) provide employment growth dispersion data used
to calibrate productivity volatility $$\sigma$$. Even under this
worst-case interpretation, large declines in young-firm employment imply
only moderate long-run declines in aggregate consumption, because a substantial
part of the decline in new-firm output is an amplified transitory factor
adjustment rather than a permanent loss.

## Theory / model

The economic setting is monopolistic competition with heterogeneous
intermediate-goods producers (Section II, pp. 3746-3750). Time is
continuous. A representative final-goods firm assembles intermediate inputs
using a Kimball (1995) aggregator (eq. 7, p. 3747):

$$
1 = \int_{i \in I} \mathcal{Y}\!\left(\frac{y_{it}}{Y_t}\right) di \tag{7}
$$

where $$y_{it}$$ is the quantity of intermediate good $$i$$ at time $$t$$, $$Y_t$$ is
aggregate output, and $$\mathcal{Y}(\cdot)$$ satisfies $$\mathcal{Y}(1)=1$$,
$$\mathcal{Y}'>0$$, $$\mathcal{Y}''<0$$. In the Dixit-Stiglitz special case
$$\mathcal{Y}(x) = x^{\xi}$$ with $$\xi \in (0,1)$$, this reduces to the familiar
CES aggregator (eq. 8, p. 3747):

$$
Y_t = \left(\int_{i \in I} y_{it}^{\xi}\, di\right)^{1/\xi}. \tag{8}
$$

Each intermediate producer uses a Cobb-Douglas production function
(eq. 12, p. 3748):

$$
y_{it} = Z_{it} k_{it}^{1-\alpha_i} l_{it}^{\alpha_i}, \tag{12}
$$

where $$l_{it}$$ is labor, $$k_{it}$$ is capital, $$\alpha_i \in (0,1)$$ is the
labor intensity (allowed to vary across firms), and $$Z_{it}$$ is
firm-specific productivity following a geometric Brownian motion
(eq. 13, p. 3748):

$$
dZ_{it} = \mu Z_{it}\, dt + \sigma Z_{it}\, dW_{it}. \tag{13}
$$

Firms must also pay a fixed overhead labor cost $$\bar{l}$$ per unit time (the
"operating leverage" that drives endogenous bankruptcy). They face an
exogenous death shock at rate $$\lambda > 0$$. The first-order conditions for
capital and labor equate marginal revenue products to factor prices
(eqs. 15-16, p. 3749):

$$
(1-\alpha_i)\!\left(1 - \frac{1}{\varepsilon_i}\right) p_{it} Z_{it}
  k_{it}^{-\alpha_i} l_{it}^{\alpha_i} = r_t^K, \tag{15}
$$

$$
\alpha_i\!\left(1 - \frac{1}{\varepsilon_i}\right) p_{it} Z_{it}
  k_{it}^{1-\alpha_i} l_{it}^{\alpha_i - 1} = w_t, \tag{16}
$$

where $$\varepsilon_i \equiv -\mathcal{Y}^{\prime}\!\left(\frac{y_{it}}{Y_t}\right) /
\!\left[\frac{y_{it}}{Y_t} \mathcal{Y}^{\prime\prime}\!\left(\frac{y_{it}}{Y_t}\right)\right]$$
is the absolute demand elasticity. In the Dixit-Stiglitz case,
$$\xi_i = 1 - \frac{1}{\varepsilon_i}$$ is the sum of factor shares, and the
markup is $$\frac{1}{\xi_i} - 1$$. The ARPL-to-MRPL ratio for firm $$i$$
follows directly from (16) (eq. 17, p. 3749):

$$
\frac{\frac{p_{it} y_{it}}{l_{it}}}{\frac{\partial(p_{it} y_{it})}{\partial l_{it}}} = \frac{1}{\xi_i \alpha_i}. \tag{17}
$$

A high ARPL-to-MRPL ratio therefore corresponds to a small $$\xi_i$$ (large
rent share) or a small $$\alpha_i$$ (low labor intensity), or both.

**Representative household.** The household maximizes (p. 3750):

$$
\mathrm{E}_t\!\left(\int_t^{\infty} u(C_s) e^{-\rho(s-t)}\, ds\right),
\quad u(C) = \frac{C^{1-\gamma}}{1-\gamma},
$$

with discount rate $$\rho > 0$$ and inverse IES $$\gamma > 0$$. Goods market
clearing requires $$C_t + I_t = Y_t$$; capital evolves as
$$\dot{K}_t = -\delta K_t + I_t$$.

**Identification and counterfactual.** Section II.D (pp. 3751-3754) considers
a transition experiment where from time $$t_0$$ onward a fraction
$$1 - e^{-\vartheta(t-t_0)}$$ of arriving firms are "new-type" with
$$\xi^* < \xi$$. The model is calibrated to match the labor share and employment
volatility of young firms from $$t_0 = 1983$$ onward (Table V, p. 3753).

## Method

The paper combines two methods: a descriptive-empirical analysis of cohort
contributions (applied to Compustat, Pitchbook, and NETS) and a dynamic
structural model solved analytically in steady state plus numerically along
the transition path.

The paper builds on `dynamic-general-equilibrium` (the Kimball-aggregator
model with Cobb-Douglas production and GBM productivity) and `panel-regression`
(the NETS establishment-level switcher regressions). The analytical steady-state
solution exploits a time-age-cohort decomposition and the Kimball demand
structure to obtain closed-form expressions for wages, output, and the
cross-sectional distribution of productivity (eqs. 19-22, pp. 3750-3751).

**ARPL-to-MRPL decomposition.** Equation (6) (p. 3746) expresses the
ARPL-to-MRPL ratio as a function of two elasticities:

$$
\frac{\text{ARPL}_i}{\text{MRPL}_i} = \frac{1}{1 + \frac{y_i}{p_i}\frac{\partial p_i}{\partial y_i}}
  \times \frac{1}{\frac{l_i}{y_i}\frac{\partial y_i}{\partial l_i}} = \frac{1}{1+\text{markup}^{-1}} \times \frac{1}{\text{labor intensity}}. \tag{6}
$$

**Lemma 1 (cohort divergence, p. 3739).** Under a time-age-cohort
decomposition, the discrepancy between the log change in market-value
contribution and the log change in employment contribution of young firms
equals the log change in the (employment-weighted) ARPL-to-MRPL ratio of
young firms versus the whole economy (eq. 2, p. 3739):

$$
\Delta\!\log\!\left(\frac{P_{t,t}}{\sum_{s \leq t} P_{t,s}}\right)
  - \Delta\!\log\!\left(\frac{l_{t,t}}{\sum_{s \leq t} l_{t,s}}\right)
  = \Delta\!\log\!\left(\frac{\sum_{i,t} \omega^i_{t,t}(\delta^i_{t,t}-1)}{\sum_{i,s \leq t}\omega^i_{t,s}(\delta^i_{t,s}-1)}\right), \tag{2}
$$

where $$\delta^i_{t,s} \equiv \frac{p^i_{t,s} y^i_{t,s} / l^i_{t,s}}{w_t}$$
is the ARPL-to-MRPL ratio of firm $$i$$ born at time $$s$$.

**Proposition 1 (aggregate implications, p. 3760).** In the single-sector
economy, as the discount rate $$\rho \to 0$$, the percentage change in
steady-state consumption is (eq. 23, p. 3760):

$$
\frac{dC^{SS}}{C^{SS}} = \frac{1}{\alpha}\frac{1-\xi}{\xi}\frac{dY^{\text{new}}}{Y^{\text{new}}}
  - \left(\frac{1-\xi}{1-\xi(1-\alpha)}\right)\frac{d\alpha}{\alpha}
  + \frac{1}{\alpha}(G - D)\frac{d\xi}{\xi}, \tag{23}
$$

where $$G \equiv \int_{z^*}^{\infty}(\tilde{g}(z) - \tilde{m}(z))\log y(z)\,dz$$
and $$D \equiv (1 + \frac{1}{\xi}) + \alpha\frac{\xi(1-\alpha)}{1-\xi(1-\alpha)} - (1-\alpha)$$,
with $$\tilde{g}(z)$$ and $$\tilde{m}(z)$$ the stationary and entering revenue
distributions. The factor $$\frac{1-\xi}{\alpha\xi}$$ in the first term is
approximately 12% under baseline parameters, quantifying why a large drop in
new-firm output passes through only modestly to consumption.

**Proposition 2 (multisector, p. 3765)** generalizes (23) to $$N^S$$ sectors
with sector-specific $$\xi^S$$ and $$\alpha^S$$ (eq. 30, p. 3765), adding a
fourth term that captures factor reallocation across sectors.

## Empirical specifications

Three separate empirical exercises establish the employment-value divergence.

**Compustat cohort contributions (Section I.A, pp. 3731-3734).** The paper
forms five-year IPO cohort bins for all Compustat nonfinancial U.S. public
firms founded within 10 years of their IPO, 1985 to 2014 (7,565 firms, 82,823
firm-year observations). Employment, sales, and market-value contributions of
each cohort are expressed as shares of aggregate totals among all Compustat
firms in the same year and summed within bins (Figure 1). No regression
estimator is used here; the results are descriptive cohort-share series.

**Pitchbook exit values (Section I.B, pp. 3735-3738).** For firms exiting by
IPO or M&A, post-money valuations are deflated by aggregate stock market
capitalization and traced by cohort age. This is also a descriptive exercise,
not a regression.

**NETS establishment switchers (Section I.D, pp. 3740-3745).** The headline
regression is:

$$
\log L_{t+1} - \log L_{t-1} = \beta_1\,\text{YoungAcquirer}_{it}
  + \beta_2\,(\text{YoungAcquirer}_{it} \times \text{Post-2005})
  + \gamma_t + \delta_{j \times k \times s} + \varepsilon_{it}, \tag{OLS}
$$

where the dependent variable is the log change in employment from year $$t-1$$
(before acquisition) to year $$t+1$$ (after acquisition) for a target
establishment. YoungAcquirer equals one if the acquiring firm is less than
eight years old. Fixed effects include year $$\gamma_t$$ and year
$$\times$$ four-digit SIC $$\times$$ state $$\delta_{j \times k \times s}$$. Standard
errors are clustered by year $$\times$$ SIC4 $$\times$$ state. Sample: 213,792
acquisitions, 1998 to 2014 (Table I, p. 3743).

The first specification (Table II) pools the full sample. The main specification
(Table III) adds the YoungAcquirer $$\times$$ Post-2005 interaction term. Table IV
adds establishment age bin dummies and log-employment in year $$t-1$$ as
controls to address differential selection by young acquirers.

Results: YoungAcquirer coefficient = -0.024`**` to -0.039`***`; Post-2005
interaction = -0.148`***` to -0.161`***` (Tables II-III); results are stable
with age and size controls (Table IV).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Compustat (via WRDS) | U.S. public firm employment, sales, and market value by IPO cohort bin, 1985-2014 | [WRDS](/wiki/commercial/wrds/) (licensed) |
| PitchBook | Exit valuations (IPO and M&A) for private and public U.S. firms by founding-year cohort, 1990-2019 | [PitchBook](/wiki/commercial/pitchbook/) (licensed) |
| National Establishment Time Series (NETS) | Establishment-level employment and ownership changes (acquirer age), 1998-2014; 213,792 acquisitions | [NETS](/wiki/commercial/nets/) (licensed) |
| Census Business Dynamics Statistics (BDS) | Aggregate employment share of young firms (ages 0-5), firm-size distribution, job creation and destruction rates, 1983-2019 | [U.S. Census Bureau public data products](/wiki/datasets/census/) |
| BEA National Accounts (GDP-by-Industry, Fixed Asset Tables) | Labor share, investment share, and value added by sector for model calibration and multisector analysis | [NIPA](/wiki/datasets/nipa/) |

Sample for Compustat analysis: 7,565 firms, 82,823 firm-year observations,
1985-2014, annual frequency. NETS sample: 213,792 acquisitions, 1998-2014.
BDS and BEA coverage: 1983-2019 (model transition period).

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13505) if you are:
studying the connection between business dynamism and the labor share and want
the analytical propositions; examining the welfare implications of rising
markups or declining labor intensity; building a calibrated model of firm
dynamics with heterogeneous markups; or using the NETS switcher design to
measure young-firm employment behavior. The Internet Appendix contains proofs,
the elastic labor supply extension, the numerical algorithm for the transition
path, and sector-level parameter estimates.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(6), December 2025. This
distillation was extracted by an LLM on 2026-06-03 and is **not
human-verified or independently reproduced**. The CC BY 4.0 licence permits
mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Barkai, Simcha, and Stavros Panageas.
> "Value without Employment."
> *The Journal of Finance* 80, no. 6 (December 2025): 3725-3770.
> DOI: 10.1111/jofi.13505. Copyright 2025 The Author(s).
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Interlocking Directorates and Competition in Banking: Barone, Schivardi & Sette (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/barone-interlocking-directorates-competition-banking-2025/
# Distilled: Exploiting Italy's 2011 Monti Decree, which unexpectedly banned shared board members among competing banks, the paper finds that prohibiting interlocks reduced corporate loan interest rates by 14 to 32 basis points on previously interlocked relationships, with larger drops for high-quality firms and high-market-share networks, and a subsequent increase in price dispersion consistent with competitive pricing. J. Finance 2025, CC BY 4.0. Seven core results with source locators, datasets used, the identification strategy, and the estimating specifications with equations.
# Tags: paper-summary, banking, competition, corporate-lending, interlocking-directorates
==============================================================================

**What this is.** The paper's core results, the identification strategy, and the estimating specifications with equations: enough to understand what the reform caused and how it was measured, without reading all 54 pages. To replicate or extend, read the full source at [doi.org/10.1111/jofi.13464](https://doi.org/10.1111/jofi.13464).

## TL;DR

The paper exploits Italy's "Save Italy" Monti Decree of December 2011 (Article 36), which unexpectedly banned shared board members (SBMs) among competing banks, as a natural experiment to identify the causal effect of interlocking directorates (IDs) on corporate loan pricing. Using a difference-in-differences design on 3.5 million firm-bank-quarter observations from the Italian Credit Register (2011Q1-2014Q4), and including a full set of firm-quarter and bank-quarter fixed effects to control for all time-varying unobserved heterogeneity at both the firm and bank level, the paper finds that prohibiting IDs reduced corporate loan interest rates by 14 to 32 basis points on treated relationships relative to controls. The effect grows to 29.2 bps at the peak (three years after reform), price dispersion on treated loans increases post-reform (consistent with a shift away from collusive uniform pricing), high-quality firms benefit most, networks with larger market share show the largest drops, and firms more exposed to interlocked banks improve investment, employment, and sales after the reform.

## Core results

Magnitudes and significance are as reported; `\*p < 0.10`, `\*\*p < 0.05`, `\*\*\*p < 0.01`. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Prohibiting IDs reduced corporate loan rates on treated relationships by 14-32 bps** (depending on fixed-effect saturation level) | Table V, p. 1985 | Preferred specification (col. 5, firm-bank + firm-quarter + bank-quarter FE): -0.139\*\*\* (0.041); baseline (col. 2, firm and bank controls + FE): -0.317\*\*\* (0.079) |
| R2 | **Effect builds over time; rate drop reaches 29.2 bps at the peak** three years after reform; no pre-trend detected | Figure 4, p. 1987 | Quarterly DiD coefficients significant from Q4 post-reform onward; cumulative drop by quarter 12 of approximately -0.30 percentage points |
| R3 | **Price dispersion on treated relationships increases post-reform**: consistent with a competitive shift; dispersion was 25 bps lower on treated loans pre-reform | Table VIII, p. 1993 | Treated\*Post = 0.067\*\* (col. 1) to 0.100\*\*\* (col. 2); pre-reform dispersion convergence confirmed (F-test p=0.000) |
| R4 | **High-quality (creditworthy) firms benefit disproportionately** from the reform: low-leverage firms see 16 bps larger rate drop | Table IX Panel A, p. 1995 | Treated\*Post\*Low interaction for leverage: -0.155\*\*\* (0.036); for ROA: +0.125\*\*\* (0.033); for liquidity: +0.116\*\*\* (0.036); for Z-score: +0.127\*\*\* (0.047) |
| R5 | **Networks with higher market share show larger rate drops**, consistent with stronger pre-reform collusion | Table IX Panel B, p. 1995 | Treated\*Post = -0.232\*\*\* (0.053) for high-market-share; Treated\*Post\*Low = +0.177\*\*\* (0.057) |
| R6 | **Investment rate increases for firms more exposed to interlocked banks**: 1.056 pp per unit share of treated credit | Table XII, p. 2003 | Treated\*Post = 1.056\*\* (0.492) without controls; 0.843\* (0.490) with controls; aggregate investment rate of sample firms up by ~0.3 pp/year |
| R7 | **Sales and employment grow for firms more exposed to treated banks** post-reform: 1.5 pp more sales growth; 0.8 pp more wage bill growth | Table XII, p. 2003 | Sales: Treated\*Post = 1.538\*\*\* (0.514), 1.593\*\*\* (0.513) with controls; wage bill: 0.757\* (0.413), 0.779\* (0.415) |

**Overall (paper's conclusion).** Prohibiting IDs among competing banks has competition-enhancing effects: it reduces the interest rates firms pay on corporate loans, increases price dispersion consistent with a breakup of collusive pricing, and translates into better real outcomes for borrowers. The effect is stronger for networks with greater market power and for firms that can better exploit improved outside options. The reform had no negative effect on credit supply, ruling out the credit-rationing mechanism. These results support stricter enforcement of ID bans.

## Theory / model

The paper has no formal mathematical model. It uses standard collusion theory as the organizing framework: IDs can facilitate collusion among competing banks by creating a direct channel for information exchange about prices and market strategies, making deviations from a collusive strategy more easily detectable and punishment more credible (Stigler (1964), Green and Porter (1984)). Under this view, shared board members (SBMs) are particularly effective collusion devices because SBMs in executive positions have direct access to pricing information and control over lending policies (p. 1968-1969).

The paper relates to three contemporaneous studies that also document anticompetitive effects of firm interconnections. Geng, Gopalakrishna, and Huang (2021) exploit staggered U.S. state-level ID law changes to show IDs reduce competition, particularly for R&D-intensive firms. Colombo (2022) studies the U.S. airline market and finds that shared appointments in a third company reduce supply and raise prices. Azar, Schmalz, and Tecu (2018) document that common ownership among airlines raises ticket prices 3-7%, providing a benchmark for effect-size comparison. The paper also uses the firm-quarter fixed-effect strategy from Khwaja and Mian (2008) to isolate bank loan supply. Faia, Mayer, and Pezone (2022) study the same Save Italy decree's effect on Italian listed firms' stock returns. Crawford, Pavanini, and Schivardi (2018) show that credit rationing via Stiglitz and Weiss (1981) can intensify in more competitive markets, motivating the paper's credit-quantity analysis.

The theory generates several comparative statics that the paper tests:

1. If collusion raises prices uniformly, the rate drop post-reform should be larger for banks with greater combined market share (stronger market power pre-reform).
2. If collusion reduces price dispersion (to prevent detection of deviations), dispersion should increase after the ban.
3. If more creditworthy firms have better outside options in a competitive market, they should gain more from the reform.
4. If IDs facilitate information sharing that reduces adverse selection, credit quantity should not fall post-reform; if anything, the results show credit is stable or slightly increasing.

The paper uses the Italian antitrust authority's practice of defining markets at the NUTS 3 provincial level (110 provinces, average radius 30 km, broadly comparable to U.S. counties) and requires that a network of interlocked banks jointly hold above 20% market share in a province to define treatment (p. 1972-1973).

## Method

The identification strategy exploits the Monti Decree (Law Decree 201/2011 of December 6, 2011), passed unexpectedly by the newly formed Monti government in the context of Italy's sovereign debt crisis. The decree was unanticipated because it was drafted under acute fiscal pressure and overcame intense banking-sector lobbying that had previously blocked similar regulation. Article 36 prohibited individuals from simultaneously holding governing-body positions in two competing banking groups, and required compliance by April 2012. This constitutes an exogenous breakup of network connections not driven by bank pricing strategies.

The two key features that make identification credible are (p. 1974-1977):

1. Firms typically borrow from multiple banks, so a given firm can have loans that are "treated" (from a bank in an interlocked network in that province) and "control" (from banks in the same province outside any network) within the same quarter. This allows the inclusion of full firm-quarter fixed effects that absorb all time-varying observed and unobserved firm-level heterogeneity.
2. A given bank can have treated loans in provinces where it belongs to a network and control loans in provinces where it does not. This allows the inclusion of full bank-quarter fixed effects that absorb all time-varying bank-level heterogeneity, including any effects of the Eurozone sovereign debt crisis or other monetary policy measures that affect banks differentially.

The treatment indicator $$TR_{ijp}$$ is time-invariant, equal to one if bank $$j$$'s loans to firms in province $$p$$ satisfy three conditions in all four quarters before the decree (pp. 1972-1974): (i) bank $$j$$ belongs to an interlocked network in province $$p$$; (ii) bank $$j$$ and at least one connected counterpart each have at least 1% individual market share in $$p$$; and (iii) the network's combined market share in $$p$$ exceeds 20%.

## Empirical specifications

**Baseline DiD (equation 1, p. 1975).** The estimating equation is

$$
r_{ijpt} = \alpha_0 + \alpha_1 POST_t + \alpha_2 TR_{ijp} + \alpha_3 TR_{ijp} \times POST_t + \alpha_4' \mathbf{X}_{ijt} + D_{ijpt} + \epsilon_{ijpt}, \tag{1}
$$

where $$r_{ijpt}$$ is the gross interest rate on revolving credit lines that bank $$j$$ charges firm $$i$$ in province $$p$$ in quarter $$t$$; $$POST_t$$ is a dummy equal to one from 2012Q1 onward; $$TR_{ijp}$$ is the time-invariant treatment dummy; $$\mathbf{X}_{ijt}$$ is a vector of time-varying firm and bank characteristics (lagged one period); and $$D_{ijpt}$$ denotes combinations of fixed effects. The causal parameter of interest is $$\alpha_3$$: the within-firm-quarter, within-bank-quarter drop in rates on treated relationships post-reform.

The preferred specification (Table V, column 5) includes firm-bank, firm-quarter, and bank-quarter fixed effects, so $$\alpha_3$$ is identified purely from within-firm-bank variation over time. Standard errors are clustered at the bank-province level.

Sample: 3,561,068 firm-bank-quarter observations, 2011Q1-2014Q4 (four preperiod quarters, twelve postperiod quarters). 29% of observations are treated.

**Price dispersion regression (equation 2, p. 1992).** To test the collusion-consistent prediction that treated loans show less price dispersion before the reform and more after, the paper constructs the province-quarter standard deviation of rates separately for treated and control loans:

$$
\sigma_{pct} = \sqrt{\frac{1}{n_{pct}} \sum_{i,j \in pc} (r_{ijpt} - \bar{r}_{pct})^2},
$$

and estimates:

$$
\sigma_{pct} = \gamma_0 + \gamma_1 POST_t + \gamma_2 TR_{pc} + \gamma_3 TR_{pc} \times POST_t + Dum_{pct} + e_{pct}, \tag{2}
$$

where $$TR_{pc}$$ is an indicator for treated province-category cells. Results reported in Table VIII (p. 1993).

**Heterogeneity specification (equation 3, p. 1993).** To test comparative statics on firm and network characteristics:

$$
r_{ijpt} = \beta_0 + \beta_1 TR_{ijp} \times POST_t + \beta_2 TR_{ijp} \times POST_t \times HET_{ij} + D_{it} + D_{jt} + D_{ij} + \eta_{ijpt}, \tag{3}
$$

where $$HET_{ij}$$ is a low-dummy (equal to one for values at or below the median) for a firm or network characteristic measured before the reform. Firm characteristics include size, leverage, ROA, liquidity, and Z-score; network characteristics include market share, number of connected banks, HHI, market share gap, and multimarket contacts (Table IX, p. 1995).

**Spillover regression (equation 5, p. 1998).** To test for within-province and within-bank spillovers from treated to control loans:

$$
r_{ijpt} = \theta_0 + \theta_1 POST_t + \theta_2 Z_{jp} + \theta_3 Z_{jp} \times POST_t + \theta_4' \mathbf{X}_{ijt} + D_{ijpt} + u_{ijpt}, \tag{5}
$$

where $$Z_{jp}$$ is a province-level or bank-level share of interlocked credit. No evidence of spillovers is found (Table A.VI).

**Firm-level rate and credit-quantity regressions (equations 6-8, pp. 1999-2000).** The firm-level average interest rate is the credit-weighted average across lenders:

$$
r_{it} = \sum_j \frac{loan_{ijt}}{\sum_j loan_{ijt}} r_{ijt}, \tag{6}
$$

and the share of treated credit at the firm level is:

$$
ShTr_i = \frac{\sum_j TR_{ij} \times loan_{ij,2011Q4}}{\sum_j loan_{ij,2011Q4}}. \tag{7}
$$

The estimating equation at the firm level is:

$$
r_{it} = \beta_0 + \beta_1 Post_t + \beta_2 ShTr_i + \beta_3 ShTr_i \times POST_t + \beta_4' \mathbf{X}_{it} + Dum_i + Dum_t + \eta_{it}. \tag{8}
$$

A firm with all treated credit records a ~30 bps drop in interest rates (Table X, p. 2000), confirming the relationship-level results are not offset by reallocation across lenders. The same structure is used with log-granted credit as the dependent variable for quantity analysis (Table XI) and with investment rate, wage bill growth, and sales growth for real effects (Table XII).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Italian Credit Register (Bank of Italy) | Quarterly firm-bank loan quantities and gross interest rates on revolving credit lines (overdraft facilities) for relationships with total lending above 75,000 euros; 2011Q1-2014Q4; 3.5 million firm-bank-quarters | No page yet |
| Or.So. Database (Bank of Italy) | Register of bank board members at quarterly frequency; identifies SBMs holding multiple appointments on competing bank boards | No page yet |
| Supervisory Reports (Bank of Italy) | Unconsolidated and consolidated balance-sheet data for individual banks and banking groups; provincial-level market shares; quarterly | No page yet |
| Cerved | Firm balance sheets and income statements for all Italian incorporated companies; provides ROA, leverage, liquidity, size, Z-score for firm controls and firm-level real-outcome regressions | [no page yet](/wiki/datasets/) |

Sample period: 2011Q1-2014Q4 (16 quarters, quarterly frequency). 192,732 unique firms; 604 banks at group level. Treatment defined in the four quarters before the reform (2011Q1-Q4). The Italian Credit Register is a proprietary administrative database with access restricted to researchers affiliated with the Bank of Italy.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13464) if you are: studying the competitive effects of board interconnections (the paper applies a natural-experiment DiD design in the corporate lending setting); extending the analysis to other lending markets, countries, or forms of firm interconnection; interested in the robustness tables (Tables VI, VII) testing alternative market-share thresholds, alternative pre-period lengths, net vs. gross rates, and closed samples; or seeking the collusion-mechanism evidence (Section V: price dispersion in Table VIII; heterogeneity by network characteristics in Table IX Panel B; and the bargaining-power analysis in Table A.V and equation 4).

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(4). This distillation was extracted by an LLM on 2026-06-05 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Barone, Guglielmo, Fabiano Schivardi, and Enrico Sette.
> "Interlocking Directorates and Competition in Banking."
> *The Journal of Finance* 80, no. 4 (August 2025): 1963-2016.
> DOI: 10.1111/jofi.13464. © 2025 The Author(s).
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Uncovering the Hidden Effort Problem: Ben-Rephael, Carlin, Da & Israelsen (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/ben-rephael-uncovering-hidden-effort-problem-2025/
# Distilled: Uses minute-by-minute Bloomberg platform activity to construct a new measure of executive workday length (AWL) and shows that greater effort by CEOs and CFOs predicts positive earnings surprises, higher cumulative abnormal returns of 25-50 bps per one-hour AWL increase (persisting 4-10 weeks), and lower CDS spreads. Weather-based IV confirms causality. J. Finance 2025, CC BY 4.0. Eight core results with source locators, datasets used, the model, and the method.
# Tags: paper-summary, corporate-governance, executive-compensation, moral-hazard
==============================================================================

**What this is.** The paper's core results, the effort measure it constructs, and the main empirical specifications: enough to know what it found and how, without reading all 51 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13429).

## TL;DR

The paper hand-collects minute-by-minute Bloomberg online status for 252 named executives (CEOs, CFOs, and other top executives) at public U.S. companies, 2017 to 2020. From this it constructs the Average Workday Length (AWL), a measure of workday span derived via an EM-based Gaussian mixture model of intraday platform activity. The paper then shows that higher AWL predicts better firm outcomes across multiple dimensions: positive earnings surprises (SUE), higher cumulative abnormal stock returns (25-50 bps per one-hour AWL increase, persisting 4-10 weeks), and lower credit default swap spreads. A weather-based instrumental variable confirms the causal direction. The paper also revisits classic agency questions: executives near their bonus EPS targets increase effort, and peer-firm sales growth (not own sales growth) drives subsequent effort, consistent with competition and peer pressure motivating harder work.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Higher AWL predicts higher SUE**: a 1-SD increase in AWL raises SUE by 0.11 SD | Table IV, p. 1283; text p. 1282 | AWL coefficient = 0.069\*\* to 0.086\*\*\* across all specifications; specs 1-6 (all industries): 0.075\*\* to 0.086\*\*\*; spec 7 (nonfinancial firms only): 0.069\*\* |
| R2 | **AWL predicts higher CARs** around earnings announcements: a 1-hour increase raises the 1-day post-announcement CAR by 27.35 bps, with the effect plateauing at 30-50 bps over 4-10 weeks | Table V, Panel A, p. 1284-1285 | 1-Day coefficient = 27.35\* (SE=14.01); 4-10 Week coefficients range from 32 to 49 bps\*\* |
| R3 | **Effect is larger for nonfinancial firms**: 1-hour increase in AWL associated with 43.91 bps at 1-day, rising to 80-100 bps over 7-10 weeks | Table V, Panel B, pp. 1285-1286 | 1-Day = 43.91\* (SE=19.33); 7-Week = 91.33\*\* (SE=27.33); 10-Week = 104.38\*\*\* (SE=26.61) |
| R4 | **Calendar-time long-short portfolio on extreme AWL changes yields 7.33 bps/day** risk-adjusted (37 bps over five days), statistically significant | Table VI, p. 1287 | Risk-adjusted H-L = 7.330\*\*\* (SE=3.129); High-Effort alpha = 4.579\*\*\* (SE=1.569) |
| R5 | **Higher AWL reduces the firm's CDS spread**: a 1-hour increase in AWL is associated with a -0.879 to -1.50 bps reduction in the next quarter's CDS spread | Table VII, p. 1289 | Spec (1) coefficient = -0.879\*\* (SE=0.380); spec (4) = -1.504\*\*\* (SE=0.600) including executive FE |
| R6 | **Weather-instrumented AWL (2SLS) confirms causality**: predicted AWL coefficient on SUE is positive and significant across all IV specifications; CAR effects grow over time and are significant from week 2 onward | Tables X-XI-XIII, pp. 1295-1303 | 2SLS second stage on SUE (Table X Panel B): range 0.058\*\*\* to 0.097\*\* across 8 specs; spec (1) = 0.093\*\*\* (SE=0.028), specs (7)-(8) = 0.067\*\*\* (SE=0.022/0.025); 2SLS CAR at 15-Day horizon = 61.67\*\*\* bps (SE=18.52), 4-Week = 81.78\*\*\* bps (SE=23.90) (Table XIII) |
| R7 | **Locus of control matters**: executives near bonus EPS targets significantly increase effort in H2 of the fiscal year; when the bonus is far outside their locus of control, effort declines | Table XV, p. 1307 | Interaction Pct\_cash\_perf \* Target\_1\_pct = 21.07\*\* (SE=7.25) to 22.72\*\* (SE=7.72) hours change in AWL; consistent with Healy (1985) |
| R8 | **Peer competition drives effort**: a 10% increase in peer-firm sales growth raises executive AWL by 0.25-0.45 hours/day over the next quarter; own-firm sales growth has no significant effect | Table XVI, p. 1308 | Lag1\_%Chng\_PeerSales coefficient = 0.025\*\* (SE=0.011) in full sample (spec 2); 0.045\*\*\* (SE=0.015) in nonfinancial firms (spec 8); Lag1\_%Chng\_Sales coefficient insignificant |

**Overall (paper's conclusion).** Executive effort, as measured by Bloomberg workday length, has a positive and statistically significant effect on earnings surprises and cumulative abnormal stock returns around earnings announcements. This effect is not anticipated by equity market participants, since it is not embedded in prices prior to announcements. Executive effort is also associated with lower credit default swap spreads. Weather-based IV supports a causal interpretation. Agency analysis shows that effort responds to compensation incentive structures and to competition, confirming the relevance of the classic principal-agent framework for executive behavior.

## Theory / model

The paper has no formal structural model; it tests a principal-agent hypothesis about hidden effort. The motivation is drawn from the standard moral hazard framework: because executive effort is unobservable to outsiders, financial markets cannot price it directly, and investigators cannot use standard regression tools to test whether effort raises firm value (Murphy (1999)). The paper's identifying insight is that Bloomberg platform activity is publicly observable and provides a plausible proxy for the work habits of executives who spend most of their day doing activities other than using Bloomberg.

The paper positions itself as a complement to Yermack (2014) and Biggerstaff, Cicero, and Puckett (2017), who study the flip side of executive effort by examining leisure activities (vacation travel, golf habits) and their association with firm underperformance. Rather than measuring absence, this paper measures presence via workday length. Bandiera et al. (2020) measure CEO time use via direct monitoring and diary-based methods; this paper instead exploits publicly observable platform activity to avoid observer effects.

The paper's agency hypotheses (Section III) are:

1. **Locus of control (Healy (1985))**: executives increase effort when earning a bonus is within their locus of control (near the target) and decrease effort when it is far outside (Degeorge, Patel, and Zeckhauser (1999); Murphy (2000)). The within-executive design (changes in AWL from H1 to H2 within a fiscal year) makes this causal.

2. **Peer competition**: an increase in market share of competing firms should reduce the focal firm's relative standing and motivate more effort. The paper tests whether peer-firm sales growth (lagged one and two quarters) predicts subsequent executive AWL while own sales growth does not.

The identification strategy for the main results (Sections II.A-II.B) is a weather-based IV: day-level "feels like" temperature near the executive's headquarters (from Weather Underground, 2017 Q3 to 2019 Q4) is used to classify good-weather and bad-weather days within each quarter and location. Good weather in warm months (Q2-Q3) is associated with shorter workdays (approximately 12 minutes less per day, Table VIII), providing exogenous variation in AWL that is uncorrelated with the firm's business activity.

## Method

**AWL construction (Section I.C, pp. 1277-1280).** The effort measure is derived from minute-by-minute Bloomberg online status. For each executive-year, the paper observes the probability $$P^j_{\min}$$ that the executive is logged in at minute $$j \in J \equiv \{12{:}00\,\text{am},\ldots, 11{:}59\,\text{pm}\}$$. A probability density function is constructed as

$$
p^i_{\min} = \frac{P^i_{\min}}{\sum_J P^j_{\min}} \tag{1}
$$

This pdf is modeled as a mixture of two normal distributions (morning and afternoon sessions), with means $$\mu_1, \mu_2$$ ($$\mu_2 > \mu_1$$) and variances $$\sigma_1^2, \sigma_2^2$$, mixing weight $$q$$:

$$
\mu_{1,2} = q\mu_1 + (1-q)\mu_2
$$

$$
\sigma_{1,2}^2 = q\sigma_1^2 + (1-q)\sigma_2^2 + q(1-q)(\mu_2 - \mu_1)^2
$$

An EM algorithm (sklearn GaussianMixture, convergence threshold 0.001) estimates all five parameters $$(\hat{q}, \hat{\mu}_1, \hat{\mu}_2, \hat{\sigma}_1^2, \hat{\sigma}_2^2)$$ for each executive-year. The average workday length is then:

$$
AWL = (\hat{\mu}_2 - \hat{\mu}_1) + \hat{\sigma}_1 + \hat{\sigma}_2 \tag{2}
$$

The mean AWL across 520 executive-year observations is 9.47 hours (SD 2.10). The measure is validated via: (i) Bloomberg activity patterns consistent with a 9am-5pm workday, (ii) near-zero activity during firm events (analyst days, investor days: 100% of executives inactive), (iii) cell phone geolocation data for a subset of three executives (AWL from Bloomberg and AWL from geolocation agree closely: 8.0 vs 7.88 hours for one executive).

**Weather instrument (Section II.B, pp. 1288-1294).** The first-stage regression is:

$$
AWL_{j,y,q} = \alpha + \beta\,\text{WeatherAWL}_{j,y,q} + \vartheta_{j,y,q} \tag{3}
$$

where

$$
\text{WeatherAWL}_{j,y,q} = \left[W_{\text{Good},j,y,q}\,AWL(\text{good})_{j,q} + W_{\text{Bad},j,y,q}\,AWL(\text{bad})_{j,q}\right]
$$

is the weighted average of good-weather and bad-weather AWLs across all years in the sample. The second-stage regression is:

$$
Y_{j,y,q} = \delta + \varphi\,\widehat{AWL}_{j,y,q} + \varepsilon_{j,y,q} \tag{4}
$$

where $$\widehat{AWL}_{j,y,q} = \hat{\alpha} + \hat{\beta}\,\text{WeatherAWL}_{j,y,q}$$ is the fitted value from the first stage and $$Y_{j,y,q}$$ is the outcome variable (SUE, CAR, etc.). The method draws on `instrumental-variables` (2SLS), with `panel-regression` for the OLS baseline.

**Portfolio construction (Section II.A, p. 1287).** Calendar-time portfolios are formed around earnings announcements. The High-Effort portfolio on a given day includes stocks whose executives' AWL change (relative to four quarters prior) is in the top 10% across all executives with the same fiscal quarter-end, and whose earnings announcement occurred within the past five trading days. The Low-Effort portfolio is defined analogously (bottom 10%). Portfolio returns are value-weighted using market capitalization; Fama-French three-factor alphas are computed using a rolling year of past daily returns.

## Empirical specifications

**SUE regressions (Tables IV, X; pp. 1283, 1295-1296).** The OLS estimating equation is:

$$
SUE_{j,q} = \alpha_j + \beta\,AWL_{j,q} + \gamma\,\log\_\text{purchase}_{j,q} + \delta\,\log\_\text{sell}_{j,q} + \theta\,\mathbf{X}_{j,q} + \varepsilon_{j,q} \tag{5}
$$

where $$SUE_{j,q}$$ is standardized unexpected earnings, $$AWL_{j,q}$$ is measured during the fiscal quarter, $$\alpha_j$$ is an individual executive fixed effect, insider trading controls $$(\log\_\text{purchase}, \log\_\text{sell}, \log\_\text{purchase\_all}, \log\_\text{sell\_all})$$ capture private information, and $$\mathbf{X}_{j,q}$$ includes firm characteristics (size, leverage, productivity, Tobin's Q). Standard errors are clustered by executive. N = 980 (full sample); N = 459 (nonfinancial only). Data sources: Bloomberg (AWL), I/B/E/S (EPS), SEC Edgar (insider trading), Fama-French (industry definitions), CRSP/Compustat (firm characteristics).

**CAR regressions (Tables V, XIII; pp. 1284-1285, 1302-1303).** Cumulative abnormal returns are computed using the Fama-French three-factor model over rolling 50-trading-day windows from day 1 through day 50 post-announcement (1 through 10 weeks). The regression is:

$$
CAR_{j,q,d} = \alpha_j + \beta\,AWL_{j,q} + \gamma\,SUE_{j,q} + \delta\,\text{InsiderTrading}_{j,q} + \varepsilon_{j,q,d} \tag{6}
$$

Individual executive fixed effects are included. N = 1,128 executive-quarter observations. Standard errors are clustered by executive. For 2SLS (Table XIII), $$AWL_{j,q}$$ is replaced by $$\widehat{AWL}_{j,q}$$ from the weather first stage.

**CDS spread regressions (Table VII; p. 1289).** An AR(1)-style specification regresses next-quarter CDS spread on current AWL and current spread:

$$
\text{Spread}_{j,q+1} = \alpha_j + \beta\,AWL_{j,q} + \gamma\,\text{Spread}_{j,q} + \delta\,\text{InsiderTrading}_{j,q} + \theta\,\mathbf{X}_{j,q} + \varepsilon_{j,q} \tag{7}
$$

N = 574 observations from 89 executives at 57 firms (those with active five-year CDS contracts). Executive and year-fixed effects included in some specifications. Standard errors clustered by executive. CDS data from DataStream.

**Agency regressions (Tables XV-XVI; pp. 1307-1308).** Locus of control: change in AWL from H1 to H2 is regressed on the interaction between $$\text{Pct\_cash\_perf}$$ (fraction of cash bonus based on accounting metrics) and $$\text{Target\_1\_pct}$$ (indicator for H1 EPS within 1% of the annual EPS bonus target), plus firm characteristics and fixed effects. N = 91 executives. Competition: AWL in quarter $$t$$ is regressed on lagged own-firm and peer-firm sales growth (quarterly, four-quarter growth rate), with executive and year fixed effects, N = 1,256.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Bloomberg Professional (hand-collected platform activity) | Primary effort measure: minute-by-minute online status for 2,734 executives (2017-2020), matched to 252 named executives at public firms | No page yet |
| CRSP monthly stock returns and market data | Stock returns for CAR computation; market capitalization for value-weighting | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| Compustat quarterly fundamentals | Firm characteristics (size, leverage, Tobin's Q, productivity, EPS) | [WRDS / Compustat](/wiki/commercial/wrds/) (licensed) |
| I/B/E/S (IBES) earnings estimates | EPS actuals and forecasts for SUE construction | [WRDS / IBES](/wiki/commercial/wrds/) (licensed) |
| SEC EDGAR (insider trading filings) | Executive open-market purchases and sales (Form 4) for insider trading controls | [SEC EDGAR](/wiki/datasets/edgar/) |
| DataStream (CDS spreads) | Five-year CDS spread data for 57 firms | No page yet |
| ISS Incentive Lab | Compensation contract data (proxy statements) for 252 executives at 174 firms | No page yet |
| Weather Underground | Historical daily weather ("feels like" temperature) for executive HQ locations, 2017 Q3-2019 Q4, used as instrument | No page yet |
| Reveal Mobile (geolocation) | Cell phone geolocation data for validation of AWL measure (anecdotal, 3 executives) | No page yet |
| Factiva (event transcripts) | Identification of executive presence at analyst days, investor days, and conferences | No page yet |
| Kenneth French Data Library | Fama-French 3-factor portfolios for CAR and alpha computation; FF-12 industry definitions | [Ken French library](/wiki/datasets/ken-french/) |
| Bloomberg corporate events calendar (EVTS function) | Event dates and types (earnings calls, analyst days, investor days, etc.) for personal-use validation | No page yet |

Sample: September 2017 to December 2019 (main effort sample; COVID period used only for validation). 252 executives, 520 executive-year observations; 1,128 executive-quarter observations for CAR regressions; 574 CDS observations.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13429) if you are: measuring executive effort using alternative proxies (the AWL construction algorithm is fully described in Section I.C); studying compensation contract design and incentive effects (Section III.A); interested in peer-competition effects on managerial behavior (Section III.B); or building on the Bloomberg platform data to study executive attention. The Internet Appendix contains extensive robustness tables (winsorized AWL, financial-only subsample, alternative weather windows).

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(2). This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Ben-Rephael, Azi, Bruce I. Carlin, Zhi Da, and Ryan D. Israelsen.
> "Uncovering the Hidden Effort Problem."
> *The Journal of Finance* 80, no. 2 (April 2025): 1261-1311.
> DOI: 10.1111/jofi.13429. © 2025 The Authors.
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Investor Factors: Betermier, Calvet, Knupfer & Kvaerner (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/betermier-investor-factors-2025/
# Distilled: pricing factors built from individual investor holdings (Norway 1997-2017); a two-factor model of the market plus a combined age-wealth portfolio prices the cross section of Norwegian equities out-of-sample and absorbs established firm factors. J. Finance 2025, paywalled. Eight core results with source locators, datasets used, the model, and the method with defining equations.
# Tags: paper-summary, asset-pricing, factors, household-finance, equities, portfolio-sort, panel-regression, peer-reviewed, unreplicated, data:titlon-ose, data:vps-norway, data:statistics-norway
==============================================================================

**What this is.** The paper's core results, datasets, and theory: enough to
know what it found without reading all 42 pages. For replication or extension,
read the full source at the [original](https://doi.org/10.1111/jofi.13474)
(paywall).

## TL;DR

Using complete administrative stockholdings of Norwegian individual investors
(308,000 investors/month, Feb 1997 to Dec 2017, 535 OSE stocks), the paper
derives theoretical conditions under which investor portfolio holdings reveal
pricing factors for the cross section of equity returns. It then constructs
Investor Pricing Factors (IPFs) by sorting investors into 90 groups by age,
wealth, and other characteristics. A two-factor model consisting of the market
(MKT) and a combined age-wealth portfolio (AW) prices both Norwegian equities
and established firm-based factors, while IPFs outperform firm-factor models
out-of-sample. Portfolio tilts toward the age-wealth factor are positively
linked to financial sophistication and negatively linked to debt and
macroeconomic income risk, consistent with joint hedging and sentiment
channels.

## Core results

Magnitudes and significance are as reported; `**`/`***` = 5%/1%. Locators
point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Two PCs explain 80% of cross-sectional variation** in group portfolio holdings; PC1 tracks the market (R2 = 0.62), PC2 tracks the combined age-wealth portfolio (R2 = 0.55) | Table I, p. 2807 | PC1 alone: 72% of variance; PC1+PC2: 80%; market R2=0.62 on PC1, AW R2=0.55 on PC2 |
| R2 | The combined age-wealth factor (AW) earns a **significant CAPM alpha** of 32 bps/month (3.8%/yr) after controlling for the market | Table II col. (2), p. 2810 | alpha = 0.32%, t = 3.16; CAPM beta on AW = -0.12, t = -6.96 |
| R3 | AW **spans firm factors**: its alpha remains 24 bps/month (sig. 5%) even after controlling for all five FF factors (size, value, momentum, profitability, investment) | Table III col. (5), p. 2813 | alpha = 0.24, t = 2.55; adj. R2 rises from 0.16 to 0.29 but AW alpha never insignificant |
| R4 | IPF\* **prices established firm factors**: adding AW to the market renders momentum, profitability, and investment alphas statistically insignificant and reduces them by ~40% | Table IV, p. 2814 | MOM alpha: 0.77% (CAPM) vs. 0.43% (IPF\*); RMW: 0.73% vs. 0.46%; CMA: 0.52% vs. 0.32%; all three IPF\* alphas statistically insignificant (MOM t=1.04, RMW t=1.42, CMA t=0.99) |
| R5 | **Out-of-sample Sharpe ratio** of IPF\* (0.45) exceeds all firm-factor models (0.19-0.40 range) and is 45% above the market (0.31) | Table V, p. 2817 | IPF\* OOS SR = 0.45; 3-factor age+wealth model = 0.51; best firm model (FIRM-6) = 0.40; market = 0.31; OS/IS ratio for IPF\* = 0.67 vs. 0.43 for all-firm model |
| R6 | Factor tilts **increase monotonically with age and wealth**: tilt rises from -0.3 (investors under 30) to +0.1 (70-75), equivalent to ~1.2%/yr average return difference | Figure 2 + p. 2821 | Tilt range [-0.3, +0.1] over life cycle; 0.4 x 3% = 1.2%/yr gap; holds for new entrants mimicking experienced investors within cohort |
| R7 | **Debt and income beta reduce tilts** (hedging channel) while finance occupation, stock market experience, and female gender raise tilts (sophistication/sentiment channel) | Table VII, p. 2823 | Income beta coef = -0.051 (t = -6.40); debt = -0.047 (t = -5.55); finance occup. = 0.627 (t = 34.60); stock mkt. experience = 0.026 (t = 7.58); male dummy = -0.156 (t = -15.00) |
| R8 | Stocks in the **long leg of AW have higher market cap, book-to-market, and profitability** than short-leg stocks; short-leg stocks have higher CAPM beta (1.02 vs. 0.73), volatility (0.18 vs. 0.08), and turnover | Table VIII, p. 2825 | Long-leg median mkt cap 973M NOK vs. 483M NOK; BtM 0.90 vs. 0.66; profitability 0.06 vs. 0.05; CAPM beta 0.73 vs. 1.02; volatility 0.08 vs. 0.18 |

**Overall (paper's conclusion).** Individual investor portfolios contain
recoverable pricing information. The market and the combined age-wealth
portfolio (long mature/wealthy, short young/less-wealthy investors) form a
parsimonious two-factor model that prices the Norwegian cross section, absorbs
firm factors, and outperforms firm-factor models out-of-sample. Hedging and
sentiment jointly drive investor tilts toward the pricing factor.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Titlon (Oslo Stock Exchange database) | Stock prices, returns, shares outstanding for 535 OSE stocks, 1997-2017 | [Titlon (OSE)](/wiki/commercial/titlon-ose/) (licensed) |
| VPS (Norwegian Central Securities Depository) | Complete individual investor stockholdings at monthly frequency, 300,000+ investors | no page yet |
| Statistics Norway (Statistisk sentralbyra) | Investor demographics, balance sheets, income, wealth from tax records; annual 1997-2017 | no page yet |
| OSE market index (Norwegian market portfolio) | Benchmark factor; market-cap-weighted portfolio of OSE stocks | no page yet |

Sample: 308,000 individual investors per month on average; 535 unique stocks; 251 months (Feb 1997 to Dec 2017).

## Theory / model

The paper's central theoretical object is the tangency portfolio, which prices
the cross section of excess stock returns. For J stocks with excess return
vector $$R^e$$, expected return vector $$\mu$$, and variance-covariance matrix
$$\Sigma$$, the tangency portfolio has weights (eq. 1, p. 2795):

$$
\tau = \frac{1}{\phi} \Sigma^{-1} (\mu - R_f \mathbf{1}), \qquad \phi = \mathbf{1}' \Sigma^{-1} (\mu - R_f \mathbf{1}) > 0
$$

Every stock's risk premium satisfies $$\mu_j - R_f = \phi \, (\Sigma \tau)_j = b_{j,\tau} (\mu_\tau - R_f)$$, so pricing the tangency portfolio is equivalent to
pricing all stocks (eq. 2, p. 2795).

**Spanning condition.** The key insight is that the researcher can recover the
tangency portfolio from investor portfolio holdings when Assumption 1 holds
(p. 2796): there exist N long-short investor portfolios $$\pi^1, \ldots, \pi^N$$
extracted from the sample such that

$$
\tau \in \operatorname{Span}[ m, \pi^1, \ldots, \pi^N ],
$$

where $$m$$ is the market portfolio and $$\operatorname{Span}[\cdot]$$ denotes the set of all linear
combinations. When this holds, the tangency portfolio is a linear combination
(eq. 4, p. 2796):

$$
\tau = m + \sum_{n=1}^{N} \eta_n \, \pi^n,
$$

and every stock's risk premium satisfies a multifactor pricing equation
(Proposition 1, eq. 6, p. 2797):

$$
\mu_j - R_f = \beta_{j,M} (\mu_M - R_f) + \sum_{n=1}^{N} \beta_{j,n} \, E(p_n),
$$

where $$p_n = (\pi^n)' R^e$$ is the return on the n-th IPF and
$$(\beta_{j,M}, \beta_{j,1}, \ldots, \beta_{j,N})'$$ is the vector of OLS regression
coefficients of stock j's return on the (N+1) factors.

Proposition 1 also implies that a stock's CAPM alpha satisfies (eq. 7, p. 2798):

$$
a_{j,M} = \phi \sum_{n=1}^{K} \eta_n (b_{j,n} - b_{j,M} \, b_{M,n}) \sigma_n^2,
$$

where $$b_{j,n} = \operatorname{cov}(R^e_j, p_n) / \sigma_n^2$$ and $$b_{M,n} = \operatorname{cov}(\text{MKT}, p_n) / \sigma_n^2$$.
Stocks in low demand (positive net exposure $$b_{j,n} - b_{j,M} \, b_{M,n}$$ with
$$\eta_n > 0$$) are underpriced relative to CAPM and tend to have low market betas.

**Theoretical foundations for age and wealth as IPF characteristics.** The paper
derives the spanning condition under two complementary models (Section I.D,
pp. 2802-2804):

1. ICAPM (Merton 1973, Breeden 1979) with heterogeneous investors: each investor
   i has CRRA utility and holds a portfolio deviating from the tangency by
   hedging demands. Under a Taylor approximation, the portfolio factor structure is
   (eq. 17, p. 2803):

   $$
   \omega^i_t = \tau_t - (T - 1 - A^i_t) \, d^1_t - (L^i_t / W^i_t) \, d^2_t,
   $$

   where $$A^i_t$$ is investor age, $$L^i_t / W^i_t$$ is the income-to-wealth ratio,
   and $$d^1_t, d^2_t$$ are deviation portfolios. Mature and wealthy investors hold
   portfolios closer to $$\tau$$ and therefore earn higher CAPM alphas.

2. Sentiment model (Fedyk, Heyerdahl-Larsen, and Walden 2013): sentiment covaries
   with age and wealth, yielding a reduced-form factor structure (eq. 18, p. 2804):

   $$
   \omega^i_t = \tau_t - f_1(A^i_t) \, d^1_t - f_2(W^i_t) \, d^2_t.
   $$

   Both frameworks predict age and wealth as natural IPF sorting characteristics.

## Method

The construction has two parts: grouping investors into a factor structure and
extracting priced long-short portfolios. It builds on `sdf-projection` (the
tangency-spanning condition) and `portfolio-sort` (investor sorting by
socioeconomic characteristics).

**Step 1: Factor structure of investor portfolios.** The strong factor structure
in individual investor portfolios documented by Balasubramaniam, Campbell,
Ramadorai, and Ranish (2023) motivates the PCA grouping approach. $$G = 90$$ investor groups are
formed annually by age (12 groups), wealth (12 groups), permanent real income
(12 groups), gender (2), education (3), region (9), industry (17), and
occupation (9). For group g with investor equity-wealth weights $$w^g_i$$, the
group portfolio is (eq. 14, p. 2801):

$$
\omega^g = \sum_{i \in I_g} w^g_i \, \omega^i, \qquad w^g_i = E^i \Big/ \sum_{i' \in I_g} E^{i'}.
$$

PCA is applied to the $$G \times G$$ variance-covariance matrix of the $$G = 90$$ group
portfolio holdings, $$\Omega'_t \Omega_t / J_t$$, to obtain principal components
$$PC_{k,t}$$ (eq. 19-20, p. 2806). The first two PCs explain 80% of the
cross-sectional variance in group holdings.

**Step 2: Extracting IPFs as long-short portfolios.** An IPF is a zero-investment
long-short portfolio $$\pi^n$$ formed as a weighted average of group portfolios
(eq. 15, p. 2802):

$$
\pi^n = \sum_{g=1}^{G} z^g_n \, \omega^g, \qquad \sum_{g=1}^{G} z^g_n = 0.
$$

The age portfolio $$\pi_{\text{AGE},t}$$ is long investors aged 70-75 and short investors
aged 18-30 (equal weights -1/2 on groups 1 and 2). The wealth portfolio
$$\pi_{\text{WEALTH},t}$$ is long the top 1% wealthiest investors and short the bottom
10%-30% of wealth (text, p. 2808). The combined age-wealth portfolio is:

$$
\pi_{\text{AW},t} = \tfrac{1}{2} \left( \pi_{\text{AGE},t} + \pi_{\text{WEALTH},t} \right).
$$

Returns on the IPFs are computed as $$AW_t = (\pi_{\text{AW},t-1})' R^e_t$$ (net of the
1-month Norwegian Interbank Offered Rate, NIBOR, as risk-free rate; p. 2810).

**Out-of-sample Sharpe ratio evaluation.** The bootstrap procedure (eq. 24,
p. 2815) follows Fama and French (2018): 100,000 bootstrap draws of $$T = 251$$
months from the factor return panel are used. The in-sample covariance matrix is
shrunk as $$\hat{\Sigma}_p = \Sigma_p + \gamma I$$ ($$\gamma = \operatorname{tr}(\Sigma_p) / (T \, E[\text{SR}]^2)$$, with $$E[\text{SR}] = 0.5$$ selected) following Kozak, Nagel, and Santosh (2020), and the tangency portfolio
is $$\hat{\tau} = \hat{\Sigma}_p^{-1} \mu_p / (\mathbf{1}' \hat{\Sigma}_p^{-1} \mu_p)$$. Out-of-sample SR
is computed on the hold-out months not drawn in each simulation.

## Empirical specifications

**PC factor-structure regression (R1; Table I, p. 2807).** The stock weight in
PC k is regressed monthly on the market, age, and wealth portfolio weights (OLS,
eq. 22, p. 2809):

$$
PC_{j,k,t} = a^k_t + \lambda^k_{\text{MKT},t} \, m_{j,t} + \lambda^k_{\text{AGE},t} \, \pi_{\text{AGE},j,t} + \lambda^k_{\text{WEALTH},t} \, \pi_{\text{WEALTH},j,t} + \epsilon^k_{j,t}, \qquad j = 1,\ldots,J_t.
$$

Time-average $$R^2$$ reported over 2005-2017. Identifies PC1 with market ($$R^2 = 0.62$$)
and PC2 with combined age-wealth factor ($$R^2 = 0.55$$).

**IPF alpha and beta regressions (R2, R3; Tables II-III, pp. 2810-2813).** Monthly
OLS spanning regressions of IPF returns on the market and/or firm factors,
February 1997 to December 2017 (T = 251):

$$
AW_t = \alpha + \beta_{\text{MKT}} \, \text{MKT}_t + [\beta_{\text{SMB}} \, \text{SMB}_t + \beta_{\text{HML}} \, \text{HML}_t + \beta_{\text{MOM}} \, \text{MOM}_t + \beta_{\text{RMW}} \, \text{RMW}_t + \beta_{\text{CMA}} \, \text{CMA}_t] + v_t.
$$

Five Fama-French firm factors (SMB, HML, RMW, CMA) and momentum (MOM) are
constructed from Norwegian equities using standard accounting and price data
(Internet Appendix Section III.B). Newey-West standard errors are not mentioned;
heteroskedasticity-robust t-statistics are reported.

**Firm-factor alpha regressions under IPF* (R4; Table IV, p. 2814).** Each firm
factor $$F_t \in \{\text{SMB}, \text{HML}, \text{MOM}, \text{RMW}, \text{CMA}\}$$ is regressed on the market and AW:

$$
F_t = \alpha_F + \beta_{F,\text{MKT}} \, \text{MKT}_t + \beta_{F,\text{AW}} \, AW_t + v_t.
$$

AW absorbs ~40% of momentum, profitability, and investment alphas, rendering
them statistically insignificant at the 5% level (MOM t = 1.04, RMW t = 1.42,
CMA t = 0.99; Table IV columns 2, 6, 8, 10).

**IPF* stock-level multifactor model (eq. 23, p. 2811).** The preferred two-factor
model (labelled IPF*) for each stock j:

$$
R^e_{j,t} = \alpha_j + \beta_{j,\text{MKT}} \, \text{MKT}_t + \beta_{j,\text{AW}} \, AW_t + v_{j,t},
$$

where $$\alpha_j = 0$$ for all j if IPF* is correctly specified.

**Portfolio tilt regression (R6, R7; Table VII, p. 2823).** This extends the
life-cycle links between demographics and value-factor tilts that Betermier,
Calvet, and Sodini (2017) documented for Swedish households into a full IPF
extraction framework. The panel regression of
investor i's tilt toward IPF AW is run at annual frequency, 2004-2017 (N = 911,432
investor-years):

$$
\text{TILT}^i_{\text{AW},t} = \delta_t + \theta' \chi^i_t + \xi^i,
$$

- $$\text{TILT}^i_{\text{AW},t} = \sum_{j=1}^{J_i} \omega^i_{j,t} \, D_{j,\text{AW},t}$$ with $$D_{j,\text{AW},t} = +1 / -1$$ if stock j is in the long / short leg of AW (eq. 25, p. 2819).
- $$\chi^i_t$$ is a vector of investor characteristics (income beta, debt indicator, stock market experience, finance occupation, gender, Oslo residence, top management dummy).
- $$\delta_t$$ are year fixed effects with additional age-group and wealth-group fixed effects.
- Standard errors clustered by calendar year x investor level.

## When to read the full paper

Read the full source if you are: constructing IPFs for other markets or asset
classes; extending the spanning-condition theory to institutional holdings;
using the bootstrap out-of-sample Sharpe methodology (Section III.C) for
factor evaluation; auditing specific coefficients in Tables III-VIII; or
reviewing the Internet Appendix robustness tests (alternative age/wealth
cutoffs, institutional portfolio pricing), where IPF* also prices the
institutional investor portfolio held on the OSE, the pricing question
studied by Koijen and Yogo (2019). The locators above point to the
exact tables.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(5), October 2025.
Copyright 2025 the American Finance Association. This distillation was
extracted by an LLM on 2026-05-31 and is **not human-verified or independently
reproduced**. The paper is paywalled; only extracts are reproduced here under
fair use for scholarly commentary.

> Betermier, Sebastien, Laurent E. Calvet, Samuli Knupfer, and Jens Soerlie
> Kvaerner. "Investor Factors." *The Journal of Finance* 80, no. 5 (October
> 2025): 2789-2830. DOI: 10.1111/jofi.13474. Copyright 2025 the American
> Finance Association. All rights reserved. This page contains an
> extract-only distillation by the Institute for Automated Research; the
> verbatim article is available at the publisher site.


==============================================================================
# How Much Does Racial Bias Affect Mortgage Lending: Bhutta, Hizmo & Ringo (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/bhutta-much-racial-bias-affect-2025/
# Distilled: Using confidential HMDA data for 2018-2019, this paper finds that standard underwriting factors explain most racial denial disparities, leaving a residual 1 to 2 percentage point excess denial gap that is itself at least partially explained by unobserved risk factors rather than discrimination. J. Finance 2025, U.S. Government work (public domain). Seven core results with source locators, datasets used, the model, and the empirical specifications.
# Tags: paper-summary, household-finance, mortgage-lending, discrimination, fair-lending
==============================================================================

**What this is.** This is a machine-distilled skeleton of the paper. Read the [original (DOI 10.1111/jofi.13444)](https://doi.org/10.1111/jofi.13444) to replicate or extend.

## TL;DR

Using confidential expanded HMDA data for 2018-2019 (nearly 9 million applications), the paper finds that observable applicant risk factors (credit score, LTV, DTI, AUS recommendation) explain most of the racial and ethnic gaps in mortgage denial rates. The residual "excess denial" gap is 2 percentage points for Black applicants and roughly 1 pp for Hispanic and Asian applicants, substantially smaller than the 8 pp gaps found by Munnell et al. (1996) or the 7-10 pp estimated by Bartlett et al. (2022) without controlling for credit score and other underwriting factors. Giacoletti, Heimer and Yu (2025) estimate a similar 7 pp raw gap and argue at least half reflects discrimination; this paper's evidence points more to unobserved risk. Cross-sectional evidence on lender strictness shows that stricter lenders have larger excess minority denial rates, consistent with tighter overlays on unobserved risk factors rather than discriminatory intent. Indirect tests (fintech lenders, market competition, regional racial animus) do not yield clear evidence of discrimination. The paper also revisits findings from Bhutta and Hizmo (2020) on minority mortgage pricing, and extends them by examining denial disparities using the newly expanded HMDA data. A separate analysis using NSMO survey data finds that minority borrowers report substantially worse service quality, suggesting a dimension of disparate treatment in service delivery that is not captured by denial statistics.

## Core results

| # | Result | Locator | Magnitude as reported |
|---|---|---|---|
| R1 | Raw Black-White denial gap before any controls | Table I, p.1474 | Black denial rate 18%, White 8%; raw gap 10 pp |
| R2 | Excess denial after full FICO-LTV-DTI-AUS-lender controls | Table II col.(3), p.1475 | Black 2.0 pp\*\*, Hispanic 0.9 pp\*\*, Asian 1.4 pp\*\* (s.e. 0.001) |
| R3 | Lender strictness correlation with excess minority denials | Figure 2 right panels, p.1480 | r = 0.63 (Black), 0.50 (Hispanic), 0.65 (Asian) vs. lender strictness for Whites |
| R4 | AUS excess denial for Black applicants (unobserved risk signal) | Table II col.(5), p.1475 | 1.5 pp\*\* (s.e. 0.001), suggesting racial gaps in AUS-observed-but-HMDA-unobserved risk |
| R5 | Racial animus correlation test for discrimination | Table IV col.(3) vs col.(6), p.1488 | Racially charged search rate interaction: 0.002\*\* for lender and AUS excess denials alike, suggesting unobserved risk rather than discrimination |
| R6 | Minority borrower service quality: processing and closing delays | Table V cols.(1)-(2), p.1493 | Black: 4.5 pp more likely to report processing delays\*\*\*, 9.6 pp more likely to have postponed closing\*\*\* |
| R7 | Minority borrower satisfaction with lender | Table V col.(6), p.1493 | Black 7.1 pp less likely to be very satisfied with lender\*\*\*; Asian 11.3 pp less\*\*\* |

**Overall (paper's conclusion).** The paper concludes that disparate treatment plays a much smaller role in generating mortgage denial disparities than the 1990s benchmark of Munnell et al. (1996) suggests, implying significant progress in fair lending over 30 years. The 1-2 pp excess denials overstate actual discrimination because unobserved risk factors that vary by race and ethnicity (and that stricter lenders screen for) explain at least part of the residual gap. Service quality disparities, however, are a documented and underexplored dimension of differential treatment.

## Theory / model

The paper has no formal economic model. It defines disparate treatment structurally as a difference in expected credit decisions across race/ethnicity for otherwise identical applicants.

Let the binary AUS recommendation be:

$$D_{AUS} = g(X, u), \tag{1}$$

where $$g(\cdot)$$ is a deterministic function of risk characteristics $$X$$ observable in HMDA and other risk characteristics $$u$$ unobserved in HMDA (see Section I, p.1472 for the full DU factor list).

Lender $$i$$'s binary denial decision is:

$$D^{i}_{Lender} = h_i(X^*, u, w, r) + e, \tag{2}$$

where $$h_i(\cdot)$$ may differ from $$g(\cdot)$$; $$X^*$$ is a potentially updated value of $$X$$ after verification; $$w$$ is lender-specific overlays beyond AUS; $$r$$ is race/ethnicity; and $$e$$ is idiosyncratic human error. Lender $$i$$ engages in disparate treatment against Black relative to White applicants if (p.1473):

$$\int h_i(X^*, u, w, \text{Black})\, dF_B(X^*, u, w) > \int h_i(X^*, u, w, \text{White})\, dF_B(X^*, u, w), \tag{3}$$

where $$F_B(\cdot)$$ is the joint CDF of underwriting factors for the Black applicant population. The identification challenge is separating $$r$$ (illegal discrimination) from $$u$$ and $$w$$ (unobserved but potentially race-correlated risk factors).

## Method

The key methodological innovation is the **lender strictness measure**, constructed as the lender fixed effect from a denial regression run exclusively on White applicants (equation (2) controls, p.1478). This measure isolates lender-specific overlay policies from any differential treatment of minorities by construction. The paper then correlates lender strictness with lender-specific excess minority denial rates (estimated with lender-varying race/ethnicity coefficients) as an indirect test of whether unobserved risk drives excess denials. The approach is analogous to the judge-specific propensity-to-release design of Arnold, Dobbie and Yang (2018) and Arnold, Dobbie and Hull (2022).

Fintech identification follows Fuster et al. (2019): a lender is coded as fintech if it appears on that paper's fintech list. Market concentration is proxied by the top-4 lenders' county market share. Racial animus is measured by the racially charged search rate from Stephens-Davidowitz (2014), standardized to mean zero and unit variance. These three cross-sectional dimensions are interacted with race/ethnicity in equation (2) to test whether excess denials are systematically higher in settings where discrimination would be easier or more prevalent (Table IV).

## Empirical specifications

**Main denial regression (Table II).** The estimating equation is a linear probability model regressing an indicator of lender denial on race/ethnicity dummies and controls (p.1475-1476):

$$D^{i}_{Lender,j} = \alpha_r \cdot \mathbf{1}[\text{race}_j = r] + \beta' X_j + \delta_l + \varepsilon_j, \tag{4}$$

where $$j$$ indexes applications; $$r$$ indexes race/ethnicity relative to non-Hispanic White; $$X_j$$ includes the FICO-LTV-DTI grid (interactions of credit score bins, LTV bins, and DTI bins; see Table II notes for exact bin definitions), AUS denial recommendation (interacted with loan purpose and program), county-by-month fixed effects, loan amount bins, co-applicant indicator, and income bins (all covariates interacted with program and loan purpose); $$\delta_l$$ is a lender fixed effect. Standard errors are clustered at the lender and county levels. Columns (1) to (3) vary the control set progressively; column (3) is the preferred full specification.

**AUS denial regression (Table II, cols. 4-5).** The dependent variable switches to the AUS denial indicator $$D_{AUS,j}$$, with the same right-hand side. Because AUS is color-blind by design, residual racial gaps in AUS recommendations reflect unobserved risk factors $$u$$ correlated with race (p.1477-1478).

**Lender-specific excess denials and strictness correlation (Figure 2, Table III).** For the 100 largest lenders, the race/ethnicity coefficients are allowed to vary by lender (i.e., the full col.(3) specification with lender-varying race/ethnicity slopes). These lender-specific excess denial estimates are plotted against lender strictness (lender FE from White-only denial regression). The correlation coefficient is reported (p.1481).

**Loan performance validation (Figure 4).** For 48 Ginnie Mae issuers matched to HMDA, the paper regresses 60-day delinquency within one year of origination on lender strictness. Residual riskiness is the lender FE from a delinquency regression controlling for flexible functions of DTI, LTV, credit score, and month dummies. Both raw and residual riskiness are negatively correlated with strictness, validating that strictness captures real overlay policies (p.1484-1486).

**Indirect tests for discrimination (Table IV).** The baseline specification (col.(3) of Table II) is augmented with interactions between race/ethnicity and (i) fintech indicator, (ii) top-4 lender county market share, and (iii) racially charged Google search rate, estimated separately for lender denials and AUS denials (p.1486-1489).

**Service quality regressions (Table V).** OLS regressions using NSMO individual-level data with controls including loan type, loan purpose, loan amount, credit score, income, LTV, self-employment status, co-applicant status, loan term, 11 LTV categories, 6 credit score categories, and 8 loan amount categories; all fully interacted with program and loan purpose. Survey fixed effects and county fixed effects included. Heteroskedasticity-robust standard errors. Outcomes are binary (processing delays, closing date postponement, satisfaction dummies), p.1492-1493.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| HMDA (confidential, expanded 2018-2019) | Main denial analysis: ~9 million applications with credit score, LTV, DTI, AUS recommendation | [/wiki/datasets/hmda/](/wiki/datasets/hmda/) |
| National Survey of Mortgage Originations (NSMO) | Survey component of NMDB; borrower satisfaction and service quality analysis (N=35,162) | [NSMO](/wiki/datasets/nsmo/) |
| National Mortgage Database (NMDB) | Provides inquiry data for pre-application discouragement analysis; links HMDA to credit bureau records | no page yet |
| Ginnie Mae securitization pool data | Loan performance (delinquency) validation for 48 matched issuers, 2018-2019 originations | no page yet |

**Sample:** First-lien, 30-year fixed-rate mortgages on owner-occupied single-family properties; 2018-2019; applications through one of three main AUS (DU, LPA, TOTAL); excludes jumbo loans and withdrawn/incomplete applications. Final AUS-processed sample: ~8.9 million applications. NSMO subsample: 35,162 respondents for service quality regressions.

## When to read the full paper

Read the paper if you are studying racial disparities in mortgage credit, the role of automated underwriting in fair lending, or methods for detecting discrimination in lending outcomes. Table II is the key excess denial table; Figure 2 and Table III document the lender strictness channel; Table IV presents the indirect discrimination tests; Table V covers service quality. The Internet Appendix (referenced throughout) contains additional robustness checks, the pre-application discouragement analysis, and the interest rate gap replication.

## Attribution and rights

This article is a U.S. Government work and is in the public domain in the USA (per PDF p.1463 copyright notice: "Published 2025. This article is a U.S. Government work and is in the public domain in the USA."). The Wiley/JF Crossref record carries the Wiley terms-and-conditions URL rather than a CC licence. Extract-only: not reproduced or human-verified here. LLM-distilled; not reproduced.

> Bhutta, Neil, Aurel Hizmo, and Daniel Ringo (2025). "How Much Does Racial Bias Affect Mortgage Lending? Evidence from Human and Algorithmic Credit Decisions." *The Journal of Finance* 80(3): 1463-1496. DOI: 10.1111/jofi.13444.


==============================================================================
# Lucky Survivor: Van Binsbergen, Hua, Peeters & Wachter (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/binsbergen-united-states-lucky-survivor-2025/
# Distilled: Using a cross-section of 55 countries from 1920 to 2020, the paper quantifies survivorship bias in U.S. equity market performance via a hierarchical Bayesian model that cross-learns crash risk across countries, finding that survivorship bias explains about one-third of the 6% historical U.S. equity premium, with luck and learning jointly accounting for roughly 2 percentage points. J. Finance 2025, CC BY-NC-ND 4.0. Five core results with source locators, datasets used, the model (hierarchical Beta-Bernoulli crash-belief model), and the method (Hamiltonian Monte Carlo MCMC).
# Tags: paper-summary, asset-pricing, equity-premium, survivorship-bias, disaster-risk
==============================================================================

**What this is.** The paper's core results, the hierarchical Bayesian model it proposes, and the HMC estimation method with defining equations: enough to know what it found and how, without reading all 34 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13452).

## TL;DR

The paper asks whether U.S. stock market outperformance is partly the result of survivorship bias: the U.S. happened to survive and thrive over the past century, and an early-1920 investor could not have known this would happen. Goetzmann and Jorion (1999) established that U.S. equity outperformance appears exceptional compared with other countries; this paper quantifies the bias by amount. Using a hierarchical Bayesian model fitted to annual total returns for 55 countries from 1920 to March 2020, the authors infer a subjective crash-belief series for each market by cross-learning from the full cross-section, applying the hierarchical-learning approach of Jones and Shanken (2005) to the cross-country setting. The U.S. crash probability, corrected for survivorship bias, is 6.54% at the end of the sample, versus its 5% in-sample frequency, and lies 2.6 percentage points below the global mean of 9.16%, a gap that widens persistently over the second half of the 20th century. The secular downward drift in perceived U.S. crash risk is attributed to positive return surprises (luck) and learning-induced valuation increases. Mapping crash risk to an Epstein-Zin equilibrium model using the rare-disaster framework of Barro (2006) as extended by Wachter (2013) to time-varying disaster risk, luck and learning jointly explain roughly 2 percentage points of the 6% historical equity premium as identified by Mehra and Prescott (1985), implying the premium has been overstated by about one-third. The declining equity premium is consistent with the trend documented by Lettau, Ludvigson, and Wachter (2008).

## Core results

Magnitudes and significance are as reported. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | The hierarchical model raises the U.S. subjective crash risk above its in-sample frequency; the survivorship correction increases U.S. crash risk by 30.8% | Figure 4, p. 2374; §III.A | U.S. in-sample crash frequency = 5%; posterior mean at end of sample = 6.54%; adjustment = 1.54 pp (30.8%) |
| R2 | U.S. subjective crash risk diverges persistently and increasingly below the global average crash risk over the second half of the 20th century | Figure 5, p. 2376; §III.A | U.S. = 6.54%, global mean = 9.16% as of March 2020; gap = 2.62 pp; gap widens after ~1960 |
| R3 | The realized U.S. equity premium decomposes into ex ante expected return (~7.3%), luck (~1.0-1.3%), and learning (~0.9-1.1%), robust across all three prior specifications | Table IV, p. 2383; §IV.B | Average realized return 9.11-9.70%; expected + residual 7.26-7.42%; luck 0.96-1.27%; learning 0.81-1.08% |
| R4 | Survivorship bias (luck and learning jointly) explains about one-third of the measured U.S. equity premium, making the premium less of a puzzle | §V, p. 2385-2386 | ~2 pp of the measured ~6% equity premium attributable to luck + learning; true ex ante expected return ~4 pp |
| R5 | The model captures actual fluctuations in U.S. stock returns at business-cycle frequency; average cross-country correlation between model-implied and realized returns = 50% | Figure 7, p. 2384; §IV.B | 10-year rolling average of model-implied returns tracks realized U.S. returns well in levels; cross-country correlation = 50% |

**Overall (paper's conclusion).** Measuring the U.S. equity premium based on realized stock market returns leads to an overestimation of the ex ante expected return by approximately one-third. After correcting for survivorship bias via cross-learning from 55 international equity markets, the U.S. equity premium puzzle becomes substantially smaller. The secular decline in perceived U.S. crash risk implies a declining equity premium going forward, a slow-moving upward trend in valuation ratios, and reduced global diversification benefits as markets become more integrated.

## Theory / model

The paper sets up a hierarchical Bayesian framework in which a representative investor infers country-specific crash risk by cross-learning from the full cross-section of countries. Crash risk for country $$i$$ is modeled as a latent constant $$p_i$$, drawn from a common Beta distribution (Section I.A, p. 2361):

$$
p_i | \alpha, \beta \sim \text{Beta}(\alpha, \beta) \quad \forall i. \tag{Prior}
$$

Each period the investor observes crash realizations $$D_{i,t} \in \{0,1\}$$ (1 if annual return below -30%), which are i.i.d. across time and countries conditional on $$p_i$$:

$$
D_{i,t} | p_i \overset{\text{i.i.d.}}{\sim} \text{Bernoulli}(p_i). \tag{Likelihood}
$$

Let $$Y_{i,\tau} = [D_{i,1}, D_{i,2}, \ldots, D_{i,\tau}]$$ be the crash history of country $$i$$ up to time $$\tau$$. The joint posterior of all parameters follows Bayes rule (eq. 1, p. 2361):

$$
f(\{p_i\}_{i=1}^n, \alpha, \beta | \{D_{j,t}\}_{j=1,t=1}^{n,\tau}) \propto \prod_{i=1}^n p_i^{Y_{i,\tau} \mathbf{1}^\top + \alpha - 1}(1-p_i)^{\tau - Y_{i,\tau} \mathbf{1}^\top + \beta - 1} f(\alpha, \beta). \tag{1}
$$

Conditional on $$\alpha$$ and $$\beta$$, the country-specific crash probability has a conjugate Beta posterior (eq. 2, p. 2362):

$$
p_i | \alpha, \beta, \{D_{j,t}\}_{j=1,t=1}^{n,\tau} \sim \text{Beta}(Y_{i,\tau}\mathbf{1}^\top + \alpha, \; \tau - Y_{i,\tau}\mathbf{1}^\top + \beta) \quad \forall i. \tag{2}
$$

The posterior mean of $$p_i$$ (eq. 5, p. 2362) is a weighted average of the country's own crash frequency and the global mean $$\alpha/(\alpha+\beta)$$:

$$
\mathbb{E}(p_i | \alpha, \beta, \{D_{j,t}\}_{j=1,t=1}^{n,\tau}) = \frac{Y_{i,\tau}\mathbf{1}^\top}{\tau} \cdot \frac{\tau}{\tau + \alpha + \beta} + \frac{\alpha}{\alpha+\beta} \left(1 - \frac{\tau}{\tau+\alpha+\beta}\right). \tag{5}
$$

The effective sample size $$\alpha+\beta$$ controls the strength of the global shrinkage; for large $$\tau$$ the estimate converges to the country-specific frequency. The hyperpriors on $$\phi = \alpha/(\alpha+\beta)$$ (global mean crash risk) and $$\lambda = \alpha+\beta$$ (effective sample size) span three specifications: uninformative ($$\phi \sim \text{Uniform}[0,1]$$, $$\lambda \sim \text{Pareto}(1, 0.5)$$), semi-informative ($$\phi \sim \text{Uniform}[0, 0.35]$$, $$\lambda \sim \text{Pareto}(1, 2.5)$$), and informative ($$\phi \sim \text{Beta}(2,98)$$, $$\lambda \sim \text{Pareto}(1, 2.5)$$; mean crash risk 2%).

**Asset pricing.** In Section IV (pp. 2377-2384), the paper maps the declining crash belief into an equilibrium equity premium. Log consumption growth follows a random walk with drift $$\mu$$, standard deviation $$\sigma$$, and a Poisson jump $$v_{t+1}$$ that drops by $$b$$ upon a crash (with subjective probability $$p_t$$):

$$
\Delta c_{t+1} = \mu + \sigma \epsilon_{t+1} + v_{t+1}, \qquad v_{t+1} = \begin{cases} \log(1-b) & \text{w.p. } p_t \\ 0 & \text{w.p. } 1-p_t \end{cases}.
$$

The investor has Epstein-Zin utility with risk aversion $$\gamma$$, EIS $$\psi$$, and discount rate $$\delta$$. With CRRA ($$\gamma = 1/\psi$$), the equity premium increases in subjective disaster risk (eq. 7, p. 2378):

$$
\log \frac{\mathbb{E}_t(1 + R_{t+1})}{1+R_f} = \gamma\sigma^2 + b[(1-b)^{-\gamma} - 1]p_t. \tag{7}
$$

The realized excess return decomposes into ex ante expected return, luck, learning, and a residual (eq. 8, p. 2381):

$$
R_{t+1}^D - \mathbb{E}_t[R_{t+1}^D] = \underbrace{\mathbb{E}_t\left[\frac{D_{t+1}}{D_t}\right]\left(\frac{Z_{t+1}^D+1}{Z_t^D} - \frac{Z_t^D+1}{Z_t^D}\right)}_{\text{Learning}} + \underbrace{\frac{Z_t^D+1}{Z_t^D}\left(\frac{D_{t+1}}{D_t} - \mathbb{E}_t\left[\frac{D_{t+1}}{D_t}\right]\right)}_{\text{Luck}} + \underbrace{\left(\frac{Z_{t+1}^D+1}{Z_t^D} - \frac{Z_t^D+1}{Z_t^D}\right)\left(\frac{D_{t+1}}{D_t} - \mathbb{E}_t\left[\frac{D_{t+1}}{D_t}\right]\right)}_{\text{Residual}}. \tag{8}
$$

Parameters are calibrated at $$\gamma=3$$, $$\beta=0.98$$, $$\mu=0.0252$$, $$\sigma=0.02$$, $$\lambda=2.6$$, $$\psi=1$$, $$b=0.3$$.

## Method

The posterior is computed using Hamiltonian Monte Carlo (HMC), specifically the No-U-Turn Sampler of Hoffman and Gelman (2011) (Section II.B, p. 2373). The algorithm is initialized with four Markov chains at randomly selected starting values; each chain draws 10,000 samples with the first 2,000 as burn-in. Chains are pooled for the posterior distribution.

The model is estimated with an expanding window: starting from an initial period, the investor updates her beliefs each year as new data arrive across the cross-section. This replicates real-time belief formation and avoids look-ahead bias. The crash indicator for each country-year is defined as an annual return below -30% (Section II, p. 2373); robustness checks use -20% and -35% thresholds (Internet Appendix). Standard MCMC diagnostics compare within- and between-chain variation via the Gelman and Rubin (1992) potential scale reduction factor.

The hierarchical Beta-Bernoulli structure (`hierarchical-bayesian-mcmc`) provides a conjugate update: integrating out $$\alpha$$ and $$\beta$$ via equation (4) (p. 2362) yields the marginal posterior for country-specific crash risk, with cross-learning strength governed by $$\alpha + \beta$$. For long-history markets (such as the U.S. with 100 years), cross-learning is attenuated relative to short-history emerging markets (e.g., Estonia with 23 years), where the global prior exerts stronger influence.

## Empirical specifications

The key empirical exercise is computing the posterior distribution of crash beliefs for each of the 55 countries at each point in time, then mapping those to equity premium components. No standard panel regression with fixed effects is run; the model is entirely Bayesian.

**Cross-sectional comparison (R1, R2).** At the end of the sample (March 2020), the posterior mean $$\mathbb{E}[p_i | \{D_{j,t}\}]$$ is compared to the in-sample crash frequency $$p^c/N$$ for each country (Figure 4, p. 2374). For the United States: in-sample frequency = 5% (6 crashes out of 100 years), posterior mean = 6.54%, correction = 30.8%. For Turkey: posterior mean shrinks by roughly half relative to the in-sample frequency of 15.5%.

**Time-series evolution (R2).** Figure 5 (p. 2376) plots the U.S. posterior mean crash risk against the global average $$\alpha/(\alpha+\beta)$$ from 1921 to 2020. The U.S. crash belief starts near the global mean, departs upward during the Great Depression (1931-1937), then drifts steadily downward relative to the global average through the remainder of the century.

**Equity premium decomposition (R3, R4).** Using equation (8) with the calibrated Epstein-Zin model parameters, the paper computes the time-average of each component for the United States over 1931-2020. This is repeated for all three prior specifications; Table IV (p. 2383) reports the results. The decomposition is not identified econometrically but derived from the model: luck is the term involving realized minus expected dividend growth, learning is the valuation-ratio revision term.

**Model fit (R5).** Model-implied realized returns are compared to actual returns in the data (Figure 7, p. 2384). The correlation between model-implied and data-realized returns, averaged across all 55 countries, is 50% over the 100-year period.

**Robustness.** Results are repeated for all three hyperprior specifications (uninformative, semi-informative, informative); all converge to similar values after 10 years of data accumulate (Figure 2, p. 2366). The crash threshold is varied from -20% to -35% in the Internet Appendix with similar results.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Global Financial Data (GFD) | Annual stock total return indices (price level + monthly reinvested dividends) for 55 countries, 1920-2020; the main source for all return data | [no page yet](/wiki/datasets/) |
| CRSP (U.S.) | U.S. total return series (replaces GFD U.S. series with CRSP proxy for S&P 500 universe) | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| Robert Shiller historical data | Used to calibrate the informative hyperprior: one crash in U.S. data 1870-1920 | no page yet |

Sample: 55 countries, unbalanced annual panel from 1920 to March 2020. Crashes defined as annual return below -30%. Several markets experience total loss episodes (nationalization, wartime closure) treated as -100% returns following Bialkowski and Ronn (2016).

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13452) if you are: building a model of the equity premium that corrects for international survivorship bias; studying how Bayesian cross-country learning disciplines rare-event probability estimates; extending the hierarchical crash-risk model to time-varying crash probabilities or multi-period disasters; or evaluating the magnitude of the equity premium puzzle in a broad international sample. The Internet Appendix contains the correlated-crash extension (Section I), additional prior robustness (Section II), and full proofs.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(4). This distillation was extracted by an LLM on 2026-06-05 and is **not human-verified or independently reproduced**. Licensed CC BY-NC-ND 4.0; non-commercial, no-derivatives; extract-only redistribution.

> van Binsbergen, Jules, Sophia Hua, Jonas Peeters, and Jessica Wachter.
> "Is the United States a Lucky Survivor? A Hierarchical Bayesian Approach."
> *The Journal of Finance* 80, no. 4 (August 2025): 2355-2388.
> DOI: 10.1111/jofi.13452. © 2025 The Author(s).
> Licensed under [CC BY-NC-ND 4.0](https://creativecommons.org/licenses/by-nc-nd/4.0/).
> This page is a distillation by the Institute for Automated Research: LLM-extracted summary; extract-only; no PDF mirrored.


==============================================================================
# Dynamic Banking and the Value of Deposits: Bolton, Li, Wang & Yang (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/bolton-dynamic-banking-value-deposits-2025/
# Distilled: A continuous-time structural model shows that banks cannot fully control deposit flows under leverage regulation, so deposit inflows can hurt shareholder value when equity capital is low, the deposit marginal q turns negative, and lending falls. J. Finance 2025, paywalled. Six core results with source locators, the model (HJB with deposit-dynamics state variable), and the method (ODE solution with boundary conditions).
# Tags: paper-summary, banking, bank-regulation, deposits, leverage, monetary-policy
==============================================================================

**What this is.** The paper's core model, its key propositions on the value and management of bank deposits under leverage regulation, and the numerical findings: enough to know what it derives and why, without reading all 43 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13454).

## TL;DR

Bolton, Li, Wang, and Yang (2025) propose a continuous-time structural model of a bank that maximizes risk-neutral shareholder value while facing two uncontrollable random processes: asset return shocks and deposit flow shocks. The central state variable is the ratio of equity capital to deposits, $$k_t = K_t / X_t$$. Because depositors freely move money in and out, the bank cannot perfectly control $$X_t$$, distinguishing it from nondepository intermediaries and nonfinancial firms. Under equity issuance costs and leverage regulations (the supplementary leverage ratio, SLR), deposit inflows can lower $$k_t$$ and push it toward a costly equity issuance boundary. As a result, the marginal value of deposits (the deposit marginal $$q$$) turns sharply negative when $$k_t$$ is low, the bank reduces its deposit rate toward the zero lower bound (ZLB), and lending falls rather than rises. The model relates to Merton (1969) in casting the bank's problem as a portfolio choice, extends the dynamic capital structure approach of Leland (1994a) and Brunnermeier and Sannikov (2014) to include stochastic deposit liabilities, and departs from Diamond and Dybvig (1983) by focusing on deposit inflow risk (not runs) under equity issuance costs. Unlike Drechsler, Savov, and Schnabl (2021), who treat deposits as long-duration liabilities without interest-rate risk, the paper incorporates deposit flow risk as the central friction. The model explains why (i) banks did not expand lending during the COVID-19 deposit surge, (ii) the SLR relaxation stimulated lending only in the short run, (iii) a low risk-free rate compresses banks' deposit management flexibility, and (iv) the deposit rate and loan growth co-move positively with bank capitalization.

## Core results

Magnitudes and qualitative characterizations are as reported from the numerical solution. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Deposit marginal q is positive for well-capitalized banks but turns sharply negative near the equity issuance boundary | Figure 4 Panel A, p. 2087; Figure 5 Panel A, p. 2088 | Deposit marginal q ~0.11 for 80% of stationary distribution of k; drops to ~-0.18 as k approaches the equity issuance lower boundary (~0.052) |
| R2 | Loan-to-capital ratio is procyclical in bank capitalization; capital requirement binds about 7% of the time | Figure 1 Panel B, p. 2084; Figure 3 Panel B, p. 2086 | A/K rises from ~0 at the lower boundary to ~14 (capital requirement ceiling) as k increases; capital requirement binds 7% of the time; SLR binds far more frequently |
| R3 | Marginal value of equity capital is sharply elevated near the equity issuance boundary, casting a long shadow | Figure 1 Panel A, p. 2084; Figure 3 Panel A, p. 2086 | v'(k) reaches ~7 at k = underbar (0.052); stays between 1.022 and 1.029 for 25% of stationary time; exceeds 1.08 for 5.5% of stationary time |
| R4 | Relaxing the SLR raises lending immediately but reduces long-run risk-taking per unit of equity, contrary to conventional wisdom | Figure 7 Panel A and B, p. 2091 | Under 4% SLR, A/K is higher in Panel A (given k) but lower in Panel B (against stationary c.d.f.) than under 5% SLR; tightening SLR generates reach-for-yield over the long run |
| R5 | Relaxing the SLR raises deposit marginal q and deposit rates, stimulating deposit-taking; but deposit marginal q becomes more negative near the new lower equity issuance boundary | Figure 8 Panel A and B, pp. 2092-2093 | Deposit rate jumps up under 4% SLR; region where ZLB binds shrinks; deposit marginal q is negative before the regulatory change, turns even more negative near the new lower boundary |
| R6 | Lower risk-free rate reduces bank lending (counterintuitive): bank reduces A/K because it has less room to manage deposit risk via the deposit spread | Figure 9 Panel A and B, p. 2094 | At r=1% vs r=2%, the deposit rate is less than 0.8 pp higher under r=2%; A/K is lower at all quantiles of k under r=1% than r=2% |

**Overall (paper's conclusion).** Deposit-taking is a double-edged sword: it lowers funding costs in normal times, but deposit inflows under leverage regulation drive up leverage and can trigger costly equity issuance. When the bank is close to its equity issuance boundary, it reduces deposit rates to the ZLB, cuts lending, and holds safe assets. Banks in low-interest-rate environments have less flexibility to control deposit flows, amplifying this mechanism and explaining patterns observed after the Global Financial Crisis and the COVID-19 pandemic.

## Theory / model

The model features a single bank maximizing risk-neutral shareholder value. Two state variables govern the bank's balance sheet: the deposit stock $$X_t$$ and equity capital $$K_t$$. The bank controls five variables: the risky loan book $$A_t$$, bond issuance $$B_t$$, deposit rate $$i_t$$, dividends $$dU_t$$, and equity issuance $$dF_t$$. The resource constraint is $$A_t = K_t + X_t + B_t$$ (p. 2072, eq. 3).

**Deposit dynamics** (p. 2070, eq. 1): The deposit stock evolves as a diffusion process partially controlled by the bank via the deposit rate:

$$
dX_t = -X_t(\delta_X \, dt - \sigma_X \, d\mathcal{W}_t^X) + X_t \, n(i_t) \, dt \tag{1}
$$

where $$\mathcal{W}_t^X$$ is a standard Brownian motion, $$\delta_X$$ is the drift of payment flows out of the bank, $$\sigma_X$$ is the deposit flow volatility, and $$n(i_t) = \omega_0 + \omega_1(i_t - r)$$ is the deposit demand function (eq. 21, p. 2079): raising the deposit rate above the risk-free rate $$r$$ attracts deposits, lowering it repels them. The deposit rate is bounded below by zero: $$i_t \geq 0$$.

**Bank equity dynamics** (p. 2072, eq. 2): Bank equity evolves as

$$
dK_t = A_t \left[(r + \alpha_A) \, dt + \sigma_A \, d\mathcal{W}_t^A\right] - B_t r \, dt - X_t i_t \, dt - C(n(i_t), X_t) \, dt - dU_t + dF_t \tag{2}
$$

where $$\alpha_A$$ is the bank's excess return on lending, $$\sigma_A$$ is asset return volatility, $$\phi \, dt$$ is the instantaneous covariance between deposit and asset shocks ($$d\mathcal{W}^X$$ and $$d\mathcal{W}^A$$), and $$C(n(i_t), X_t) = c(n(i_t)) X_t$$ is the cost of maintaining the deposit franchise.

**Homogeneity and the HJB equation.** The functional forms imply the shareholder value function is homogeneous of degree one: $$V(X, K) = v(k) X$$ where $$k \equiv K/X$$ (eq. 9, p. 2075). Within the dividend/issuance boundaries $$[\underline{k}, \overline{k}]$$ the HJB equation for the scaled value function $$v(k)$$ is (eq. 10, p. 2075):

$$
\rho v(k) = \max_{\pi^A, i} \left\{ [v(k) - v'(k)k]\left[-\delta_X + n(i)\right] + \tfrac{1}{2} v''(k) k^2 \sigma_X^2 + v'(k)(1+k)\left(r + \pi^A \alpha_A\right) + \tfrac{1}{2} v''(k)(1+k)^2 (\pi^A \sigma_A)^2 - v'(k)[i + c(n(i))] - v''(k) k(1+k) \pi^A \sigma_A \sigma_X \phi \right\} \tag{10}
$$

where $$\pi^A = A/(X+K)$$ is the portfolio weight on risky assets, $$\rho > r$$ is the shareholders' discount rate. The deposit marginal $$q$$ equals $$V_X(X,K) = v(k) - v'(k)k$$ and the equity marginal $$q$$ equals $$V_K(X,K) = v'(k)$$.

**Regulatory constraints.** The capital requirement (eq. 6, p. 2073) restricts the risky asset-to-equity ratio: $$A_t / K_t \leq \xi_K$$ (baseline $$\xi_K = 14.3$$). The supplementary leverage ratio (SLR, eq. 7-8, pp. 2073-2076) imposes a lower bound on $$k$$:

$$
k \geq \underline{k} \equiv \frac{1}{1 - \xi_L^{-1}} - 1 \tag{13}
$$

with $$\xi_L = 20$$ in the baseline (implying $$\underline{k} \approx 0.05$$). This is the equity issuance boundary; hitting it requires the bank to raise costly external equity.

**Equity issuance costs.** Equity issuance costs are $$dH_t = \psi_1 dF_t + \psi_0 X_t \, dt$$, where $$\psi_1 = 5\%$$ is the proportional cost per dollar issued and $$\psi_0 = 0.14\%$$ is a fixed flow cost proportional to the deposit stock size (governing how frequently equity issuance is triggered). The bank maximizes (eq. 5, p. 2073):

$$
V_0 = \max_{\{A,B,i,U,F\}} \mathbb{E}\left[\int_{t=0}^{\tau} e^{-\rho t}(dU_t - dF_t - dH_t)\right] \tag{5}
$$

where $$\tau$$ is the stochastic closing time (when regulatory constraints are violated).

## Method

The model is solved numerically as an ODE boundary value problem. The homogeneity reduction to the one-dimensional $$v(k)$$ transforms the two-dimensional HJB into the ordinary differential equation (10), which is solved over the interval $$[\underline{k}, \overline{k}]$$ using shooting/iteration.

**Optimal risky asset allocation.** The first-order condition from (10) for $$\pi^A$$ yields the optimal loan-to-capital ratio (eq. 18, p. 2077):

$$
\frac{A}{K} = \frac{\alpha_A}{\gamma(k) \sigma_A^2} + \frac{\sigma_X}{\sigma_A} \phi \tag{18}
$$

where $$\gamma(k) \equiv -v''(k)k / v'(k)$$ is the bank's endogenous relative risk-aversion (eq. 19, p. 2078), derived from the curvature of the value function. Even though shareholders are risk-neutral, $$\gamma(k) > 0$$ because equity issuance costs make the bank endogenously risk-averse. The hedging term $$(\sigma_X / \sigma_A)\phi$$ reflects the deposit risk as a natural hedge for the asset-side shock when $$\phi > 0$$.

**Optimal deposit rate.** The first-order condition for $$i$$ from (10) yields the q-theory formula for the optimal deposit rate (eq. 23, p. 2079):

$$
i = r + \frac{V_X(X,K)/V_K(X,K) - 1/\omega_1}{\omega_1 \theta} - \frac{\omega_0}{\omega_1} = r + \frac{(v(k) - v'(k)k)/v'(k) - 1/\omega_1}{\omega_1 \theta} - \frac{\omega_0}{\omega_1} \tag{23}
$$

The deposit rate rises in the ratio of deposit marginal q to equity marginal q. When $$V_X / V_K$$ is high (deposits are more valuable than equity), the bank sets a high $$i_t$$ to attract deposits. When $$V_X$$ falls near $$\underline{k}$$, the bank reduces $$i_t$$ toward zero.

**Boundary conditions.** At the equity issuance boundary $$\underline{k}$$: value-matching $$v(\underline{k} + m) = 1 + \psi_1$$ (eq. 14) and smooth-pasting $$v'(\underline{k}) = 1 + \psi_1$$. At the dividend boundary $$\overline{k}$$: $$v'(\overline{k}) = 1$$ (eq. 16) and supercontact condition $$v''(\overline{k}) = 0$$ (eq. 17, p. 2077). The paper builds on `hjb-optimal-stopping` techniques from Leland (1994a) and `value-function-iteration` for the numerical ODE solution.

## Empirical specifications

This is a pure-theory paper with a calibrated numerical solution; there are no econometric regressions or estimated equations. The authors calibrate the model to match empirical moments of U.S. banking data (Table I, p. 2082), including:

- Average return on assets of 1.04% (matching $$\alpha_A = 0.2\%$$ at baseline parameters, FRED data).
- Average deposit-to-assets ratio of 92% (matching $$\phi = 0.8$$, consistent with Drechsler, Savov, and Schnabl (2017)).
- Average equity issuance frequency of once every four years (matching $$\psi_0 = 0.14\%$$, consistent with Baron (2020)).
- Average bank deposit growth rate of 1.9% per quarter (matching $$\omega_0 = 0.06$$, consistent with Lin (2019)).
- Average return on equity of 11% (matching $$\theta = 0.5$$).

**Comparative statics and applications.** The paper analyzes two applications numerically:

1. **Leverage regulation (Section IV.A):** Compares model solutions under SLR of 5% vs. 4%, tracing the immediate (given $$k$$) and long-run (stationary distribution of $$k$$) effects on $$A/K$$, deposit marginal $$q$$, deposit rate, and equity issuance frequency (Figures 7 and 8, pp. 2091-2093).

2. **Low-interest-rate environment (Section IV.B):** Compares model solutions under $$r = 1\%$$ vs. $$r = 2\%$$ (adjusting $$\rho$$ by 1% to control for the wedge $$\rho - r$$), tracing effects on deposit rate and $$A/K$$ against the stationary c.d.f. of $$k$$ (Figure 9, p. 2094). Drechsler, Savov, and Schnabl (2017) document that a lower $$r$$ compresses the deposit spread $$r - i$$ and constrains banks; this paper shows the mechanism via imperfect deposit flow control.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Federal Reserve Economic Data (FRED) | Calibration targets: average Fed funds rate, return on assets of U.S. banks | [FRED](/wiki/datasets/fred/) |

The paper is theoretical with a calibrated numerical solution. All key moments (average ROA 1.04%, deposit-to-asset ratio 92%, equity issuance frequency, deposit growth 1.9%/quarter) are cited from published empirical papers (Baron (2020), Lin (2019), Drechsler, Savov, and Schnabl (2017)) and FRED aggregate statistics; no proprietary microdata are used.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13454) if you are: (i) building a model of bank balance-sheet management where deposit risk matters alongside equity issuance costs; (ii) analyzing the short-run vs. long-run effects of leverage regulation (SLR) on bank lending and risk-taking; (iii) studying why low interest rates reduce bank lending (the deposit-management-flexibility channel); (iv) studying the COVID-19 episode of massive deposit inflows and the SLR exemption. The Internet Appendix contains the extension with reserve requirements, the jump-risk model, and the negative deposit rate case.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(4), August 2025. Published by Wiley on behalf of the American Finance Association. This distillation was extracted by an LLM on 2026-06-05 and is **not human-verified or independently reproduced**. The publisher licence is Wiley VOR terms (paywalled); extract-only reproduction applies.

> Bolton, Patrick, Ye Li, Neng Wang, and Jinqiang Yang.
> "Dynamic Banking and the Value of Deposits."
> *The Journal of Finance* 80, no. 4 (August 2025): 2063–2105.
> DOI: [10.1111/jofi.13454](https://doi.org/10.1111/jofi.13454).
> © 2025 the American Finance Association.
> This page is an extract-only distillation by the Institute for Automated Research; no reproduction of the full text.


==============================================================================
# CEO Stress, Aging, and Death: Borgschulte, Guenzel, Liu & Malmendier (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/borgschulte-ceo-stress-aging-death-2025/
# Distilled: Managerial stress from industry distress shocks accelerates visible aging by roughly one year and raises CEO mortality hazard by ~15%, equivalent to 1.1 years of additional chronological age; antitakeover laws that reduce monitoring intensity imply a two-year longevity gain. J. Finance 2025, CC BY 4.0. Eight core results with source locators, datasets used, the empirical strategy (DiD apparent-aging + stratified Cox hazard), and the identifying variation.
# Tags: paper-summary, corporate-governance, health-economics, executive-compensation
==============================================================================

**What this is.** The paper's core results, the two empirical strategies (DiD
apparent-aging and stratified Cox mortality hazard), and the datasets. Enough to
know what was found and how, without reading the full 42 pages. To replicate or
extend, read the original at [doi.org/10.1111/jofi.13497](https://doi.org/10.1111/jofi.13497).

## TL;DR

Using two quasi-experimental sources of variation in CEO job demands, the paper
documents that managerial stress causes accelerated visible aging and higher
mortality. First, in a difference-in-differences design applied to 3,002 facial
images of Fortune 1000 CEOs during the Great Recession, industry distress
exposure makes CEOs look roughly one year older (eventually 1.1 to 1.2 years
after 2012). Second, in a stratified Cox hazard model on 1,900 CEOs from Forbes
Executive Compensation Surveys (1975 to 1991), industry distress raises the
mortality hazard by about 15%, equivalent to 1.1 fewer years of chronological
life. Third, the staggered passage of antitakeover (Business Combination) laws
across U.S. states in the mid-1980s, which reduced monitoring intensity, is
associated with a 16 to 21% lower mortality hazard per year of BC law
exposure, implying roughly a two-year longevity gain for the average
protected CEO. The effects are of similar magnitude across both stress proxies
and consistent with a causal interpretation: neither compensation nor CEO tenure
fully accounts for the health costs, suggesting the market does not price them in.

## Core results

Magnitudes and significance as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Industry distress raises CEO apparent age (DiD, post-2006) | Table III col. (1), p. 3420 | +0.806 years (SE 0.382, `\*\*`) |
| R2 | Distress-induced apparent aging grows over time, reaching 1.0-1.2 years at post-2012 horizon | Table III cols. (3)-(4), p. 3420; Figure 4, p. 3418 | 0.634 (insig.) in 2007-2011; 1.049`\*\*` to 1.183`\*\*\*` from 2012 onward |
| R3 | Industry distress raises CEO mortality hazard by ~15% | Table IV (cols. 1-6), p. 3426 | Average hazard coefficient 0.136; hazard ratio exp(0.136) = 1.145; equivalent to 1.1 years older |
| R4 | BC antitakeover law exposure (binary) reduces CEO mortality hazard | Table V cols. (1)-(4), p. 3432 | Coefficients -0.198 to -0.234 (`\*\*` to `\*\*\*`); average -0.217 |
| R5 | Each additional year of BC law exposure reduces mortality hazard ~3.8% | Table V cols. (5)-(8), p. 3432 | Coefficients -0.037 to -0.040 (`\*\*\*`); average -0.039 |
| R6 | BC law protection equivalent to being ~2 years younger; no compensating pay differential found | p. 3433; Internet Appendix Table IA.XXIV | Hazard ratio shift corresponds to mortality rate of a CEO 2 years younger; pay effect insignificant and positive |
| R7 | Kaplan-Meier survival curves: ~67% of distressed CEOs die within 30 years of appointment vs. ~32 years for nondistressed | Figure 5, p. 3425 | Visually left-shifted survival curve; 1-year mortality at median CEO age pushed from 1.337% to 1.532% |
| R8 | BC law Kaplan-Meier: 1980s cohorts with BC exposure right-shifted vs. same-era no-BC cohorts; 1970s and 1980s no-BC curves nearly identical | Figure 6, p. 3431 | 25% cumulative mortality reached ~25 years (no BC) vs. ~28-30 years (BC) after appointment |

**Overall (paper's conclusion).** Heightened job demands in the form of
industry-wide distress and stricter corporate monitoring impose significant
personal health costs on CEOs: faster visible aging and shorter lives. The
effects are of similar magnitude whether identified by economic distress shocks
or by variation in governance intensity from antitakeover laws. For context,
Sullivan and Von Wachter (2009) estimate that job displacement raises the
mortality hazard by 10 to 15% and reduces life expectancy by 1 to 1.5 years
in a general male population; the industry-distress estimate here is of
comparable magnitude, but operates through an opposite channel (more, not
less, work effort). The absence of a compensating pay differential suggests
the market does not fully account for these costs, pointing to an
underappreciated private cost of CEO service.

## Theory / model

The paper has no formal structural model. The economic framework builds on the
notion that work-related stress arises when job demands exceed available coping
resources (Lazarus and Folkman (1984), p. 3403). In the CEO context, this is
operationalized through two contrasting shocks: industry-wide distress
(temporary demand increase) and antitakeover law protection (permanent demand
decrease). Both shocks affect the intensity of CEO job demands without directly
imposing financial hardship on the CEO, which allows identification to isolate
health effects from income effects that confound most stress-and-health
studies. Prior work by Bertrand and Mullainathan (2003) introduced
antitakeover-law variation as a proxy for CEO monitoring intensity; this paper
re-deploys that variation to study health outcomes rather than managerial
behavior.

The biological mechanism is that chronic stress triggers cortisol and other
hormonal responses, causing cellular damage that manifests as visible aging
(p. 3423). Apparent age is validated as a clinical biomarker for mortality
(Christensen et al. (2004), Christensen et al. (2009)): differences between
apparent and chronological age predict short-term and long-term mortality even
when physicians know the chronological age, and correlate with physical
functioning, cognitive performance, and leucocyte telomere length (p. 3413).
The paper's identification tests parallel pre-trends in both the DiD and the
Kaplan-Meier analysis (Figures 4 and 6) and rule out picture-management and
image-selection confounds through a battery of robustness checks.

**Industry distress identification.** An industry is distressed in year $$t$$ if
the median firm's forward-looking two-year stock return falls below $$-30\%$$
(Babina (2020)). The distress indicator for CEO $$j$$ equals 1 if the CEO's
firm was in a distressed industry in 2007, 2008, or both (the Great Recession
crisis years); it does not update after the CEO departs. Treatment status is
orthogonal to pre-crisis aging trends (Figure 4, p. 3418) and to image
sharpness (Internet Appendix Table IA.III).

**Antitakeover law identification.** Business Combination (BC) laws passed
staggered across 33 U.S. states between 1985 and 1997 (Figure 1, p. 3413).
Laws apply by state of incorporation, not state of headquarters, reducing
concern that local economic conditions drive the results. The constitutionality
of BC laws was established by a 1989 federal ruling, strengthening the
exogeneity argument (p. 3412).

## Method

**Part 1: Apparent-age estimation.** Apparent age is estimated from CEO facial
images using the deep CNN of Antipov et al. (2016), trained on more than
250,000 images and winner of the 2016 ChaLearn Looking At People competition
(p. 3414). The model is an ensemble of 11 sub-networks (bagging in the style of
Breiman (1996)) and outputs a $$100 \times 1$$ probability vector over ages 0 to 99;
the apparent-age point estimate is the expected value of this distribution.
The software is validated within the CEO context by comparing 250 random pairs
of CEO images to human assessments; agreement is ~70% overall and ~90% when the
software-estimated age gap is in the top tercile (p. 3414).

The outcome variable is the apparent-age gap:

$$
\text{Apparent Age Gap}_{i,j,t} = \widehat{\text{Apparent Age}}_{i,j,t} - \text{Chronological Age}_{j,t}
$$

where $$i$$ indexes an image, $$j$$ a CEO, and $$t$$ a time bin (p. 3418).

**Part 2: Cox proportional hazards model.** Mortality is estimated using
stratified Cox (1972) proportional hazards models. CEOs enter the risk set
when they take office and exit at death or the October 1, 2017 censoring date.
The baseline hazard $$\lambda_{0,j}(t)$$ is allowed to vary across Fama and French
(1997) 49 industries (p. 3424).

## Empirical specifications

**Apparent-aging DiD (R1, R2).** The pre-versus-post graphical test uses
time-bin indicators interacted with the distress indicator (equation 1, p. 3418):

$$
\text{Apparent Age Gap}_{i,j,t} = \beta_0 + \sum_{\substack{t \in T \\ t \neq 2005\text{-}06}} \beta_{1,t} \cdot \text{Industry Distress}_j \times \mathbb{1}_t + \boldsymbol{\beta}_2' \mathbf{X}_{i,j,t} + \delta_t + \theta_j + \varepsilon_{i,j,t} \tag{1}
$$

The main regression collapses post-crisis to a single indicator (equation 2,
p. 3419):

$$
\text{Apparent Age Gap}_{i,j,t} = \beta_0 + \beta_1 \cdot \text{Industry Distress}_j \times \mathbb{1}_{[t > 2006]} + \boldsymbol{\beta}_2' \mathbf{X}_{i,j,t} + \delta_t + \theta_j + \varepsilon_{i,j,t} \tag{2}
$$

where $$\mathbf{X}_{i,j,t}$$ includes image-level controls for smile, mood,
self-confidence, style, side face, logo, glasses, magazine quality, lighting,
natural pose, pre-2007 industry shock experience, and pre-2007 CEO tenure.
CEO fixed effects $$\theta_j$$ absorb time-invariant facial characteristics.
Standard errors are clustered at the three-digit SIC level. Observations are
weighted by image sharpness (Laplacian). Sample: 3,002 images of 453 CEOs.

**Mortality hazard, industry distress (R3, R7).** The Cox hazard model
stratified by FF49 industry is (equation 3, p. 3424):

$$
\ln \lambda(t \mid \text{Industry Distress}_{i,t}, \mathbf{X}_{i,t}) = \ln \lambda_{0,j}(t) + \beta \cdot \text{Industry Distress}_{i,t} + \boldsymbol{\delta}' \mathbf{X}_{i,t} \tag{3}
$$

where $$\text{Industry Distress}_{i,t}$$ equals 1 if CEO $$i$$ has experienced industry
distress (forward-looking 2-year median firm return $$< -30\%$$) in year $$t$$ or any
prior year. Controls include chronological age, linear or fixed time effects,
and state-of-headquarters location fixed effects. Sample: 1,900 CEOs, 58,034
CEO-year observations; standard errors clustered at 3-digit SIC.

**Mortality hazard, BC law indicator (R4, R6, R8).** The BC law binary
specification is (equation 4, p. 3429):

$$
\ln \lambda(t \mid BC_{i,t}, \mathbf{X}_{i,t}) = \ln \lambda_{0,j}(t) + \beta \cdot I(BC_{i,t}) + \boldsymbol{\delta}' \mathbf{X}_{i,t} \tag{4}
$$

**Mortality hazard, BC cumulative exposure (R5).** The cumulative-exposure
specification counts years of BC law coverage until year $$t$$ (equation 5,
p. 3430):

$$
\ln \lambda(t \mid BC_{i,t}, \mathbf{X}_{i,t}) = \ln \lambda_{0,j}(t) + \beta \cdot BC_{i,t} + \boldsymbol{\delta}' \mathbf{X}_{i,t} \tag{5}
$$

The BC analyses restrict to CEOs appointed before the BC laws were enacted
(1,605 CEOs) to address selection; standard errors are clustered at the
state-of-incorporation level. Both BC specifications add a first-generation
antitakeover law exposure control following Karpoff and Wittry (2018).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CEO Apparent Aging Data Set: 3,002 Getty Images / Google Images photos of 453 Fortune 1000 CEOs (2006 cohort), dated images, 2000-2016 | Outcome (apparent-age gap); identified by ML apparent-age CNN | No page yet |
| CEO Mortality Data Set: Forbes Executive Compensation Surveys 1975-1991 (Gibbons and Murphy (1992)); hand-collected birth and death dates from Ancestry.com for 2,361 of 2,720 CEOs at 1,501 firms; tenure from Execucomp and NYT "Business People" | Outcome (mortality/longevity); treatment (industry distress, BC law exposure) | [Forbes exec comp](/wiki/datasets/forbes-executive-compensation/) |
| CRSP (via WRDS): stock returns, PERMNO identifiers, historical SIC codes; used to construct annual industry-distress panel | Industry distress definition (median 2-year forward return < -30%) and sample restriction | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| Compustat (via WRDS): assets, employees; Comphist / Compustat Snapshot for historical state of incorporation | Firm controls; historical state-of-incorporation for BC law assignment | [WRDS / Compustat](/wiki/commercial/wrds/) (licensed) |
| BC law passage dates, by state: Cheng, Nagar, and Rajan (2004); Cain, McKeon, and Solomon (2017); Karpoff and Wittry (2018) | Treatment variable (BC law indicator and cumulative exposure) | No page yet |
| Antipov et al. (2016) deep CNN apparent-age software (Oxford VGG architecture) | Apparent-age estimation from facial photos | No page yet |
| Human Mortality Database (2019) | Benchmark mortality rates for economic significance comparisons | No page yet |

Sample periods: apparent-aging analysis, 2000-2016 (images) / Fortune 1000 cohort 2006; mortality analysis, 1975-2017 (CEO-year panel, censoring October 1, 2017).

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13497) if you are:
extending the apparent-aging ML approach to other executive samples or
professional groups (see Section II and Internet Appendix Section II.A for the
CNN architecture and image-processing details); replicating the mortality
analysis (full robustness tables are in Internet Appendix Sections III-IV, with
22+ additional specifications); studying the BC law / antitakeover
identification in detail (Sections IV.D and Internet Appendix Tables
IA.XIX-IA.XXIII follow Karpoff and Wittry (2018) exhaustively); or examining
the pay-and-health compensating-differential calibration (Section IV.E and
Internet Appendix Section IV).

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(6), December 2025,
pp. 3401-3442. This distillation was extracted by an LLM on 2026-06-03 and is
**not human-verified or independently reproduced**. The CC BY 4.0 licence
permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Borgschulte, Mark, Marius Guenzel, Canyao Liu,
> and Ulrike Malmendier. "CEO Stress, Aging, and Death."
> *The Journal of Finance* 80, no. 6 (December 2025): 3401–3442.
> DOI: 10.1111/jofi.13497. © 2025 The Author(s).
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Does Floor Trading Matter: Brogaard, Ringgenberg & Roesch (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/brogaard-floor-trading-matter-2025/
# Distilled: Using the COVID-19 suspension of NYSE floor trading on March 23, 2020 as a natural experiment, this paper finds that human floor traders significantly improve market quality: their removal raises proportional effective spreads by roughly 9 basis points (more than 70% of the pre-closure mean) and increases Hasbrouck pricing errors by approximately 6%. J. Finance 2025, CC BY 4.0. Seven core results with source locators, datasets used, the DiD identification design, and the mechanism tests.
# Tags: paper-summary, market-microstructure, liquidity, algorithmic-trading
==============================================================================

**What this is.** The paper's core results, identification design, estimating
equations, and mechanism tests: enough to know what it found and how, without
reading all 40 pages. To replicate or extend it, read the full source at the
[original](https://doi.org/10.1111/jofi.13401).

## TL;DR

On March 23, 2020, the NYSE suspended floor trading because of COVID-19, moving
to fully electronic trading. Using this abrupt, exogenous shock in a
difference-in-differences framework, the paper shows that human floor traders
improve market quality. Removing them raises proportional effective spreads by
roughly 9 basis points (more than 70% relative to the pre-closure mean) and
increases Hasbrouck pricing errors by about 6%, with the effects largest
immediately after the open when information arrival is highest. Two partial
reopenings (May 26 and June 17, 2020) allow the authors to distinguish between
two non-mutually-exclusive mechanisms: (i) in-person information transfer
between floor brokers and designated market makers (DMMs), and (ii) floor-only
D-order types. Continuous trading quality improves only after the second
reopening (DMMs return in person), while auction quality improves after the
first (D orders resume). Floor trading matters in the age of algorithms.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%.
Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Effective spreads increase significantly** for NYSE stocks after floor closure, under both identification strategies | Table III Panel A col. 2 (p. 394); Table III Panel B col. 2 (p. 395) | NYSE vs. NASDAQ: +8.92 bps (t=7.90); within-stock NYSE vs. off-NYSE: +1.46% (t=2.68); stable across all fixed-effect specs |
| R2 | **Pricing errors increase** for NYSE stocks after floor closure, indicating worse price efficiency | Table IV Panel A cols. 1-2 (p. 396); Table IV Panel B col. 1 (p. 396) | NYSE vs. NASDAQ: +6.28\*\*\* (t=4.68) in log pricing error; within-stock: +2.39\* (t=1.84); approx. 6% and 2% increases respectively |
| R3 | **Effects are strongest immediately after the opening auction** (9:30-10:00 am) and decline monotonically through the trading day | Table V Panels A-B (pp. 398-399) | 9:30-10:00: +8.16\*\*\* (t=5.38); 10:00-10:30: +5.69\*\*\* (t=6.69); coefficients near zero from 12:00 onward |
| R4 | **Floor traders matter more when stock-specific information is higher**: triple interaction with information complexity is positive and significant | Table V Panel C (p. 399) | Complex x Treated x After = +17.45\*\*\* (t=7.33); confirms information-transfer channel |
| R5 | **Opening and closing auction quality deteriorates** after floor closure; price deviations increase for both auction types | Table VI (pp. 400-401) | Opening: +34 to +47 bps (Treated x After, all significant); Closing: +32 bps (approx. 100% increase relative to unconditional mean) |
| R6 | **Continuous trading spreads improve only at the second reopening** (DMMs return in person), not the first (D orders resume); D orders are not the continuous-trading mechanism | Table VII Panel A col. 2 (p. 404); Table VII Panel B col. 2 (p. 405) | Treated x Open1 = +0.27 (t=0.72, insignificant); Treated x Open2 = -4.85\*\*\* (t=-6.69) |
| R7 | **Auction quality improves at the first reopening** (when D orders resume) but not the second; D orders are the auction mechanism | Table VIII (pp. 407-408) | Treated x Open1 = -20.50\*\* to -20.68\*\* for opening auction (significant); Treated x Open1 for closing auction = -3.74\*\*\* to -4.00\*\*\*; second reopening insignificant for auctions |

**Overall (paper's conclusion).** The NYSE floor suspension degraded market
quality across multiple dimensions: liquidity (spreads), price efficiency
(pricing errors), and auction quality. The partial reopening evidence shows two
non-mutually-exclusive mechanisms are at work. During continuous trading, the
floor facilitates in-person information transfer between floor brokers and DMMs
that electronic markets cannot replicate; this channel reverses with the second
reopening, consistent with theoretical predictions by Benveniste, Marcus and
Wilhelm (1992) and the empirical evidence in Battalio, Ellul and Jennings
(2007) that personal relationships lower informational asymmetries. During
auctions, D orders (accessible only to floor traders) improve auction outcomes;
this channel reverses with the first reopening, consistent with Jegadeesh and
Wu (2022). The results address the debate sparked by Hu and Murphy (2020), who
found NYSE closing auction quality worse than NASDAQ and attributed it to floor
traders; here, floor traders are shown to improve auction quality and liquidity.
Bessembinder, Hao and Zheng (2020) and Clark-Joseph, Ye and Zi (2017) document
the value of DMM participation; this paper separately identifies floor brokers
as an additional, distinct channel. The price-discovery role of designated
dealers confirmed here is consistent with Madhavan and Panchapagesan (2000).
Overall, human floor traders remain valuable intermediaries even in the
algorithmic trading era.

## Theory / model

The paper has no formal structural model. The identification argument begins
with an additive model of market quality (eq. 1, p. 387):

$$
E[\text{MarketQuality}_{0,i,e,t} \mid i, e, t] = \rho_e + \lambda_t + \Gamma_i + \alpha_{i,t} \tag{1}
$$

where $$\rho_e$$ captures the exchange effect, $$\lambda_t$$ an aggregate time
effect, $$\Gamma_i$$ a time-invariant firm effect, and $$\alpha_{i,t}$$ time-varying
firm-level shocks. Observed market quality is then (eq. 2, p. 387):

$$
\text{MarketQuality}_{i,e,t} = \rho_e + \lambda_t + \Gamma_i + \alpha_{i,t} + \beta \cdot \mathbf{1}_{\text{FloorTrading}} + \xi_{i,e,t} \tag{2}
$$

where $$\mathbf{1}_{\text{FloorTrading}}$$ equals 1 if floor trading is suspended
and 0 otherwise; $$\beta$$ is the treatment effect of interest. The canonical
DiD estimator compares treated minus control before and after (eqs. 3-4,
p. 387-388), recovering $$\beta$$ when the parallel trends assumption holds,
i.e., when time-varying shocks to control firms ($$\alpha_{j,t}$$) evolve as in
treatment firms ($$\alpha_{i,t}$$).

The two tested mechanisms follow from three observable changes caused by the
floor closure: (i) loss of in-person interaction between DMMs and floor brokers,
(ii) loss of manual auction operation, and (iii) loss of special floor-only
order types (D orders). The staggered partial reopenings isolate each channel:
first reopening restores D orders but not in-person interaction; second
reopening restores in-person interaction and manual auctions.

## Method

The paper applies two difference-in-differences strategies using the same
additive model, differing in the control group.

**Approach 1: NYSE vs. matched NASDAQ stocks** (eq. 5, p. 389). The control
group is NASDAQ-listed stocks matched on price, trading volume, market
capitalization, and Fama-French 48-industry classification using one-to-one
nearest-neighbor propensity score matching (PSM) without replacement. The
treatment indicator is 1 if the stock is listed on NYSE after March 23, 2020:

$$
y_{i,e,t} = \beta \mathbf{1}_{i,e,t} + \lambda_t + \Gamma_i + \gamma C_{i,e,t} + \epsilon_{i,e,t} \tag{5}
$$

where $$y_{i,e,t}$$ is the market quality measure, $$\mathbf{1}_{i,e,t}$$
equals 1 for NYSE-listed stocks after March 23, $$\lambda_t$$ and $$\Gamma_i$$ are
date and firm fixed effects, and $$C_{i,e,t}$$ is a vector of controls (log
trading volume, parity trading volume). Standard errors are clustered by firm.

**Approach 2: Within-stock variation across exchanges** (eq. 6, p. 391). The
same NYSE-listed stocks are compared across trading venues, using trades in
the same stock on other exchanges (IEX, NASDAQ, BZX, etc.) as the control.
This includes firm x date fixed effects ($$\kappa_{it}$$) to absorb time-varying
firm-level shocks:

$$
y_{i,e,t} = \beta D_{i,e,t} + \kappa_{i,t} + \gamma C_{i,e,t} + \epsilon_{i,e,t} \tag{6}
$$

where $$D_{i,e,t}$$ equals 1 if firm $$i$$ is NYSE-listed, trading on the NYSE,
after March 23, 2020. The firm x date fixed effects absorb any time-varying
firm-level response to COVID-19, so confounders must differentially affect
NYSE vs. off-NYSE trading in the same stock around March 23 to bias the
estimate.

Both strategies use `difference-in-differences` and `panel-regression`
building blocks. The matching step uses `matching` (PSM). The market quality
measures are: proportional effective spread (PESPR), Hasbrouck (1993) pricing
error, and auction price deviation (eq. 7, p. 400):

$$
|\text{Deviation\%}| = 2 \times |\log(\text{trade}) - \log(\text{mid})| = 2 \times |\log(\text{trade}/\text{mid})| \tag{7}
$$

where log(trade) is the log auction trade price and log(mid) is the last
midpoint price from continuous trading. Higher values indicate worse auction
quality. Information complexity (Panel C of Table V) is measured as
$$1 - R^2$$ from a regression of each stock's return on the contemporaneous SPY
return within each half-hour interval (from Morck, Yeung, and Yu (2000)).

## Empirical specifications

**Main closure window** (Tables III-VI): two-week window from March 16 to March
27, 2020 (one week before, one week after the March 23 floor closure). Sample:
approximately 1,600 NYSE-listed stocks (PSM approach) or 552-553 NYSE-listed
stocks with sufficient off-exchange trading (within-stock approach). Results
are robust to a larger window encompassing all of 2020-2021 (Internet Appendix
Table IA.II).

**Spread regression (R1, Table III):** Dependent variable is PESPR in
percentage of midpoint. PESPR on NYSE is computed from NYSE quotes and trades
only; PESPR off-NYSE is from NBBO excluding NYSE. Specifications range from
stock fixed effects only (col. 1) to stock + day fixed effects (col. 2) to
adding price and log volume controls (col. 3-4 for Panel A; col. 3-5 for
Panel B). The treatment coefficient of 8.92 bps in Panel A and 1.46% in
Panel B are stable across all specifications, supporting identification
assumptions.

**Pricing error regression (R2, Table IV):** Dependent variable is
$$\log|\text{PricingError}_{i,e,t}|$$, the Hasbrouck (1993) measure estimated
following Rosch, Subrahmanyam, and Van Dijk (2017). Specifications mirror
Table III. Panel A coefficient of 6.28 (t=4.68) implies approximately 6%
increase in pricing errors; Panel B coefficient of 2.39 (t=1.84) implies
approximately 2% increase.

**Intraday spread regression (R3-R4, Table V):** Equation mirrors eq. 6 with
PESPR computed within each 30-minute interval from 9:30 to 16:00. Panel A
(morning) and Panel B (afternoon) show results by half-hour. Panel C includes
a triple interaction with information complexity (1 - R2), with stock x day
fixed effects.

**Auction deviation regression (R5, Table VI):** Dependent variable is
$$|\text{Deviation\%}|$$ for opening and closing auctions. Equation follows eq.
5 with stock and day fixed effects; additional control for D-order volume in
cols. 3-4 (opening) and cols. 7-8 (closing). Specification confirms D orders
are not the primary driver of auction deterioration.

**Reopening regressions (R6-R7, Tables VII-VIII):** Two-week window centered
around each of the two partial reopenings (May 26 and June 17, 2020). The
estimating equation for continuous trading (Table VII) includes two treatment
indicators (Open1, Open2) interacted with Treated; stock x event fixed effects
absorb different aggregate conditions across reopenings:

$$
\text{PESPR}_{i,e,t} = \beta_1(\text{Treated}_{i,e,t} \times \text{Open1}_{i,e,t}) + \beta_2(\text{Treated}_{i,e,t} \times \text{Open2}_{i,e,t}) + \beta_3 \text{Treated} + \beta_4 \text{Open1} + \beta_5 \text{Open2} + \gamma C_{i,e,t} + FE + \epsilon_{i,e,t}
$$

The analogous form applies to auctions (Table VIII). A negative and significant
$$\beta_2$$ (Open2) with insignificant $$\beta_1$$ (Open1) for continuous trading
identifies in-person interaction as the mechanism. A negative and significant
$$\beta_1$$ (Open1) with insignificant $$\beta_2$$ for auctions identifies D orders
as the auction mechanism.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CRSP (Center for Research in Security Prices) | Stock prices, dollar trading volume (thousands USD), market capitalization; used to construct the matched sample and identify NYSE vs. NASDAQ stocks | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| NYSE TAQ (Trade and Quotes) via WRDS Intraday Indicators | Proportional quoted spreads (PQSPR), proportional effective spreads (PESPR), and Hasbrouck pricing errors; the main market quality measures for both identification strategies | [TAQ](/wiki/commercial/taq/) (licensed) |
| OneMarket Data (OneTick software) | Intraday TAQ data for within-exchange comparisons (PESPR computed separately by exchange for each 30-minute interval); NYSE data provided directly | no page yet |

Sample: main closure window March 16-27, 2020 (10 trading days); reopening windows May 18-June 1 and June 10-23, 2020 (10 trading days each). The sample starts with 3,447 US common stocks (CRSP share code 10 or 11) listed on NYSE (1,256) or NASDAQ (2,191) in December 2019, excluding stocks with more than one share class (431) or market cap below $500 million (1,412), yielding approximately 1,600 equities (Table I, p. 386).

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13401) if you are:
comparing market quality across trading mechanisms (hybrid vs. fully electronic
exchanges); designing or evaluating floor-trading regulation; studying the role
of human intermediaries and in-person information exchange in modern markets;
or extending the DiD design to other exchange structure changes. The partial
reopening evidence (Tables VII-IX) is especially useful for separating
information-transfer from order-type mechanisms. The Internet Appendix (20
robustness measures from WRDS Intraday Indicators) provides a comprehensive
falsification suite.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(1), February 2025. This
distillation was extracted by an LLM on 2026-06-06 and is **not human-verified
or independently reproduced**. The CC BY 4.0 licence permits mirroring; the
verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Brogaard, Jonathan, Matthew C. Ringgenberg,
> and Dominik Roesch. "Does Floor Trading Matter?" *The Journal of Finance*
> 80, no. 1 (February 2025): 375-414. DOI: 10.1111/jofi.13401.
> (C) 2024 The Author(s). Licensed under
> [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Forest through the Trees: Bryzgalova, Pelger & Zhu (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/bryzgalova-forest-cross-sections-2025/
# Distilled: Asset Pricing Trees (AP Trees) use decision-tree conditional sorts with global SDF-spanning pruning to build interpretable cross-sections of stock returns that achieve out-of-sample Sharpe ratios up to three times higher than conventional double and triple sorts. J. Finance 2025, CC BY 4.0. Eight core results with source locators, datasets used, the model (SDF projection), and the method (AP Trees + AP Pruning) with its defining equations.
# Tags: paper-summary, asset-pricing, factors, anomalies, cross-section, portfolio-sort, decision-trees, factor-models, machine-learning, open-access, cc-by, peer-reviewed, unreplicated, data:wrds, data:ken-french
==============================================================================

**What this is.** The paper's core results, the model it builds on (the SDF
projected on stock returns), and the method it contributes (AP Trees and AP
Pruning) with the defining equations: enough to know what it found and how,
without reading all 60 pages. To replicate or extend it, read the full source
at the [original](https://doi.org/10.1111/jofi.13477).

## TL;DR

The paper proposes Asset Pricing Trees (AP Trees): a decision-tree method
that groups individual stocks into managed portfolios by conditional
characteristic splits, then selects a sparse subset via global SDF-spanning
pruning (LASSO + robust mean/variance shrinkage). Applied to 10
CRSP/Compustat characteristics (Jan 1964 to Dec 2016, 53 years of monthly
data), AP Trees deliver small cross-sections of 10 to 40 long-only
portfolios whose combined out-of-sample Sharpe ratio is up to three times
higher than conventional double/triple sorts, and whose SDF alpha is
significant against all leading factor models. The key drivers are
(i) modeling characteristic interactions via conditional splits and
(ii) optimizing the SDF-spanning objective rather than return prediction.

## Core results

Magnitudes and significance are as reported; `**`/`***` = 5%/1%. Locators
point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | AP Trees deliver **out-of-sample Sharpe ratios up to three times higher** than conventional triple sorts across all 36 three-characteristic cross-sections | Figure 6, Panel A, p. 2474; Table B.II, p. 2497 | AP Tree SR ranges from ~0.24 to ~0.65 (monthly); triple-sort SR rarely exceeds 0.4; difference is up to 3x for cross-sections including investment, idiosyncratic volatility, or profitability |
| R2 | **AP Tree SDF alphas are large and significant** against FF3, FF5, XSF, and FF11 models; triple-sort SDFs are routinely spanned | Table I, p. 2482; Figure 6, Panel B, p. 2474 | For size/OP/investment: AP Trees (10) α = 0.94 vs FF3 (t=10.11), 0.81 vs FF5 (t=8.76); triple sorts α = 0.75 vs FF3 (t=7.40), 0.47 vs FF5 (t=5.57) |
| R3 | **Cross-sectional fit (XS-R²) is far lower for AP Trees**, confirming they contain pricing information not in standard factor models | Table I, p. 2482 | AP Trees (10) XS-R² vs FF5 = 11%; triple sorts (32) XS-R² vs FF5 = 91%; AP Trees (40) = 64% vs 91% for TS(32) |
| R4 | **Interaction nodes account for roughly half of AP Trees' SR**; removing interactions halves the out-of-sample Sharpe ratio | Figure 8, p. 2478 | AP Tree (10) monthly SR ~0.4–0.65 with interactions; dropping interactions reduces SR to roughly that of XSF long-short factors (~0.2); similar pattern for AP Trees (40) |
| R5 | **SDF-spanning objective drives the result**: V-Trees (same splits, variance criterion only) have 2–3x lower SR; ML return-prediction portfolios (deep learning, random forest) have SR at best half that of AP Trees | Figure 9, p. 2481; Figure 15, p. 2494 | AP Trees monthly SR ~0.4–0.65; V-Trees ~0.2–0.3; DL-MV, RF-MV ~0.2–0.35 across 10-characteristic cross-sections |
| R6 | **10 pruned AP Tree portfolios retain ~90% of the SR** of 40-portfolio cross-sections, showing the quality of the cross-section is not driven by size | Figure 7, p. 2477; Table I, p. 2482 | AP Trees (10) SR = 0.65 vs AP Trees (40) SR = 0.69 for size/OP/investment; 10 portfolios already span ~90% of the SDF |
| R7 | **In the large-dimension setting (10 characteristics), AP Trees raise monthly SR by ~0.1 over the best 25x9 double-sort cross-section** (roughly 20% gain), and 80–100% over anomaly-based deciles | Figure 14, p. 2493; §V | AP Trees (10/40) achieve SR ~0.5; best DS25 combination reaches ~0.4–0.45; decile-sort SR ~0.3; quintile SR ~0.25 |
| R8 | **Microcaps do not drive results**: excluding small caps (size quantile below 0.4) or restricting to the top 600 stocks by market cap, AP Trees still roughly double the SR of triple sorts | Table III, p. 2490 | Top-600-stocks AP Trees (10): SR = 0.30, alpha vs FF3 = 0.70 (t=5.04), alpha vs FF5 = 0.32 (t=3.87); triple sorts SR = 0.17 (TS32) / 0.17 (TS64) |

**Overall (paper's conclusion).** Conventional single/double/triple sorts and
their stacked combinations, including the size and value sorts of Fama and
French (1993), do not span the conditional SDF and so provide an unreliable
and often misspecified benchmark for asset pricing models. AP Trees resolve
this by finding a small,
interpretable, well-diversified cross-section that genuinely spans the
conditional SDF projected on characteristics, enabling better model
evaluation and construction of tradable risk factors.

## Theory / model

The economic object is the stochastic discount factor (SDF) that prices
individual stocks. Under no-arbitrage there is a unique minimum-variance SDF
spanned by individual stock excess returns. Given a set of firm characteristics
$$C_{t-1}$$ (an $$N \times K$$ matrix for $$N$$ stocks, $$K$$ characteristics) as the
conditioning information, the conditional SDF is its projection on individual
stock excess returns $$R_t$$ (p. 2455):

$$
M_t^C = 1 - \sum_{i=1}^N b_{t-1,i} \left( R_{t,i} - \mathbb{E}_{t-1}[R_{t,i}] \right), \qquad b_{t-1,i} = f(C_{t-1,i})
$$

with $$f(\cdot)$$ a general, potentially nonlinear and nonseparable function.
Reduced-form models approximate this dependence with $$J$$ basis functions
$$f_j(\cdot)$$, so $$f(C_{t-1,i}) \sim \sum_j f_j(C_{t-1,i}) w_j$$. That turns the
conditional problem into an unconditional one over $$J$$ managed portfolios
(equation 1, p. 2455):

$$
M_t^C = 1 - \sum_{j=1}^J w_j \left( R^{\text{man}}_{t,j} - \mathbb{E}[R^{\text{man}}_{t,j}] \right), \qquad R^{\text{man}}_{t,j} = \sum_{i=1}^N f_j(C_{t-1,i}) R_{t,i}
$$

There is a one-to-one mapping between the basis functions $$f_j$$ and the managed
portfolios $$R^{\text{man}}$$. Managed portfolios **span** the projected SDF exactly when
their mean-variance-efficient combination achieves the highest Sharpe ratio
(p. 2456); pricing a spanning set is then equivalent to pricing the SDF itself.

The paper's theoretical contribution is a misspecification result. Suppose a
researcher uses only a subset $$R^{\text{select}}$$ of the spanning managed portfolios and
omits $$R^{\text{omit}}$$, then proposes a $$K$$-factor model $$F$$ that prices $$R^{\text{select}}$$
(intercept $$\alpha^{\text{select}} = 0$$). Proposition 1 (p. 2457) bounds the mispricing of
the omitted assets by a Sharpe-ratio gap:

$$
\text{SR}^2(R^{\text{select}}, R^{\text{omit}}) - \text{SR}^2(F)
   \leq {\alpha^{\text{omit}}}' (\Sigma^{\text{omit}})^{-1} \alpha^{\text{omit}}
   \leq \text{SR}^2(R^{\text{select}}, R^{\text{omit}}) - \text{SR}^2(R^{\text{select}}),
$$

with equalities if $$R^{\text{select}}$$ spans the factors, that is, if
$$\text{SR}(R^{\text{select}}) = \text{SR}(R^{\text{select}}, F)$$. In words: a model that perfectly explains the
chosen test assets can still be a grossly misspecified model for individual
stocks if those test assets do not span the SDF. The proof follows the spanning
arguments of Barillas and Shanken (2016) and is given in the Internet Appendix.
This is why the choice of test assets, not just the candidate model, matters,
and it motivates constructing a cross-section that provably spans the SDF.

**Identification.** Fully out-of-sample evaluation with a train / validate /
test split (Figure 5, p. 2472): portfolio selection and tuning are fixed on the
first two blocks before the test block, so there is no look-ahead bias.

## Method

The method has two parts: building the tree-based managed portfolios (AP Trees)
and selecting a sparse spanning subset of them (AP Pruning). It builds on
`decision-trees` and `conditional-sorts` for the portfolios, and on `lasso`,
`ridge-shrinkage`, `sdf-projection`, and `robust-mean-variance-optimization`
for the selection.

**AP Trees.** Stocks are grouped by a sequence of conditional consecutive
splits (median splits at each node, without loss of generality), so each final
and intermediate node is a managed portfolio that traces back to firm
fundamentals (Figure 1, p. 2449). Trees are grown to depth four. With $$M$$
candidate splitting characteristics and depth $$d$$, this yields $$M^d \times 2^d$$
overlapping portfolios that capture up to $$d$$-way interactions. Equivalently,
AP Trees are a nonparametric estimator of the SDF mapping (p. 2463):

$$
f(C_{t-1,i}) = \sum_{j=1}^J w_j \cdot \mathbf{1}\{ C_{t-1,i} \in A_j \}, \qquad \text{regions } A_j \text{ given by the recursive tree nodes}
$$

**AP Pruning.** The naive SDF weights solving $$\mathbb{E}[R^{\text{man}} M^C] = 0$$ are
$$\omega = \Sigma^{-1} \mu$$, whose sample version $$\hat{\omega}_{\text{naive}} = \hat{\Sigma}^{-1}
\hat{\mu}$$ overfits in high dimension. AP Pruning instead selects a sparse set of
tree nodes that span the SDF, with robust moments. Definition 1 (p. 2466), step
one, estimates robust SDF weights on the training data:

$$
\min_{\omega} \; \tfrac{1}{2} \left( \hat{\mu}^{\text{robust}} - \hat{\Sigma}^{\text{robust}} \omega \right)' \left(\hat{\Sigma}^{\text{robust}}\right)^{-1} \left( \hat{\mu}^{\text{robust}} - \hat{\Sigma}^{\text{robust}} \omega \right) + \lambda_1 \|\omega\|_1
$$

$$
\text{with} \quad \hat{\Sigma}^{\text{robust}} = \hat{\Sigma} + \lambda_2 I_N, \qquad \hat{\mu}^{\text{robust}} = \hat{\mu} + \lambda_0 \mathbf{1}
$$

where $$\|\omega\|_1 = \sum_i |w_i|$$, $$\mathbf{1}$$ is a vector of ones, and $$N$$ is the number
of assets. The tuning parameters $$(\lambda_0, \lambda_1, \lambda_2)$$ are chosen on
the validation block to maximize the robust-SDF Sharpe ratio, then performance
is evaluated only on the untouched test block.

Proposition 2 (p. 2468) shows AP Pruning is equivalent to a robust tangency
portfolio (Definition 2, equation 2):

$$
\min_{\omega} \; \tfrac{1}{2} \omega' \hat{\Sigma} \omega + \lambda_1 \|\omega\|_1 + \tfrac{1}{2} \lambda_2 \|\omega\|_2^2
$$

$$
\text{subject to} \quad \omega' \mathbf{1} = 1, \qquad \omega' \left( \hat{\mu} + r_f \mathbf{1} \right) \geq \mu_0 + r_f
$$

The mapping is exact: the LASSO term $$\lambda_1$$ induces sparsity (a small set of
AP Tree basis assets), the target return $$\mu_0$$ corresponds to shrinking the
mean toward its cross-sectional average, and the ridge term $$\lambda_2$$
corresponds to variance shrinkage of the covariance matrix. This generalizes the
robust SDF recovery of Kozak, Nagel, and Santosh (2020), which shrinks only the
covariance, by also shrinking the mean (Proposition 3). Proposition 4 gives a
robust-control reading: the shrinkage solves a minimax problem over joint
estimation uncertainty in means and variances.

## Empirical specifications

Estimation runs on monthly CRSP/Compustat data, January 1964 to December 2016,
split into training (first 20 years), validation (10 years), and out-of-sample
testing (last 23 years). Each headline result is the test-block value of one of
the following constructions, not an OLS regression with fixed effects:

- **Out-of-sample Sharpe ratio (R1, R6, R7, R8).** For a given cross-section,
  the SDF / tangency portfolio is estimated on training, tuned on validation
  (Definition 1-2), then its realized monthly SR is computed on the test block.
  AP Trees are compared against triple sorts, double sorts (25x9), deciles, and
  quintiles of the same characteristics.
- **SDF alpha (R2).** The candidate factor models FF3, FF5, XSF
  (cross-section-specific factors), and FF11 are confronted with each
  cross-section; the SDF alpha is the pricing error of the cross-section's
  implied SDF against the model, with $$t$$-statistics reported (Table I,
  p. 2482; Figure 6 Panel B). Spanning is assessed by whether one
  cross-section's SDF is priced by another's factors.
- **Cross-sectional fit, XS-R² (R3).** The share of cross-sectional variation in
  average managed-portfolio returns explained by the candidate model (Table I).
- **Channel decompositions (R4, R5).** Re-running the construction with
  interaction nodes removed (R4, Figure 8), with a variance-only split criterion
  (V-Trees, R5, Figure 9), and against ML return-prediction portfolios built
  from Gu, Kelly, and Xiu (2020) tools (random forest, neural nets; Figure 15),
  isolating the contribution of interactions and of the SDF-spanning objective.

Robustness reported in the paper: excluding microcaps or restricting to the top
600 stocks by market cap (Table III, p. 2490), and rolling-window time-varying
SDF weights (Figure C.6, p. 2502).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CRSP monthly stock returns and market data | Portfolio returns, market cap, momentum, short-term reversal, idiosyncratic volatility | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| Compustat annual fundamentals | Book-to-market, investment, operating profitability, accruals, turnover, long-term reversal | [WRDS / Compustat](/wiki/commercial/wrds/) (licensed) |
| Kenneth French Data Library | 10 firm-specific characteristic definitions (Table A.I); Fama-French factor returns for benchmarking (FF3, FF5) | [Ken French library](/wiki/datasets/ken-french/) |
| One-month Treasury bill rate | Proxy for the risk-free rate | No page yet |

Sample: January 1964 to December 2016 (53 years, monthly). Training sample: first 20 years; validation: 10 years; testing (out-of-sample): last 23 years.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13477) if you are:
replicating (the Internet Appendix contains formal proofs and additional
robustness); building new cross-sections beyond the 10 characteristics
studied; evaluating whether a candidate factor model spans the SDF; or
extending AP Trees to other asset classes or conditional models. The
locators above point to the exact tables and figures.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(5). This distillation
was extracted by an LLM on 2026-05-31 and is **not human-verified or
independently reproduced**. The CC BY 4.0 licence permits mirroring; the
verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Bryzgalova, Svetlana, Markus Pelger, and Jason Zhu.
> "Forest through the Trees: Building Cross-Sections of Stock Returns."
> *The Journal of Finance* 80, no. 5 (October 2025): 2447–2506.
> DOI: 10.1111/jofi.13477. © 2025 The Author(s).
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Privacy and Team Incentives: Buffa, Liu & White (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/buffa-privacy-team-incentives-2025/
# Distilled: When compensation contracts are bilateral and private, principals contracting with complementary-effort teams face a commitment problem that depresses incentive pay. Delegating contracting authority to the most skilled agent (team leader) mitigates the problem via an observability effect, and dominates centralized contracting when effort intensity is high enough or agents are sufficiently asymmetric. The Journal of Finance 2025, paywalled. Seven core results with source locators, no estimation, pure theory with a banking-syndicate application.
# Tags: paper-summary, contract-theory, team-incentives, moral-hazard, organizational-design
==============================================================================

**What this is.** The core propositions and their economic logic from a pure
theory paper on team contracting under private contracts, with an
application to banking syndicates: enough to know what it proves and how,
without reading all 55 pages. To replicate or extend it, read the full
source at [https://doi.org/10.1111/jofi.13496](https://doi.org/10.1111/jofi.13496).

## TL;DR

When compensation contracts are bilateral (observed only by the two parties
who sign them), a principal contracting with two complementary-effort agents
cannot commit to paying her agents enough: any promise of a high bonus to
one agent can be secretly reneged on, and rational agents anticipate this,
so equilibrium effort falls below the second-best (public-contracts)
optimum. The paper shows that delegating contracting to the most skilled
agent (the "Agent") who then sub-contracts with the less skilled agent (the
"Subagent") partially solves this commitment problem via an *observability
effect*: the Agent now observes the Subagent's contract and so is not
afraid of the principal reducing the Subagent's incentives. The cost is a
*self-interest effect*: the Agent skews the budget toward himself. Delegation
dominates centralized private contracting when the project's effort
intensity $$\rho = \theta/\gamma$$ is above a threshold
$$\bar{\rho}(\alpha)$$ that decreases with the skill gap $$\alpha - 1/2$$
between the agents. Applied to banking syndicates, the theory predicts when
sole mandates, fee concentration, and hierarchical structures are optimal.

## Core results

Magnitudes and significance are as reported (pure theory; all results are
propositions or lemmas). Locators point into the source PDF.

| # | Result | Locator | Magnitude / statement |
|---|---|---|---|
| R1 | Under **public contracts** (second best), the optimal compensation budget equals effort intensity and the optimal allocation equals relative skill. | Proposition 1, p. 3453 | $$b^* = \rho,\; \phi^* = \alpha$$ exactly; four structural parameters collapse to two: $$\rho \equiv \theta/\gamma$$ and $$\alpha$$. |
| R2 | Under **centralized private contracts**, the budget is distorted downward and the allocation skewed toward the more skilled agent. | Proposition 2, p. 3456 | $$b^C = \rho - \frac{\alpha(1-\alpha)(2-\rho)\rho^2}{1-\alpha(1-\alpha)\rho^2} < b^*$$; $$\phi^C = \alpha + \frac{(\alpha-1/2)2\alpha(1-\alpha)\rho}{1-2\alpha(1-\alpha)\rho} \geq \alpha = \phi^*$$. Distortions grow with $$\alpha(1-\alpha)$$ (skill heterogeneity) and $$\rho$$. |
| R3 | Under **delegated private contracts** (principal contracts with one Agent, who sub-contracts), the budget distortion is smaller but the allocation distortion may be larger or smaller. | Proposition 3, p. 3460 | $$b^D = \rho - \alpha_A(1-\alpha_A)\rho^2 < b^*$$; $$\phi^D_A = \alpha_A + (1-\alpha_A)(1-\rho)$$. Budget always closer to second best than under centralized: $$b^C < b^D < b^*$$ (Lemma 1). |
| R4 | The principal **prefers to delegate to the more skilled agent**; delegation to the less skilled agent entails a larger allocation distortion that dominates. | Proposition 4, p. 3462 | For any $$\alpha > 1/2$$, $$v^D_{A=1} > v^D_{A=2}$$, where $$v^D_{A=i}$$ is the principal's expected payoff when agent $$i$$ is the Agent. The result follows because $$g(\alpha,\Delta) > 1$$ for all $$\alpha \in (1/2,1)$$ and $$\Delta \in (0,1-\alpha)$$ (Appendix eq. A11, p. 3489). |
| R5 | Delegation **dominates centralized contracting iff effort intensity is high enough**: $$\rho > \bar{\rho}(\alpha)$$; it is also Pareto-improving iff $$\rho > \tilde{\rho}(\alpha)$$, with $$1/2 < \tilde{\rho}(\alpha) < \bar{\rho}(\alpha) < 1/(2\alpha)$$. Both thresholds decrease with $$\alpha$$. | Proposition 5, p. 3468; Figure 4, p. 3469 | Delegation preferred when observability effect (from $$\rho$$ large) overcomes self-interest effect; delegation is Pareto-improving for a wider parameter region than where principal strictly prefers it. |
| R6 | With **partial transparency** (agents observe each other's contracts with probability $$\lambda$$), more transparency raises the compensation budget and reduces the skew toward the more skilled agent under centralized contracting but leaves the allocation unchanged under delegation. Delegation is optimal iff $$\lambda < \bar{\lambda}$$ for a unique threshold $$\bar{\lambda} \in (0,1)$$. | Proposition 6, pp. 3474-3475; Figure 5, p. 3476 | $$b^C(\lambda) = \rho - \frac{(1-\lambda)\alpha(1-\alpha)\rho^2(1-\rho)(2-\rho)}{(1-\rho)(1-\alpha(1-\alpha)\rho^2)+\lambda\alpha(1-\alpha)\rho^2(2-\rho)}$$; $$\phi^D_A(\lambda) = \alpha_A\rho + (1-\rho)$$ (invariant in $$\lambda$$). |
| R7 | When agents' efforts are **more substitutable** (CES probability function with $$\nu > 0$$), the delegation region *expands*: sole mandates are more likely to be awarded as bank efforts become more substitutable. | Section IV.C, Figure 7, p. 3479 | For baseline parameters $$\alpha=0.75, \rho=0.55$$: centralized contracting preferred when $$\nu < 0.35$$; delegation optimal for $$\nu > 0.35$$. |

**Overall (paper's conclusion).** With bilateral private contracts, the
principal faces a credibility problem that distorts team incentives
downward. Delegating contracting to the most skilled team member can
restore efficiency when effort intensity is high. The theory delivers novel,
testable predictions for banking syndicates: sole mandates (delegation) are
more likely for firm-commitment deals, colder markets, less well-known
issuers, larger skill gaps between underwriters, and when private
compensation components are relatively more important.

## Theory / model

The economic environment (Section I, p. 3448) has two dates and three
risk-neutral players with limited liability. A principal hires two agents
to implement a risky project. Agent $$i = 1, 2$$ exerts unobservable effort
$$e_i \geq 0$$. Project output $$X$$ is Bernoulli (p. 3448, eq. 1):

$$
X(e_1, e_2) = \begin{cases} 1 & \text{with prob. } \pi(e_1, e_2) \\ 0 & \text{with prob. } 1 - \pi(e_1, e_2) \end{cases} \tag{1}
$$

The success probability follows a Cobb-Douglas team-effort function (p. 3449, eq. 2):

$$
\pi(e_1, e_2) = \left(e_1^\alpha e_2^{1-\alpha}\right)^\theta \tag{2}
$$

where $$\theta > 0$$ is the elasticity of expected output to team effort
$$e_1^\alpha e_2^{1-\alpha}$$, and $$\alpha \geq 1/2$$ captures the relative
skill of agent 1 (more skilled). The product $$\alpha(1-\alpha)$$ is an
inverse measure of skill heterogeneity. The effort cost is (p. 3449, eq. 3):

$$
c(e_i) = \kappa e_i^\gamma, \quad \kappa, \gamma > 0, \quad \gamma > \theta, \quad \kappa \geq 1 \tag{3}
$$

The principal's payoff (if the project succeeds) net of the total compensation
budget $$b$$ is $$v = (1-b)\pi(e_1,e_2)$$ (p. 3451). Each agent's payoff is
expected compensation minus effort cost: $$u_1 = \phi b (e_1^\alpha e_2^{1-\alpha})^\theta - \kappa e_1^\gamma$$ and $$u_2 = (1-\phi)b(e_1^\alpha e_2^{1-\alpha})^\theta - \kappa e_2^\gamma$$.

The key ratio $$\rho \equiv \theta/\gamma \in (0,1)$$ captures *effort intensity*:
how elastic expected output is to team effort, relative to the cost
elasticity. Proposition 1 shows this is the only determinant of the optimal
second-best compensation budget when contracts are public.

**Two contracting schemes** (Section III, p. 3454; Figure 1, p. 3450):

- *Centralized contracting*: principal offers contracts to both agents
  privately. Each agent observes only his own offer.
- *Delegated contracting*: principal offers a total budget $$b$$ to the Agent
  (the more skilled agent), who then sub-contracts with the Subagent. The
  Agent observes both contracts; the Subagent observes only his own offer.

**Commitment problem.** With public contracts, Proposition 1 establishes the
second-best optimum $$(b^*, \phi^*) = (\rho, \alpha)$$ as the benchmark. When
contracts are private, the principal can secretly renege on the promised
high-incentive contract for one agent: agent $$i$$ cannot observe agent $$j$$'s
contract, so he cannot verify whether the indirect effort externality he
expects is actually being provided. This destroys the indirect-incentive
channel and depresses the equilibrium budget (Proposition 2, p. 3456).

## Method

The paper's method is theoretical (pure theory, no estimation). Equilibria
are solved by backward induction in a two-period game, using the
**Perfect Bayesian Equilibrium (PBE)** with *passive beliefs* (agents do not
revise beliefs about the other agent's effort when receiving an out-of-equilibrium
offer, p. 3455). The solution procedure is:

1. Given the compensation budget $$b$$ and allocation $$\phi$$, solve each agent's
   incentive-compatibility (IC) constraint for optimal effort (equations 4-5
   in the public case, 10-12 in the centralized private case, 15-18 in the
   delegated case).
2. Impose equilibrium: each agent's conjecture about the other's effort equals
   the equilibrium effort level.
3. Solve the principal's program for $$(b^*, \phi^*)$$ (or the Agent's
   allocation program for $$\phi_A^D$$ in the delegated case).

The paper builds on `principal-agent` and `promotion-contest` frameworks.
The key technical contribution is formalizing the *observability effect* vs.
the *self-interest effect* of delegation, both deriving from the same
bilateral-privacy assumption.

For the banking-syndicate application, the model is extended to **partial
transparency** via a mixing parameter $$\lambda \in [0,1]$$ (Proposition 6,
p. 3474): agents observe each other's contracts with probability $$\lambda$$.
The fully private and fully public cases are nested at $$\lambda = 0$$ and
$$\lambda = 1$$ respectively. A **CES probability function** (eq. 33, p. 3478)
$$\pi(e_1,e_2) = (\alpha e_1^\nu + (1-\alpha)e_2^\nu)^\theta$$ with
substitutability parameter $$\nu$$ nests the Cobb-Douglas as $$\nu \to 0^+$$.

## Empirical specifications

This is a pure theory paper; there is no econometric estimation. The
paper's empirical content is a set of *qualitative comparative-statics
predictions* for banking syndicates (Section IV, pp. 3470-3482), which
can be taken to data. The key mappings from model to data are:

- **Degree of centralization / delegation**: fraction of banks in the top tier
  of a syndicate hierarchy. More banks in the top tier = more centralized
  (the issuer deals directly with each rather than routing through a lead bank).
  A more concentrated distribution of underwriting fees (high HHI among top-tier
  banks) is an alternative delegation measure (p. 3471).
- **Relative skill** $$\alpha$$: proxied by standard underwriter reputation
  measures: Megginson and Weiss (1991) market-share rank; Carter and Manaster
  (1990) tombstone-based rank.
- **Effort intensity** $$\rho$$: harder-to-sell deals have higher $$\rho$$.
  Proxies include: firm-commitment vs. best-efforts underwriting; market
  "coldness" (volume of deals in the quarter); issuer credit quality / cash flows;
  issuer name recognition (p. 3472).
- **Degree of transparency** $$\lambda$$: higher when publicly-disclosed fees
  or spreads dominate compensation; lower when side benefits (e.g., future
  business from the issuer, allocation of underpriced shares) are a large
  share of total compensation (p. 3475).

**Testable predictions** from Propositions 5-6 and the CES extension:

1. Sole mandates (delegation) are more likely when the issue is firm-commitment,
   in colder markets, from less well-known issuers, and for lower-rated debt.
2. Fee income is more concentrated among a few top-tier banks when
   (a) underwriters' skill is more asymmetric, (b) the deal is harder to place,
   and (c) private compensation components are relatively more important.
3. More pay transparency (higher $$\lambda$$) increases total underwriting
   spreads and reduces the share of the highest-reputation bank(s) under
   centralized (joint-mandate) structures.
4. Sole mandates become *more* likely as bank effort substitutability increases
   (CES result, Figure 7, p. 3479).

## Datasets used

This is a theoretical paper. No dataset is used for estimation. The
application to banking syndicates references the following empirical
literature for operationalizing model parameters:

| Reference / proxy | Role in paper | Wiki page |
|---|---|---|
| Megginson and Weiss (1991) underwriter reputation (market-share rank) | Proxy for relative skill $$\alpha$$ in syndicate application | No page yet |
| Carter and Manaster (1990) tombstone rank | Alternative proxy for relative skill $$\alpha$$ | No page yet |
| Syndicate structure data (fraction of banks in top tier; HHI of fees) | Observable proxy for degree of delegation | No page yet |

No quantitative empirical exercise is conducted in the paper itself.

## Relation to prior work

The paper builds on several strands of the literature. Holmstrom (1982)
establishes that public contracts with team moral hazard can in principle
achieve first-best outcomes, which this paper uses as a conceptual
benchmark (p. 3482). Segal (1999) analyzes the principal's incentive to
deviate from an efficient trade profile when contract offers are privately
observed, and characterizes the optimal mechanism when agents' messages to
the principal can be made contingent on other agents' messages; this paper
complements that analysis by showing delegation can solve the commitment
problem (p. 3483). Aghion and Tirole (1997) study a double-sided moral
hazard problem where delegation encourages a single agent's effort; the
key difference here is that two agents' efforts are complements and the
principal makes no direct effort contribution, so delegation operates
through a different channel (p. 3484).

On pay transparency, Halac et al. (2021) analyze a model where the
principal can commit to the public distribution but keeps the realization
of pay packages private, ruling out bad equilibria; their model differs
in that the principal can commit to non-discriminatory pay, which she
cannot in the present paper (p. 3481). Cullen and Pakzad-Hurson (2023)
show that full pay transparency lowers pay inequality by reducing the
principal's bargaining power; this contrasts with the present paper's
finding that transparency raises pay levels and reduces inequality only
when efforts are highly substitutable (p. 3481). DeMarzo and Kaniel
(2023) build a model where agents have "keeping up with the Joneses"
(KUJ) preferences and private contracts worsen externalities; in
equilibrium agents' KUJ preferences result in less negative optimal
compensation on peer output, providing a rationale for "payment for luck"
(p. 3484).

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13496) if you are:
building a model of team contracting with private contracts; interested in
the formal proofs of the propositions (Appendix pp. 3486-3494 and Internet
Appendix); extending the theory to endogenous privacy, dynamic contracts, or
more than two agents; or calibrating the banking-syndicate predictions to
data (the comparative-statics section, IV.B, pp. 3475-3480 maps model
parameters to observables). The Internet Appendix derives CES equilibria in
full generality and provides robustness under non-passive beliefs.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(6), December 2025, pp.
3443-3497. DOI: 10.1111/jofi.13496. Copyright 2025 the American Finance
Association. This article is paywalled; no CC licence was found in Crossref
metadata. This distillation is **extract-only** under fair-use principles:
core results and equations reproduced for research commentary purposes.
Distilled by an LLM (claude-sonnet-4-6) on 2026-06-03. **Not
human-verified. Not independently reproduced.**

> Buffa, Andrea M., Qing Liu, and Lucy White. "Privacy and Team Incentives."
> *The Journal of Finance* 80, no. 6 (December 2025): 3443-3497.
> DOI: 10.1111/jofi.13496. Copyright 2025 the American Finance Association.


==============================================================================
# Pockets of Predictability (Replication): Cakici, Fieberg, Neumaier, Poddig & Zaremba (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/cakici-pockets-predictability-replication-2025/
# Distilled: Cakici et al. replicate Farmer-Schmidt-Timmermann (2023) and find a critical one-sided vs two-sided kernel lookahead error in the original code; correcting it collapses average integral R-squared by roughly 20-fold and invalidates most FST conclusions about exploitable pockets of predictability. J. Finance 80(6), December 2025, CC BY 4.0. Eight core results with source locators, datasets used, and the identification strategy.
# Tags: paper-summary, return-predictability, replication, market-timing, time-series, panel-regression, open-access, cc-by, peer-reviewed, data:wrds, data:ken-french
==============================================================================

**What this is.** The paper's core results, datasets, and identification
strategy: enough to know what it found without reading all 20 pages. To
replicate or extend it, read the full source at the
[original](https://doi.org/10.1111/jofi.13484) (open access).

## TL;DR

Farmer, Schmidt, and Timmermann (2023, FST) claimed that U.S. aggregate stock
market returns exhibit "pockets of predictability" identifiable ex ante via
one-sided kernel regressions. Cakici et al. audit the FST replication package
and find that the pocket-identification step in FST's code uses a two-sided
kernel, not the one-sided kernel described in the paper. A two-sided kernel
draws on data both before and after the forecast date, making the procedure
in-sample rather than out-of-sample and leaking future information into the
model. Correcting this single error reduces average integral R-squared by a
factor of roughly 20 (e.g., from 1.51-3.70% to 0.09-0.28% for daily
forecasts). The in-pocket vs out-of-pocket return predictability difference
largely disappears, and market-timing alphas become mostly insignificant.
Economic restrictions on forecasts offer partial improvement but cannot
restore the original conclusions.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%. Locators
point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Original (two-sided) code reproduces FST exactly**: in-pocket average integral R-squared ranges from 1.48% to 3.70% (daily), with strong in-pocket/out-of-pocket asymmetry | Table II Panel A, p. 3779; Figure 1 p. 3773 | Mean integral R² (daily): dp 1.51%, tbl 1.70%, tsp 2.92%, rvar 2.77% (two-sided kernel) |
| R2 | **Corrected (one-sided) code collapses predictability**: pockets become roughly 20x more frequent, 10x shorter, and far less predictable | Table II Panel B, p. 3779 | Mean integral R² (daily): dp 0.18%, tbl 0.09%, tsp 0.09%, rvar 0.28% (one-sided kernel) |
| R3 | **In-pocket CW t-statistics vanish with the one-sided kernel**: under the two-sided code in-pocket CW t-stats commonly exceed 3-4; under the corrected code they are insignificant for nearly all 27 model-predictor combinations | Table III Panel A.1 vs A.2, pp. 3781-3782 | Two-sided in-pocket CW (unrestricted): dp 3.00\*\*\*, tbl 4.75\*\*\*, tsp 3.04\*\*\*; one-sided in-pocket CW (unrestricted): dp -0.47, tbl 0.10, tsp -1.06 |
| R4 | **In-pocket alphas drop sharply**: unrestricted in-pocket annualised alphas fall from 0.76-6.38% (two-sided) to -0.44-2.51% (one-sided); only one out of nine individual/composite predictors exceeds 1% significance under the one-sided kernel | Table III Panel B.1 vs B.2, pp. 3782-3783 | Average Sharpe ratio drops from 0.71 (two-sided) to 0.44 (one-sided), below the prevailing-mean benchmark 0.46 |
| R5 | **Benchmark model beats kernel models out-of-pocket (two-sided code)**: out-of-pocket CW t-stats are significantly negative (at 10%) for most individual predictors under the original code | Table III Panel A.1, p. 3781 | Out-of-pocket CW: dp -1.62†, tbl -1.33†, tsp -1.52†, rvar -1.77†† (two-sided, unrestricted) |
| R6 | **Alternative bandwidth robustness: corrected code always fails to identify in-pocket predictability** across 2-, 2.5-, and 3-year estimation windows and 6-, 12-, 18-month SED windows; not a single significant in-pocket CW t-stat in Panel B | Table IV Panel B, pp. 3785-3786 | All Panel B in-pocket CW entries insignificant across all bandwidth/window combinations |
| R7 | **Monthly data confirms the result**: monthly in-pocket CW t-stats are strong under the two-sided kernel (e.g., tbl 3.55\*\*\*, tsp 2.44\*\*\*) but mostly insignificant under the one-sided kernel | Table V Panels A and B, p. 3788 | One-sided in-pocket monthly CW: dp 0.90, tbl 1.22, tsp 0.57, rvar 1.01 |
| R8 | **Partial exception: factor returns (SMB, HML) retain some time-varying predictability** even with the corrected one-sided kernel, though weaker than FST documented with the two-sided approach | §II.E, p. 3788-3789 (Internet Appendix Section IV) | Qualitative finding: significant CW stats and market-timing gains persist for factor portfolios; aggregate equity market is the null result |

**Overall (paper's conclusion).** The FST pocket-of-predictability evidence
is an artefact of in-sample kernel estimation. Once the identification step
is restricted to information available before the forecast date, the pockets
shrink to one-day artefacts, predictability inside and outside pockets
becomes statistically indistinguishable, and market-timing strategies based
on the pockets offer no reliable abnormal returns.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CRSP U.S. stock market excess return (daily and monthly, 1926-2016) | Dependent variable: aggregate market excess return (CRSP return minus short T-bill) | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| Dividend-price ratio (dp), 1926-2016 | Predictor variable (sourced from FST replication package) | no page yet |
| 3-month T-bill rate (tbl), 1954-2016 | Predictor variable (sourced from FST replication package) | no page yet |
| Term spread (tsp), 1962-2016 | Predictor variable (sourced from FST replication package) | no page yet |
| Realized variance (rvar), 1927-2016 | Predictor variable (sourced from FST replication package) | no page yet |
| Ken French Data Library (SMB, HML factor returns) | Used in Section II.E robustness for factor-level predictability | [Ken French](/wiki/datasets/ken-french/) |

Study period: 1926-2016 (predictor-dependent; see Table I, p. 3777). Data
sourced directly from the FST replication package ("Replication-code
20190881.zip", Journal of Finance website).

## Theory / model

This paper has no original structural economic model. It is a methodological
audit and replication of Farmer, Schmidt, and Timmermann (2023, FST). The
theoretical object under scrutiny is the claim that aggregate equity market
return predictability is time-varying and can be identified ex ante using
one-sided kernel regressions. The testable hypothesis is:

- **Null**: market-timing strategies built on FST's "pockets" offer no reliable
  abnormal returns once a correctly out-of-sample pocket-identification kernel
  is used.
- **Identification**: the two-framework comparison is the entire identification
  strategy. Every FST analysis is run twice with all other parameters held
  fixed; only the kernel type in the second estimation stage differs. Any
  difference in results is attributed to the kernel type (one-sided vs
  two-sided), since that is the sole deviation from the FST code.

**The FST return prediction model** (eq. 1, p. 3774):

$$
r_{t+1} = x_t' \beta_t + \epsilon_{t+1} \tag{1}
$$

- $$r_{t+1}$$ is the excess U.S. stock market return
- $$x_t$$ is a vector of predictor variables (dp, tbl, tsp, rvar)
- $$\beta_t$$ are time-varying regression coefficients
- $$\sigma_t^2 = E[\epsilon_{t+1}^2 \mid x_t]$$ allows for conditional heteroskedasticity

The $$\beta_t$$ are estimated by the local constant model (eq. 2, p. 3774):

$$
\hat{\beta}_t = \operatorname*{arg\,min}_{\beta_0} \sum_{s=1}^{T} K_{hT}(s-t) \cdot [r_{s+1} - x_s' \beta_0]^2 \tag{2}
$$

with kernel weights $$K_{hT}(u) = K(u/hT)/(hT)$$ and bandwidth $$h$$. FST use a
2.5-year bandwidth in this step with a one-sided Epanechnikov kernel (eq. 3,
p. 3775):

$$
K(u) = \tfrac{3}{2}(1 - u^2) \cdot \mathbf{1}\{-1 < u < 0\} \tag{3}
$$

Only data from before time $$t$$ receives positive weight under the one-sided
kernel, making the $$\beta_t$$ estimation genuinely out-of-sample. The discrepancy
arises in the second stage.

## Method

The method builds on `kernel-regression` for both the return-prediction
estimation and the pocket-identification step, and on `time-series-forecasting`
for evaluating out-of-sample performance against the prevailing-mean benchmark.
The paper applies these techniques, it does not propose a new one.

**Squared error differential (SED)** (eq. 4, p. 3775) measures whether the
kernel model outperforms the prevailing-mean benchmark at each date $$t$$:

$$
\text{SED}_t = (r_t - \bar{r}_{t|t-1})^2 - (r_t - \hat{r}_{t|t-1})^2 \tag{4}
$$

- $$\bar{r}_{t|t-1}$$ is the prevailing-mean forecast
- $$\hat{r}_{t|t-1}$$ is the kernel model forecast
- Positive $$\text{SED}_t$$ means the kernel model has smaller forecast error that period

**Pocket identification** (eq. 5, p. 3775): a pocket begins when the fitted
SED trend is positive:

$$
\widehat{\text{SED}}_t = \gamma_{0,t} + \gamma_{1,t} \cdot t > 0 \tag{5}
$$

- $$\gamma_{0,t}$$ and $$\gamma_{1,t}$$ should be estimated with a one-sided Epanechnikov kernel and one-year bandwidth
- FST's published code instead uses a two-sided kernel with a 24-month symmetric window (12 months before and 12 months after day $$t$$), so the identification draws on data that is unavailable at forecast time

**Two-framework comparison design**: the paper runs the complete FST analysis
twice, in parallel, changing only this kernel choice. Panel A results use the
original (two-sided) code; Panel B results use the corrected (one-sided) code.
All other parameters, bandwidth choices, predictor series, and performance
metrics are identical. This clean design means the contrast of Panel A vs
Panel B isolates the kernel-type effect.

**Performance metrics** (§II.B, pp. 3780-3783):
- Clark-West (2007) t-statistic comparing kernel-model forecasts to the
  prevailing-mean benchmark
- Annualised alpha from a stock/T-bill timing strategy (holding stocks when the
  kernel model predicts positive returns, T-bills otherwise), with Newey-West
  (1987) t-statistics
- Annualised Sharpe ratio of the timing strategy
- Three forecast restriction variants following Campbell and Thompson (2008):
  unrestricted, non-negative excess return forecasts only, and sign restrictions
  on both forecasts and slope coefficients

## Empirical specifications

All regressions use daily U.S. excess stock market returns as the dependent
variable unless noted (monthly robustness in Table V). The sample varies by
predictor (Table I, p. 3777): dp starts November 5, 1926; tbl starts January
4, 1954; tsp starts January 2, 1962; rvar starts January 15, 1927; all end
December 2016.

**Constant-coefficient replication regressions (Table I, p. 3777).**
Univariate OLS of daily excess return on each lagged predictor, separately
for full sample, in-pocket, and out-of-pocket subsamples. Slope coefficients
and Newey-West (1987) adjusted t-statistics and R-squared reported. The pocket
partition is the only thing that differs between Panel A and Panel B in this
table; the regression itself is identical.

**Pocket statistics (Table II, p. 3779).**
For each predictor and kernel type, counts the number of pockets, their
fraction of sample, duration (min/mean/max days), and integral R-squared
(min/mean/max). Integral R-squared $$IR^2$$ is computed as in FST (p. 1289):
the average within-pocket R-squared, weighting by pocket length. Results
reported separately for daily and monthly data.

**Clark-West prediction performance (Table III, pp. 3781-3782).**
CW t-statistic comparing kernel model to the prevailing-mean benchmark
$$\bar{r}_{t+1} = (1/t)\sum_{s=1}^{t} r_s$$. Reported separately for:
- full sample, in-pocket subperiod, out-of-pocket subperiod
- Panel A.1 (in-sample / two-sided kernel), Panel A.2 (one-sided kernel)
- nine predictor/composite specifications (dp, tbl, tsp, rvar, pc, mv,
  comb1, comb2, comb3)
- three forecast restriction variants
Significance under one-tailed test (positive direction) for alphas; two-tailed
for CW tests.

**Economic significance (Table III Panel B, pp. 3782-3783).**
Annualised alpha, Newey-West t-statistic, and Sharpe ratio of the asset
allocation strategy (stocks when forecast is positive, T-bills otherwise).
Reported for the same nine predictor/composite specifications and three
restriction variants, separately for the two kernel approaches.

**Bandwidth robustness (Table IV, pp. 3785-3786).**
CW t-statistics for coefficient estimation windows of 2-, 2.5-, 3-year and
SED estimation windows of 6-month, 12-month, 15-month. Panel A uses the
two-sided kernel; Panel B uses the corrected one-sided kernel. Nine predictor
and composite specifications, daily data.

**Monthly data robustness (Table V, p. 3788).**
Reproduces Table III (Panel A and Panel B) using monthly returns, following
FST Table VII Panel A. Estimation period is 2.5 years for monthly forecasts.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13484) (open access) if you
are: auditing the FST replication package directly; extending the
kernel-regression methodology to other predictors or markets; or assessing
whether the partial improvements from economic restrictions (sign constraints
on slope coefficients and forecasts) restore the FST conclusions. The
locators above point to the exact tables. For "what did this paper find,"
the table above is sufficient.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(6), December 2025,
pp. 3771-3790. This distillation was extracted by an LLM on 2026-05-31 and
is **not human-verified or independently reproduced**. The article is open
access under CC BY 4.0; PDF mirroring is permitted by the licence but the
PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Cakici, Nusret, Christian Fieberg, Tobias
> Neumaier, Thorsten Poddig, and Adam Zaremba. "Pockets of Predictability:
> A Replication." *The Journal of Finance* 80, no. 6 (December 2025):
> 3771-3790. DOI: 10.1111/jofi.13484. © 2025 The Author(s). Licensed under
> [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Social Security and Trends in Wealth Inequality: Catherine, Miller & Sarin (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/catherine-social-security-trends-wealth-2025/
# Distilled: When Social Security wealth is properly included, top wealth shares in the United States have not meaningfully increased since 1989, overturning the finding of large inequality growth based on marketable-wealth-only measures. Social Security grew from $7.2 trillion in 1989 to $40.6 trillion in 2019 and now represents nearly 50% of the wealth of the bottom 90%. J. Finance 2025, paywalled. Eight core results with source locators, datasets used, the model, and the empirical method.
# Tags: paper-summary, household-finance, wealth-inequality, social-security
==============================================================================

**What this is.** The paper's core results, the measurement model (Social Security wealth valuation), and the empirical method (SCF wealth-share construction plus earnings simulation): enough to understand what it found and how, without reading all 35 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1111/jofi.13440).

## TL;DR

Recent work by Saez and Zucman (2016) and others documents large increases in U.S. wealth inequality over the past three decades based on measures that exclude Social Security. This paper shows that when Social Security is properly included, top wealth shares have not meaningfully changed since 1989. Social Security wealth grew from $7.2 trillion in 1989 to $40.6 trillion in 2019 and now constitutes nearly 50% of the total wealth of the bottom 90%. The result is robust to potential benefit cuts, liquidity discounts, heterogeneous discount rates, and alternative accrued-benefit definitions. The main driver of Social Security's growth is falling interest rates, which disproportionately raised the present value of the long-duration, wage-indexed cash flows that low- and middle-class households hold. The paper builds on Feldstein (1974) and Feldstein (1976), who showed that total wealth inclusive of Social Security is more equally distributed than marketable wealth alone, and extends that insight to document how the difference has grown over 30 years. It also directly challenges the finding in Greenwald et al. (2021) that falling interest rates drove rising marketable wealth inequality, showing that once Social Security's own long-duration assets are included the inequality trend is substantially reversed.

## Core results

Magnitudes and significance are as reported. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Top 10% wealth share rises only 1.0 pp (risk-free) or 1.7 pp (risk-adjusted) once Social Security is included**, vs. 9.5 pp for marketable wealth alone | Figure 3 Panel A, p. 1512; Table III Panel A, p. 1518 | Top 10% share: marketable wealth +9.5 pp; risk-free +1.0 pp; risk-adjusted +1.7 pp (1989-2019) |
| R2 | **Top 1% wealth share rises only 1.5 pp (risk-free) or 1.9 pp (risk-adjusted)**, vs. 6.4 pp for marketable wealth | Figure 3 Panel B, p. 1512; Table III Panel A, p. 1518 | Top 1% share: marketable wealth +6.4 pp; risk-free +1.5 pp; risk-adjusted +1.9 pp |
| R3 | **Aggregate Social Security wealth grew from $7.2 trillion to $40.6 trillion (1989-2019)**, a 5.6x increase; largest contributor is the falling yield curve (45.8-48.3% of log growth) | Figure 2, p. 1511; Table I, p. 1515 | Log total growth 1.746; yield-curve change accounts for 0.843 (risk-free) of that log growth |
| R4 | **Social Security wealth rose from 26.0% to 49.8% of the total wealth of the bottom 90%** between 1989 and 2019 | Figure 5, p. 1517 | Bottom 90% SS share: 26.0% (1989) to 49.8% (2019) under risk-adjusted valuation |
| R5 | **Even under the most conservative policy-risk scenario (40% across-the-board benefit cut)**, top 10% and top 1% shares rise by only 4.3 and 3.4 pp, less than half the marketable-wealth trend | Figure 7, p. 1520; Table III Panel B, p. 1518 | Top 10% +4.3 pp, top 1% +3.4 pp under high-cost benefit cut vs. marketable +9.5/+6.4 pp |
| R6 | **A 3% liquidity discount halves aggregate SS wealth** but the attenuation of inequality trends persists: top 10% +4.7 pp, top 1% +3.6 pp | Figure 8, p. 1521; Table III Panel C, p. 1518 | Top 10% +4.7 pp, top 1% +3.6 pp even after 3% liquidity premium discount |
| R7 | **Heterogeneous discount rates** (borrowing rates for constrained households) raise the top 10% and top 1% increases to 4.2 and 3.2 pp, still approximately half the marketable-wealth trend | Figure 11, p. 1525; Table III Panel C | Top 10% +4.2 pp, top 1% +3.2 pp under heterogeneous private discounting |
| R8 | **Under the NPV wealth concept** (benefits minus future taxes), the top 10% and top 1% shares actually declined by 3.0 and 0.2 pp between 1989 and 2019, because NPV SS wealth was low or negative for young workers in the high-rate 1989 environment | Figure 12, p. 1526 | Top 10% NPV change: -3.0 pp; top 1% NPV change: -0.2 pp |

**Overall (paper's conclusion).** Prior studies find large increases in U.S. wealth inequality based on marketable wealth measures. When Social Security is incorporated, top wealth shares have not increased since 1989. The top wealth estimates may still be overstated because the paper excludes programs like disability insurance and Medicare, which accrue disproportionately to the bottom of the wealth distribution. Public transfer programs make the U.S. economy more progressive, and inequality estimates need to reflect this (p. 1529).

## Theory / model

The paper has no formal equilibrium model; instead it derives a valuation framework for Social Security wealth and tests how its inclusion changes measured wealth inequality. The tested hypotheses are:

1. When Social Security's accrued benefit value is included in household wealth, top wealth shares are substantially lower and their trend since 1989 is much smaller than estimates based on marketable wealth alone.
2. Social Security grew disproportionately relative to marketable wealth primarily because of falling interest rates, which inflate the present value of its long-duration wage-indexed cash flows that constitute the bulk of low- and middle-class wealth.

**Retiree Social Security wealth** (p. 1505, equation 1). For a retiree observed in year $$t$$, Social Security wealth $$S_{it}$$ is the present value of future nominal benefits $$B_{it}$$ adjusted for CPI-indexed growth and survival probabilities $$m_{itk}$$:

$$
S_{it} = \sum_{s=t}^{T} \left( \prod_{k=t}^{s-1} (1 - m_{itk}) \right) \frac{B_{it}}{(1 + r_{ts})^{s-t}} \frac{\mathbb{E}[P_s]}{P_t} \tag{1}
$$

**Accrued Social Security wealth for workers** (p. 1507, equation 7). The accrued benefits concept values benefits proportional to past tax contributions:

$$
S_{it} = \frac{\text{Past Taxes}_{it}}{\text{Past Taxes}_{it} + \text{Future Taxes}_{it}} \sum_{s=t+1}^{T} \frac{\mathbb{E}[B_{is}]}{(1 + r_{ts})^{s-t}} \tag{7}
$$

where future and past taxes are present-valued at the appropriate discount rate (equations 8-9, p. 1507). Past taxes are grossed up using the realized return on 30-year Treasury bonds to convert them to present-value terms.

**Alternative NPV concept** (p. 1508, equation 10):

$$
S_{it} = \sum_{s=t+1}^{T} \frac{\mathbb{E}[B_{is} - T_{is}]}{(1 + r_{ts})^{s-t}} \tag{10}
$$

This values Social Security as the stream of net expected transfers, which can be negative for young workers in high-rate environments.

## Method

**Earnings simulation for workers.** Earnings are modeled as the product of an aggregate wage index $$L_{1,t}$$ and an idiosyncratic component $$L_{2,it}$$ (p. 1505, equation 2):

$$
L_{it} = L_{1,t} \cdot L_{2,it} \tag{2}
$$

The idiosyncratic component evolves via a rich income process (equations 3a-3g, pp. 1505-1506) calibrated to Guvenen et al. (2021). The process has a persistent component $$z_t^i$$ following an AR(1):

$$
z_t^i = \rho z_{t-1}^i + \eta_t^i \tag{3b}
$$

with innovations from a mixture of normals, transitory shocks from a second mixture, and nonemployment shocks with exponentially distributed duration. 10,000 earnings paths are simulated per survey year-gender-age cell and matched to SCF respondents.

**Indexed taxable earnings and benefits** (equations 4-6, pp. 1506-1507). Payroll taxes are 10.6% of earnings up to the Social Security wage base $$SSWB_t$$. The Average Indexed Monthly Earnings (AIYE) is the average of the best 35 years of indexed earnings. Benefits are a piecewise-linear concave function of AIYE reflecting Social Security's progressive design:

$$
B_{it} = \frac{P_{t}}{P_{t_i+60}} \begin{cases} 0.9 \cdot \text{AIYE}_i & \text{if } \text{AIYE}_i < b_{1,c_i} \\ 0.9 \cdot b_{1,c_i} + 0.32(\text{AIYE}_i - b_{1,c_i}) & \text{if } b_{1,c_i} \le \text{AIYE}_i < b_{2,c_i} \\ 0.9 \cdot b_{1,c_i} + 0.32(b_{2,c_i} - b_{1,c_i}) + 0.15(\text{AIYE}_i - b_{2,c_i}) & \text{if } b_{2,c_i} \le \text{AIYE}_i \end{cases} \tag{6}
$$

**Risk adjustment for macroeconomic risk.** Because Social Security benefits are wage-indexed, they are exposed to aggregate labor market risk. Assuming cointegration of labor and stock markets (Benzoni, Collin-Dufresne, and Goldstein (2007)), the market beta of a cash flow proportional to $$L_{1,t+n}$$ is (p. 1513, equation 14):

$$
\beta_t^{L_{1,n}} = \left(1 - \frac{\phi}{\kappa}\right)\left(1 - e^{-\kappa n}\right) \tag{14}
$$

and the expected return on this cash flow under no-arbitrage is (equation 15):

$$
\mathbb{E}_t\left[r_t^{L_{1,n}}\right] = \beta_t^{L_{1,n}}(\mu - r) + r \tag{15}
$$

Parameters are calibrated at $$\kappa = 0.16$$, $$\phi = 0.08$$ (from Benzoni, Collin-Dufresne, and Goldstein (2007)), and equity premium $$\mu - r = 0.06$$. The risk-adjusted discount factor for a cash flow proportional to $$L_{1,n}$$ paid in year $$k$$ is (equation 16, p. 1513):

$$
\chi_{t,n,k} \approx \left[\prod_{s=t}^{n} \left(1 + \beta_s^{L_{1,n}}(\mu - r) + r_{ts}\right) \prod_{s=n+1}^{k}(1 + r_{ts})\right]^{-1} \tag{16}
$$

**Heterogeneous discount rates** (equations 17-18, p. 1523). Unconstrained households (no debt, liquid or illiquid assets above thresholds) are discounted at the risk-adjusted forward rate. Constrained households face their opportunity cost of debt, estimated via Tobit regressions of balance-weighted interest rate spreads by income quintile, age, and year:

$$
f_{h,a,q,t}^{\text{constrained}} = \begin{cases} f_{t,h}^{\text{risk-adj}} & \text{with prob. } p_{a+h-1,q,t} \\ f_{t,h}^{\text{risk-free}} + s_{a+h-1,q,t} & \text{with prob. } 1 - p_{a+h-1,q,t} \end{cases} \tag{18}
$$

## Empirical specifications

**Top wealth share construction.** Marketable wealth shares are constructed from the SCF using the net worth variable (assets minus liabilities), supplemented by Forbes 400 data for the top 0.01% following Saez and Zucman (2016), and augmented with DFA data on defined benefit pension obligations. Total wealth shares add the simulated Social Security wealth $$S_{it}$$ to each household's net worth. The share of total wealth held by the top 10% and top 1% is computed after ranking by the relevant total-wealth concept (not by marketable wealth, so the ranked groups can differ by specification).

**Decomposition of Social Security wealth growth** (Table I, p. 1515). The log per-capita Social Security wealth change is additively decomposed by:
1. Change in the yield curve (holding age distribution, survival, and policy at 1989 values)
2. Shift in the age distribution (holding yield curve at 2019 values)
3. Change in life expectancy
4. Social Security expansion and other factors (scope of taxable earnings, benefit formulas)

Each contribution is identified by sequential substitution of 2019 for 1989 parameters.

**Robustness (Table III, p. 1518):** Results are checked under (i) benefit cuts calibrated to SSA actuarial cost scenarios (low, intermediate, high), (ii) tax hikes on bottom 90% or bottom 99%, (iii) liquidity premiums of 1-3%, (iv) heterogeneous household discount rates, (v) declining wage growth, (vi) alternative accrued-benefits definitions (pro-rata by age, stop-working), and (vii) the NPV valuation. The headline attenuation of inequality trends is unchanged across all specifications.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Survey of Consumer Finances (SCF), triennial 1989-2019 | Marketable wealth shares; Social Security wealth for retirees; earnings-match base for workers | [no page yet] |
| Forbes 400 list | Supplement to extend wealth distribution to the top 0.01% following Saez and Zucman (2016) | [no page yet] |
| Distributional Financial Accounts (DFA), Federal Reserve Board | Aggregate value of defined benefit pension obligations by wealth group | [DFA](/wiki/datasets/dfa/) |
| Federal Reserve zero-coupon yield curve (Treasury notes, up to 30 years) | Discount rates for Social Security cash flows; forward rate extrapolation beyond 30 years | [no page yet] |
| SSA Annual Reports and actuarial projections | Calibration of Social Security parameters (bend points, wage base, benefit formulas, cost scenarios) | [no page yet] |
| Human Mortality Database (HMD), 1989-2017 | Survival probabilities by gender, calibrated and adjusted for income-based life expectancy differences | [no page yet] |
| Health Inequality Project (HIP) | Income-based life expectancy differences used to adjust survival probabilities | [no page yet] |
| Guvenen et al. (2021) income process estimates | Calibration of idiosyncratic earnings dynamics for the 10,000-path simulation | [no page yet] |

Sample: SCF waves 1989, 1992, 1995, 1998, 2001, 2004, 2007, 2010, 2013, 2016, 2019 (triennial). Social Security wealth simulated using 10,000 earnings paths per survey year-gender-age cell.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13440) if you are: (i) replicating wealth inequality estimates that include public programs; (ii) assessing how changes in interest rates affect the distribution of total household wealth; (iii) designing or evaluating Social Security reform scenarios (the robustness section covers benefit cuts, tax hikes, liquidity premiums, and heterogeneous discounting); or (iv) comparing U.S. wealth inequality across studies that use different wealth concepts (Figure 14 directly overlays this paper's SS-inclusive estimates on SCF, Saez and Zucman (2016), and Smith, Zidar, and Zwick (2020) series). Table III and Figures 7-13 contain the full robustness battery.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(3). This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The paper is paywalled; only extracts are reproduced here consistent with fair use.

> Catherine, Sylvain, Max Miller, and Natasha Sarin. "Social Security and Trends in Wealth Inequality." *The Journal of Finance* 80, no. 3 (June 2025): 1497-1531. DOI: 10.1111/jofi.13440. © 2025 the American Finance Association.


==============================================================================
# Too Much, Too Soon, for Too Long: Chemla, Rivera & Shi (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/chemla-too-much-too-soon-2025/
# Distilled: In a general equilibrium model with dynamic moral hazard and endogenous outside options, competitive executive compensation is inefficiently high, front-loaded, and associated with excessive managerial tenure. J. Finance 2025, CC BY 4.0. Six core results with source locators, the model, and the method.
# Tags: paper-summary, executive-compensation, corporate-governance, agency, moral-hazard
==============================================================================

**What this is.** The paper's core results, the model it builds on (dynamic principal-agent contracting embedded in a general equilibrium with endogenous outside options), and its main propositions: enough to know what it found and how, without reading all 50 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13470).

## TL;DR

The paper embeds a continuous-time dynamic moral hazard problem (in the spirit of DeMarzo and Sannikov (2006)) into a general equilibrium economy where outside options for managers and liquidation values for firms are endogenously determined by equilibrium compensation. Firms compete for managers by promising deferred pay ("carrots") backed by termination threats ("sticks"). The framework builds on the two-period binary setup of Bolton and Scharfstein (1990) and the continuous-time limit of Biais et al. (2007). The welfare criterion follows Dicks (2012) (maximize shareholder value). In contrast to static governance externality models such as Acharya and Volpin (2010) and frictionless assortment theories of executive compensation such as Gabaix and Landier (2008), the central finding here is that competitive markets generate overcompensation even without rent extraction. Evidence that executive pay rises sharply when CEOs move to a new firm (Falato, Li, and Milbourn (2015)) motivates the focus on endogenous outside options. The resulting competitive equilibrium is inefficient: firms fail to internalize the effect of their compensation packages on the outside options of managers at other firms. This "compensation externality" leads to executives being paid too much (overcompensation), too soon (insufficient deferral), and keeping their jobs for too long (excessively low turnover), while the associated capital structure features excessively low credit line limits and long-term debt.

## Core results

Magnitudes and significance are as reported. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Equilibrium compensation exceeds the social optimum** (overcompensation): firms fail to internalize that raising pay increases the manager's outside option, reducing the effectiveness of termination for all other firms | Corollary 2, p. 2940; Figure 4 Panel B, p. 2946 | In the baseline calibration, equilibrium initial compensation W0\* is approximately 15.4 vs the planner's W0^p = kA = 5.3; equilibrium compensation roughly 3x the planner level |
| R2 | **Insufficient deferral**: overcompensation is front-loaded; managers receive their first payment sooner in equilibrium than under the social optimum | Proposition 4, eq. (19)-(20), p. 2941; p. 2946 | Front-loading proxy S\*(W0\*) = 0.78 > 0.65 = S^p(W0^p); manager paid roughly 20% sooner in equilibrium than optimal |
| R3 | **Excessively long tenure**: overcompensation causes the manager's continuation value to drift upward faster, reducing the likelihood of hitting the termination threshold | Proposition 4, eq. (20), p. 2941; p. 2946 | T\*(W0\*) = 0.17 < 0.21 = T^p(W0^p); equilibrium forced-turnover rate 2.2% per year vs the planner's higher rate |
| R4 | **Capital structure distortion**: the equilibrium compensation contract is implemented with excessively low credit line limits and long-term debt relative to the social optimum | Proposition 5, p. 2942 | CL\* < CL^p and D\* < D^p; debt and credit lines are lower in equilibrium because high, front-loaded compensation requires low debt instruments to remain incentive-compatible |
| R5 | **Managerial bargaining power amplifies overcompensation**: as managers capture a larger share of the surplus, equilibrium compensation more than doubles and shareholder value falls disproportionately | Lemma 6, eq. (23), p. 2955; Figure 9, p. 2955 | Raising manager bargaining power beta from 0 to 0.25 more than doubles W0\*; shareholder value decreases disproportionately because higher pay raises manager outside options, further undermining termination threats |
| R6 | **Moral hazard severity and cash-flow volatility amplify the distortions**: industries with higher lambda or sigma see a larger gap between equilibrium and optimal compensation, more front-loading, and less turnover | Section IV.C, Figures 5-7, pp. 2947-2950 | The shareholder value gap F(W0^p; R^p, L^p) - F(W0\*; R\*, L\*) is increasing in lambda and sigma; compensation W0\* rises steeply in lambda while the planner's W0^p stays near kA |

**Overall (paper's conclusion).** The compensation externality arises because firms are price-takers with respect to the equilibrium outside option: when an individual firm raises pay to maximize its own shareholders' value, it inadvertently increases the outside option for all managers, making termination less effective as an incentive device across the economy. The resulting equilibrium is inefficient even when firms hold all bargaining power and are well-intentioned. A benevolent planner can achieve Pareto improvements by coordinating future compensation down, restoring termination effectiveness without harming managers.

## Theory / model

The model has two parts: an illustrative two-period binary setup (Section I, pp. 2926-2930) and the full continuous-time infinite-horizon model (Section II, pp. 2931-2937).

**Two-period setup (Section I).** The economy has a continuum of risk-neutral firms and managers. Each period, a project generates a binary cash flow: high ($$y > 0$$) with probability $$p$$ or zero with probability $$1-p$$. The manager privately observes realized cash flows and can divert them, receiving a fraction $$\lambda \in (0, 1]$$ of diverted funds. Limited liability requires all compensation payments to be nonneg.

The one-period (static) optimal contract is $$\Gamma^S = \{\tilde{x}, c_H, c_L\}$$ where the firm pays $$c_H = \lambda y$$ if reported cash flow is high and $$c_L = 0$$ otherwise. Outside values satisfy (p. 2927, eq. (1)):

$$
R = \tilde{x}\delta\lambda\mu - \kappa_A \quad \text{and} \quad L = \tilde{x}(1-\lambda)\mu - \kappa_P. \tag{1}
$$

For the two-period dynamic contract $$\Gamma^D = \{x, c\}$$ (continuation probability in the low state $$x$$, period-1 high-state compensation $$c$$), the principal maximizes shareholder value subject to the incentive-compatibility constraint (p. 2929, eq. (2)):

$$
\text{IC-1}: \quad c + \delta\lambda\mu \geq x\delta\lambda\mu + (1 - x)R + \lambda y. \tag{2}
$$

This IC shows that termination threats (setting $$x < 1$$) reduce the cost of incentive provision in the high state by an amount $$\delta\lambda\mu - R$$ that depends critically on the manager's outside option $$R$$. **When $$R$$ is high (outside options are lucrative), termination becomes less effective**, so firms must compensate more.

**Lemma 1** (p. 2929): When agent termination cost $$\kappa_A > \tfrac{1-p}{p}\kappa_P$$, agents are terminated after poor performance and expected compensation is $$(\delta + \delta^2)\lambda\mu - \delta\kappa_A$$; shareholder value is $$2(1-\lambda)\mu + p\kappa_A - (1-p)\kappa_P$$.

**Lemma 2** (p. 2930): If moral hazard is sufficiently severe ($$p\delta\lambda > (1-p)(1-\lambda)$$), the equilibrium features overcompensation. The planner sets outside options to zero by shutting down new matches; shareholders gain up to $$\Delta\mu$$ where $$\Delta \equiv p\delta\lambda - (1-p)(1-\lambda)$$.

**Full continuous-time model (Section II).** Time is continuous and infinite, $$t \in [0, \infty)$$. Cash flows follow

$$
dY_t = \mu\, dt + \sigma\, dB_t,
$$

where $$B_t$$ is a standard Brownian motion. The manager privately observes cumulative cash flows $$Y = \{Y_t\}_{t \geq 0}$$ while the firm relies on reported cash flows $$\hat{Y}$$. The manager can divert $$dY_t - d\hat{Y}_t$$ and receive a fraction $$\lambda$$ of diverted funds. Firms discount at rate $$r$$; managers discount at $$\gamma > r$$ (managers are impatient).

The firm's initial value under contract $$\Gamma = (C, \tau)$$ (cumulative compensation process and termination time) is (p. 2932):

$$
F_0(\hat{Y}; \Gamma) \equiv \mathbb{E}\!\left[\int_0^\tau e^{-rt}(d\hat{Y}_t - dC_t) + e^{-r\tau}L\right].
$$

The manager's initial value is:

$$
W_0(\hat{Y}; \Gamma) \equiv \mathbb{E}\!\left[\int_0^\tau e^{-\gamma t}\!\left(dC_t + \lambda(dY_t - d\hat{Y}_t)\right) + e^{-\gamma\tau}R\right].
$$

The optimal contract solves (p. 2933, eqs. (3)-(5)):

$$
\max_{W_0, \Gamma} F_0(Y; \Gamma) \tag{3}
$$

subject to the promise-keeping constraint $$W_0(Y; \Gamma) = W_0$$ and the incentive-compatibility constraint $$W_t(Y; \Gamma) \geq W_t(\hat{Y}; \Gamma)$$ for all $$t \in [0, \tau]$$.

The equilibrium conditions pin down the endogenous outside option $$R^*$$ and liquidation value $$L^*$$ (p. 2934, eqs. (6)-(7)):

$$
R^* = W_0^* - \kappa_A, \tag{6}
$$

$$
L^* = F_0^* - \kappa_P. \tag{7}
$$

**Proposition 1** (p. 2937, eq. (16)): Under Assumption 1, the unique equilibrium compensation level $$W_0^*$$ satisfies

$$
F'(W_0^*; R^*, L^*) = 0. \tag{16}
$$

The firm maximizes its value function at the interior point where the marginal value of promised compensation is zero, taking outside options as given.

**Proposition 3** (Social Optimum, p. 2939, eq. (18)): The socially optimal compensation $$W_0^p$$ satisfies

$$
F'(W_0^p; R^p, L^p) + \frac{\partial}{\partial R}F(W_0^p; R^p, L^p) \leq 0. \tag{18}
$$

The second term, $$\frac{\partial}{\partial R}F < 0$$, is the general equilibrium effect: a $1 increase in compensation raises managers' outside options by $1, reducing firm value. Firms in equilibrium set $$F'(W_0^*) = 0$$, ignoring this negative externality, so $$W_0^* > W_0^p$$ (Corollary 2).

## Method

This is a theory paper. The solution method combines:

1. **Optimal contract characterization via the HJB/ODE.** Following DeMarzo and Sannikov (2006), the firm's value function $$F(W; R, L)$$ is characterized by an ODE (Corollary 1, p. 2936, eqs. (11)-(13)):

$$
rF(W; R, L) = \mu + \gamma W F'(W; R, L) + \tfrac{1}{2}\lambda^2\sigma^2 F''(W; R, L), \quad R \leq W < \bar{W}, \tag{11}
$$

$$
F'(W; R, L) = -1, \quad W \geq \bar{W}, \tag{12}
$$

with boundary conditions $$F(R; R, L) = L$$ and $$rF(\bar{W}; R, L) = \mu - \gamma\bar{W}$$. The optimal contract (Lemma 3, p. 2935) specifies:

- **(i) Pay-for-performance:** $$dW_t = \gamma W_t\, dt - dC_t + \lambda(dY_t - \mu\, dt)$$.
- **(ii) Deferral:** payments only when $$W_t \geq \bar{W}$$.
- **(iii) Termination:** $$\tau = \min\{t \mid W_t = R\}$$.

The ODE is solved numerically; existence and uniqueness of the equilibrium are established analytically (Propositions 1, Appendix D-E, pp. 2961-2965).

2. **Equilibrium fixed-point.** The equilibrium $$(R^*, L^*)$$ is found as a fixed point of eqs. (6) and (7) given the solution to the firm's contracting problem. Proposition 1 shows that the FOC $$F'(W_0^*) = 0$$ pins down the unique interior equilibrium.

3. **Calibration.** Parameters are calibrated to data moments (Table I, p. 2943): $$r = 0.04$$ (annual interest rate), $$\gamma = 0.09$$ (manager discount rate), $$\mu = 10$$ (normalization), $$\sigma = 9$$ (matching 10-15% fraction with operating losses), $$\lambda = 0.29$$ (moral hazard; Ward (2023)), $$\kappa_P = 15$$ (6% CEO replacement cost; Taylor (2010)), $$\kappa_A = 5.3$$ (2.2% forced turnover rate; Taylor (2010)).

The model is extended in Section V to incorporate: (i) noncompete clauses (Lemma 4), (ii) endogenous termination costs via a search framework (Lemma 5), (iii) Nash bargaining (Lemma 6, eq. (22)-(23)), and (iv) forward-looking firm liquidation values (Lemma 7).

## Empirical specifications

This is a pure theory and calibration paper with no regression analysis. There are no panel regressions, no instrumental variable designs, and no event studies. The "empirical" content consists of:

- **Calibration targets** matched to observed moments: fraction of firms with operating losses, annual interest rates, manager discount rate (Ward (2023), Chen et al. (2023)), forcing-turnover rate (Taylor (2010), Eisfeldt and Kuhnen (2013), Jenter and Kanaan (2015)), and CEO replacement costs (Taylor (2010)).
- **Quantitative comparative statics** (Figures 5-9, pp. 2947-2955): the model computes equilibrium versus planner outcomes as each parameter ($$\lambda$$, $$\kappa_A$$, $$\kappa_P$$, $$\beta$$) varies, holding others at calibrated values from Table I.
- **Welfare comparisons**: the gap $$F(W_0^p; R^p, L^p) - F(W_0^*; R^*, L^*)$$ is computed numerically (Figures 3-4) and reported as the gain from planner intervention.

The paper generates the following testable empirical predictions (Section IV.C, pp. 2947-2950): (1) CEO overcompensation is more severe in industries with more mobile managers (lower $$\kappa_A$$) or higher CEO replacement costs for firms (higher $$\kappa_P$$); (2) overcompensation, insufficient deferral, and excessive tenure are most pronounced in industries with high cash-flow volatility or severe moral hazard; (3) the credit line limits and long-term debt of firms are excessively low in equilibrium.

## Datasets used

This is a pure theory paper. It does not use empirical datasets directly. Calibration relies on parameter estimates reported in the literature:

| Source | Role in paper | Wiki page |
|---|---|---|
| Ward (2023) estimates of CEO discount rate and moral hazard parameter | Calibrate $$\gamma = 0.09$$ and $$\lambda = 0.29$$ (Table I) | No page yet |
| Taylor (2010) structural estimates of CEO replacement costs and turnover | Calibrate $$\kappa_P = 15$$ (6% replacement cost) and $$\kappa_A = 5.3$$ (2.2% forced turnover) | No page yet |
| Chen et al. (2023) manager discount rate estimates | Cross-check for $$\gamma$$ (11% estimate; paper uses 9%) | No page yet |
| Eisfeldt and Kuhnen (2013); Jenter and Kanaan (2015) | Cross-check for forced turnover rate lower bound (1.6-2.8%) | No page yet |

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13470) if you are: (a) building or extending a dynamic contracting model in a general equilibrium setting; (b) analyzing the policy implications of executive compensation externalities (noncompete clauses, pay transparency mandates); (c) studying the connection between optimal incentive contracts and capital structure in the spirit of DeMarzo and Sannikov (2006); or (d) replicating the quantitative calibration and comparative statics (Figures 2-9). The Internet Appendix contains proofs of renegotiation-proofness and the tax-implementation of the social optimum as an equilibrium.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(5). This distillation was extracted by an LLM on 2026-06-05 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Chemla, Gilles, Alejandro Rivera, and Liyan Shi.
> "Too Much, Too Soon, for Too Long: The Dynamics of Competitive Executive Compensation."
> *The Journal of Finance* 80, no. 5 (October 2025): 2921-2970.
> DOI: 10.1111/jofi.13470. © 2025 The Author(s).
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Venture Capital and Startup Agglomeration: Chen & Ewens (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/chen-venture-capital-startup-agglomeration-2025/
# Distilled: Using the Volcker Rule as a natural experiment, Chen and Ewens show that bank withdrawal from VC limited partnerships caused fewer and smaller VC funds in high-bank-exposure states, reduced startup financing and valuations, and induced startups to migrate to VC hubs (CA, MA, NY), directly implicating the local supply of venture capital in startup geographic concentration. J. Finance 2025, paywalled. Eight core results with source locators, datasets used, the identification strategy (DiD on Volcker Rule exposure), and the estimating specifications.
# Tags: paper-summary, venture-capital, entrepreneurship, agglomeration, geography
==============================================================================

**What this is.** The paper's core results, identification strategy, and estimating equations: enough to understand what it found and how, without reading all 46 pages. To replicate or extend, read the original at [https://doi.org/10.1111/jofi.13451](https://doi.org/10.1111/jofi.13451). Replication data are at [https://github.com/michaelewens/Banks-In-VC](https://github.com/michaelewens/Banks-In-VC).

## TL;DR

The paper asks whether local VC supply drives the geographic concentration of high-growth startups in the United States. The identification lever is the Volcker Rule (implemented December 2013): banking entities were prohibited from investing in VC funds as limited partners (LPs), and this restriction fell disproportionately on Midwestern and Southern states where banks had historically been a larger share of VC fund capital (bank exposure up to 25% of LP capital in some states, below 5% in CA). Using a difference-in-differences design, the paper shows that states more exposed to the bank LP shock experienced: (1) fewer and smaller VC funds; (2) smaller startup financing rounds, lower pre-money valuations, and increased pre-VC financing; and (3) a 24-30% increase in startups migrating to VC hubs (CA, MA, NY), but no increase in migration to non-VC-hub states. The migration evidence directly links local VC supply to startup agglomeration.

## Core results

Magnitudes and significance as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%. "1-SD increase in bank exposure" corresponds to moving from a low-exposure state (e.g., New York) to a high-exposure state (e.g., Wisconsin or Missouri). Regression coefficients are multiplied by 100 in the migration tables (Table VIII). Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Fewer VC funds** in high-exposure states after Volcker Rule | Table III Panel A col.(1), p. 2173 | Bank Expo x Post = -0.036\*\* (SE 0.013); 11% fewer VC funds per 1-SD bank exposure increase |
| R2 | **Less total VC capital raised** in high-exposure states | Table III Panel A col.(5), p. 2173 | Bank Expo x Post = -0.013\*\*\* (SE 0.006); 9% less total VC capital per 1-SD increase |
| R3 | **Smaller VC funds** (intensive margin) in high-exposure states | Table III Panel B col.(1)-(2), p. 2174 | Bank Expo x Post = -0.281\* to -0.305\*\*; 22% smaller fund size per 1-SD increase (NY to MO) |
| R4 | **Lower probability of raising a follow-on fund** for pre-Volcker VC firms | Table III Panel C, p. 2174 | Bank Expo = -0.030\*\*\* to -0.055\*\*; by 2018, 10pp lower probability against mean of 51.4% |
| R5 | **Startups raise 7% smaller first VC rounds** in high-exposure states | Table VI Panel A col.(1), p. 2180 | Bank Expo x Post = -0.089\*\*\* (SE 0.027); 7% smaller first-round VC financing per 1-SD increase |
| R6 | **Startup pre-money valuations 9% lower** (approx. $1.3M) in high-exposure states | Table VI Panel B col.(1), p. 2180 | Bank Expo x Post = -0.112\*\*\* (SE 0.040); 9% lower pre-money valuation per 1-SD increase |
| R7 | **Startups 30% more likely to migrate to CA** from high-exposure states post-Volcker | Table VIII Panel A col.(1), p. 2184 | Bank Expo x Post = 0.117\*\*\* (SE 0.018); null for non-VC-hub states (placebo Panel C) |
| R8 | **24% more likely to migrate to any VC hub** (CA, MA, NY) from high-exposure states | Table VIII Panel B col.(1), p. 2185-2186 | Bank Expo x Post = 0.175\*\*\* (SE 0.041); differential vs. non-VC-hub confirmed significant |

**Overall (paper's conclusion).** The Volcker Rule created an unintended natural experiment that reveals a causal role for local VC supply in startup geographic clustering. The loss of bank LPs reduced VC fundraising and startup financing in non-hub regions; startups responded by migrating to VC hubs, exacerbating existing geographic disparities in entrepreneurship. VC funding constraints rather than industry alignment or geographic distance explain the migration direction. VCs outside traditional hubs are financially constrained, and nonlocal VCs cannot fill the gap because of information asymmetry and local bias in investing.

## Theory / model

The paper has no formal structural model. The economic logic rests on two documented facts combined into a causal chain, with testable predictions:

**Fact 1 (LP home bias).** All LP types, including banks, exhibit significant in-state overweighting when investing in VC funds, following the finding of Hochberg and Rauh (2013) on LP home bias in private equity. Table V Panel A (p. 2178) shows that bank LPs allocate 23.2% of investments to in-state VCs, which is 11.5% above the share of all VC investments in the state (benchmark BM1) and 11.9% above the share of all out-of-state investments in the state (BM2). This surpasses all other LP types (pension funds: 20.8%; endowments: 12.6%).

**Fact 2 (geographic heterogeneity in bank LP exposure).** States outside the major VC hubs had historically higher bank LP shares, consistent with the geographic VC concentration documented by Chen et al. (2010). The Midwest and South had bank exposure ratios (bank-years with VC revenue per VC fund raised) of 2.44 and 2.11 respectively, versus 0.84 for the Northeast and 0.57 for the West (Figure 2 Panel A, p. 2168; Table I, p. 2166).

**Causal chain.** If bank LPs are locally biased and the Volcker Rule removes bank LPs disproportionately from non-hub states, then: (1) local VC supply falls more in those states; (2) startups there receive less and cheaper-priced capital; (3) startups facing local VC shortfalls migrate to where VC is available. The approach is analogous to Kortum and Lerner (2000), who assess VC contributions to innovation using supply-side variation, and Gonzalez-Uribe (2020), who uses LP supply shocks to trace VC effects. The migration channel resolves the ex ante ambiguity: VCs could conceivably substitute distant LPs, or nonlocal VCs could fill the gap. The results rule both out, attributing the failure to information asymmetry (a 10% increase in distance between an out-of-state VC and a startup is associated with a 0.5% higher likelihood of requiring a local co-investor; Table XII Panel A, p. 2193). The changes in startup valuations mirror findings in Gompers and Lerner (2000), who show that VC inflows create demand pressure and drive valuation changes.

The paper extends Samila and Sorenson (2011), who instrument VC supply with endowment LP returns at the MSA level; this paper adds a cleaner regional shock via the Volcker Rule.

**Identification assumption.** Parallel trends: high- and low-bank-exposure states evolved similarly in pre-Volcker VC activity (2010-2012). The dynamic estimation in Table IV (p. 2176) confirms pre-trend coefficients for Bank Expo x 2010, 2011, 2012 are all insignificant across outcome variables, with treatment effects emerging only from 2014 onward.

## Method

The primary estimator is a difference-in-differences regression (equation 1, p. 2170). The unit of analysis varies by outcome: state-year (VC fundraising aggregates), VC-fund level (fund size and follow-on), or startup level (financing and migration). The baseline specification is:

$$
Y_{it} = \beta_1 \, \text{Bank Expo}_i \times \text{Post}_t + \beta_2 X_i + \gamma_t + \epsilon_{it} \tag{1}
$$

where $$\text{Bank Expo}_i$$ is the state-level treatment variable measuring VCs' pre-Volcker reliance on banks as LPs (the ratio of aggregate bank-years with VC revenue to the number of VC funds raised in the state over 2001-2013); $$\text{Post}_t = 1$$ for 2014-2018 (the post-Volcker period); $$X_i$$ is a vector of entrepreneurial firm characteristics, state fixed effects, founding year fixed effects, and industry fixed effects; and $$\gamma_t$$ is year fixed effects. The main coefficient of interest is $$\beta_1$$.

The treatment variable is continuous, capturing cross-state variation richer than a binary High/Low split. It builds on `difference-in-differences` and `panel-regression`. Identification relies on the differential shock to LP capital across states: the Volcker Rule's scope for VC funds was unexpected (Congress did not intend to include VC funds), so pre-2014 bank LP distribution was not strategically adjusted in anticipation.

The paper also runs a within-VC-firm analysis (Panel C of Table III): a single-difference regression of whether a pre-Volcker VC firm raised a follow-on fund by year $$t \in \{2014, ..., 2018\}$$ on Bank Expo, controlling for the last pre-Volcker fund vintage year fixed effects.

## Empirical specifications

**VC fundraising (R1-R4).** State-year DiD with state and year fixed effects (Table III Panel A). Dependent variables: ln(1 + number of VC funds), ln(1 + total VC capital raised). State-year-level controls follow Gompers and Lerner (1998): GDP growth, log GDP per capita, house price growth, STEM employment growth. Errors clustered by state. Robustness: exclude California, narrow to 2011-2017, Poisson regression, IHS transformation.

**VC fund size (R3).** VC-fund-level DiD with VC firm fixed effects, vintage year fixed effects, and fund-sequence fixed effects (Table III Panel B). Dependent variable: ln(VC fund size). Errors clustered by state. Sample: 1,617 VC funds, 2010-2018.

**Startup financing (R5-R6).** Startup-year DiD with HQ state fixed effects, financing year fixed effects, founding year fixed effects, Series A or Seed fixed effects, and industry fixed effects (Table VI). Dependent variables: ln(capital raised in first VC round), ln(pre-money valuation), equity sold, ln(syndication size), indicator for pre-VC financing. Sample: 11,048 startups (5,903 for valuation). Errors clustered by startup HQ state.

**Startup migration (R7-R8).** Startup-year DiD at startup level (Table VIII). Dependent variables: dummy for moving HQ to CA (Panel A), to VC hubs CA/MA/NY (Panel B), or to non-VC-hub states (Panel C, placebo). Fixed effects: incorporation state, origin state, year, founding year, industry. Errors clustered by startup initial HQ state. Sample: 56,487 startup-year observations. All regression coefficients multiplied by 100.

$$
Y_{ist} = \beta_1 \, \text{Bank Expo}_s \times \text{Post}_t + \beta_2 X_{is} + \text{FE} + \epsilon_{ist}
$$

The triple-difference specifications (Tables IX-XI) add a third interaction (industry alignment, geographic distance, or VC financing constraints) to explore heterogeneity in migration responses. The geographic distance specification (Table X) finds distance does not explain migration choices post-shock; the VC-funding-constraint specifications (Table XI, Panels A-C) show biotech startups, older startups, and startups in high-VC-funded industries are more likely to migrate.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Call Reports (FFIEC) + FR Y-9Cs (BHC filings) | Construct bank VC revenue series 2001-2013; identify banking entities investing in VC; build treatment variable Bank Expo | [no page yet] |
| VentureSource (CB Insights / Dow Jones) | VC fund characteristics, startup financing rounds, startup HQ state over time; primary VC data 2010-2018 | [no page yet] |
| Pitchbook | VC fund and startup data (robustness checks); LP commitment information | [PitchBook](/wiki/commercial/pitchbook/) (licensed) |
| SEC EDGAR (Form D filings) | Identify startup migration via consecutive business-address changes 2002-2018; 56,487+ startup-year observations | [EDGAR](/wiki/datasets/edgar/) |
| Preqin | LP commitment data for robustness checks (LP home bias analysis, Table V) | [Preqin](/wiki/commercial/preqin/) (licensed) |

Sample: state-year panel covers 35 U.S. states, 2010-2018 (315 state-year observations). VC fund sample: 1,617 funds. Startup sample: 11,048 startups (first Seed or Series A round, $100M); 1,700 identified as having ever moved to a different state.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13451) if you are: studying causal drivers of startup geographic concentration; evaluating the regional impact of financial regulation (Dodd-Frank, Volcker Rule); interested in how LP capital constraints propagate through the VC intermediation chain to real activity; or building on the startup migration literature. The Internet Appendix contains additional robustness tests (Tables IA.I-IA.XIV), including placebo tests on banking sector outcomes, alternative treatment variable constructions, and the LP home-bias theoretical exercise. Replication data and code are at [https://github.com/michaelewens/Banks-In-VC](https://github.com/michaelewens/Banks-In-VC).

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(4), August 2025. This distillation was extracted by an LLM on 2026-06-05 and is **not human-verified or independently reproduced**. The paper is paywalled (Wiley VOR terms, not CC-licensed); only core results are extracted here.

> Chen, Jun, and Michael Ewens. "Venture Capital and Startup Agglomeration."
> *The Journal of Finance* 80, no. 4 (August 2025): 2153-2198.
> DOI: 10.1111/jofi.13451. © 2025 the American Finance Association.
> Extract-only: the Wiley VOR licence does not permit redistribution.


==============================================================================
# Bank Funding Risk, Reference Rates, and Credit Supply: Cooperman, Duffie, Luck, Wang & Yang (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/cooperman-bank-funding-risk-reference-2025/
# Distilled: Credit-sensitive reference rates like LIBOR mitigate banks' debt-overhang cost from revolving credit commitments; the transition to risk-free SOFR increases expected draw costs by about 15 bps and reduces equilibrium credit line commitments by roughly 6%, with effects concentrated at high-debt-overhang banks. J. Finance 2025, paywalled. Six core results with source locators, datasets used, the equilibrium model of credit line provision, and the empirical method.
# Tags: paper-summary, banking, credit-supply, bank-funding, reference-rates, libor-sofr
==============================================================================

**What this is.** The paper's core results, the equilibrium model of credit line provision, and the empirical evidence on how banks fund credit drawdowns: enough to know what it found and how, without reading all 52 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13411).

## TL;DR

Revolving corporate credit lines create a debt-overhang cost for bank shareholders: because borrowers draw more heavily when funding markets are stressed, banks must sometimes fund drawdowns with expensive unsecured wholesale debt, diluting legacy shareholders. Credit-sensitive reference rates like LIBOR partially offset this by reducing borrower incentives to draw precisely when bank funding costs are high. The LIBOR-to-SOFR transition removes this offset. Using a calibrated equilibrium model and confidential Federal Reserve data (FR 2052a, FR Y-14Q), the authors show that SOFR-linked lines lead to higher spreads (about 15 bps), smaller credit line commitments (roughly 6% lower for a baseline bank), and a 2.6% welfare loss. The adverse impact is offset at banks that expect drawn funds to be left on deposit, which happened at large universal banks during COVID but not at regional banks during the GFC.

The paper relates to Kashyap, Rajan, and Stein (2002), who identify the liquidity coinsurance synergy between deposits and credit lines; the authors add a new debt-overhang complementarity channel to this synergy. It also extends Jermann (2019), who shows LIBOR-linked loan revenues insure banks against funding risk, to an equilibrium credit provision setting. Empirically, it builds on Ivashina and Scharfstein (2010) on GFC drawdowns, Acharya and Mora (2015) on bank funding pressures, and Gatev and Strahan (2006) on flight-to-safety deposit inflows. In the conclusion, the authors discuss the AXI credit-sensitive reference rate proposed by Berndt, Duffie, and Zhu (2023) as a potential LIBOR successor.

## Core results

Magnitudes and significance are as reported; `\*`, `\*\*`, `\*\*\*` = 10%, 5%, 1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | During COVID, each dollar of credit line drawdown is associated with **94 cents deposited at the same bank**, insulating shareholders from funding costs | Table IV Panel A, col. 1, p. 24-25 | Coeff. on Delta Draws x COVID = 0.94\*\*\* (0.34); baseline coeff. on Delta Draws = 0.07 (0.04); total = ~1.01 during COVID |
| R2 | During the GFC (post-Lehman), each dollar of C&I loan growth required **~$4 of new wholesale funding** (no deposit offset) | Table V col. 3-4, p. 28 | Coeff. on DeltaC&ILoans x Lehman = 4.33\*\*\* (0.41) in levels (col. 3); normalized Lehman interaction = 0.90\* (0.47) per unit assets (col. 4) |
| R3 | LIBOR-to-SOFR transition reduces equilibrium **aggregate credit line commitments by 5.9%** for baseline bank | Table VI col. Baseline, p. 38; Figure 5 right, p. 34 | -5.87% (baseline theta=1.0, D=0.2); range: +4.64% (low-overhang) to -11.36% (high-overhang) |
| R4 | Transition reduces **expected drawn credit by 2.77%** for baseline bank | Table VI, p. 38 | -2.77% (baseline); +1.12% (low-overhang); -5.20% (high-overhang) |
| R5 | The equilibrium **spread on drawn credit rises ~15 bps** and the drawn-rate spread rises 51.7 bps over the reference rate for baseline bank | Table VI, p. 38; p. 6 | Change in drawn spread s\* = +51.7 bps; expected cost of drawn credit rises ~15 bps; SOFR lines ~24 bps more expensive than LIBOR lines in normal times (Figure 5 left, p. 34) |
| R6 | **Welfare falls 2.62%** for baseline bank; the welfare-maximizing reference rate has ~72% of LIBOR's credit sensitivity | Table VI, p. 38; p. 39 | Welfare change: -2.62% baseline, -4.98% high-debt-overhang, +1.16% low-debt-overhang; optimal lambda\* = 0.72 for baseline |

**Overall (paper's conclusion).** The transition from LIBOR to SOFR increases the ex-ante expected cost banks must charge on revolving credit, reducing the equilibrium supply of credit lines. The magnitude depends on how much of future drawdowns banks expect to receive back as deposits. The adverse effect is concentrated at high-funding-spread, low-deposit-retention banks (predominantly large regional banks), while well-capitalized banks with strong deposit franchises may gain market share. The welfare-maximizing reference rate is more credit-sensitive than risk-free SOFR but need not be as credit-sensitive as LIBOR.

## Theory / model

The model has a three-date structure (Figure 1, p. 11). At time 0, a bank offers a menu of credit line contracts $$\{(L, s(L)) : L \geq 0\}$$ distinguishing line size $$L$$ and contractual spread $$s(L)$$ over the floating reference rate $$R$$. The credit-sensitive rate is $$R = r + W$$ (LIBOR form, for credit-spread benchmark $$W$$); the risk-free rate is $$R = r$$ (SOFR form). At time 1, the reference rate $$R$$, the bank's unsecured wholesale credit spread $$S$$, and the borrower's liquidity shock $$\psi$$ are realized. The borrower draws $$q \leq L$$ and deposits $$d \leq q$$ at the same bank. At time 2, the bank is solvent or not.

The borrower's optimal draw quantity solves (eq. 1, p. 13):

$$
Q(L) = \underset{0 \leq q \leq L}{\sup} \; b(q, \psi) - q\delta(1 + R + s(L)) \tag{1}
$$

where $$b(q, \psi)$$ is a reduced-form liquidity benefit with marginal benefit $$b_x(x, y)$$ increasing, differentiable, and strictly concave in $$x$$. The optimal draw has a threshold structure: $$Q(L) = L$$ if the marginal liquidity benefit is high enough, and $$Q(L) = B(\delta(1+R+s(L)), \psi)$$ where $$B(\cdot, y)$$ is the inverse of $$b_x(\cdot, y)$$.

The market value of credit lines to **legacy bank shareholders** is (Proposition 1, eq. 7, p. 14):

$$
G(L) = p_1\bigl(\delta Q(L)(1+R+s(L)) - Q(L)\bigr) - p_1\delta(1-\varphi-C)Q(L)S - (1-p_1)CQ(L) \tag{7}
$$

where $$p_1 = P_1(X \geq 0)$$ is the bank's solvency probability at time 1, $$\varphi = d/q$$ is the fraction of drawn funds deposited at the bank, $$C$$ is the regulatory capital ratio, and $$S$$ is the bank's unsecured wholesale credit spread. The second term is the **debt-overhang cost**: the bank must fund $$(1-\varphi-C)Q(L)$$ of drawdowns with expensive new wholesale debt at spread $$S$$. Under Bertrand competition ($$E[G(L)] = 0$$), the equilibrium spread is (eq. 8, p. 14):

$$
s(L) = \frac{E\bigl[p_1 Q(L)\bigl(1 - \delta(1+R) + \delta(1-\varphi-C)S\bigr) + (1-p_1)CQ(L)\bigr]}{E\bigl[\delta p_1 Q(L)\bigr]} \tag{8}
$$

This spread rises in $$\text{Cov}(S, Q(L))$$: the key debt-overhang wedge. A credit-sensitive reference rate ($$R = r+W$$ with $$W \approx S$$) reduces drawdowns exactly when $$S$$ is high, cutting this covariance. A risk-free reference rate removes this attenuation.

**Identification.** The empirical section uses the GFC (Ivashina and Scharfstein (2010)) and COVID recession as quasi-natural experiments that shift the LIBOR-OIS spread to historically extreme levels, allowing the authors to compare how banks funded drawdowns under each episode. The calibrated model is then identified by fitting the distribution of credit line utilization in the FR Y-14Q data across normal times and the COVID recession.

## Method

The paper proceeds in three steps. First, the equilibrium model of Section II identifies the theoretical mechanism. Second, confidential FR 2052a and FR Y-14Q data are used empirically to pin down bank funding cost composition and drawdown funding, building on `panel-regression` and `difference-in-differences`. Third, the model is calibrated via nonlinear least squares to historical LIBOR-linked credit line behavior and the model is used to simulate the LIBOR-to-SOFR transition.

**Calibration of borrower demand.** The liquidity benefit is specified as (eq. 11, p. 30):

$$
b(q, \psi) = \frac{\psi^\alpha q^{1-\alpha}}{1-\alpha} \tag{11}
$$

for price elasticity $$1/\alpha$$, baseline $$\alpha = 1/25$$ (elasticity of 25). The optimal draw is therefore (eq. 13, p. 33):

$$
Q(L) = \min\!\bigl((K(W) + \epsilon)^+(1+R+s)^{-1/\alpha}, L\bigr) \tag{13}
$$

where $$K(W)$$ is the common component of the borrower liquidity shock (a nonlinear increasing function of LIBOR-OIS spread $$W$$) and $$\epsilon$$ is an idiosyncratic shock. The deposited fraction of drawn funds is a logistic function of LIBOR-OIS (eq. 12, p. 32):

$$
\Phi(x) = \frac{D}{1+e^{-m(x-w_0)}} \tag{12}
$$

with baseline parameters $$D = 0.2$$, $$m = 0.1$$, $$w_0 = 146.1$$ bps. The risk-neutral probability of a GFC-like crisis is set to $$p = 4\%$$, calibrated to daily LIBOR-OIS observations from January 2005 to April 2021.

**Aggregate equilibrium quantities** in the continuum-of-borrowers model satisfy (eq. 14-15, p. 33):

$$
M \cdot E[Q(L) \mid W] = M \cdot E\!\left[\min\!\left((K(W)+\epsilon)^+(1+R+s)^{-1/\alpha}, L\right) \;\middle|\; W\right] \tag{14}
$$

$$
P[Q(L) = L \mid W] = P\!\left[(K(W)+\epsilon)^+(1+R+s)^{-1/\alpha} \geq L \;\middle|\; W\right] \tag{15}
$$

## Empirical specifications

**Funding of drawdowns during COVID (eq. 9, p. 24).** Using monthly FR 2052a data (20 largest BHCs, July 2017 to April 2022, $$N = 1{,}111$$, 20 banks), the authors estimate:

$$
\Delta y_{bt} = \tau_t + \gamma_b + \beta_1 \Delta\text{Drawdowns}_{bt} + \beta_2 \Delta\text{Drawdowns}_{bt} \times \text{COVID}_t + \epsilon_{bt} \tag{9}
$$

where $$y_{bt}$$ is the change in corporate deposits, FHLB advances, unsecured wholesale funding, or total deposits. $$\text{COVID}_t$$ takes the value one during March and April 2020. Bank and time fixed effects are included. Robust standard errors are clustered at the bank level. The coefficient $$\beta_1 + \beta_2$$ measures the marginal funding response per dollar of drawdown during COVID; $$\beta_1$$ measures the non-COVID response (Table IV, p. 25).

**Funding of drawdowns during GFC (eq. 10, p. 27).** Using weekly FR 2416 data (30 commercial banks, December 2007 to May 2009, $$N = 1{,}765$$), the authors estimate:

$$
y_{bt} = \tau_t + \gamma_b + \beta_1 \Delta C\&I\text{Loans}_{bt} + \beta_2 \times \Delta C\&I\text{Loans}_{bt} \times \text{Lehman}_t + \epsilon_{bt} \tag{10}
$$

where $$\text{Lehman}_t$$ takes value one from September 15, 2008 through end of 2008. The dependent variable is either total deposits or wholesale short-term funding (WHSLE). Month and bank fixed effects. Robust standard errors (Table V, p. 28).

Both regressions test whether drawdowns generated deposit inflows (low marginal funding cost) or required new wholesale funding (high marginal funding cost). During COVID, banks needed essentially no new external funding because drawdowns were left on deposit; during the GFC, the same drawdowns forced banks to raise costly wholesale debt.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| FR 2052a (Federal Reserve confidential) | Daily/monthly bank balance sheet, funding composition by counterparty and product type; main panel for COVID empirics (20 BHCs, July 2017-April 2022) | no page yet |
| FR Y-14Q Schedule H1 (corporate loans) and H2 (CRE) | Loan-level credit commitments and utilization; calibration target for borrower demand and line draw distribution | [FR Y-14Q](/wiki/confidential/fr-y14q/) (confidential) |
| FR 2416 (Federal Reserve, weekly) | Weekly balance sheet for 30 banks; GFC drawdown-funding regression (Dec 2007-May 2009) | no page yet |
| FR 2420 (Federal Reserve, transaction-based) | Corporate deposit and wholesale funding rates (overnight rates, CDs, fed funds); sensitivity of funding costs to LIBOR-OIS | no page yet |
| FRED | LIBOR-OIS spread, SOFR, effective federal funds rate | [FRED](/wiki/datasets/fred/) |
| FR Y-9C / bank call reports | Public balance sheet data merged with FR 2052a for context | no page yet |
| FHLB Des Moines historical fixed-rate advance file | FHLB advance rates for funding cost benchmark | no page yet |
| Bloomberg | Long-term bank debt floating-rate and LIBOR-referenced shares | no page yet |

Sample: Empirical analysis spans December 2007 to April 2022; calibration uses LIBOR-OIS observations from January 2005 to April 2021.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13411) if you are: (i) pricing revolving credit facilities under SOFR to account for debt-overhang costs, and want the calibrated equilibrium spread formula (Section IV.B); (ii) studying bank balance sheet dynamics during the GFC vs. COVID using FR 2052a or FR 2416 data (Sections III.E-F and Tables III-V); (iii) evaluating alternative credit-sensitive reference rates such as AXI, BSBY, or Ameribor from a welfare-maximizing standpoint (Section V); or (iv) studying how bank capitalization and deposit franchise strength mediate the impact of reference-rate transition on credit supply heterogeneity (Section IV.C and Table VI).

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(1), February 2025, pp. 5-56. Accepted March 4, 2024; published online December 20, 2024. This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The paper is paywalled (Wiley/AFA); only textual extracts are reproduced here.

> Cooperman, Harry, Darrell Duffie, Stephan Luck, Zachry Wang, and Yilin (David) Yang.
> "Bank Funding Risk, Reference Rates, and Credit Supply."
> *The Journal of Finance* 80, no. 1 (February 2025): 5-56.
> DOI: [10.1111/jofi.13411](https://doi.org/10.1111/jofi.13411).
> © 2024 the American Finance Association. All rights reserved.
> This page presents distilled extracts only; redistribution of the verbatim article is not permitted.


==============================================================================
# Repo over the Financial Crisis: Copeland & Martin (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/copeland-repo-financial-crisis-2025/
# Distilled: Using new confidential data covering all four segments of the U.S. repo market (bilateral and tri-party, interdealer and dealer-to-client), this paper documents that the 2008 decline in repo activity was largest in bilateral (MIX) segments and disproportionately concentrated in Treasury-backed repos, and was driven by a pullback in securities-driven market-making trades rather than by counterparty credit concerns. J. Finance 2025, U.S. Government work / public domain. Six core results with source locators, datasets used, and the empirical specifications.
# Tags: paper-summary, repo-markets, financial-crisis, market-microstructure
==============================================================================

**What this is.** The paper's core results, the data sources it introduces, and
the empirical specifications it runs: enough to know what it found and how,
without reading all 26 pages. To replicate or extend it, read the original at
[doi.org/10.1111/jofi.13406](https://doi.org/10.1111/jofi.13406).

## TL;DR

Using new confidential data that cover all four segments of the U.S. repo
market (interdealer and dealer-to-client, general collateral and mixed), the
paper documents three facts about the 2007-2009 Global Financial Crisis: (1)
the decline in repo activity was far larger in bilateral (MIX) segments than
in tri-party (GC) segments; (2) more than half the decline was
concentrated in Treasury-backed repos, the safest and most liquid asset class,
contradicting the flight-to-quality narrative; and (3) the decline was
not correlated with measures of dealer counterparty credit risk (CDS spreads,
CP rates), but was instead driven by a pullback in securities-driven
market-making trades by large securities dealers. The paper uses a
securities-dealer panel and a seemingly unrelated regression (SUR) design
to trace the repo decline to a reduction in back-to-back securities-driven
trades entered by large dealers, particularly after the Lehman Brothers
bankruptcy.

## Core results

Magnitudes and significance are as reported; `\*` = 10%, `\*\*` = 5%, `\*\*\*` = 1%. Locators
point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Treasury repo fell more than other asset classes** despite the flight-to-quality narrative | Table I, p. 920 | Total repo -17% (-$798B pre-crisis to crisis); Treasury -20% (-$470B); All others -15% (-$328B); agency MBS only -1% |
| R2 | **MIX (bilateral) segments contracted sharply; GC (tri-party) segments were stable or grew** | Table II, p. 921 | MIX ID: -24%; MIX DtC: -31%; GC ID: +7%; GC DtC: -10%. For Treasuries: MIX ID: -23%; MIX DtC: -37%; GC ID: +52%; GC DtC: +7% |
| R3 | **CDS and CP spreads have no statistically significant association with MIX repo activity**; counterparty credit concerns do not explain the decline | Table IV, p. 925; Table V, p. 927 | CDS coeff = -3.757 (SE 12.733), p>0.10; R² = 0.029. CP coeff = 0.035 (SE 0.055), p>0.10 |
| R4 | **Post-Lehman, the association between MIX repo and MIX reverse repo falls, while MIX reverse repo and DtC GC repo become correlated** for large dealers, indicating a pivot to cash-driven trades | Table VI, p. 929 | PL × delta MIX Revr coeff in MIX repo: -0.17 (insig.); PL × delta MIX Revr coeff in DtC GC repo: 0.20\*\* (SE 0.09) |
| R5 | **Large dealers drive the change**: the post-Lehman drop in securities-driven repo is concentrated among above-median dealers | Table VII, p. 931 | PL × Large × delta MIX Revr in MIX repo: -0.60\* (SE 0.33); in DtC GC repo: +0.70\*\* (SE 0.30); no significant change for small dealers |
| R6 | **On-the-run Treasury repo, a direct proxy for securities-driven trades, fell $155 billion** (54%) from Q2 to Q4 2008; this decline is concentrated at large dealers | Figure 4, p. 933 | From $286B (Q2 2008) to $131B (Q4 2008); distribution narrows, with the 75th percentile falling more than the median or 25th percentile |

**Overall (paper's conclusion).** The massive decline in repo over the
financial crisis was driven primarily by a pullback in securities-driven trades
that support Treasury market-making by securities dealers, not by a generalized
disruption in funding conditions or by clients fleeing dealers out of
counterparty credit concerns. Copeland and Martin (2025) and Copeland, Martin,
and Walker (2014) jointly show that the GC (funding) segment remained stable,
so funding disruptions did occur but were institution-specific (Bear Stearns,
Lehman Brothers), not market-wide. The extent of true funding disruption was
considerably smaller than the decline in overall repo activity implies.

## Theory / model

The paper has no formal model. Its identification strategy rests on the
institutional structure of the repo market: the U.S. repo market is segmented
into four segments that differ in clearing platform and in whether they
accommodate securities-driven versus cash-driven trades (pp. 914-915).

**Key institutional distinction.** The general collateral (GC) segment
settles on the tri-party platform. By design, all GC trades are
cash-driven: parties agree only on an asset class, not a specific security,
so the GC segment cannot support securities-driven price-discovery trades.
The mixed (MIX) bilateral segment accommodates both cash-driven and
securities-driven trades. This segmentation means that changes in
MIX activity capture changes in both motives, while changes in GC activity
capture only changes in cash-driven (funding) motives.

**Tested hypotheses.** Given the segmentation above, the paper tests three
hypotheses:

1. Whether the decline in repo was disproportionately in MIX (bilateral)
   versus GC (tri-party) segments. Gorton and Metrick (2012) and Gorton,
   Metrick, and Ross (2020) document large aggregate haircut increases and
   net-repo declines at banks; Copeland, Martin, and Walker (2014) and
   Krishnamurthy, Nagel, and Orlov (2014) find that DtC GC (tri-party)
   activity held up. This paper unifies both views by showing MIX segments
   drove most of the decline.
2. Whether the decline was correlated with measures of dealer counterparty
   credit risk (a "run on repo" hypothesis in the spirit of Gorton and
   Metrick (2012)).
3. Whether, post-Lehman, the statistical association between dealers' MIX
   repo and MIX reverse repo weakened while the MIX reverse repo and DtC GC
   repo association strengthened (a pivot-to-cash-driven-trades hypothesis).
   Musto, Nini, and Schwarz (2018) document the corresponding price effects
   in Treasuries; this paper documents the associated quantity decline in
   securities-driven Treasury repo.

## Method

The paper applies two estimation strategies: aggregate descriptive comparisons
across periods and segments, and a dealer-level panel regression.

**Aggregate analysis.** Using FR 2004C data (a weekly Federal Reserve survey of
primary dealers) plus FICC GCF Repo and FICC DVP service data, the paper
constructs average daily repo outstanding by segment and asset class for the
pre-crisis period (July 1 to September 13, 2008) and the crisis period
(October 15 to December 17, 2008), then computes differences. This is a
descriptive before-after comparison with no causal identification design
(p. 919).

**Dealer-level panel: counterparty credit risk.** The paper estimates four
OLS regressions (equations 1-4, p. 924) relating changes in MIX repo to
contemporaneous or lagged changes in CDS spreads, plus Treasury-related
controls. Standard errors are clustered at the dealer level (18 dealers, 21
weeks, 378 observations). The specifications are (pp. 924-926):

$$
\Delta \text{MIXRepo}_{i,t} = \alpha_0 + \alpha_1 \Delta \text{CDS}_{i,t} + \Delta X_t \beta_0 + \varepsilon^1_{i,t} \tag{1}
$$

$$
\Delta \text{MIXRepo}_{i,t} = \alpha_2 + \alpha_3 \Delta \text{CDS}_{i,t-1} + \Delta X_t \beta_1 + \varepsilon^2_{i,t} \tag{2}
$$

and log-level variants (equations 3 and 4). The CP spread regressions
(equations 5 and 6, p. 926) are:

$$
\log(\text{MIXRepo})_{i,t} = \alpha_0 + \beta_0 \text{CP}_{i,t} + \log X_t \gamma_0 + \eta_i + \varepsilon^1_{i,t} \tag{5}
$$

$$
\Delta \log(\text{MIXRepo})_{i,t} = \alpha_1 + \beta_1 \Delta \text{CP}_{i,t} + \Delta \log X_t \gamma_1 + \varepsilon^2_{i,t} \tag{6}
$$

**Dealer-level panel: repo flow analysis.** The paper estimates a seemingly
unrelated regression (SUR) system (equations 7-9, p. 928) with three
left-hand side variables: change in MIX repo, change in DtC GC repo,
and change in ID GC repo. The key right-hand side variables are MIX
reverse repo and ID GC reverse repo, interacted with a post-Lehman
dummy (PL) to allow strategies to differ across the two periods:

$$
\Delta \text{MIXRepo}_{j,t} = \alpha_0 + \alpha_1 \Delta \text{MIXRevr}_{j,t} + \alpha_2 PL_t + \alpha_3 PL_t \cdot \Delta \text{MIXRevr}_{j,t} + \Delta X_t \beta_0 + \varepsilon^0_{j,t} \tag{7}
$$

$$
\Delta \text{DtC\text{-}GCRepo}_{j,t} = \alpha_4 + \alpha_5 \Delta \text{MIXRevr}_{j,t} + \alpha_6 \Delta \text{ID\text{-}GCRevr}_{j,t} + \alpha_7 PL_t + \alpha_8 PL_t \cdot \Delta \text{MIXRevr}_{j,t} + \alpha_9 PL_t \cdot \Delta \text{ID\text{-}GCRevr}_{j,t} + \Delta X_t \beta_1 + \varepsilon^1_{j,t} \tag{8}
$$

$$
\Delta \text{ID\text{-}GCRepo}_{j,t} = \alpha_{10} + \alpha_{11} \Delta \text{MIXRevr}_{j,t} + \alpha_{12} \Delta \text{ID\text{-}GCRevr}_{j,t} + \alpha_{13} PL_t + \alpha_{14} PL_t \cdot \Delta \text{MIXRevr}_{j,t} + \alpha_{15} PL_t \cdot \Delta \text{ID\text{-}GCRevr}_{j,t} + \Delta X_t \beta_2 + \varepsilon^2_{j,t} \tag{9}
$$

SUR allows correlated error terms across the three equations. Standard
errors are clustered at the dealer level. The 288-observation sample
reflects 18 dealers over the full weekly sample. Table VII repeats the
cross-sectional analysis with dealer-size and domestic/foreign dummy
interactions.

## Empirical specifications

**Sample.** Dealer-level weekly data covering July 1 to December 17, 2008
(excluding two quarter-end weeks and the week of Lehman's bankruptcy),
leaving 21 weeks and 18 primary dealers (378 observations for the CDS panel,
288 for the SUR after excluding Lehman Brothers). The pre-crisis period is
July 1 to September 13; the crisis period is October 15 to December 17.

**Counterparty-risk regressions (R3).** OLS with four specifications
(level and log, contemporaneous and lagged CDS; level and change in CP).
Controls: net Treasury bill issuance, net coupon issuance, Federal Reserve
SOMA operations, total UST holdings by the Fed. Standard errors clustered
at the dealer level. The R-squared statistics are very low (0.013-0.031),
and all CDS and CP coefficients are statistically insignificant, confirming
no systematic relationship between dealer risk measures and MIX repo
(Tables IV-V, pp. 925, 927).

**SUR repo-flow regressions (R4, R5).** The identifying variation is the
differential post-Lehman change in the association between a dealer's
repo and reverse repo activity across segments. The PL interaction terms
capture this change. The null hypothesis tested is that these associations
did not change after Lehman; rejection implies dealers shifted strategies.
Table VII adds size (large vs small) and location (domestic vs foreign)
dummies to capture heterogeneous responses (p. 930).

**On-the-run Treasury repo (R6).** Using the FR 2004SI survey (a
special-purpose data collection), the paper tracks repo involving
specifically identified on-the-run Treasuries, which are almost never used
in cash-driven transactions. The time series of aggregate on-the-run
Treasury repo and the cross-sectional distribution across dealers are
plotted in Figure 4 (p. 933) as direct evidence of the securities-driven
trade decline.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| FR 2004C (Federal Reserve weekly survey of primary dealers) | Primary source for repo and reverse repo outstanding by asset class; covers all primary dealers at weekly frequency, 2004-2013 | No page yet |
| FICC GCF Repo Service data (dealer-level daily, licensed via FRBNY) | Interdealer GC segment repo and reverse repo by asset class at dealer level | No page yet |
| FICC DVP Service data (aggregate-level daily) | Interdealer MIX segment repo; aggregate only | No page yet |
| DtC TPR (tri-party repo) data (confidential, from FRBNY) | Dealer-to-client GC segment repo by dealer and asset class | No page yet |
| Markit Group CDS spreads (five-year modified restructuring, USD) | Proxy for dealer counterparty credit risk; matched to 13 of 18 dealers | [Markit CDS](/wiki/commercial/markit-cds/) (licensed) |
| DTCC commercial paper interest rates (confidential) | Dealer-level weekly CP interest rates; proxy for dealer counterparty risk | No page yet |
| FR 2004SI (special-purpose survey) | On-the-run Treasury repo activity by primary dealer; direct proxy for securities-driven trades | No page yet |

Sample: second half of 2008 (July-December) for dealer-level analysis;
2004-2013 for aggregate FR 2004C time series. Weekly frequency. 18-19
primary dealers (Lehman excluded from panel analysis).

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13406) if you are: analyzing
the drivers of repo market stress during the 2007-2009 GFC; studying the
comparative resilience of tri-party versus bilateral repo markets; investigating
the role of securities dealers as intermediaries in Treasury market-making;
or examining whether the March 2020 Treasury market disruptions share
mechanisms with the 2008 crisis (the paper's conclusion draws this parallel
explicitly). The Internet Appendix contains additional robustness tables
(IA.II) and a map of the U.S. repo market (Section III).

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(2), April 2025, pages 911-936.
DOI: 10.1111/jofi.13406. This article is a U.S. Government work and is in the
public domain in the USA (as stated on p. 911 of the artifact). Crossref
records Wiley standard terms; the artifact's public-domain notice takes
precedence for U.S. readers. This distillation was extracted by an LLM on
2026-06-06 and is **not human-verified or independently reproduced**.
Extract-only: the verbatim PDF is not hosted here.

> **Citation.** Copeland, Adam, and Antoine Martin. "Repo over the Financial
> Crisis." *The Journal of Finance* 80, no. 2 (April 2025): 911-936.
> DOI: 10.1111/jofi.13406.


==============================================================================
# Term Structure in a Heterogeneous Monetary Union: Costain, Nuno & Thomas (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/costain-term-structure-interest-rates-2025/
# Distilled: Costain, Nuno, and Thomas build an arbitrage-based affine term structure model for a two-country monetary union with sovereign default risk, showing that the credit risk premium accounts for roughly three-quarters of the Italy-Germany sovereign spread, and that ECB PEPP asset purchases compressed Italian yields primarily through a default risk extraction channel rather than the standard duration risk channel. J. Finance 2025, paywalled. Six core results with source locators, datasets used, the model, and the method.
# Tags: paper-summary, fixed-income, term-structure, sovereign-debt, monetary-policy
==============================================================================

**What this is.** This is an LLM-distilled skeleton of the paper; read the
[original](https://doi.org/10.1111/jofi.13463) to replicate or extend. Not
human-verified; not reproduced.

## TL;DR

Costain, Nuno, and Thomas build an arbitrage-based affine term structure model
(ATSM) for a two-country monetary union (Core = Germany, Periphery = Italy)
where the Peripheral sovereign faces rollover-crisis default risk. The model
extends Vayanos and Vila (2021) to a multicountry setting with sovereign
default risk, building on Duffie and Singleton (1999) for defaultable bond
pricing. It decomposes yields into four components: expectations, term premium,
expected default loss, and credit risk premium. Calibrated to German and Italian
zero-coupon yields from 1999 to 2022, the model finds that the credit risk
premium accounts for roughly three-quarters of the long-run Italy-Germany
sovereign spread. For the ECB PEPP announcement of March 18, 2020, which
drove a ~71 bps fall in Italian 10-year yields, the model attributes about
four-fifths of the sovereign spread compression to a reduction in the credit
risk premium via a *default risk extraction channel*, not the conventional
duration risk channel identified in Greenwood and Vayanos (2014) and
Hamilton and Wu (2012).

## Core results

| # | Result | Locator | Magnitude as reported |
|---|---|---|---|
| R1 | Credit risk premium dominates the long-run IT-DE 10Y sovereign spread | Table II, p. 2419; Figure 5, p. 2420 | Model spread = 79 bps; expected default loss = 22 bps (28%); credit risk premium = 62 bps (78%); data spread = 126 bps |
| R2 | Model matches German 10Y term premium and Sharpe ratio | Table II, p. 2419; Figure 4, p. 2418 | Model TP = 142 bps vs. data term premium ~125 bps; model Sharpe 0.57, data 0.51 |
| R3 | PEPP announcement: 81% of IT yield decline attributed to credit risk premium | Figure 6, p. 2424; Table III, p. 2427 | Observed IT 10Y decline = 71 bps (March 18-20, 2020); model matches at calibration; credit risk premium share = 81% across robustness checks |
| R4 | PEPP announcement: small effect on German yields, consistent with asymmetry | Figure 6, p. 2424 | German 10Y: roughly -10 bps (model); nonmonotonic, small observed shift; asymmetry explained by default risk channel operating only on Peripheral bonds |
| R5 | Credit risk premium dominance is robust across 12 alternative calibrations | Table III, p. 2427 | CRP share: 74% long-run, 81% PEPP impact; low-gamma and low-delta calibrations worsen model fit substantially |
| R6 | Pandemic shock (out-of-sample): model predicts large upward shift in Italian yields | Figure 9, p. 2431 | Model-predicted shift = 120-160 bps across maturities; broadly matches observed Feb-to-Mar 2020 data; mainly driven by credit risk premium increase |

**Overall (paper's conclusion).** The paper proposes a model in which default
risk opens a novel *default risk extraction channel* for central bank asset
purchases, allowing large-scale parallel yield curve shifts like those observed
in Italy during the Covid-19 pandemic. Asset purchases compress both the credit
risk premium (by extracting defaultable bonds from private markets) and the
expected default loss (endogenously, by relieving fiscal pressure). The
duration risk channel is secondary in the euro-area context. The model
identifies that credit risk premia, not expected losses, dominate sovereign
spreads, and that the flexible PEPP design substantially enhanced its impact
relative to the earlier, rigid APP.

## Theory / model

The economy has two countries (Core and Periphery) in continuous time with an
infinite horizon. Core issues risk-free bonds; Periphery may partially default
on its obligations (Section I, pp. 2395-2408).

**Short-rate process.** The instantaneous riskless rate $$r_t$$ follows an
Ornstein-Uhlenbeck process (eq. 1, p. 2396):

$$dr_t = \kappa(\bar{r} - r_t)dt + \sigma dB_t \tag{1}$$

**Bond prices.** Conjecture that bond prices are log-affine:

$$P_t(\tau) = e^{-[A_t(\tau)r_t + C_t(\tau)]}, \quad P_t^*(\tau) = e^{-[A_t^*(\tau)r_t + C_t^*(\tau)]} \tag{8}$$

where $$\tau$$ is maturity, $$P_t(\tau)$$ is the Peripheral bond price, and
$$P_t^*(\tau)$$ is the Core bond price (p. 2398).

**Yield decomposition (Proposition 1, p. 2400, eq. 16-17).** Peripheral
yields $$y_t(\tau)$$ decompose into four affine components:

$$y_t(\tau) = \underbrace{\frac{1}{\tau}\mathbb{E}_t\int_0^\tau r_{t+s}ds}_{y_t^{EX}(\tau)} + \underbrace{\frac{1}{\tau}\mathbb{E}_t\int_0^\tau \left\{A_{t+s}(\tau-s)\lambda_{t+s} - \frac{\sigma^2}{2}[A_{t+s}(\tau-s)]^2\right\}ds}_{y_t^{TP}(\tau)} + \underbrace{\frac{1}{\tau}\mathbb{E}_t\int_0^\tau \delta\psi_{t+s}ds}_{y_t^{DL}(\tau)} + \underbrace{\frac{1}{\tau}\mathbb{E}_t\int_0^\tau \xi_{t+s}ds}_{y_t^{CR}(\tau)} \tag{16}$$

where $$\psi_t$$ is the Peripheral default arrival rate, $$\delta$$ is the
haircut, and $$\xi_t = \gamma\psi_t\delta^2\int_0^\infty X_t(\tau)d\tau$$
is the credit risk premium price (eqs. 13, 15, pp. 2399-2400).

**Proposition 2 (low default risk, p. 2401).** When $$\psi \to 0$$, term
premia are equalized across countries and depend only on the *aggregate* net
bond supply in the monetary union, so Core and Peripheral term premia move
symmetrically with ECB purchases.

**Proposition 3 (default risk shifts short yields, p. 2402-2403).** In the
ergodic distribution with constant default rate, the Peripheral yield is:

$$y_t(\tau) = (\psi\delta + \bar{\xi}) + \frac{(1+\Xi)(1-e^{-\hat\kappa\tau})}{\tau\hat\kappa}r_t + \frac{\int_0^\tau [A(u)(\kappa\bar{r}+\bar\lambda) - \frac{1}{2}\sigma^2[A(u)]^2]du}{\tau} \tag{\text{Prop. 3}}$$

where $$\lim_{\tau\to 0}y_t(\tau) = (1+\Xi)r_t + (\psi\delta + \bar\xi)$$,
so the default-related constant $$\psi\delta + \bar\xi$$ shifts the entire
Peripheral yield curve in parallel, including the short end, independent of
changes in the short-term riskless rate (p. 2403).

**Endogenous default (Section II, pp. 2403-2408).** Following Corsetti and
Dedola (2016), the paper models rollover crises as in Calvo (1988) and Cole
and Kehoe (2000). In a rollover crisis (arrival rate $$\eta$$), the Peripheral
government compares the cost of repayment $$V_0^R$$ (discounted emergency
taxation) to the cost of default $$V_0^D$$ (fixed restructuring cost $$\chi$$)
and defaults if $$V_0^R > V_0^D$$:

$$\mathbb{P}(\text{default at time }0|\text{crisis}) \approx \Phi(V_0^R) \tag{25}$$

The unconditional default rate is $$\psi_t = \eta\Phi_t$$, where fiscal
pressure $$F_t$$ (eq. 29) aggregates primary deficits, bond redemptions, and
central bank remittances. Central bank purchases reduce $$F_t$$ (via
remittances), hence reduce $$\psi_t$$, reinforcing both duration and default
risk extraction (eq. 26-29, pp. 2405-2407).

## Method

The paper solves for equilibrium in two stages.

**One-factor analytical solution (Section I).** Market clearing (eq. 7,
p. 2397) combined with the arbitrageurs' first-order conditions (eqs. 10-13)
yields two integral equations for the factor loadings $$\lambda_t$$ (price of
interest-rate risk) and $$\xi_t$$ (price of default risk):

$$\lambda_t = \gamma\sigma^2\int_0^\infty\left[(S_t(\tau)-Z_t(\tau))A_t(\tau) + (S_t^*(\tau)-Z_t^*(\tau))A_t^*(\tau)\right]d\tau \tag{14}$$

$$\xi_t = \gamma\psi_t\delta^2\int_0^\infty (S_t(\tau)-Z_t(\tau))d\tau \tag{15}$$

When $$\psi_t$$ is a deterministic function of time, both sides of (15) are
affine in $$r_t$$, yielding time-varying affine solutions $$\lambda_t = \Lambda_t r_t + \bar\lambda_t$$ and $$\xi_t = \Xi_t r_t + \bar\xi_t$$.

**Multifactor quantitative model (Section III, pp. 2408-2416).** For
calibration, the model adds two mean-zero PH demand shifters $$\varepsilon_t^h$$
and $$\varepsilon_t^{h*}$$ to capture PH demand fluctuations:

$$Z_t(\tau) = h(\tau) - \varsigma(\tau)\varepsilon_t^h + \tau\alpha(\tau)(y_t(\tau) - \hat\delta\psi_t) \tag{30}$$

The factor vector $$q_t \equiv [r_t, \varepsilon_t^h, \varepsilon_t^{h*}]^\top$$
follows a multivariate Ornstein-Uhlenbeck process:

$$dq_t = -K(q_t - \bar{r}\mathcal{E}_1)dt + \Sigma dB_t \tag{31}$$

Bond prices remain log-affine in $$q_t$$: $$P_t(\tau) = e^{-[A_t(\tau)^\top q_t + C_t(\tau)]}$$, $$P_t^*(\tau) = e^{-[A_t^*(\tau)^\top q_t + C_t^*(\tau)]}$$ (eq. 32, p. 2409).

**Calibration.** Two-step procedure: (i) directly calibrate observable
parameters including the OU parameters of $$r_t$$ (mean $$\bar{r} = 1.22\%$$,
$$\kappa = 0.062$$, $$\sigma = 63$$ bps), fiscal variables, and Eurosystem
purchase paths from ECB data; (ii) estimate the remaining 10 parameters
(Table I, p. 2416) by minimizing a distance criterion over long-run yield
moments (1999-2022) and the two-day PEPP announcement window shift. The
numerical solution uses finite-difference methods for the time-varying PDE
system.

## Empirical specifications

**Long-run moments target.** The distance criterion minimizes the sum of
squared deviations between model ergodic distribution and data for (Section
III.C, pp. 2413-2414):

- Mean yields on 1m, 1Y, 5Y, 10Y German bonds and 1Y, 5Y, 10Y Italian bonds
- Standard deviations of yields at same maturities (both countries)
- Within-country correlations between 1Y and 10Y yields (Germany and Italy)
- Cross-country correlation of 1Y yields and cross-country correlation of 10Y yields

All yields expressed in annualized percentage points. Sample: January 1999 to
December 2022 (monthly zero-coupon yields from Datastream).

**PEPP announcement target.** The second component of the distance criterion
matches the observed yield curve shifts (March 18 to March 20, 2020) for 1m,
1Y, 5Y, 10Y, and 20Y German and Italian bonds. The model compares equilibrium
yields under the pre-PEPP and post-PEPP fiscal/purchase scenarios (Section
III.C, p. 2414). The scale of the Italian yield shift pins down the elasticity
parameter $$\theta$$ (slope of the default rate with respect to fiscal
pressure).

**Robustness (Table III, p. 2427).** For each of 12 alternative
parameterizations (varying $$\kappa$$, $$\sigma$$, $$\delta$$, $$\gamma$$,
$$\psi$$, $$\theta$$, $$\hat{r}+\phi$$, $$\sigma_h$$, $$\alpha_h$$, $$\zeta=0$$
cases), the paper reports model fit (sum of squared deviations), credit risk
premium share of the total default compensation in the long run and in the
PEPP episode, and mean 10Y yields for Germany and Italy. The benchmark credit
risk premium share of 74% (long run) and 81% (PEPP) is robust across all
plausible calibrations.

**Pandemic out-of-sample test (Figure 9, p. 2431).** The model compares
average weekly German and Italian zero-coupon yields from February 13-19, 2020
to those from March 12-18, 2020, four weeks later. These observations are not
used in estimation and serve as an out-of-sample check.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| German and Italian zero-coupon sovereign bond yields (Datastream) | Main estimation targets: yields at 1m, 1Y, 5Y, 10Y, 20Y maturities, Jan 1999-Dec 2022; also two-day PEPP announcement window | [no page yet](/wiki/datasets/) |
| ECB and Eurosystem bond holdings (ECB Data Portal / official ECB publications) | Calibrate net bond supply and Eurosystem purchase paths for APP and PEPP; maturity distribution of Eurosystem holdings as of July 2021 | [ECB Data Portal](/wiki/datasets/ecb-data-warehouse/) |
| Banco de Espana fiscal projections (in-house debt sustainability model) | Long-run fiscal forecasts for Italy and Germany (primary deficits, debt, interest) extended from two-year-ahead Eurosystem projections | [no page yet](/wiki/datasets/) |
| Eser et al. (2023) PH demand estimates | Calibrate the fraction of sovereign debt held by preferred-habitat investors (44.2% of net debt for each country) | [no page yet](/wiki/datasets/) |
| Cruces and Trebesch (2013) haircut evidence | Fix haircut parameter $$\delta = 0.25$$ consistent with international evidence on sovereign defaults | [no page yet](/wiki/datasets/) |

**Sample.** Monthly data, January 1999 to December 2022. For the PEPP
announcement, the estimation window is two trading days (March 18-20, 2020).

## When to read the full paper

Read the original if you are:

- Modeling the term structure of sovereign spreads in the euro area and need
  a tractable affine framework with credit risk.
- Studying how ECB asset purchase programs (APP, PEPP) transmit to yields in
  heterogeneous monetary unions via the default risk channel.
- Seeking the yield decomposition into expectations, term premium, expected
  default loss, and credit risk premium (Proposition 1, Table II for numbers).
- Interested in the identification strategy: Germany's negligible default risk
  identifies risk aversion from the term premium; Italy adds the spread to
  identify the long-run default rate and credit risk premium (pp. 2392, 2417-2421).
- Working on fiscal-monetary interactions with rollover crises, building on
  Calvo (1988) and Cole and Kehoe (2000) in a dynamic ATSM setting.

The key quantitative tables are Table I (estimated parameters), Table II
(10Y yield decomposition and Sharpe ratios), Table III (robustness), and
Table IV (conditional Sharpe ratios around the PEPP). Figures 4-9 document
model fit to long-run moments, the PEPP announcement, and the pandemic shock.

## Attribution and rights

Costain, James, Galo Nuno, and Carlos Thomas. "The Term Structure of Interest
Rates in a Heterogeneous Monetary Union." *The Journal of Finance* 80(4),
August 2025, 2389-2434. DOI: 10.1111/jofi.13463.

Published under Wiley VOR terms (paywalled). Rights held by the American
Finance Association. This page contains only distilled extracts (tables,
equations, and findings with locators). LLM-distilled; not human-verified;
not reproduced.


==============================================================================
# Financial Education of Executives: Custodio, Mendes & Metzger (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/custodio-impact-financial-education-executives-2025/
# Distilled: An RCT with 92 medium and large Mozambican firms shows that an 18-hour MBA-style corporate finance course for top executives causes firms to reduce working capital by 0.4 to 0.5 standard deviations (driven mainly by shorter accounts receivable collection periods), generating cash that is channeled into capital expenditure and raising ROA by 0.6 to 1.1 standard deviations. J. Finance 2025, CC BY 4.0. Eight core results with source locators, datasets used, the identification strategy, and the estimating equations.
# Tags: paper-summary, managerial-capital, corporate-finance, financial-education
==============================================================================

**What this is.** The paper's core results, the RCT identification strategy, and the estimating equations from a field experiment on financial education for top executives in Mozambique: enough to know what was found and how, without reading all 46 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13476).

## TL;DR

The paper reports an RCT with 92 medium and large Mozambican firms: 45 were randomly assigned to receive a free 18-hour MBA-style corporate finance course (covering capital budgeting, capital structure, working capital management, and risk management) in May 2017; 47 were assigned to a control group that received the same course later. Using difference-in-differences on panel accounting data (2008-2017) and a 15-month follow-up survey, the authors find that treated firms significantly reduced working capital (by 0.4 to 0.5 standard deviations), primarily through faster collection of accounts receivable (0.4 to 1 standard deviation reduction). The freed-up cash was invested: capital expenditures increased by 9 to 13 percentage points of assets. Firm performance improved substantially: ROA rose by 0.67 to 1.09 standard deviations, with effects persisting over two years. No negative effect on sales is found. The learning channel (rather than networking) is the most plausible mechanism: executives without prior finance experience benefit most.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Working capital falls significantly** in treated firms relative to controls | Table II, p. 2895 | Coefficient on Treatment x Post ranges from -0.157\* (col. 2) to -0.298\*\*\* (col. 4, external sample); full-sample preferred spec: -0.222\*\* (col. 3). Corresponds to -0.4 to -0.5 SD (SD = 0.45) |
| R2 | **Accounts receivable drives the working capital reduction** (collection period falls by 39 to 97 days) | Table V, p. 2902 | A/R coefficient: -0.106\* (col. 1, OLS) to -0.266\*\*\* (col. 2, firm FE); corresponds to -0.39 to -0.88 SD (SD = 0.27). Inventory coefficient: -0.037 to -0.058, less robust |
| R3 | **Capital expenditure rises significantly**: freed-up cash is invested | Table IV, p. 2900, cols. 7-9 | Coefficient 0.095\*\* (col. 7) to 0.133\*\* (col. 8); approximately 0.45 to 0.63 SD increase (SD = 0.21) |
| R4 | **Firm ROA increases by 0.14 to 0.23** in the post-treatment period | Table IV, p. 2900, cols. 10-12 | Coefficient 0.140\*\* (col. 11, firm FE) to 0.229\*\* (col. 10, OLS); significant at 5%; equivalent to 0.67 to 1.09 SD of ROA (SD = 0.21) |
| R5 | **Survey evidence confirms policy changes**: 55% of treated vs 7% of control firms implemented at least one financial policy change; working capital gap 25.6 pp (significant at 1%) | Table III, p. 2897 | Open-ended responses: 29% treated vs 4% control changed working capital management (diff. 25.6\*\*\*). Combined open- and close-ended: 77% vs 52% (diff. 25.7\*\*) |
| R6 | **Working capital effects persist two years** after the intervention | Table VI, p. 2906, col. 1 | Post-1-year coefficient: -0.249\*\* (full sample); post-2-year: -0.165\*\*; post-3-year: not significant (due to control group being treated) |
| R7 | **ROA effects persist two years** and are directionally positive through year three | Table VI, p. 2906, col. 4 | Post-1-year: 0.280\*\*; post-2-year: 0.324\*\*; post-3-year: 0.138 (not significant) |
| R8 | **Heterogeneous effects**: firms with smaller size, lower leverage, and executives without prior finance experience benefit most from the program | Figure 3, p. 2904 | Point estimates for ROA uniformly positive across all subgroups; significantly more pronounced for small firms vs large and for executives without prior finance experience |

**Overall (paper's conclusion).** An 18-hour MBA-style executive finance course for top managers of medium and large private firms in Mozambique caused measurable improvements in financial decision making, primarily through better working capital management (faster A/R collection). The resulting cash was channeled into long-term investment, and firm ROA increased substantially and persistently. The evidence suggests that managerial financial expertise is an important constraint on firm performance in environments with severe financial frictions, and that low-cost educational interventions can build managerial capital and relax these constraints.

## Theory / model

The paper has no formal structural model. The theoretical framework is based on the observation that in a frictionless world financial decisions are irrelevant (Modigliani-Miller), but in environments with financial frictions, the ability to make optimal financial decisions can positively affect firm value. Mozambique is chosen as the setting because financial frictions are severe and there is large heterogeneity in financial expertise among top executives.

The paper tests two hypotheses:

1. Providing financial education to top executives causes changes in firm financial policies (the education-to-policy channel).
2. These policy changes improve firm performance, because the pre-treatment policies were suboptimal due to constraints on managerial capital (the policy-to-performance channel).

The paper situates itself relative to two literatures. First, the CEO-characteristics literature (Bertrand and Schoar (2003); Custódio and Metzger (2014)) documents in non-experimental settings that CEOs' financial expertise correlates with more sophisticated financial policies. This paper provides experimental evidence for the same link. Second, the management-practices RCT literature (Bruhn and Zia (2013); Bloom et al. (2013)) shows training effects for microentrepreneurs and small firms; this paper extends it to top executives of medium and large firms and to financial (rather than operational) practices.

The competing explanations the design rules out are:

- **Signaling (Spence (1973))**: if education only signals pre-existing ability, no real change in policies should occur. The RCT design, keeping CEOs matched to their firms while randomly varying education, breaks the endogenous matching that drives signaling effects.
- **Networking**: a separate networking event for the control group (organized around the treatment dates) allows the authors to show that networking per se does not replicate the working capital or ROA changes observed in treated firms (p. 2878).

The identification strategy exploits the staggered delivery of the course: the treatment group (cohort 1) received the course in May 2017; the control group (cohort 2) received it in November 2018/April 2019. The post-treatment window is the year-end 2017 financial data for the treatment group, before the control group was treated. Firm-level randomization was stratified by industry and conducted at the business-group level to minimize contamination.

## Method

The primary estimator is a difference-in-differences (DID) intention-to-treat (ITT) regression. Standard errors are clustered at the business-group level (the level of randomization). The approach builds on `difference-in-differences` and `panel-regression`.

The ITT design avoids post-randomization selection bias: all 92 firms are included regardless of whether the assigned participant actually attended (4 of 45 treatment firms did not attend for idiosyncratic reasons; participation rate 91%).

Two data sources are used:

- **External data** (KPMG "Top 100 Companies in Mozambique" annual reports): externally obtained, independent of study participation, available for 83 of 92 firms. Used for main outcomes (working capital, leverage, ROA, sales).
- **Hand-collected data**: directly requested from firms for supplementary outcomes (A/R, A/P, cash, CapEx). Available for 65 of 92 firms (71%), collected in 2018 and again in 2023.

A pseudo-external sample of supplementary outcomes is constructed by restricting hand-collected data to firms for which external data are also available, to address attrition concerns (p. 2894). Lee bounds (Lee (2009)) are computed for all outcomes to bound treatment effects under nonrandom sample selection (Table VII, p. 2909).

As a complementary estimator for the heterogeneous-effects analysis (Section III.E), ANCOVA is used instead of DID: the post-treatment outcome is regressed on the treatment indicator and the pre-treatment mean of the outcome variable. This approach increases power in settings of low autocorrelation and small samples (McKenzie (2012)).

## Empirical specifications

**Main DID specification (eq. 1, p. 2894).** For a firm-level outcome $$Y_{it}$$:

$$
Y_{it} = \alpha + \beta_1 \times \text{Treatment}_i \times \text{Post}_t + \beta_2 \times \text{Treatment}_i + \beta_3 \times \text{Post}_t + X_{it} + \gamma_i + \gamma_t + \varepsilon_{it} \tag{1}
$$

where $$\text{Treatment}_i$$ is a dummy for assignment to the treatment group (randomized), $$\text{Post}_t$$ equals one in 2017 (the first post-treatment year-end), $$\gamma_i$$ are firm fixed effects, $$\gamma_t$$ are year fixed effects, and $$X_{it}$$ are optional controls (GDP growth interacted with $$\text{Treatment}$$; lagged log size, Mozambican nationality, accounting background, and finance background interacted with $$\text{Post}$$). The coefficient of interest is $$\beta_1$$, the ITT effect. Standard errors are clustered at the business-group level.

The sample period is 2008 to 2017 (full sample) for main outcomes and 2013 to 2017 for supplementary outcomes (hand-collected). Specification (1) in Table II is a cross-sectional OLS without firm or year FE; specification (2) adds firm and year FE; specification (3) adds controls; specification (4) repeats specification (3) on the external-only sample. These four variants are run for each outcome throughout the paper.

**Persistence specification (Table VI, p. 2906).** Specification (3) from Table II is re-estimated replacing $$\text{Post}_t$$ with $$\text{Post}_{t=1\text{yr}}$$, $$\text{Post}_{t=2\text{yr}}$$, and $$\text{Post}_{t=3\text{yr}}$$, corresponding to year-end 2017, 2018, and 2019. Only firms for which data exist in all post-treatment years are included.

**Working capital components (Table V, p. 2902).** The same DID specification is applied to accounts receivable (A/R / Sales), accounts payable (A/P / Assets), and inventories (Inventory / Assets), scaled by one-year lagged sales or assets. The headline result is the A/R effect: coefficient -0.106\* (OLS, col. 1) to -0.266\*\*\* (firm FE + controls, col. 2), corresponding to a reduction in the average collection period of 39 to 97 days from a pretreatment mean of 179 days for treatment firms.

**Heterogeneous effects (Figure 3, p. 2904).** ANCOVA regressions of post-treatment ROA on the treatment indicator, pre-treatment ROA mean, and subgroup indicators (separately for below/above median assets, employment, cash, leverage; Masters-or-higher dummy; prior CFO experience dummy; prior accounting/finance background dummy; discretion-over-policies dummy). Results show larger point estimates for smaller firms, lower-leverage firms, and executives without prior finance experience.

**Robustness.** Randomization-t p-values (Young (2019), Table IA.V); Lee bounds for nonrandom attrition (Table VII); winsorized outcome variables (Figure IA.8); ANCOVA instead of DID (Table IA.VI); bootstrap standard errors following Bloom et al. (2013) (Table IA.VI Panel B); results on balanced panel (Table IA.VII); exclusion of CEO-turnover firms (Table IA.X); alternative scaling of working capital (Table IA.XI).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| KPMG "Top 100 Companies in Mozambique" annual reports (2008-2019) | External accounting data: working capital, sales, leverage, ROA for 83 of 92 firms | No page yet |
| Hand-collected firm financial statements (2013-2019) | Supplementary accounting data: A/R, A/P, cash, CapEx for 65 of 92 firms; collected via face-to-face meetings and email | No page yet |
| Follow-up survey (November 2018) | Survey outcomes: stated policy changes in working capital, valuation, capital structure, and risk management; 15 months post-intervention | No page yet |
| Post-learning exit surveys (May 2017; November 2018/April 2019) | Intentions to change financial policies immediately after attending the course | No page yet |
| KPMG reports (nonparticipating firms) | External-validity benchmark: nonparticipating Mozambican firms for parallel-trend checks | No page yet |

Sample: 92 firms (45 treatment, 47 control) enrolled; up to 89 with at least one year of financial data. Accounting data spans 2008-2019 for external sources and 2013-2019 for hand-collected data. The outcome window for the main test is year-end 2017.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13476) if you are: (i) replicating the ITT estimates or running the Lee bounds / robustness variants (the Internet Appendix contains all supplementary tables); (ii) designing a similar executive education RCT and need the exact course content, recruitment procedure, or survey instruments; (iii) interested in the mechanisms section (V.B) and policy considerations (V.C) for external-validity arguments; or (iv) studying heterogeneous treatment effects by firm size, leverage, or executive background.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(5). This distillation was extracted by an LLM on 2026-06-05 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Custódio, Cláudia, Diogo Mendes, and Daniel Metzger.
> "The Impact of the Financial Education of Executives on the Financial Practices of Medium and Large Enterprises."
> *The Journal of Finance* 80, no. 5 (October 2025): 2875–2920.
> DOI: 10.1111/jofi.13476. © 2025 The Author(s).
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# ESG News, Future Cash Flows, and Firm Value: Derrien, Kruger, Landier & Yao (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/derrien-esg-news-future-cash-2025/
# Distilled: Using RepRisk ESG incident data and IBES analyst forecasts across 9,737 firms in 49 countries from 2008 to 2019, the paper shows that negative ESG news causes analysts to significantly downgrade earnings forecasts at short and longer horizons, driven primarily by expected sales declines rather than higher costs, and that forecast revisions can account for most of the negative impact of ESG incidents on firm value. J. Finance 2025, paywalled. Ten core results with source locators, datasets used, the model (Gordon / dividend discount decomposition), and the empirical specifications.
# Tags: paper-summary, esg, corporate-finance, analyst-forecasts, cash-flows
==============================================================================

**What this is.** The paper's core results, the valuation model used to
decompose ESG shocks into cash-flow and discount-rate components, and the
regression specifications behind each finding: enough to know what was found
and how, without reading the full 56 pages. To replicate or extend, read the
original at [doi.org/10.1111/jofi.13498](https://doi.org/10.1111/jofi.13498).

## TL;DR

Using RepRisk ESG incident data matched to IBES consensus analyst forecasts
across 9,737 firms in 49 countries (2008-2019, 81,749 ESG incidents), the
paper shows that negative ESG news triggers a significant, approximately
parallel downward shift in analyst EPS forecasts over all horizons from one
quarter to three years. The revision reflects primarily expected sales declines
(anticipated customer withdrawal) rather than higher costs, consistent with
the customer-demand channel emphasized in Servaes and Tamayo (2013) and
corroborated by retail store evidence in Duan, Li, and Michaely (2024). A
dividend discount decomposition following the framework of Hommel, Landier,
and Thesmar (2023) shows that forecast revisions can account for essentially
all of the negative stock-price reaction to ESG incidents, while implied
discount rates do not change significantly. This is consistent with Berk and
van Binsbergen (2024), who argue theoretically that ESG divestment has no
detectable cost-of-capital effect. The paper uses ESG news events from
RepRisk rather than ESG ratings (which suffer from the disagreement documented
by Berg, Koelbel, and Rigobon (2022)) because news events provide cleaner
identifiable shocks. Glosner (2021) documents that negative ESG shocks predict
negative future returns, suggesting underreaction; this paper studies the
analyst expectations channel that underlies that pattern. Analysts who
downward-adjust EPS forecasts following ESG incidents reduce their forecast
errors relative to those who do not, confirming the revisions are rational.
The paper also shows that the ESG effect on earnings forecasts persists over
longer horizons than effects from other negative events, consistent with the
permanent-shock interpretation suggested by Pedersen, Fitzgibbons, and Pomorski
(2021) for the cash flow channel of ESG information.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%.
Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Negative ESG incidents cause significant **parallel downward revision in EPS forecasts at all horizons** (Q1 through three years); the term structure is approximately flat | Table III Panel A, p. 3512; Figure 2, p. 3521 | Q1: -0.142\*\*\* (t=-2.09); one-year: -0.130\*\*\* (t=-3.08); two-year: -0.148\*\*\* (t=-3.76); three-year: -0.157\*\*\* (t=-4.18) |
| R2 | ESG incidents cause a significant **decline in stock returns and analyst-implied price target revisions** of similar magnitude to EPS revisions | Table III Panel A cols. 9-10, p. 3512 | PTG: -0.168\*\*\* (t=-6.20); Return: -0.177\*\*\* (t=-5.08) |
| R3 | **Multiple incidents amplify the effect**: firms with at least two incidents in months [t-6,t] see EPS declines roughly twice as large as firms with one incident | Table III Panel B, p. 3513 | 1 incident: -0.001 to -0.119 across horizons; 2+ incidents: -0.113 to -0.277 across horizons |
| R4 | **Social incidents have the strongest and most persistent effect** on EPS forecasts; environmental incidents are less significant; governance incidents also significant but smaller | Table IV Panels A-C, pp. 3514-3515 | S incidents one-year: -0.175\*\*\* (t=-4.23); G incidents one-year: -0.150\*\*\* (t=-3.13); E incidents one-year: -0.100 (t=-1.70, n.s.) |
| R5 | ESG incidents have a **longer-lived term structure** than other types of negative corporate events: the three-year impact is 21% higher than the one-year impact, while for other KD negative events the three-year impact is 42% lower than the one-year impact | Figure 2 p. 3521; Table VII p. 3523 | Three-year/one-year ratio: 1.21 for ESG vs. 0.58 for average KD events; F-test rejects equal term structures (p < 0.01) |
| R6 | **ESG-induced EPS revisions are driven primarily by expected sales declines** (not cost increases): sales forecast revisions are consistently negative across all horizons; gross margin revisions are smaller and less significant | Table VIII Panel A, p. 3524-3525 | Sales one-year: -0.036\*\*\* (t=-3.81); two-year: -0.055\*\*\* (t=-4.75); avg. decline ~0.051% per year; gross margin one-year: -0.027\*\* (t=-2.53) |
| R7 | **Cash flow changes, not discount rate changes, account for observed firm value declines**: a dividend discount decomposition shows the EPS-forecast-implied value change covers the stock return; implied discount rates do not change significantly | Table IX p. 3529 | At [t,t+3]: market return -0.30% (t=-1.84), forecast-implied value change -0.41% (t=-2.16); discount rate change -0.01% (t=-0.11, n.s.) |
| R8 | **Effect is stronger for smaller firms** and stronger (though not significantly so) for B2C industries with high advertising intensity | Table XI p. 3539; Table XII p. 3541 | Large-firm interaction coefficient 0.670\*\*\* (t=5.39) in EPS regressions, implying small-firm effect roughly 0.67 p.p. larger; B2C interaction negative and economically meaningful at one- and two-year horizons |
| R9 | **Realized firm earnings and sales decline after ESG incidents**: net income decreases by 8.8% to 11.8% and sales by 1.2% to 4.2% in the year(s) following incidents, confirming analysts are correct | Table XIII Panels A-B, pp. 3543-3544 | Earnings [t-1 to t]: -0.088\*\*\* (t=-4.79); [t-1 to t+1]: -0.118\*\*\* (t=-4.79); Sales [t-1 to t+1]: -0.026\*\*\* (t=-6.66) |
| R10 | **Analysts who downward-adjust EPS forecasts after ESG incidents reduce their forecast errors**: interaction of downward adjustment and ESG incident is negative and significant for annual and two/three-year horizons | Table XIV p. 3547 | One-year horizon: -0.002\*\* (t=-2.84); two-year: -0.003\*\*\* (t=-3.56); three-year: -0.004\*\*\* (t=-3.98) |

**Overall (paper's conclusion).** Negative ESG news generates a permanent
downward shift in analyst earnings expectations driven primarily by anticipated
customer withdrawal (reduced future sales), not by higher costs. This cash flow
effect can quantitatively account for most of the negative stock-price response
to ESG incidents, while no significant change in cost of capital is detected.
The downward revisions are rational: realized earnings and sales drop after ESG
incidents, and analysts who revise downward make more accurate forecasts.

## Theory / model

The paper has no structural economic model. The theoretical framework is a
dividend discount decomposition to separate cash-flow from discount-rate
effects of ESG shocks on firm value.

**Hypotheses tested.** ESG information can affect firm value through two
channels: (1) a *cash flow channel*, where ESG incidents signal lower future
earnings because customers avoid firms with poor ESG profiles, or because the
firm cannot instantaneously adjust its production technology; and (2) a
*discount rate channel*, where divestment by ESG-conscious investors raises
the cost of capital.

**Gordon growth-formula pass.** As a first pass (Section IV.A, p. 3527), the
equity value of firm $$i$$ at time $$t$$ follows Gordon's formula for a growing
perpetuity:

$$
PV_{it} = \frac{b_i F_t\text{EPS}_{i,t+1}}{r_{it} - g_{it}} \tag{3}
$$

where $$b_i$$ is the payout ratio, $$F_t\text{EPS}_{i,t+1}$$ is the one-year
earnings forecast, $$r_{it}$$ is the discount rate, and $$g_{it}$$ is the
expected earnings growth rate. The theoretical firm-level return induced by an
ESG information shock is

$$
\frac{\Delta PV_{it}}{PV_{it}} = \frac{\Delta F_t\text{EPS}_{i,t+1}}{F_t\text{EPS}_{i,t+1}} - \frac{\Delta r_{it} - \Delta g}{r_{it} - g_{it}}
$$

Because Table III shows that the ESG impact on LTG is economically and
statistically insignificant (column (8)), the growth term drops out and
changes in earnings forecasts should equal changes in firm value. The
similarity between the EPS revision coefficient and the stock return
coefficient (Table III Panel A, cols. (5)-(7) vs. (10)) confirms this.

**Discounted dividends decomposition.** The more formal approach (Section IV.B,
pp. 3527-3528) uses the present value of near-term earnings payouts:

$$
\frac{PV_{it}(r_{it})}{b_i} = \frac{F_t\text{EPS}_{i,t+1}}{(1+r_{it})^{\theta_{it}}} + \frac{F_t\text{EPS}_{i,t+2}}{(1+r_{it})^{\theta_{it}+1}} + \frac{F_t\text{EPS}_{i,t+3}}{(1+r_{it})^{\theta_{it}+2}} + \frac{1}{(1+r_{it})^{\theta_{it}+2}} \cdot \frac{(1+g_t)F_t\text{EPS}_{i,t+3}}{r_{it}-g_t} \tag{4}
$$

where $$\theta_{it}$$ is the fraction of the fiscal year remaining, $$b_i$$ is the
rolling-average industry payout ratio, and $$g_t$$ is the expected long-run
nominal GDP growth from macro forecasters. The implied discount rate $$r_{it}$$ is
the solution to

$$
PV_{it}(r_{it}) = P_{it} \tag{5}
$$

where $$P_{it}$$ is the observed stock price. The authors compute the forecast-
implied value change when EPS forecasts are updated at each post-event window,
holding $$r_{it}$$ fixed, and compare it to the actual market return and to
changes in the implied discount rate (Table IX, p. 3529).

## Method

The paper applies a standard panel fixed-effects framework augmented with an
event-study design for the valuation decomposition. It does not propose a new
method. The approach builds on `panel-regression` and `event-study`.

**Baseline analyst reaction regression.** For each forecast horizon $$h$$,
equation (1) (p. 3509) is estimated:

$$
\frac{\Delta F_t\text{EPS}_{i,t+h}}{\text{abs}(F_{t-1}\text{EPS}_{i,t+h})} = \alpha + \beta \, \mathbf{1}[\text{ESG incidents in } [t-6,t]] + \gamma_{\text{Country}\times\text{Industry}\times t} + \sigma_i + \epsilon_{i,t} \tag{1}
$$

The dependent variable is the month-over-month change in consensus EPS
forecast scaled by the absolute value of the prior month's consensus forecast.
The main independent variable is an indicator equal to one if RepRisk reports
at least one ESG incident in the six months prior to month $$t$$. The
specification includes firm fixed effects ($$\sigma_i$$) and industry-by-country-
by-month fixed effects ($$\gamma_{\text{Country}\times\text{Industry}\times t}$$).
Standard errors are double-clustered at the firm and month levels.

**Mechanism regressions.** The same specification as equation (1) is estimated
replacing the EPS forecast change with the change in consensus sales forecasts
($$\frac{\Delta F_t\text{Sales}_{i,t+h}}{F_{t-1}\text{Sales}_{i,t+h}}$$) and
the change in gross margin forecasts
($$\frac{\Delta F_t\text{GrossMargin}_{i,t+h}}{F_{t-1}\text{GrossMargin}_{i,t+h}}$$)
to separate the sales from the cost channel (Table VIII, p. 3524).

**Term structure comparison.** To compare ESG incidents with other negative
corporate events, equation (2) (p. 3521) is estimated pooling horizons and
testing whether the slope of the term structure differs:

$$
\frac{\Delta F_t\text{EPS}_{i,t+h}}{\text{abs}(F_{t-1}\text{EPS}_{i,t+h})} = \alpha + \beta \, \mathbf{1}[\text{ESG incidents in } [t-6,t]] + \eta \, \mathbf{1}[\text{KD Negative Events in } [t-6,t]] + \gamma_{\text{Country}\times\text{Industry}\times t} + \sigma_i + \epsilon_{i,t} \tag{2}
$$

**Event-study for the valuation decomposition.** Equation (6) (p. 3528)
estimates how the forecast-implied value change, actual return, and implied
discount rate evolve over months $$s = 0, 1, \ldots, 6$$ following an ESG
event month:

$$
y_{t,t+s} = \alpha + \beta \, \mathbf{1}[\text{ESG incidents in month } t] + \gamma_{\text{Country}\times\text{Industry}\times t} + \text{Controls} + \epsilon_{i,t} \tag{6}
$$

Controls include firm size and book-to-market quintile dummies. Standard errors
are double-clustered at the firm and month level.

**Realized fundamentals regression.** To test whether analysts are correct,
equation (7) (p. 3538) uses annual data:

$$
\frac{Y_{i,t+h} - Y_{i,t-1}}{Y_{i,t-1}} = \alpha + \beta \, \mathbf{1}[\text{ESG incidents between year } t-1 \text{ and } t] + \gamma_{\text{Country}\times\text{Industry}\times t} + \sigma_i + \epsilon_{i,t} \tag{7}
$$

where $$Y_{i,t}$$ denotes realized annual earnings, sales, or gross margin.

**Analyst accuracy regression.** Equation (8) (p. 3545) uses the analyst-firm
panel to compare forecast accuracy for analysts who do versus do not downward-
adjust following ESG incidents:

$$
\frac{|FEPS_{i,e,j,t} - EPS_{i,e}| - |FEPS_{i,e,j,t-1} - EPS_{i,e}|}{|EPS_{i,e}|} = \alpha + \eta \, \text{DownwardAdj}_{i,e,j,t} + \beta \, \text{DownwardAdj}_{i,e,j,t} \times \mathbf{1}[\text{ESG incidents of firm } i \text{ in } [t-6,t]] + \gamma_{i,e,t} + \epsilon_{i,e,j,t} \tag{8}
$$

where $$FEPS_{i,e,j,t}$$ is analyst $$j$$'s EPS forecast for firm $$i$$ earnings
announcement $$e$$ in month $$t$$, and $$\gamma_{i,e,t}$$ indicates firm-by-earnings-
announcement-by-month fixed effects.

## Empirical specifications

All regressions use panel data at the firm-month level (or firm-year for the
realized-fundamentals tests). The headline specifications are:

- **EPS/Sales/GrossMargin forecast regressions (R1, R3, R4, R5, R6, R8)**:
  Equation (1) with firm FE and industry-by-country-by-month FE, double-
  clustered standard errors, horizon $$h$$ = Q1, Q2, Q3, Q4, one-year,
  two-year, three-year. Sample: 2008-2019, global, 9,737 firms in 49 countries.
  EPS observations: 2,976,889; sales: 2,831,931; gross margin: 1,442,110.
  The indicator for ESG incidents is cumulated over months $$[t-6,t]$$ (the
  main specification; Internet Appendix Tables IA.III-IV show robustness to
  $$[t-3,t]$$ and $$[t-9,t]$$). The ESG incident variable has 10.44% of
  firm-months with at least one incident (6.57% exactly one, 3.87% at least two).

- **Term structure test (R5)**: Equation (2) pools one-, two-, and three-year
  horizons; the F-test in columns (4)-(5) of Table VII (p. 3523) rejects equal
  slopes for ESG vs. average KD events (p < 0.01).

- **Valuation decomposition event study (R7)**: Equation (6), US firms only
  (needed for the payout-ratio and growth rate computation). Table IX (p. 3529)
  reports cumulative event-window coefficients for windows $$[t,t]$$ through
  $$[t,t+6]$$. Standard errors double-clustered by firm and month.

- **Realized outcomes (R9)**: Equation (7) at the firm-year level, annual data,
  firm FE and country-by-industry-by-year FE. Dependent variable is percentage
  change in earnings, sales, or gross margin over one- and two-year windows.

- **Analyst accuracy (R10)**: Equation (8), analyst-firm panel. Firm-by-
  earnings-announcement-by-month FE. Observations: 2.6M (Q1) to 3.2M (three
  years). Results are negative and significant for annual and longer horizons.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| RepRisk ESG incident data | Main independent variable: daily negative ESG incidents at firm level, 2007-2019, 28 ESG issue categories, novelty/severity/reach scores | [RepRisk](/wiki/commercial/reprisk/) (licensed) |
| IBES consensus analyst forecasts | Dependent variable: EPS, sales, gross margin, LTG, and price target consensus forecasts at firm-month level, quarterly and annual horizons | [I/B/E/S](/wiki/commercial/ibes/) (licensed) |
| CRSP / Compustat (via WRDS) | Stock returns (daily), firm fundamentals (annual), book-to-market, market cap, payout ratios | [WRDS](/wiki/commercial/wrds/) (licensed) |
| Capital IQ Key Developments | Comparison non-ESG corporate events (153 types); identify 33 types with significant negative EPS impact | No page yet |
| Refinitiv (Asset4), Morningstar Sustainalytics, MSCI ESG scores | Validation that RepRisk incidents correlate with major ESG ratings (Appendix) | No page yet |

Sample: January 2008 to December 2019 (monthly). Final sample: 744,858 unique
firm-month observations; 9,737 firms in 49 countries; 81,749 ESG incidents.
EPS forecasts: 2,976,889 observations; sales: 2,831,931; LTG: 253,735; PTG: 688,899.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13498) if you are:
studying the cash-flow versus discount-rate channel for ESG-related firm value
changes; working with RepRisk data and need to understand its term structure
and heterogeneity properties; building models of ESG-driven analyst
expectations; or extending the analysis to additional horizons, industries,
or event types beyond the 33 KD categories examined. The tables above provide
exact locators for each result.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(6), December 2025.
This distillation was extracted by an LLM on 2026-06-03 and is **not
human-verified or independently reproduced**. The paper is paywalled; only
extracts are reproduced here.

> Derrien, Francois, Philipp Kruger, Augustin Landier, and Tianhao Yao.
> "ESG News, Future Cash Flows, and Firm Value."
> *The Journal of Finance* 80, no. 6 (December 2025): 3499-3554.
> DOI: 10.1111/jofi.13498. Published by Wiley on behalf of the American
> Finance Association. All rights reserved. Extract-only reproduction.


==============================================================================
# Conflicting Priorities: Donaldson, Gromb & Piacentino (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/donaldson-conflicting-priorities-theory-covenants-2025/
# Distilled: A theory of why firms use secured debt, unsecured debt, and negative pledge covenants together, despite covenants being defeated by collateral priority. The model shows covenants and collateral are complementary tools: collateral implements efficient dilution that covenants alone cannot, while covenants commit the borrower not to use collateral when dilution is inefficient. The optimal debt structure is multilayered, consistent with observed covenant violations and waivers. J. Finance 2025, paywalled. Five core propositions with source locators, the three-date model, and the mechanism.
# Tags: paper-summary, corporate-finance, debt-structure, covenants, collateral
==============================================================================

**What this is.** The paper's five main propositions, the three-date model
with its key equations, and the mechanism linking collateral, covenants, and
investment efficiency: enough to understand what it proved and how, without
reading all 30 pages. To replicate or extend it, read the full source at the
[original](https://doi.org/10.1111/jofi.13445).

## TL;DR

The paper develops a theory of debt structure in which collateral and negative
pledge covenants are complementary tools for managing the over/underinvestment
trade-off. A borrower faces two frictions: limited pledgeability (private
benefits cannot be pledged) and nonexclusive contracting (covenants can be
violated by taking on new secured debt that retains priority). Because
collateral trumps covenants, negative pledge covenants have no teeth on their
own: a new secured debt issue retains its priority even when it violates a
covenant. Yet covenants are useful: the threat of acceleration, though
generally not credible when all debt is covenant-protected, becomes credible
when only some debt is. The optimal debt structure is multilayered, combining
secured and unsecured debt with and without covenants, and it is always
first-best efficient in equilibrium. Covenants are violated and waived on the
equilibrium path, consistent with observed practice.

## Core results

Propositions are the primary units; locators point into the source PDF.

| # | Result | Locator | Key condition |
|---|---|---|---|
| R1 | Unsecured debt implements first-best | Prop. 1, p. 1747, eq. (5)-(6) | Private benefit of low-quality project $$Y_1^L \le Y_1^*$$, or total expected cash flows exceed funding needs; else overinvestment temptation prevents efficiency |
| R2 | Secured debt implements first-best | Prop. 2, p. 1748, eq. (7)-(8) | $$X_1^H \ge X_1^L$$: high-quality project has higher pledgeable cash flow; "mild" underinvestment problem |
| R3 | Covenants are irrelevant when all debt is covenant-protected | Prop. 3, p. 1749, eq. (9) | Acceleration threat is self-defeating: forcing liquidation subsidizes secured debt without deterring overinvestment; covenants have no bite unless $$\phi \le \phi^*$$ |
| R4 | Mix of covenant-protected and unprotected debt implements first-best | Prop. 4, p. 1750, eq. (10) | $$X_1^L \ge X_1^H$$: "severe" underinvestment problem; covenant fraction $$\phi$$ must be small enough that acceleration is credible only when Q=L |
| R5 | Equilibrium debt structure is first-best efficient | Prop. 5, p. 1751 | Always achievable; instrument choice (secured vs. covenant-protected unsecured) depends on severity of underinvestment problem |

**Overall (paper's conclusion).** Collateral and covenants implement
efficiency only in concert: covenants commit the borrower not to use
collateral when dilution is inefficient (bad dilution); collateral is needed
to break that commitment and engage in good dilution when covenants would
otherwise block efficient investment. Covenant violations and waivers are
on-path, not failures of contracting. The results speak to the policy debate
on debt priority: strong priority rules are useful because they let borrowers
dilute when, but only when, it is efficient to do so.

## Theory / model

The model has three dates $$t \in \{0, 1, 2\}$$ and a borrower B with two
sequential projects (Section I, pp. 1743-1746). Closest in spirit to this
paper is Ayotte and Bolton (2011), who also study negative pledge covenants
and property versus priority rights. The paper extends that analysis by
rationalizing covenant violations and waivers, and by showing covenants and
collateral are complementary rather than substitutes. The framework follows
Hart and Moore (1995) in using hard claims (debt) to constrain investment.
The collateral-overhang problem analyzed in Donaldson, Gromb, and Piacentino
(2020a) is the baseline: this paper adds covenants as an additional instrument
alongside collateral. Policy implications connect to the debate in Bebchuk
and Fried (1996) about the efficiency of strong priority for secured creditors.
Collateral and capital structure evidence from Rampini and Viswanathan (2013)
is cited in the empirical discussion (p. 1756).

**Projects.** Project 0 costs $$I_0$$ at Date 0 and succeeds with probability
$$p$$, yielding cash flow $$X_0 > 0$$ and private benefit $$Y_0 > 0$$; its value is
positive (eq. 1, p. 1743):

$$
p(X_0 + Y_0) > I_0. \tag{1}
$$

Project 1 costs $$I_1$$ at Date 1, succeeds with probability $$p$$, and yields
cash flow $$X_1^Q > 0$$ and private benefit $$Y_1^Q > 0$$ depending on quality
$$Q \in \{H, L\}$$ revealed at Date 1. Project 1 has positive value only if
$$Q = H$$ (eq. 2, p. 1744):

$$
p(X_1^H + Y_1^H) > I_1 > p(X_1^L + Y_1^L). \tag{2}
$$

**Frictions.** Two frictions generate a role for both collateral and covenants
(Section I.B, p. 1744):
1. *Limited pledgeability*: private benefits $$Y_t$$ cannot be pledged to
   creditors; only cash flows $$X_t$$ are pledgeable.
2. *Nonexclusive contracting*: existing creditors cannot prevent B from
   contracting with new creditors at Date 1.

**Instruments.** Three financing instruments exist (pp. 1744-1745):
1. *Secured debt*: face value $$F^s$$, collateral gives absolute priority over
   unsecured claims.
2. *Unsecured debt*: face value $$F^u$$, no collateral.
3. *Covenant-protected (unsecured) debt*: unsecured but grants the right to
   accelerate if B takes on new secured debt.

**Priority rules.** Secured debt has priority over unsecured; earlier secured
debt has priority over later secured debt; earlier unsecured (or accelerated)
debt has priority over later unsecured debt (p. 1745).

**Assumptions.** Under the efficient investment policy, expected cash flows
exceed funding needs (Assumption 1, eq. 3, p. 1746):

$$
pX_0 - I_0 + q(pX_1^H - I_1) \ge 0. \tag{3}
$$

Liquidation value suffices to repay secured debt needed to finance Project 1
(Assumption 2, eq. 4, p. 1746):

$$
p\!\left(X_0 + X_1^Q\right) > \frac{I_1}{p}. \tag{4}
$$

The first-best policy is to undertake both projects and invest in Project 1
if and only if $$Q = H$$ (Lemma 1, p. 1746). The paper derives conditions on
the debt structure at Date 0 under which this first-best is achieved as a
subgame perfect equilibrium.

**Key tension.** Unsecured debt allows dilution at Date 1 by new secured
creditors, which relaxes financial constraints (good dilution when $$Q = H$$)
but also enables overinvestment (bad dilution when $$Q = L$$). Secured debt
at Date 0 limits dilution capacity, preventing bad dilution but potentially
causing underinvestment. The central result is that a multilayered structure
combining instruments can implement the first-best whenever either instrument
alone cannot.

## Method

The paper uses a three-date contracting model solved by backward induction to
subgame perfect equilibrium, with competitive creditors who earn zero profit
in equilibrium (Section I.C, p. 1745). All contracts, including covenant
violations, are observable. B has full bargaining power in renegotiations
(p. 1746). The solution method is to:

1. Characterize the Date 1 subgame equilibrium given any Date 0 debt structure
   (which instruments, what face values, what covenant fraction $$\phi$$).
2. Derive necessary and sufficient conditions on the Date 0 debt structure for
   the first-best to obtain in every subgame.
3. Show that a date-0 structure satisfying those conditions always exists
   (Proposition 5).

The model builds on `principal-agent` contracting with nonexclusivity (no
exclusivity is enforceable via covenants because new secured debt retains
priority regardless of covenant violations). It uses `real-options` logic
in the sense that the right to dilute existing debt is like an option held by
the borrower; the debt structure determines when that option is valuable and
when it should be exercised.

**Covenant irrelevance result (Proposition 3).** If all unsecured debt at
Date 0 is covenant-protected (fraction $$\phi = 1$$), the acceleration threat is
generically not credible. The covenant-protected creditor's benefit from
acceleration (leapfrogging unprotected unsecured debt) is zero when $$\phi = 1$$
because there is no unprotected unsecured debt to leapfrog. The threshold
fraction at which acceleration becomes credible is (eq. 9, p. 1749):

$$
\phi^* := 1 - \frac{(1-p)I_1/p}{p(X_0 + X_1^L - I_1/p)} \in (0,1). \tag{9}
$$

Acceleration is credible only if $$\phi \le \phi^*$$. Above $$\phi^*$$, covenants
have no bite.

**Proposition 4 condition.** For covenants (at fraction $$\phi \le \phi^*$$) to
implement the first-best when secured debt cannot, a sufficient (and under
additional conditions necessary) condition is (eq. 10, p. 1750):

$$
X_1^L \ge X_1^H. \tag{10}
$$

This says negative-value projects ($$Q = L$$) have larger pledgeable cash flows
than positive-value projects ($$Q = H$$). Intuitively, the covenant-protected
creditor has more to gain from accelerating against a bad project (larger cash
flows to grab) than a good one, making the threat selective.

## Empirical specifications

This is a pure theory paper with no empirical estimation. Section IV
(pp. 1754-1756) describes empirical relevance and new predictions.

**Consistency with stylized facts.** The model is consistent with four
documented patterns (p. 1754-1755):
- Well-capitalized/highly rated firms rely heavily on unsecured debt (consistent
  with Proposition 1 and Rauh and Sufi (2010) and Benmelech, Kumar, and
  Rajan (2024)).
- Negative pledge covenants are common in roughly 44% of debt contracts
  (consistent with Proposition 4 and Billett, King, and Mauer (2007) and
  Ivashina and Vallee (2018)).
- Covenants are frequently violated and renegotiated or waived (consistent with
  Proposition 5; citations include Beneish and Press (1993, 1995) and
  Dichev and Skinner (2002)).
- Covenants in some debt decrease the yield on other debt by reducing default
  risk (consistent with Proposition 4; Bradley and Roberts (2015)).

**Untested predictions.** Propositions 1-5 imply four new predictions not yet
directly tested (Predictions 1-4, pp. 1755-1756):
- Firms more exposed to underinvestment (growth opportunities, high fixed costs,
  nonredeployable assets) use covenants more.
- Firms more exposed to overinvestment (distressed firms, declining industries)
  use collateral more.
- Collateral use increases and covenant use decreases with asset tangibility.
- Covenant use decreases with the costs associated with asset sales (less
  redeployable, harder to value, more firm-specific assets).

## Datasets used

This is a pure theory paper. No datasets are used in the analysis. The
empirical discussion in Section IV cites existing empirical studies (Billett,
King, and Mauer (2007); Rauh and Sufi (2010); Benmelech, Kumar, and Rajan
(2024); Beneish and Press (1993, 1995)) but does not reanalyze any data.

| Dataset | Role in paper | Wiki page |
|---|---|---|
| None (theory paper) | N/A | N/A |

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13445) if you are:
building models of debt structure with multiple creditors and priority rules;
analyzing why negative pledge covenants exist despite being defeated by
collateral priority; studying the design of optimal debt contracts when
pledgeability is limited and contracting is nonexclusive; or working on the
policy debate about the efficiency of strong priority for secured creditors.
The proofs in Appendix A (Lemmas A.1-A.18, pp. 1757-1763) are the formal
foundation for all five propositions.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(3), June 2025, pp. 1739-1768.
DOI: 10.1111/jofi.13445. Published under Wiley VOR terms; no CC licence.
This distillation was extracted by an LLM on 2026-06-06 and is **not
human-verified or independently reproduced**. Extract-only: the verbatim PDF
is not hosted here.

> Donaldson, Jason Roderick, Denis Gromb, and Giorgia Piacentino.
> "Conflicting Priorities: A Theory of Covenants and Collateral."
> *The Journal of Finance* 80, no. 3 (June 2025): 1739-1768.
> DOI: 10.1111/jofi.13445. © 2025 the American Finance Association.


==============================================================================
# Sustainability or Greenwashing: Duchin, Gao & Xu (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/duchin-sustainability-greenwashing-evidence-asset-2025/
# Distilled: Firms divest pollutive plants in response to environmental pressures without any reduction in pollution levels, consistent with a greenwashing divestiture strategy. Sellers gain higher ESG ratings and lower EPA enforcement costs while buyers are firms with weaker environmental pressures and pre-existing business ties to the sellers. J. Finance 2025, CC BY 4.0. Nine core results with source locators, datasets used, the conceptual framework, and the empirical specifications.
# Tags: paper-summary, esg, greenwashing, environmental-finance, industrial-pollution
==============================================================================

**What this is.** The paper's core results, the conceptual framework it builds on (Coasean real-asset-market equilibrium with heterogeneous environmental pressures), and the empirical strategy (multilogit, DID Poisson, stacked DID, BERT text analysis, event-study CARs): enough to know what it found and how, without reading all 56 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1111/jofi.13412).

## TL;DR

The paper studies the market for pollutive industrial plants. Using 888 divestitures of EPA Toxic Release Inventory (TRI) plants from 2000 to 2020, the authors show that: (i) firms facing stronger environmental pressures (ESG ratings, pension fund holdings, Democratic-leaning counties, RepRisk incidents) are significantly more likely to divest pollutive plants; (ii) buyers face weaker environmental pressures and have preexisting supply-chain or joint-venture ties with sellers; (iii) despite these transfers, pollution levels do not decline at sold plants or across the combined buyer-seller portfolio; (iv) sellers gain substantially in ESG ratings, EPA enforcement cost reductions, and announcement returns, while disclosing improved environmental performance in conference calls; and (v) buyers capture higher gains in the most pollutive deals. The evidence is consistent with a greenwashing strategy: firms exploit information frictions to cosmetically redraw firm boundaries along their value chains without achieving real environmental improvements.

## Core results

Magnitudes and significance are as reported; `\*\*`/`\*\*\*` = 5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Environmental pressures increase the propensity to divest pollutive plants; divestiture is the dominant response relative to closure or abatement | Table III Panel A, p. 717; Table II, p. 715 | Pressure Index: 0.6 pp higher divestiture likelihood per 1-SD (32% relative); ESG rating (Rated): 1.2 pp more likely to divest; Env. Event: +141% relative magnitude vs. other responses |
| R2 | Pollution levels amplify the sensitivity: heavier polluters are significantly more likely to divest in response to environmental pressures | Table III Panel B, p. 717 | Interquartile rise in total toxic release raises sensitivity of divestitures to Pressure Index by 7.7 (= 2.569 x 3), implying 2.3 pp higher divestiture likelihood per 1-SD increase in Pressure Index |
| R3 | Quasi-exogenous RepRisk environmental incidents sharply increase divestiture likelihood without pretrends | Table IV, p. 719; Figure 2, p. 720 | Env. Event coefficient = 1.077\*\* (SE 0.542); probability of divesting rises by ~1.5 pp immediately after incident (Figure 2); social/governance events have no effect |
| R4 | Pollution does not decline at divested plants or across the combined buyer-seller portfolio following divestitures | Table VI Panels A-B, pp. 723-724; Figures 3-4, pp. 725-726 | All Divested x Post coefficients statistically indistinguishable from zero; result holds in GDID and stacked regressions, for total pollution and intensity, with and without state/industry-year FE |
| R5 | Sellers' ESG and environmental ratings increase substantially following divestitures | Table XI Panel A, p. 738 | CSR Score: +0.302\* (GDID); Environmental Score: +0.234\*\*\* (GDID, stacked); environmental score gain = ~160% of sample mean |
| R6 | EPA enforcement actions and compliance costs decline sharply for sellers following divestitures | Table XI Panel B, pp. 738-739 | Enforcement action probability: -5 pp (vs. sample mean 7 pp); average enforcement costs fall to ~3.6% of pre-divestiture level (e^-3.33) |
| R7 | Announcement CARs are higher for divestitures of more pollutive plants, consistent with market recognition of gains from offloading pollution | Table XII, p. 742 | Interquartile increase in pollution -> 3 to 4 pp higher CAR[-1,+1]; sample average CAR = 2.5 pp; results hold for market and FF benchmarks |
| R8 | In the most pollutive deals, buyers capture ~$400M more value than sellers; pattern reverses for least pollutive deals | Figure 6, p. 744 | Top-pollution-quartile divestitures: buyers earn ~$400M higher gains (market model); bottom-quartile: sellers earn $600-700M more than buyers; consistent with comparative advantage in operating pollutive assets |
| R9 | More opaque firms (higher information asymmetry) are more likely to divest pollutive plants in response to environmental pressures | Table VIII, p. 731 | Pressure Index x #Segments: 0.823\*\*; x #Industries: 2.102\*\*; x #Subsidiaries: 0.018\*; x #Layers: 1.773\*\*; x %Blockholders: -5.610\*; all significant at 10% or better |

**Overall (paper's conclusion).** The real asset market allows firms to respond to environmental pressures by divesting pollutive plants along their supply chains, thereby improving ESG ratings and reducing regulatory compliance costs without losing access to the assets and without reducing overall pollution. These findings are more consistent with a greenwashing strategy than with genuine environmental improvement: ESG rating agencies, environmental regulators, and prosocial investors fail to recognize that divestitures of pollutive assets are ineffective conduits for reducing industrial pollution.

## Theory / model

The paper has no formal mathematical model. It develops a conceptual framework built on the Coase (1937) insight that variation in environmental costs across firms will induce some firms to sell and others to buy pollutive assets. The key predictions come from three premises:

1. Firms that face stronger environmental pressures (from investors, regulators, or the public) find it optimal to divest pollutive assets; firms that face weaker pressures find it optimal to hold or acquire them. Environmental pressures stem from prosocial preferences of stakeholders: Hart and Zingales (2017) argue firms should maximize shareholder welfare, Hartzmark and Sussman (2019) show investor ESG preferences generate real firm responses, Starks, Venkat, and Zhu (2017) document the role of pension funds in pressing firms on ESG, and Broccardo, Hart, and Zingales (2022) model how prosocial investors use divestment as a mechanism. In a real-asset-market equilibrium following Jovanovic and Braguinsky (2004), the market clears by transferring pollutive assets from high-pressure to low-pressure firms.

2. Information asymmetry is a prerequisite for greenwashing. Corporate managers can exploit information frictions - outsiders' difficulty monitoring firm-level environmental performance - to divest pollutive assets cosmetically along supply chains that maintain their access to the divested assets, without actual abatement. Following Demsetz and Lehn (1985) and Duchin, Matsusaka, and Ozbas (2010), more complex ownership and organizational structures raise the information costs outsiders incur, making greenwashing divestiture more feasible (and therefore more prevalent) in such firms. Gormley and Matsa (2011) provide a precedent for firms responding to external liability risk by restructuring asset ownership.

3. Business ties (supply chain relations and joint ventures) between buyers and sellers lower counterparty risk and information asymmetry, making it rational for sellers to divest to connected buyers who can return the plant's output to the seller at lower cost. This predicts that connected firms are more likely to be buyers and that the likelihood of selling to a connected buyer increases when that buyer faces weaker environmental pressures (Table IX Panel B, p. 733).

**Identification strategy.** The primary concern is that environmental pressures are correlated with omitted firm characteristics. The paper's main causal design exploits quasi-exogenous variation from RepRisk environmental risk incidents, arguing that firms cannot fully control the annual timing of such incidents (Table IV, Figure 2). The resulting within-firm, time-series estimates confirm no pretrends and an immediate divestiture response. Pollution analyses use generalized DID (Poisson with plant-chemical FE + chemical-year FE) and stacked DID that matches each divested plant to never-divested plants in the same NAICS3 industry and state, saturating regressions with cohort-interactive FEs to address heterogeneous treatment-timing bias.

## Method

The paper applies four estimating methods:

**Multilogit for firm responses (R1).** The firm chooses among four mutually exclusive responses to environmental pressure: divestiture, plant closure, enhanced abatement, or no action. Marginal effects of each environmental pressure measure are reported, scaled by the sample mean of each outcome to produce the "Rel. Magnitude" statistic (Table II, p. 715).

**Linear probability and Poisson panel regressions for divestiture determinants (R1-R3, R9).** The baseline seller-side regression (equation 1, p. 716) is:

$$
\text{Divest}_{i,t} = \beta \, \text{Pressure}_{i,t} + \phi_{j,t} + \epsilon_{i,t} \tag{1}
$$

where $$i$$ is a publicly listed TRI-plant-owning firm, $$j$$ denotes the industry, $$t$$ the year, and $$\phi_{j,t}$$ are industry-year fixed effects. The dependent variable equals 100 if firm $$i$$ sells at least one TRI plant in year $$t$$. Standard errors are clustered by firm. Pollution-level interactions, information-asymmetry interactions, and connected-firm pressure interactions follow the same structure with additional firm fixed effects where noted.

**Generalized DID and stacked regressions for pollution changes (R4).** The plant-chemical-year DiD specification (equation 2, p. 722) is:

$$
\text{Pollution}_{i,c,t} = \beta \, \text{Divested}_i \times \text{Post}_{i,t} + \alpha_{i,c} + \tau_{c,t} + \epsilon_{i,c,t} \tag{2}
$$

where $$i$$ is the plant, $$c$$ the chemical type, $$\alpha_{i,c}$$ are plant-chemical fixed effects, and $$\tau_{c,t}$$ are chemical-year fixed effects. Poisson regressions handle skewness. Stacked regressions match each divested plant to never-divested controls in the same NAICS3-state cell, saturating with cohort-plant-chemical, cohort-chemical-year, cohort-state-year, and cohort-industry-year interactive FEs to address heterogeneous treatment-timing (De Chaisemartin and d'Haultfoeuille (2020), Sun and Abraham (2021)). A dynamic version (equation 3, p. 723) decomposes $$\text{Post}_{i,t}$$ into annual event-time dummies from $$k \geq -3$$.

**BERT text classification for conference call analysis (R5-R6 mechanism, Section IV.C).** A BERT NLP model (Devlin et al. (2019)) is trained on 1,000 manually classified sentences from conference calls using a SASB environmental keyword dictionary. The model classifies each sentence-group containing environmental keywords as expressing positive or negative environmental sentiment. The DID specification (equation 4, p. 734) regresses positive/negative disclosure indicators on $$\text{Seller(Pollutive)}_f \times \text{Post}_{f,t}$$ with firm and year FEs.

**Event-study CARs (R7-R8).** Cumulative abnormal returns are measured in a three-day window $$\text{CAR}[-1,+1]$$ relative to the market model and the Fama-French three-factor model. Buyer gains minus seller gains are computed as the product of each firm's market capitalization and its $$\text{CAR}[-1,+1]$$.

## Empirical specifications

The paper's headline empirical tests are:

**Multilogit (Table II, p. 715; R1).** Sample: all public TRI-plant-owning firms, 12,764 obs. Firm Char controls include Tobin's Q, Leverage, Cash Holdings, Tangibility. Industry-year FEs. Robust SE. Pseudo-R² = 0.017-0.026 across pressure measures.

**Seller propensity regressions (Table III, p. 717; R1-R2).** Sample: Panel A, 11,295-18,826 firm-years; Panel B, 11,742-12,084 firm-years. Industry-year FEs; firm characteristics. SE clustered by firm. R² = 0.056-0.079. The key result: Pressure Index coefficient = 2.110\*\*\* (SE 0.593) in Panel A; Pressure Index x Pollution (interquartile quantity) = 2.568\*\*\* (SE 0.615) in Panel B.

**Within-firm divestiture response to RepRisk incidents (Table IV, p. 719; R3).** Sample: 11,243-11,629 firm-years. Firm FE + Year FE + Industry-Year FE. SE clustered by firm. Environmental Event coefficient = 1.074\*-1.133\*\* (SE 0.615-0.544); social and governance events insignificant.

**Buyer vs. seller environmental pressure comparison (Table V, p. 721).** Deal-level regression; 968-1,776 observations. Buyer indicator is negative and significant for all environmental pressure measures: Public -0.055\*\*, Rated -0.047\*\*, Democratic HQ -0.054\*\*, Env. Event -0.052\*\*, Pressure Index -0.059\*\*\*.

**Pollution DiD (Table VI, p. 723-724; R4).** Sample: Panel A, 992,313-3,994,695 plant-chemical-years; Panel B, 872,163-9,431,650 plant-chemical-years. Plant-chemical FE + chemical-year FE + state-year FE + industry-year FE. All Divested x Post estimates insignificant; power analysis shows specifications can detect effects of 2-3% of sample SD (Internet Appendix Table IA.II).

**Conference call disclosure DiD (Table X, p. 735; R5 mechanism).** Sample: 33,873-237,867 firm-year conference calls. Firm FE + Year FE + Industry-Year FE. Seller(Pollutive) x Post: Positive Env. Disclosure = 0.062\*\* (stacked, SE 0.031), ~5.6 pp or 47% of sample mean (12 pp). Negative disclosure: insignificant.

**ESG ratings and enforcement costs DiD (Table XI, p. 738-739; R5-R6).** Panel A (ESG): 29,273-133,621 obs; CSR Score +0.302\* (GDID), +0.941\* (stacked); Env. Score +0.234\*\*\* (GDID), +0.245\*\*\* (stacked). Panel B (Enforcement): 5,495-129,439 obs; Enforcement Action -0.048\*\*\* to -0.073\*\*\*; Enforcement Cost (Poisson) -2.708\*\*\* to -4.376\*\*\*.

**Announcement CARs (Table XII, p. 742; R7).** Sample: 246-278 deal-level observations (public sellers). Past Pollution (Quartile) coefficient = 0.010\*\*-0.013\*\* across all four specifications. R² = 0.307-0.429.

**Business ties (Table IX, p. 733; R9 mechanism).** Panel A: Operationally Related coefficient = 0.651\*\*\* (SE 0.099) in matched-pair regression; Buyer of Pollutive Assets coefficient = 0.069\*\*\* in new-relationship regression. Panel B: Pressure Index, Connected Firms' Min = -0.007\* to -0.009\*\*.

**Nonpollutive asset placebo (Table XIV, p. 746).** All headline effects are absent for divestitures of nonpollutive assets across all five panels, confirming that findings are specific to pollutive divestiture and not generic divestiture effects.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| EPA Toxic Release Inventory (TRI) | Plant-chemical-level toxic emissions (total pollution, pollution intensity, abatement activities); 2000-2020; 1,056,361 plant-chemical-year obs | [EPA TRI](/wiki/datasets/epa-tri/) |
| SDC Mergers and Acquisitions database | Divestitures and spin-offs of industrial plants; 888 pollutive deals 2000-2020 | [no page yet] |
| KLD / MSCI ESG database | ESG ratings (CSR Score, Environmental Score); coverage of public U.S. firms | [WRDS](/wiki/commercial/wrds/) (licensed) |
| RepRisk ESG Business Intelligence | Environmental, social, governance risk incidents; starting 2007 | [RepRisk](/wiki/commercial/reprisk/) (licensed) |
| MIT Election Data and Science Lab | County-level presidential vote share for Democratic HQ classification | [no page yet] |
| Thomson Reuters Street Events (SE) | Conference call transcripts; management presentations; starting 2001 | [no page yet] |
| EPA Enforcement and Compliance History Online (ECHO) | EPA enforcement actions and compliance costs | [no page yet] |
| Compustat (WRDS) | Firm financials (Q, Leverage, Cash Holdings, Tangibility); segment data; ownership structure | [WRDS](/wiki/commercial/wrds/) (licensed) |
| CRSP (WRDS) | Equity returns for announcement CARs and market capitalization | [WRDS](/wiki/commercial/wrds/) (licensed) |
| Factset / Compustat Segment | Supply chain relationships | [WRDS](/wiki/commercial/wrds/) (licensed) |
| Orbis | Subsidiary and organizational layer data | [no page yet] |
| 13-F filings (SEC) | Institutional investor holdings (pension funds, blockholders) | [EDGAR](/wiki/datasets/edgar/) |

Sample period: 2000-2020 (annual). Plant-chemical-year sample: 1,056,361 observations. Firm-year sample: 19,459 observations.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13412) if you are: (i) studying whether environmental divestitures by public firms reduce actual pollution or are primarily cosmetic; (ii) investigating the role of ESG ratings and regulatory enforcement in incentivizing greenwashing; (iii) analyzing real-asset-market equilibria with heterogeneous environmental preferences; (iv) replicating the BERT-based conference call text analysis; or (v) studying gains from trade in pollutive-asset divestitures and the role of buyer-seller business ties. The Internet Appendix contains detailed matching procedures (Section I.B), minimum detectable effect estimates (Table IA.II), and numerous robustness tables.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(2). This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Duchin, Ran, Janet Gao, and Qiping Xu.
> "Sustainability or Greenwashing: Evidence from the Asset Market for Industrial Pollution."
> *The Journal of Finance* 80, no. 2 (April 2025): 699-754.
> DOI: 10.1111/jofi.13412. © 2024 The Author(s).
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Equilibrium Data Mining and Data Abundance: Dugast & Foucault (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/dugast-equilibrium-data-mining-data-2025/
# Distilled: A rational-expectations equilibrium model shows that data abundance (a larger data frontier) always raises price informativeness but can reduce data miners' search intensity and the capital allocated to quant funds, with asset managers' average performance being hump-shaped in both the data frontier and search costs. J. Finance 2025, CC BY-NC 4.0. Seven core results with source locators, the model equations, and the equilibrium derivation.
# Tags: paper-summary, asset-pricing, information-economics, market-microstructure
==============================================================================

**What this is.** The paper's core propositions, the model it builds on (a noisy rational-expectations equilibrium with two types of asset managers), and the mechanism it establishes (the price-informativeness effect dominates the hidden-gold-nugget effect for a large enough data frontier): enough to know what it found and how, without reading all 48 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13397).

## TL;DR

The paper builds a rational-expectations equilibrium model to study how the big data revolution affects the market for active asset management. Two types of managers coexist: "experts" (discretionary funds) with a fixed signal precision, and "data miners" (quant funds) who discover predictors through a sequential search process. The key finding is that the two dimensions of the big data revolution, lower information-processing costs ($$c$$) and a larger data frontier ($$\tau_{dm}^{\text{max}}$$, reflecting more available data sets), have asymmetric and sometimes opposite effects. Reducing search costs always raises data miners' search intensity and capital allocated to quants. But a larger data frontier can reduce quant search intensity and the allocation to quants once it is large enough, because greater price informativeness erodes the value of any given signal. Despite this, a larger data frontier always raises price informativeness. Asset managers' average gross performance is hump-shaped in both $$c$$ and $$\tau_{dm}^{\text{max}}$$, implying the big data revolution should eventually erode active managers' performance.

## Core results

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | An asset manager's optimal position is proportional to the gap between her signal and the asset price; the equilibrium price is a sufficient statistic for the aggregate demand | Proposition 1, eq. 11-13, p. 223 | Trading aggressiveness $$\beta(\tau) = \frac{\tau}{\rho \sigma_\omega^2}$$; equilibrium price $$p^* = \lambda(\tau^*) \xi$$ where $$\lambda(\tau^*) \equiv \frac{\bar{\tau}^2}{\bar{\tau}^2 + \rho^2 \sigma_\omega^4 \sigma_\eta^2}$$ |
| R2 | Price informativeness always increases with average signal quality and therefore with data miners' search intensity $$\tau^*$$ | Lemma 1, eq. 14, p. 224 | $$\mathcal{I}(\tau^*; \tau_{dm}^{\text{max}}) = \text{Var}[\omega \mid p^*]^{-1} = \sigma_\omega^{-2} + \bar{\tau}^2 / (\rho^2 \sigma_\omega^4 \sigma_\eta^2)$$; strictly increasing in $$\tau^*$$ |
| R3 | A decrease in data miners' search costs $$c$$ always increases data miners' equilibrium search intensity $$\tau^*$$, raising capital allocated to quants ($$\mu^*$$), average signal quality, and price informativeness | Proposition 4, p. 227 | $$\partial \tau^* / \partial c < 0$$; $$\tau^*$$ converges to $$\tau_{dm}^{\text{max}}$$ as $$c \to 0$$; capital to data miners $$\mu^* = \Gamma(\tau^*)$$ increases with $$\tau^*$$ |
| R4 | A larger data frontier ($$\tau_{dm}^{\text{max}}$$) reduces data miners' search intensity $$\tau^*$$ and capital allocated to quants ($$\mu^*$$) once $$\tau_{dm}^{\text{max}}$$ exceeds a threshold $$\tau^{tr}(c)$$, because the price informativeness effect dominates the hidden gold-nugget effect | Proposition 5, p. 227 | Threshold $$\tau^{tr}(c)$$ exists for all $$c > 0$$; for $$\tau_{dm}^{\text{max}} > \tau^{tr}(c)$$: $$\partial \tau^* / \partial \tau_{dm}^{\text{max}} < 0$$ and $$\partial \mu^* / \partial \tau_{dm}^{\text{max}} < 0$$ |
| R5 | Despite reducing quant search intensity when $$\tau_{dm}^{\text{max}}$$ is large, a push back of the data frontier always raises average signal quality $$\bar{\tau}$$ and therefore price informativeness $$\mathcal{I}$$; price informativeness is bounded above as $$\tau_{dm}^{\text{max}} \to \infty$$ | Proposition 5, p. 227 | $$\partial \mathcal{I} / \partial \tau_{dm}^{\text{max}} > 0$$ always; $$\mathcal{I}$$ bounded above as $$\tau_{dm}^{\text{max}} \to \infty$$ (Assumption 1 + proof) |
| R6 | Asset managers' average gross excess return is hump-shaped in search costs $$c$$ and in the data frontier $$\tau_{dm}^{\text{max}}$$; the big data revolution is predicted to first raise, then reduce, average active management performance | Corollary 1, Figure 3, pp. 232-233 | $$\mathbb{E}[\bar{R}^e(\tau)] = \frac{1}{W_0 \rho} \left( \frac{1}{\bar{\tau}} + \frac{\bar{\tau}}{\rho^2 \sigma_\omega^2 \sigma_\eta^2} \right)^{-1}$$; peaks at $$\bar{\tau} = \rho \sigma_\omega \sigma_\eta$$; hump-shaped in $$c$$ and $$\tau_{dm}^{\text{max}}$$ |
| R7 | In the fee extension (Nash bargaining, $$\kappa > 0$$), data miners charge no rents; experts' fees are set by their scarcity and decline when data miners' search intensity rises, whether from lower $$c$$ or a data-frontier increase that raises $$\tau^*$$ | Corollary 6, Section VI, eq. 39-41, pp. 237-238 | $$f_{dm}^* = 0$$; $$f_{ex}^*(\tau) = \kappa(w(\tau) - w(\tau^*))$$; experts' fees decline with a fall in $$c$$ and decline (for low-skill experts) or may rise (high-skill experts) with $$\tau_{dm}^{\text{max}}$$ |

**Overall (paper's conclusion).** The two dimensions of the big data revolution, lower data-processing costs and data abundance, have the same effect on the allocation of capital to quants when $$\tau_{dm}^{\text{max}} \leq \tau^{tr}(c)$$ (both increase $$\mu^*$$) but opposite effects when $$\tau_{dm}^{\text{max}} > \tau^{tr}(c)$$ (lower $$c$$ raises $$\mu^*$$; larger $$\tau_{dm}^{\text{max}}$$ reduces $$\mu^*$$). The model predicts that the rise of quant funds driven by both forces should eventually reverse as data abundance grows, and that average active management performance is eventually eroded by greater price informativeness. Distinguishing these two dimensions of the big data revolution is therefore essential for empirical analysis.

## Theory / model

The model has four periods (Figure 1, p. 218). Period 0: investors (mass one) allocate savings $$W_0$$ to either an expert or a data miner, as in Garleanu and Pedersen (2018). Period 1: data miners conduct sequential search for a predictor. Period 2: trading occurs. Period 3: the risky asset payoff $$\omega \sim \mathcal{N}(0, \sigma_\omega^2)$$ is realized.

**Signals.** All asset managers receive a noisy signal before trading (eq. 1, p. 218):

$$
s_{\tau_i} = \omega + \tau_i^{-1/2} \varepsilon_i, \qquad \varepsilon_i \sim \mathcal{N}(0, \sigma_\omega^2) \tag{1}
$$

where $$\tau_i$$ is signal precision ("quality"). Experts' skill $$\tau$$ is fixed and drawn from cumulative distribution $$\Gamma(\cdot)$$ (density $$\gamma(\cdot)$$) on $$[0, \tau_{ex}^{\text{max}}]$$. Data miners discover their predictor through search: each round costs $$c$$ and yields a precision draw $$\tau \sim \Phi(\cdot)$$ on $$[0, \tau_{dm}^{\text{max}}]$$ (eq. 2, p. 219):

$$
\Phi(\tau) = \Pr(\tilde{\tau} \leq \tau) = \frac{\Psi(\tau)}{\Psi(\tau_{dm}^{\text{max}})}, \qquad \tau \in [0, \tau_{dm}^{\text{max}}] \tag{2}
$$

The data frontier $$\tau_{dm}^{\text{max}}$$ captures the maximum attainable signal precision from available data sets; higher $$\tau_{dm}^{\text{max}}$$ reflects data abundance. A data miner with stopping threshold $$\tau_i^*$$ stops when her draw exceeds $$\tau_i^*$$, so the likelihood of stopping in a given round is:

$$
\Lambda(\tau_i^*; \tau_{dm}^{\text{max}}) \equiv \Pr(\tau \in [\tau_i^*, \tau_{dm}^{\text{max}}]) = 1 - \Phi(\tau_i^*) \tag{3}
$$

A higher $$\tau_i^*$$ means more demanding search (fewer stops per round on average), so $$\tau^*$$ is called the "search intensity."

**Capital allocation.** Let $$\mu$$ denote the fraction of investor capital allocated to data miners. Investors observe experts' skills and anticipate data miners' search strategy. They optimally allocate to experts with skill $$\tau \geq \underline{\tau}$$ (the marginal expert) until each expert is at capacity. The rest goes to data miners. In a stable interior equilibrium (Proposition 3, eq. 8, p. 222):

$$
\mu^* = \Gamma(\tau^*) \tag{8}
$$

where $$\tau^*$$ is data miners' equilibrium search intensity, and the marginal expert skill equals $$\tau^*$$.

**Trading.** The market for the risky asset is as in Vives (1995): noise traders with aggregate demand $$\eta \sim \mathcal{N}(0, \sigma_\eta^2)$$ trade alongside asset managers. Price informativeness is measured by the inverse of residual payoff variance, following Grossman and Stiglitz (1980) and Verrecchia (1982). Risk-neutral dealers post a price equal to their expectation of the payoff conditional on aggregate demand (eq. 4, p. 221):

$$
p^* = \mathbb{E}[\omega \mid D(p^*)] \tag{4}
$$

Asset managers have constant absolute risk aversion $$\rho$$. Asset manager $$i$$ returns to her client (eq. 5, p. 221):

$$
W_{i,j} = W_0 + x_i(s_{\tau_i}, p)(\omega - p) - (n_i c)\mathbb{1}_{\{j=dm\}} \tag{5}
$$

where $$n_i$$ is the number of search rounds conducted. Investor utility from trading with an expert of skill $$\tau$$ is (eq. 6, p. 221):

$$
H(\tau) = \mathbb{E}\left[ -\exp\left( -\rho(W_0 + x_i(s_{\tau_i}, p)(\omega - p)) \right) \right] \tag{6}
$$

and investor utility from trading with a data miner of search intensity $$\tau_i^*$$ is (eq. 7, p. 221):

$$
V(\tau_i^*) = \underbrace{\mathbb{E}\left[ -\exp\left(-\rho(W_0 + x_i(s_{\tau_i}, p)(\omega - p))\right) \right]}_{\text{Expected utility from trading}} \times \underbrace{\mathbb{E}\left[\exp(\rho(n_i c))\right]}_{\text{Expected utility cost of exploration}} \tag{7}
$$

## Method

The paper characterizes the equilibrium in three steps.

**Step 1: Trading equilibrium (Proposition 1, p. 223).** Taking $$\tau^*$$ (and hence $$\mu^*$$) as given, the paper solves for the trading equilibrium. In equilibrium, each asset manager's demand is proportional to her signal minus the price (eq. 11):

$$
x^*(s_\tau, p) = \beta(\tau)(s_\tau - p), \qquad \beta(\tau) = \frac{\tau}{\rho \sigma_\omega^2} \tag{11}
$$

The equilibrium price (eq. 12) is:

$$
p^* = \mathbb{E}[\omega \mid D(p^*)] = \lambda(\tau^*)\xi, \qquad \xi \equiv \omega + \rho \sigma_\omega^2 \bar{\tau}(\tau^*; \tau_{dm}^{\text{max}})^{-1} \eta \tag{12}
$$

where $$\bar{\tau}(\tau^*; \tau_{dm}^{\text{max}})$$ is the average signal quality across all asset managers and $$\lambda(\tau^*)$$ is defined in eq. 13. Price informativeness is (eq. 14, p. 224):

$$
\mathcal{I}(\tau^*; \tau_{dm}^{\text{max}}) \equiv \text{Var}[\omega \mid p^*]^{-1} = \frac{1}{\sigma_\omega^2} + \frac{\bar{\tau}(\tau^*; \tau_{dm}^{\text{max}})^2}{\rho^2 \sigma_\omega^4 \sigma_\eta^2} \tag{14}
$$

**Step 2: Equilibrium data mining (Proposition 2, p. 225).** The trading value of a signal of quality $$\tau$$ is (Lemma 2, eq. 16, p. 224):

$$
g(\tau, \tau^*) = -\left(1 + \frac{\tau}{\sigma_\omega^2 \mathcal{I}(\tau^*; \tau_{dm}^{\text{max}})}\right)^{-\frac{1}{2}} \tag{16}
$$

A data miner's continuation value after finding and then rejecting a predictor of quality $$\hat{\tau}_i$$ is (eq. 17-18, pp. 224-225):

$$
J(\hat{\tau}_i, \tau^*) = \frac{\exp(\rho c) \Lambda(\hat{\tau}_i; \tau_{dm}^{\text{max}})}{1 - \exp(\rho c)(1 - \Lambda(\hat{\tau}_i; \tau_{dm}^{\text{max}}))} \times \mathbb{E}_\phi\left[g(\tau, \tau^*) \mid \hat{\tau}_i \leq \tau \leq \tau_{dm}^{\text{max}}\right] \tag{18}
$$

In a symmetric equilibrium, $$\tau^*$$ solves $$g(\tau^*, \tau^*) = J(\tau^*, \tau^*)$$, which reduces to (eq. 21, p. 225):

$$
F(\tau^*) = \exp(-\rho c), \tag{21}
$$

where

$$
F(\tau^*) \equiv \int_{\tau^*}^{\tau_{dm}^{\text{max}}} r(\tau, \tau^*)\phi(\tau)d\tau + (1 - \Lambda(\tau^*; \tau_{dm}^{\text{max}})), \qquad r(\tau, \tau^*) \equiv \left(\frac{\tau^* + \sigma_\omega^2 \mathcal{I}(\tau^*; \tau_{dm}^{\text{max}})}{\tau + \sigma_\omega^2 \mathcal{I}(\tau^*; \tau_{dm}^{\text{max}})}\right)^{\frac{1}{2}} \tag{22-23}
$$

Proposition 2 establishes that this equation has a unique solution $$\tau^* \in (0, \tau_{dm}^{\text{max}})$$ whenever $$F(0) < \exp(-\rho c)$$.

**Step 3: Full equilibrium (Proposition 3, p. 226).** Combining Steps 1 and 2 with the capital-allocation condition $$\mu^* = \Gamma(\tau^*)$$ (eq. 8) yields the full equilibrium characterization. The model is solved analytically under the parameterization $$\Phi(\tau) = \frac{1-(1+\tau)^{-3/2}}{1-(1+\tau_{dm}^{\text{max}})^{-3/2}}$$ and $$\Gamma(\tau) = 1-(1+\tau)^{-3/2}$$ for numerical illustrations (Figure 2, p. 230).

The decomposition of the data-frontier effect (eq. A26, p. 250 Appendix; related text at eq. 25, pp. 228-229) separates the hidden gold-nugget effect (a larger $$\tau_{dm}^{\text{max}}$$ raises the value of the best possible predictor) from the price informativeness effect (more data raises $$\mathcal{I}$$, reducing the value of any given signal). The second term captures the price informativeness effect:

$$
\frac{\partial F}{\partial \tau_{dm}^{\text{max}}} = \underbrace{\phi(\tau_{dm}^{\text{max}})(r(\tau_{dm}^{\text{max}}, \tau^*) - \mathbb{E}_\phi[\min\{1, r(\tau, \tau^*)\}])}_{<0: \text{Gold Nugget Effect}} + \underbrace{\left(\int_{\tau^*}^{\tau_{dm}^{\text{max}}} \frac{\partial r(\tau, \tau^*)}{\partial \mathcal{I}} \phi(\tau)d\tau\right) \frac{\partial \mathcal{I}}{\partial \tau_{dm}^{\text{max}}}}_{>0: \text{Informativeness Effect}} \tag{26}
$$

When $$\tau_{dm}^{\text{max}}$$ is large enough, the positive informativeness effect dominates, so $$\partial F / \partial \tau_{dm}^{\text{max}} > 0$$ and $$\tau^*$$ decreases with $$\tau_{dm}^{\text{max}}$$ (Proposition 5).

## Empirical specifications

This is a pure theory paper. It derives no estimating equations and runs no regressions. The empirical implications are summarized in Table I (p. 243), which characterizes the directional effects of lower search costs ($$c \searrow$$) and data abundance ($$\tau_{dm}^{\text{max}} \nearrow$$) on:

- Allocation of capital to data miners ($$\mu^*$$): increases with lower $$c$$; hump-shaped in $$\tau_{dm}^{\text{max}}$$
- Price informativeness ($$\mathcal{I}$$): always increases with both shocks
- Average signal quality ($$\bar{\tau}$$): always increases with both shocks
- Data miners' relative performance ($$RP$$): increases with $$\tau_{dm}^{\text{max}}$$ (Corollary 4); ambiguous with lower $$c$$
- Within-group performance dispersion ($$\Delta R_\alpha$$): decreases with lower $$c$$; increases with $$\tau_{dm}^{\text{max}}$$ above threshold (Corollaries 2-3)
- Average performance ($$\mathbb{E}[\bar{R}^e]$$): hump-shaped in both $$c$$ and $$\tau_{dm}^{\text{max}}$$ (Corollary 1)

Section VII (pp. 242-244) suggests three types of empirical tests: (i) cross-sectional variation in quant and discretionary fund holdings to capture differential exposure to alternative data (shocks to $$\tau_{dm}^{\text{max}}$$), exploiting the finding of Abis (2022) that quant funds grew from 6.1% to 18.6% of U.S. equity AUM between 2000 and 2017; (ii) regulatory changes that reduce information processing costs (e.g., the SEC's XBLR mandate lowering $$c$$ for IT-intensive funds, consistent with evidence from Zhao (2021) that this mandate reduced the performance gap between quant and discretionary funds); and (iii) the introduction of cloud computing (e.g., Amazon Web Services in 2006) as shocks to $$c$$. Measurement of signal quality $$\tau$$ follows Proposition 1: the theoretical coefficient from regressing a fund's holdings $$x^*(s_\tau, p^*)$$ on $$\omega - p^*$$ is $$\beta(\tau) = \tau / (\rho \sigma_\omega^2)$$ (eq. 45, p. 244), which is strictly positive for informed managers and increases with skill. The paper also relates to evidence from Pastor, Stambaugh, and Taylor (2015) that the size of the active management industry is negatively related to funds' performance. Han and Sangiorgi (2018) and Banerjee and Breon-Drish (2021) model information acquisition as search but analyze different questions; the key distinction here is the simultaneous variation in both the intensive and extensive margins. Stambaugh (2020) documents a related result that improved manager skills reduce average performance, which the model endogenizes through the price informativeness channel.

## Datasets used

This paper uses no empirical datasets. All results are derived analytically from the theoretical model. Numerical illustrations use the parametric family $$\Phi(\tau) = \frac{1-(1+\tau)^{-3/2}}{1-(1+\tau_{dm}^{\text{max}})^{-3/2}}$$ and $$\Gamma(\tau) = 1-(1+\tau)^{-3/2}$$, with calibrated values for $$\rho$$, $$\sigma_\omega$$, $$\sigma_\eta$$, $$W_0$$, and $$c$$ (Figures 2-6, pp. 230-239).

| Dataset | Role in paper | Wiki page |
|---|---|---|
| No empirical data used | Theory paper only | n/a |

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13397) if you are: building a model of quant fund behavior in equilibrium; studying the welfare effects of the big data revolution on price efficiency and active management; looking for testable predictions on the cross-sectional variation in fund performance as a function of data availability; or extending the model to allow learning about signal quality (fn. 10, p. 220) or non-extreme decreasing returns to scale (Section II.E, Internet Appendix). The appendix (pp. 245-252) contains the full proofs of all propositions.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(1). This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The CC BY-NC 4.0 licence permits sharing with attribution for non-commercial purposes; the verbatim PDF is not hosted in this batch.

> **Citation.** Dugast, Jérôme, and Thierry Foucault.
> "Equilibrium Data Mining and Data Abundance."
> *The Journal of Finance* 80, no. 1 (February 2025): 211-258.
> DOI: 10.1111/jofi.13397. CC BY-NC 4.0.
> This page is an extract by the Institute for Automated Research: core results
> re-expressed for research reference; **changes were made**.


==============================================================================
# Would Order-By-Order Auctions Be Competitive: Ernst, Spatt & Sun (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/ernst-would-order-order-auctions-2025/
# Distilled: A theoretical model comparing brokers' routing (current U.S. equity market structure) to SEC-proposed order-by-order auctions for retail trades shows that auctions improve allocative efficiency but worsen retail investor welfare in illiquid stocks due to the winner's curse. J. Finance 2025, CC BY-NC-ND 4.0. Six core results with source locators, the model (inventory-cost common-value auction), and the method (linear symmetric equilibrium).
# Tags: paper-summary, market-microstructure, market-design, auction-theory
==============================================================================

**What this is.** This is a machine-distilled skeleton of the original paper. Read the full paper at [doi:10.1111/jofi.13449](https://doi.org/10.1111/jofi.13449) to replicate or extend the analysis.

## TL;DR

Ernst, Spatt, and Sun (2025) build a theoretical model comparing two mechanisms for executing segregated retail equity orders: the current system of brokers' routing (where retail brokers route to a wholesaler based on aggregate execution quality) and order-by-order auctions (the SEC's proposed Rule 615, where any market participant bids on each individual order). In the baseline model, order-by-order auctions always improve total welfare and wholesaler profits relative to brokers' routing because they ensure the lowest-cost market maker always gets the order (first-best allocation). However, the common-value nature of the auction amplifies the winner's curse: market makers bid conservatively because winning reveals that all rivals had higher cost signals. This reduces competition relative to brokers' routing for retail investors, particularly in illiquid stocks (where the common-value component of inventory cost is large) or when the number of bidders is small. The paper extends the model to analyze institutional trader entry, endogenous participation, alternative information structures, and cross-stock subsidization under heterogeneous stocks.

## Core results

| # | Result | Locator | Magnitude as reported |
|---|---|---|---|
| R1 | Total welfare is always higher under order-by-order auctions than brokers' routing | Proposition 2, p. 1889 | $$W_{total}(1) - W_{total}(p) = (1-p)\frac{N-1}{N+1}\frac{c_2}{2} > 0$$ for all $$p < 1$$ |
| R2 | Wholesalers earn strictly higher profits under order-by-order auctions | Proposition 2, p. 1889 | $$W_W(1) - W_W(p) = (1-p)\frac{c_1 + Nc_2}{N(1+N)} > 0$$ for all $$p < 1$$ |
| R3 | Retail investor welfare is higher under order-by-order auctions only if N is large enough or correlation is low enough | Proposition 2, p. 1889 | $$W_I(1) < W_I(p) \iff N(N-3) > 2\frac{c_1}{c_2}$$; investor welfare can go either way |
| R4 | With institutional traders (no information advantage), more institutional traders raise total and investor welfare but hurt wholesalers | Propositions 5-6, pp. 1892-1894 | $$\widetilde{W}^{OBO}_{total}$$ and $$\widetilde{W}^{OBO}_I$$ increasing in $$N_0$$; $$\widetilde{W}^{OBO}_W$$ decreasing in $$N_0$$ |
| R5 | With severe information asymmetry, institutional trader entry always leads to lower investor welfare | Proposition 9, Remark 1, p. 1899 | When $$\pi_I \in (\pi_1, \pi_3)$$ and $$\delta_c > \underline{\delta}$$: investor welfare strictly lower when institutional entry is allowed |
| R6 | Cross-subsidization under brokers' routing means heterogeneous investors have mixed welfare effects from switching to auctions | Lemma 4, pp. 1907-1908 | High-$$c_0$$ (illiquid stock) investors worse off; low-$$c_0$$ (liquid stock) investors better off under order-by-order auctions |

**Overall (paper's conclusion).** Order-by-order auctions improve allocative efficiency and total welfare, but the winner's curse leads market makers to bid conservatively, extracting more rent from retail investors. Investor welfare is lower under order-by-order auctions for illiquid stocks and when the number of participating bidders is small. The entry of informed institutional traders can further reduce competition and harm investors. Cross-subsidization under brokers' routing insulates investors in high-cost stocks, and removing it via auctions creates distributional losers among retail investor populations with illiquid holdings.

## Theory / model

The paper has no empirical estimation. All results are derived analytically from the following model.

**Theoretical antecedents.** The inventory cost structure and the common-value auction framework build on Menezes and Monteiro (2004) and Klemperer (2018), who develop tractable linear equilibria for common-value first-price auctions (p. 1885). The winner's curse logic draws on Milgrom and Weber (1982), who show that bidder profit is lower when new information is common rather than independent (p. 1883). Related prior work on competing market makers includes Bernhardt and Hughson (1997) (order splitting in duopoly) and Biais, Martimort and Rochet (2000) (common-value auctions with informed traders), both cited on p. 1884. The baseline motivation for retail order segmentation draws on Easley, Kiefer and O'Hara (1996) (PFOF and adverse selection, p. 1884) and Baldauf, Mollner and Yueshen (2024) (retail investors less correlated, p. 1880). Ernst, Spatt and Sun (2024) provide companion empirical evidence on retail liquidity programs (p. 1882).

**Setup (Section I, p. 1886).** Two dates: time 0 and time 1, no discounting. Three types: a retail investor, a broker, and $$N \geq 2$$ ex-ante identical risk-neutral wholesalers $$i \in \{1, 2, \ldots, N\}$$. The broker minimizes the bid-ask spread paid by the investor. At time 0, the broker receives a one-unit sell order and sends it to one wholesaler.

**Inventory cost (eq. 1, p. 1887):**

$$\zeta_i = c_0 + c_1 \frac{1}{N}\sum_{j=1}^{N} y_j + c_2 y_i \tag{1}$$

where $$c_0, c_1, c_2 > 0$$. The term $$c_0$$ is the unconditional expected inventory cost (common to all). The term $$c_1 \frac{1}{N}\sum_j y_j$$ is the common-value component (aggregate cost shock). The term $$c_2 y_i$$ is the private-value component. Each wholesaler $$i$$ receives i.i.d. cost shocks $$y_i \sim U[-\frac{1}{2}, \frac{1}{2}]$$.

**Information (Assumption 1, p. 1887).** Each wholesaler $$i$$ observes a noisy signal $$w_i$$ about $$y_i$$: with probability $$p$$, $$w_i = y_i$$; with probability $$1-p$$, $$w_i \sim U[-\frac{1}{2}, \frac{1}{2}]$$ (independent noise). Under brokers' routing: $$p < 1$$. Under order-by-order auctions: $$p = 1$$. The broker allocates to the wholesaler submitting the lowest spread.

**Welfare definitions (p. 1889).** Wholesaler expected profit is $$W_W(p) = \frac{p(c_1 + Nc_2)}{N(1+N)}$$. Investor expected welfare is $$W_I(p) = -\left[c_0 + p\frac{2c_1 - (N-3)Nc_2}{2N(1+N)}\right]$$. Total welfare is $$W_{total}(p) = W_W(p) + W_I(p) = -\left(c_0 - p\frac{N-1}{N+1}\frac{c_2}{2}\right)$$.

**Institutional traders extension (Section II, p. 1890).** With $$N_0 \geq 2$$ institutional traders who additionally observe $$\tilde{c}_0 \in \{c_0 - \delta_c, c_0 + \delta_c\}$$ (the common component), the inventory cost is:

$$\tilde{\zeta}_i = \tilde{c}_0 + c_1 \frac{1}{\tilde{N}}\sum_{j=1}^{\tilde{N}} y_j + c_2 y_i \tag{2}$$

where $$\tilde{N}$$ is total active market makers. The key result: when $$\delta_c > \underline{\delta}$$, institutional traders crowd out wholesalers from high-quality (low-cost, $$\tilde{c}_0 = c_0 - \delta_c$$) orders, as their lowest possible spread remains below the highest possible wholesaler spread. Market segmentation emerges: only institutional traders compete for low-cost orders, while both types compete for high-cost orders.

## Method

The paper solves for linear symmetric equilibria analytically throughout. There is no estimation.

**Equilibrium strategy (Proposition 1, p. 1888).** For any $$p \in [0,1]$$, there exists a linear symmetric equilibrium in which wholesaler $$i$$'s spread is:

$$s(w_i; p) = K_0(p) + K_1(p) w_i$$

where:

$$K_0(p) = c_0 + \frac{p}{2N}\left[c_1\left(\frac{1}{N} + \frac{N-1}{2} - \frac{1}{2}\right) + c_2\right]$$

$$K_1(p) = \frac{N-1}{N} p \left[c_1\left(\frac{1}{2} + \frac{1}{N}\right) + c_2\right]$$

The intercept $$K_0(p)$$ is the equilibrium spread when $$w_i = 0$$ (no private signal); the slope $$K_1(p)$$ captures the responsiveness to the signal. At $$p = 1$$ (order-by-order auctions), the slope is maximized: heterogeneous bids create more rent for market makers. At $$p \to 0$$ (no signal), all spreads converge and competition intensifies.

**Winner's curse intuition.** The common-value component $$c_1 \frac{1}{N}\sum_j y_j$$ is the same for all market makers. Winning the auction reveals that all rivals had higher signals (i.e., higher cost shocks), making the common value worse for the winner than the unconditional expectation. This winner's curse is proportional to $$\frac{c_1}{c_2}$$: when the common component is large relative to the private component, the winner's curse is severe, market makers shade bids more conservatively, and investor welfare deteriorates.

**Equilibrium with institutional traders (Propositions 4 and 7).** With no information advantage ($$\delta_c = 0$$), replacing $$N$$ by $$N + N_0$$ in Proposition 1 gives the equilibrium. With severe information advantage ($$\delta_c > \underline{\delta}$$), the equilibrium features two distinct bidding strategies: institutional traders use $$\tilde{s}^+(y;\delta_c)$$ for high-cost orders and $$\tilde{s}^-(y;\delta_c)$$ for low-cost orders, while wholesalers use $$\tilde{s}^+(y;\delta_c)$$ only (they are outbid on low-cost orders regardless of their signal).

**Heterogeneous stocks extension (Proposition 15, p. 1906).** Under brokers' routing with a distribution $$G(c_0, c_1, c_2)$$ of stock characteristics, wholesalers compete before observing individual order characteristics, so they submit a single equilibrium strategy using average characteristics $$(\bar{c}_0, \bar{c}_1, \bar{c}_2)$$ as arguments in the $$K_0, K_1$$ formulas. Under order-by-order auctions, characteristics are observed and the Proposition 1 equilibrium applies order-by-order. This generates cross-subsidization: brokers' routing taxes liquid (low-cost) stocks to subsidize illiquid (high-cost) stocks.

## Empirical specifications

This paper is purely theoretical. There are no regressions, no data, and no calibration exercises. Section IV (pp. 1908-1909) discusses empirical implications and qualitative consistency with related empirical findings (Ernst, Spatt, and Sun (2024); Dyhrberg, Shkilko, and Werner (2022)), but no formal empirical test is conducted.

The key comparative statics are:

- Total welfare difference: $$W_{total}(1) - W_{total}(p) = (1-p)\frac{N-1}{N+1}\frac{c_2}{2}$$, increasing in the private-value share $$c_2$$ and in $$N$$ (Proposition 2).
- Investor welfare favors order-by-order auctions iff $$N(N-3) > 2\frac{c_1}{c_2}$$, i.e., the number of bidders is large and the common-value fraction $$\frac{c_1}{c_2}$$ is small (Proposition 2).
- With endogenous entry under severe adverse selection (Proposition 9): allowing institutional trader entry leads to weakly fewer total liquidity providers and strictly lower investor welfare when both wholesalers and institutional traders coexist in equilibrium.
- Cross-subsidization threshold (Lemma 4): investors with order characteristics $$c_0 > \bar{c}_0$$ (illiquid stocks), $$c_1 > \bar{c}_1$$, or $$c_2 < \bar{c}_2$$ (when $$N > 3$$) are worse off switching from brokers' routing to order-by-order auctions.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| None | Theory paper; no data used | n/a |

The paper's empirical discussion references Dyhrberg, Shkilko, and Werner (2022) (SEC 605 reports on retail vs. institutional volume) and Ernst, Spatt, and Sun (2024) (retail liquidity program bidding data), but neither dataset is used in this paper's analysis.

## When to read the full paper

Read Ernst, Spatt, and Sun (2025) if you are:
- Studying the theoretical welfare implications of the SEC's Rule 615 (order-by-order auction) proposal for retail equity order flow.
- Interested in the interplay of common-value auctions, winner's curse, and market competition in financial market design.
- Analyzing cross-subsidization effects in brokers' routing and how they affect heterogeneous retail investors with different portfolio compositions.
- Working on auction theory with asymmetric bidders: the institutional-trader extension (Section II) provides a tractable linear-equilibrium model with asymmetric information.
- Evaluating policy tradeoffs between allocative efficiency (favors order-by-order auctions) and retail investor welfare (ambiguous, depends on $$N$$ and $$\frac{c_1}{c_2}$$).

The key propositions are Proposition 2 (pp. 1889-1890, baseline welfare comparison) and Proposition 9 (pp. 1898-1900, endogenous entry with adverse selection).

## Attribution and rights

Thomas Ernst, Chester Spatt, Jian Sun, "Would Order-By-Order Auctions Be Competitive?", *The Journal of Finance*, vol. 80, no. 4 (August 2025), pp. 1879-1927. DOI: 10.1111/jofi.13449.

Published under a Creative Commons Attribution-NonCommercial-NoDerivs (CC BY-NC-ND 4.0) licence. This page is an LLM-distilled extract (not human-verified, not reproduced). The original article is available at [https://doi.org/10.1111/jofi.13449](https://doi.org/10.1111/jofi.13449). Extract only; the CC BY-NC-ND 4.0 licence does not permit redistribution of modified or derivative copies.


==============================================================================
# Impediments to the Schumpeterian Process: Faccio & McConnell (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/faccio-impediments-schumpeterian-process-replacement-2025/
# Distilled: Using hand-assembled data on the 20 largest firms across up to 75 countries from circa 1910, Faccio and McConnell find that political connections are the primary impediment to the replacement of large firms, but only when accompanied by cross-border barriers to trade and capital flows. The Journal of Finance 80(6) 2025, CC BY 4.0. Eight core results with source locators, datasets used, the identification strategy, and the empirical specifications.
# Tags: paper-summary, corporate-governance, political-economy, creative-destruction
==============================================================================

**What this is.** The paper's core results, the identification strategy (including the Italy quasi-experiment), and the empirical specifications with real equations: enough to know what it found and how, without reading all 41 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13481).

## TL;DR

Faccio and McConnell ask what prevents the "destructive" part of Schumpeter's creative destruction: why do large incumbent firms so often fail to be replaced by new large firms? This extends the 44-country, 2-decade evidence of Fogel, Morck, and Yeung (2008) to 75 countries and up to a century, investigating the mechanisms behind the lack of replacement. Using hand-assembled data on the 20 largest firms in 60 countries circa 1910 and Worldscope data for 47 countries in 2000, they track which firms remain among the largest in 2018. Intrafirm innovation (new technology, patents, R&D) has at most a secondary role. Bank-board interlocks matter over the shorter 2000-2018 period but not over the century-long horizon. Political connections are the most robust predictor: a politically connected large firm circa 1910 is roughly 9.6 percentage points more likely to remain among the 20 largest in its country in 2018 than an unconnected firm of equal size, and the effect is concentrated in countries closed to cross-border trade and capital flows. This is consistent with the formal model of Akcigit, Baslandze, and Lotti (2023) predicting that connected firms remain large only when regulatory wedges are sufficiently large. A quasi-natural experiment using the fall of fascism in Italy, documented in Faccio and McConnell (2024), confirms that the connection is causal: firms whose political connections were exogenously severed experienced a large, statistically significant drop in the probability of remaining among the 20 largest. The result that politically connected Italian firms are of average or below-average quality is also consistent with Braggion and Moore (2013)'s evidence on late Victorian Britain.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Only 13.6% of the 20 largest firms circa 1910 remain among the 20 largest in their country in 2018; over the century replacement is the norm | Table I Panel A, p. 3367; p. 3366 | 13.6% unconditional survival rate across 60 countries (1,115 firms); 152 firms remain large in all three treatment groups |
| R2 | Political connections (circa 1910) increase the probability of remaining large by 9.59 percentage points vs. an unconditional rate of 10% for connected-country firms; New tech is insignificant | Table II col. (4), p. 3370 | Political connections coeff. = 0.0959\*\* (p-value 0.018); New tech coeff. = -0.0027 (p = 0.368) |
| R3 | The political-connections effect is robust to controlling for state ownership, industry, and size centile fixed effects; bank-board interlocks are insignificant over the century | Table II cols. (6)-(9), pp. 3370-3372 | Political connections coeff. ranges 0.0825\*\*\* to 0.1059\*\*\* across specifications; bank-board interlocks coeff. not statistically significant in col. (3) |
| R4 | Italy quasi-experiment: after the fall of fascism (1944), formerly politically connected firms became disproportionately less likely to remain among the 20 largest; implied counterfactual replacement rate is -18% to +36% of observed difference | Table IV col. (1), p. 3377; p. 3379 | Fascist x Post Fascism coeff. = -0.0927\*\*\* (p = 0.009); difference-in-differences sample mean Y for connected firms pre-1944 = 0.1283 |
| R5 | Politically connected firms in Italy are not the best firms: ROE is marginally significantly negative and M/B is similar to unconnected firms | Table III Panel B cols. (1)-(4), pp. 3374-3375 | ROE coeff. = -0.0267\* (p = 0.087) in col. (1); M/B coeff. = 0.0168 (p = 0.592) in col. (3) |
| R6 | In the 2000-2018 sample, political connections increase the probability of remaining among the 20 largest by 2.6 to 4.3 percentage points vs. an unconditional rate of 2.2%; patents and R&D are also significant but smaller in magnitude | Table V cols. (1)-(5), pp. 3382-3383 | Political connections coeff. = 0.0257\* to 0.0384\*\*; patents coeff. = 0.0056\*\*; R&D/TA coeff. = 0.0082\*\* |
| R7 | Political connections interact negatively and significantly with openness: the effect of connections on remaining large is attenuated and disappears in open economies (both trade and capital flows required) | Table VII col. (2), p. 3387; Table VIII col. (1), p. 3388 | Table VII: Political connections x Openness Ind. level coeff. = -0.5034\* (p = 0.050); Table VIII: Political connections x Openness coeff. = -0.0667\*\*\* (p = 0.002) |
| R8 | Back-of-envelope: absent political connections, per capita GDP growth 2001-2021 would have been 4.24 to 7.17 percentage points higher than the sample median of 26.98% | p. 3380 | Regression coeff. of GDP growth on fraction remaining large = -0.4628 (p = 0.021); counterfactual replacement rate gap = 2.47 to 4.18 percentage points |

**Overall (paper's conclusion).** Political connections are the most robust impediment to the Schumpeterian process of large-firm replacement, but only when cross-border barriers to trade and capital are in place. Intrafirm innovation plays a secondary role at most over the century-long horizon. The Italy quasi-experiment confirms causality. The implied macroeconomic cost is substantial: absent political-connection-enabled entrenchment, long-run per capita GDP and TFP growth would have been materially higher.

## Theory / model

The paper has no formal structural model. It organizes competing hypotheses from Schumpeter, Brandeis, and Steffens and tests them empirically.

**Hypothesis 1 (Schumpeter 1942).** Large firms can remain dominant by reinventing themselves through intrafirm innovation. Tested via the new-tech indicator (circa 1910), patents, R&D, and productivity (2000 sample).

**Hypothesis 2 (Brandeis 1914, 1934).** Large firms remain large through bank-board interlocking directorates that allow them to restrict access to domestic capital and suppress entry. Tested via the count of bank-board interlocks per firm.

**Hypothesis 3 (Steffens 1906).** Large firms remain large by capturing the political process to obtain regulations that suppress entry. Tested via the political connections indicator.

**The Rajan-Zingales mechanism.** Following Rajan and Zingales (2003), a necessary condition for domestic political capture to be effective is that cross-border competition is also restricted. When the economy is open to both cross-border trade flows and cross-border capital flows, it is difficult for domestic incumbents to entrench their positions even with political connections, because foreign entrants bypass domestic restrictions. The paper therefore examines a triple interaction:

$$
\text{Pr}(\text{Top 20 in 2018}) = \alpha + \beta_1 \text{PC}_{i} + \beta_2 (\text{PC}_{i} \times \text{Openness}_{c}) + \gamma X_{it} + \delta_c + \varepsilon_{it} \tag{T1}
$$

where $$\text{PC}_{i}$$ is an indicator for political connections, $$\text{Openness}_{c}$$ is the interaction of trade flows and capital flows at the country (or country-industry) level, $$X_{it}$$ includes size, age, innovation proxies, and $$\delta_c$$ are country fixed effects. The prediction is $$\beta_2 < 0$$: political connections enable entrenchment only in closed economies.

**Identification strategy (Italy quasi-experiment).** The fall of fascism in 1943-1944 plausibly exogenously severed political connections: fascist politicians became ineligible for office, and post-WWII laws prohibited companies from having politicians as officers and directors. The paper exploits this shock in a difference-in-differences framework, comparing firms with fascist political connections to firms without, before and after 1944. The subset of connections involving directors appointed as life Senators by the King provides further exogeneity (the firm did not choose to be connected; the King appointed the Senator).

## Method

The paper applies linear probability models of firm survival in the top-20 ranking and difference-in-differences estimation. It builds on `panel-regression`, `difference-in-differences`, and `probit-regression` primitives.

**Linear probability models (LPM) for the circa 1910 and 2000 samples.** The estimator is OLS. The dependent variable is a binary indicator for whether a firm that was among the 20 largest in its country in 1910 (or 2000) is also among the 20 largest in 2018. All variables other than binary indicators are standardized. Standard errors are robust (circa 1910) or double-clustered at the firm and year level (Italy panel). The specification family is:

$$
Y_{i,2018} = \alpha + \beta_1 \ln(\text{BookEquity}_{i,t_0}) + \beta_2 \ln(\text{FirmAge}_{i,t_0}+1) + \beta_3 \text{NewTech}_{i} + \beta_4 \text{PC}_{i} + \beta_5 \text{BB\_Interlocks}_{i} + \delta_c + \delta_{\text{ind}} + \varepsilon_i \tag{M1}
$$

where $$t_0 \in \{1910, 2000\}$$, $$\delta_c$$ are country fixed effects, and $$\delta_{\text{ind}}$$ are industry fixed effects (included in selected specifications).

**Difference-in-differences for the Italy panel.** The dependent variable is an indicator for being among the 20 largest Italian firms, measured 20 years ahead (p. 3376). The key regressor is the interaction of a time-invariant indicator for fascist political connections and an indicator for years 1944 and after:

$$
Y_{i,t+20} = \alpha + \beta \left(\text{Fascist}_i \times \text{Post Fascism}_t\right) + \gamma_1 \ln(\text{BookEquity}_{it}) + \gamma_2 \ln(\text{FirmAge}_{it}+1) + \delta_i + \delta_{\text{ind}} + \delta_t + \varepsilon_{it} \tag{M2}
$$

where $$\delta_i$$ are firm fixed effects, $$\delta_{\text{ind}}$$ are industry fixed effects, and $$\delta_t$$ are year fixed effects. Standard errors are double-clustered at the firm and year level. The sample runs from 1921 (first year fascists were elected) to 1971 (last year of imita.db data).

**Productivity measure.** For the 2000 sample, total factor productivity is constructed using the Olley-Pakes (1996) method following Kogan et al. (2017) and the Stata code of Imrohoroglu and Tuzel (2014) (p. 3366).

## Empirical specifications

**Circa 1910 baseline (R2, R3).** LPM of the probability of being among the 20 largest firms in 2018, on a sample of 1,115 firms from 60 countries, with country fixed effects and optionally industry fixed effects. Key regressors: standardized ln(Book Equity), standardized ln(Firm Age), New Tech indicator, Political Connections indicator, ln(Bank-Board Interlocks + 1), State-owned enterprise indicator. Robust standard errors. Specifications are progressively augmented from col. (1) to (9) of Table II (p. 3370).

**Italy 1911 selection model (R5).** LPM explaining the probability of being politically connected in 1911, on 793 Italian firms with book equity at least 1 million Italian lira. Regressors include ln(Book Equity), ln(Net Income), ln(Firm Age), New Tech, University Town, ROE, M/B, publicly traded indicator. Specifications (1)-(4) of Table III Panel B (p. 3374) add industry and location fixed effects.

**Italy difference-in-differences (R4).** Specification (M2) above. The main sample in col. (1) of Table IV (p. 3377) has 3,835 observations and 226 unique firms (publicly traded in Milan), 1921-1971. Robustness checks in cols. (3)-(6) isolate connections involving King-appointed Senators and exclude firms subject to post-fascism retaliation.

**Openness interaction (R7).** Tables VII and VIII augment the baseline specification with a triple interaction term $$\text{PC}_i \times \text{Trade}_c \times \text{Capital Flows}_c$$ where trade is (Imports + Exports)/GDP and capital flows are proxied by telephones per 100 people in 1914 (Kingsbury 1915) or by the dollar value of foreign M&A as a fraction of total M&A (Thomson ONE). Cols. (3)-(5) of both tables provide falsification tests using only trade or only capital flows. Standard errors are robust (circa 1910) or clustered at country level (2000 sample).

**2000 baseline (R6).** LPM of the probability of being among the 20 largest in 2018, on Worldscope firms from 47 countries in 2000. Regressors parallel the 1910 specifications plus patents, R&D/TA, productivity (Olley-Pakes), M/B, ROE, leverage, diversification, industry change 2000-2018, and a not-in-WS-2018 indicator for survivorship. Country, four-digit SIC industry, and country x industry fixed effects are added progressively; cols. (3)-(9) of Table V (pp. 3382-3383). Political connections coeff. = 0.0257\* to 0.0384\*\*.

**Back-of-envelope growth regression (R8).** Country-level OLS regressing per capita GDP growth (2001-2021, World Bank) on the fraction of firms remaining large (1910-2018). 57 observations. Regression coefficient = -0.4628 (p = 0.021), Table IV.IA (Internet Appendix, p. 3380 text). TFP growth (Feenstra, Inklaar, and Timmer 2015 Penn World Table 10.1) regression yields coefficient = -0.2486 (p = 0.083), 49 countries.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Hand-assembled business directories (56 sources, circa 1900-1925) | 20 largest firms per country circa 1910 (60 countries, 1,115 firms); book equity, founding year, political connections, bank-board interlocks | No page yet |
| Worldscope | 20 largest publicly traded firms per country in 2000 (47 countries) and 2018; book equity, financial data, firm identifiers | [no page yet](/wiki/datasets/) |
| Faccio (2006) political connections database | Political connections indicator for 2000 sample, 47 countries | No page yet |
| imita.db (IMprese ITAliane Data Base) | Italian publicly traded and private firms 1921-1971 (archival volumes by Credito Italiano) | No page yet |
| Il Taccuino dell'Azionista (1942, 1947 editions) | Supplement imita.db for 1941-1945 Italian firms | No page yet |
| Bena et al. (2017) patent data | Patents granted per firm in 2000 (Worldscope firms) | No page yet |
| World Bank World Development Indicators | Trade flows (imports + exports)/GDP; per capita GDP growth 2001-2021 | No page yet |
| Thomson ONE Banker M&A data | Foreign M&A fraction at country-industry level for capital flows proxy; M&A deal value 2000-2018 | No page yet |
| Penn World Table 10.1 (Feenstra, Inklaar, Timmer 2015) | TFP growth 2001-2019 for back-of-envelope calculation | No page yet |
| Capital IQ | Board member biographies to assess persistence of political connections in the 2000s for circa 1910 connected firms | No page yet |

Sample circa 1910: 1,115 firms, 60 countries, minimum 10 firms per country; book equity measured 1900-1925. Sample 2000: 30,891 firm-year observations for the full 47-country Worldscope panel (Table I Panel B, p. 3368). Italy DiD panel: 3,835 observations, 226 unique firms (main specification), 1921-1971.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13481) if you are: studying the long-run dynamics of corporate market power and the barriers that perpetuate it; building on the Rajan-Zingales framework for the interaction of political connections with trade and capital openness; using the Italy fascism quasi-experiment to identify causal effects of political connections on firm outcomes; or working on cross-country corporate governance with historical data going back to 1910. The Internet Appendix (available at the journal's online version) contains the list of 60 countries, 56 data sources, and variable definitions.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(6). This distillation was extracted by an LLM on 2026-06-03 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Faccio, Mara, and John J. McConnell.
> "Impediments to the Schumpeterian Process in the Replacement of Large Firms."
> *The Journal of Finance* 80, no. 6 (December 2025): 3359–3399.
> DOI: 10.1111/jofi.13481. © 2025 The Author(s).
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# The Stock Market and Bank Risk-Taking: Falato & Scharfstein (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/falato-stock-market-bank-risk-2025/
# Distilled: Banks that go public (IPO) increase risk as measured by confidential CAMELS supervisory ratings, relative to a matched control group of banks that filed but withdrew their IPOs. The increase in risk boosts short-term ROE but reduces it four years out, consistent with stock-market short-termism driving bank risk. J. Finance 2025, paywalled. Eight core results with source locators, datasets, the theoretical mechanism (Stein 1989 short-termism), and the difference-in-differences estimating equations.
# Tags: paper-summary, banking, risk-taking, short-termism, corporate-governance
==============================================================================

**What this is.** The paper's core results, the short-termism mechanism it tests,
and the difference-in-differences specifications with their equations: enough to
know what it found and how, without reading all 40 pages. To replicate or extend
it, read the full source at the [original](https://doi.org/10.1111/jofi.13502).

## TL;DR

Using confidential CAMELS supervisory ratings from the Federal Reserve, the paper
shows that banks increase risk after going public relative to a matched control
group of banks that filed for an IPO but withdrew it. The baseline difference-in-
differences (DD) estimate implies that going public raises the composite CAMELS
rating by 0.316 on a 1-5 scale (about half a within-bank standard deviation) and
raises the probability of a weak CAMELS rating (3 or above) by 8.8 percentage
points, roughly equal to the unconditional sample mean. The increase in risk
shows up both in observable balance sheet measures (higher risk-weighted assets,
less Tier 1 capital, more volatile liabilities) and in confidential supervisory
assessments that investors cannot observe, consistent with the Stein (1989) model
of short-termism in which managers boost unobservable hard-to-detect risks to
raise short-term earnings and stock prices. The risk increase is larger for banks
with higher institutional investor turnover, higher CEO short-term disclosure, and
larger equity option grants of shorter duration. Banks that went public before the
2007-2009 financial crisis underperformed significantly during the crisis.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%.
Locators point into the source PDF (page numbers from the journal article).

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Going public raises the composite CAMELS rating (higher = riskier) in the IPO sample | Table III Panel A col. (1), p. 3236 | DD coefficient = 0.316\*\*\* (SE 0.114); ~0.5 within-bank SD |
| R2 | Going public raises the probability of a weak CAMELS rating (3 or above) | Table III Panel A col. (2), p. 3236 | DD coefficient = 0.088\*\*\* (SE 0.024); equal to unconditional mean probability |
| R3 | 2SLS-IV using banking sector stock returns in the two months after announcement as instrument confirms the result | Table III Panel B cols. (1)-(2), p. 3237 | 2SLS coefficient = 0.343\*\* (SE 0.145) for composite CAMELS; 0.104\*\*\* (SE 0.033) for weak CAMELS |
| R4 | Asset risk (CAMELS A rating and STBL loan risk) rises after public transition | Table IV Panel A, p. 3242 | A-rating DD = 0.316\*\* (SE 0.167); STBL DD = 0.403\*\*\* (SE 0.125) in IPO sample |
| R5 | Financing risk (capital adequacy and liquidity in CAMELS; risk sensitivity) also rises after public transition | Table IV Panel B, p. 3242 | C,L-ratings DD = 0.254\*\* (SE 0.115); Risk-rating DD = 0.308\*\*\* (SE 0.111) in IPO sample |
| R6 | Observable balance sheet measures confirm risk increase: RWA/A rises and Tier 1 capital falls after IPO | Table V Panels A and B, p. 3244 | RWA/A DD = 0.371\*\*\* (SE 0.116); Tier 1 capital DD = -0.010\*\*\* (SE 0.003); Volatile Liabilities DD = 0.039\*\*\* (SE 0.008) |
| R7 | ROE rises 70 bps by quarter +4 but falls 130 bps below pre-transition level by year +4; earnings quality (CAMELS E) deteriorates; discretionary loan loss provisions fall post-IPO | Table VII Panels A and B, pp. 3247-3248 | ROE change at t+4 quarters = 0.007\*\* (SE 0.003); at t+16 quarters = -0.013\*\* (SE 0.006); E-rating DD at t+16 = 0.363\*\*\* (SE 0.157); LLP/Loans DD = -0.006\*\*\* (SE 0.001) |
| R8 | Risk increase is larger for banks with higher institutional investor turnover, more short-term CEO disclosure, higher equity option grant values, and shorter-duration option grants; banks that went public before the crisis underperformed significantly during 2007-2009 | Table VIII, p. 3250; Table X Panel B, p. 3253 | Triple-DD on CEO short-term disclosure = 0.669\*\* (SE 0.327); institutional investor turnover = 0.071\*\* (SE 0.029); crisis ROE interaction for full sample = -0.028\*\*\* (SE 0.002) |

**Overall (paper's conclusion).** Access to public equity markets causes banks to
increase risk. The mechanism is consistent with short-termism: public listing
introduces pressure from stock market investors and compensation incentives that
lead banks to boost short-run earnings by taking hard-to-observe risks, at the
expense of long-run performance. The finding has implications for both
compensation regulation and the wisdom of governance reforms that enhance
shareholder power, since good governance coupled with stock-market pressure may
increase rather than decrease risk in banking.

## Theory / model

The paper has no formal structural model. It is organized around the short-termism
channel first formalized by Stein (1989). In that model, stock market investors
rationally attribute higher current earnings to both a permanent and a transitory
shock. Because earnings embed news about long-run value, managers who care about
the short-term stock price have incentives to cut hard-to-observe long-run
investments and boost short-term earnings, even at the expense of long-run value.
In banking, the easiest way to raise short-term earnings is to take more risk:
loosen lending standards, increase loan yields, rely on cheaper but less stable
wholesale funding. These actions increase current earnings but create future credit
and rollover risks.

An alternative behavioral version due to La Porta (1996) holds that investors
overextrapolate current earnings, reinforcing the managerial incentive to boost
them. In both cases the prediction is the same: banks that place greater weight
on short-term stock price performance should take more risk after going public.

In Bolton, Scheinkman & Xiong (2006), equity compensation also leads to
short-termism because managers exploit the market's overvaluation of short-term
performance. Earlier cross-sectional work by Kwan (2004) and Nichols, Wahlen &
Wieland (2009) finds no significant risk differences between public and private
banks, but that work relies on ex-post performance measures rather than ex-ante
supervisory CAMELS ratings, which may explain why it misses the effect
documented here (p. 3252).

The two key empirical predictions are:

1. Going public raises ex-ante supervisory risk (CAMELS), especially on
   dimensions that investors cannot easily observe.
2. The risk increase should be larger for banks more subject to short-term pressure
   (higher institutional investor turnover, higher CEO short-term disclosure, more
   short-duration equity compensation) and should boost short-run ROE at the cost
   of long-run underperformance (p. 3225).

**Identification strategy.** The concern is that IPOs are endogenous: banks may go
public in response to growth opportunities that are also correlated with a riskier
environment. The paper addresses this by using a difference-in-differences (DD)
design in which the control group is banks that announced but then withdrew their
IPO filings (following Bernstein (2015) and Seru (2014)). The idea is that both
groups intended to go public for the same reasons, so comparing within-bank changes
in risk for treated banks to those of control banks differences out the selection
concern. The paper verifies that treated and control banks have parallel pre-trends
in CAMELS and are balanced on all observable characteristics except size (p. 3233,
Table II). As a further check, the paper instruments IPO completion with S&P
banking sector index returns in the two months after the announcement (Bernstein
(2015) instrument): deals announced when bank stocks are doing poorly are less
likely to be completed.

## Method

The paper applies three main estimators. It builds on `difference-in-differences`
as the primary causal design, `instrumental-variables` (2SLS) for robustness, and
`panel-regression` with `matching` for sensitivity.

**Baseline DD estimator (equation 1, p. 3232).** For each bank $$i$$ and year-quarter
$$t$$, the estimating equation is:

$$
\text{RISK}_{it} = \beta_1 \times \text{After}_{it} + \beta_2 \times \text{After}_{it} \times \text{Treatment}_i + \gamma \times Z_{it} + \mu_t + \alpha_i + \epsilon_{it} \tag{1}
$$

where $$\text{RISK}$$ is the composite CAMELS rating (1-5) or the weak CAMELS indicator;
$$\text{After}_{it}$$ equals one for all bank-quarters after the IPO announcement date;
$$\text{Treatment}_i$$ equals one for banks that completed the IPO (zero for withdrawn
filers); $$Z_{it}$$ is bank size (log total assets); $$\mu_t$$ is year-quarter fixed
effects; and $$\alpha_i$$ is bank fixed effects. Standard errors are clustered at the
BHC level. The coefficient of interest is $$\beta_2$$, the difference-in-differences
estimate.

**2SLS-IV estimator (equations 2 and 3, p. 3238).** To address residual selection
concerns, the paper instruments deal completion with banking sector stock returns
in the two months after the announcement. The second-stage is:

$$
\text{RISK}_i^{\text{Post}} = \beta_1 \widehat{\text{CompletedIPO}}_i + \gamma_1 \text{RISK}_i^{\text{Pre}} + \gamma_2 Z_i + \mu_t + \epsilon_i \tag{2}
$$

where $$\text{RISK}_i^{\text{Post}}$$ is the average risk proxy after the announcement and
$$\text{RISK}_i^{\text{Pre}}$$ is the pre-announcement average. The first stage is:

$$
\text{CompletedIPO}_i = \beta_2 \cdot S\&P\text{BankReturns}_i + \gamma_3 \text{RISK}_i^{\text{Pre}} + \gamma_4 Z_i + \mu_t + \epsilon_i \tag{3}
$$

where $$S\&P\text{BankReturns}_i$$ is the S&P bank index return in the two months after
the announcement. The exclusion restriction is that these short-window returns are
uncorrelated with longer-term bank-specific risk decisions.

**Triple-DD estimator (cross-sectional heterogeneity, p. 3250).** To test whether
the risk increase is greater for banks with stronger short-term incentives, the
paper estimates:

$$
\text{RISK}_{it} = \beta_1 \text{After}_{it} + \beta_2 \text{After}_{it} \times \text{Treatment}_i + \beta_3 \text{After}_{it} \times \text{Treatment}_i \times X_i + \beta_4 \text{After}_{it} \times X_i + \gamma Z_{it} + \gamma_1 \text{After}_{it} \times Z_{it} + \mu_t + \alpha_i + \varepsilon_{it}
$$

where $$X_i$$ is the cumulative density of a short-termism proxy (institutional
investor turnover, CEO short-term disclosure, equity option value or duration).
The term $$\text{After}_{it} \times X_i$$ cannot be identified from $$\text{After}_{it} \times
\text{Treatment}_i \times X_i$$ because $$X_i$$ does not vary within private banks, so
it drops out of the estimation.

## Empirical specifications

All main regressions use quarterly Call Report and supervisory data for U.S.
commercial banks held by BHCs, 1990-2012, restricted to a 10-year pre-crisis
window (1997-2006) for baseline tests. Standard errors are clustered at the BHC
level. The key specifications and their links to the core results are:

- **Baseline DD on composite CAMELS and weak CAMELS (R1, R2).** Spec (1) in Table
  III, Panel A: equation (1) above with bank FE, year-quarter FE, supervisor FE,
  and log total assets; IPO sample of 406 completed and 122 withdrawn banks; 8,237
  bank-quarter observations.

- **Subcomponent supervisory ratings (R4, R5).** Table IV Panels A and B: same
  specification (1) applied to the A rating (asset quality), STBL loan risk
  rating, C and L combined (capital adequacy and liquidity), and the risk
  sensitivity rating. Identifies which dimensions of the CAMELS deteriorate.

- **Balance sheet DD (R6).** Table V: equation (1) with RWA/A, residential real
  estate loans to total loans, Tier 1 capital ratio, and volatile liabilities
  as outcomes. Confirms risk increase is visible in observable data.

- **Robustness via controlling for observables (Table VI).** Adds the balance
  sheet risk measures as controls to the supervisory rating regressions; the
  treatment effect remains significant, confirming that hidden risk
  (unobservable to investors) is the source of the CAMELS deterioration.

- **Performance dynamics (R7, Table VII).** Calendar-time specification:
  $$Y_{t+N} - Y_{t-1} = \beta_1 \text{Treatment}_i + \gamma Z_{it} + \mu_t + \alpha_i + \varepsilon_{it}$$ tracking the change in
  quarterly ROE and the E rating at horizons N = 1, 4, 8, and 16 quarters
  post-announcement; and DD regressions of discretionary LLP/loans, loan loss
  provisions to delinquencies, IBES long-term EPS growth forecast, and earnings
  restatements.

- **Triple-DD on mechanism (R8, Table VIII).** Interacts the treatment effect
  with CEO short-term disclosure (Brochet, Loumioti & Serafeim (2015) frequency-
  of-short-term-horizon-words measure from earnings calls and 10-K MD&A sections),
  institutional investor turnover, equity option B-S value, and option grant
  duration.

- **M&A robustness (Table III Panel A cols. 3-4).** Replicates the main result
  using the Completed M&As Sample (1,631 banks, 10,312 observations), where
  treatment is acquisition by a publicly traded BHC and control is other
  acquisitions that do not change ownership status. Coefficients are smaller
  (0.098\*\*\* for composite CAMELS, M&A sample) but strongly significant.

- **Broader cross-section and financial crisis (Table X).** OLS and FE regressions
  of CAMELS on a public BHC dummy in the merged BHC-Commercial Bank Sample
  restricted to the pre-crisis window (90,733 bank-quarter observations);
  Public BHC dummy coefficient = 0.089\*\*\* (FE, p. 3253). Triple-DD
  on crisis ROE: After\*Treatment\*Crisis = -0.028\*\*\* (full IPO sample, p. 3253).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| NIC (National Information Center), Federal Reserve | Confidential CAMELS supervisory ratings (composite and components); BHC ownership and public listing history | no page yet |
| STBL (Survey of Terms of Business Lending), Federal Reserve | Confidential loan-level risk ratings for C&I loans (1 to 5 scale); 1997-2012; used by DellAriccia, Laeven & Suarez (2017) and others | no page yet |
| Call Reports (Reports of Condition and Income), FDIC/Federal Reserve | Balance sheet variables: total assets, RWA, Tier 1 capital, deposits, loans, volatile liabilities; 1990-2012 | no page yet |
| SDC New Issues / S&P Capital IQ / SNL Financial Capital Offerings | Lists of completed and withdrawn bank IPO filings; 1990-2012 | no page yet |
| CRSP-FRB Link (New York Fed) | Stock market listing history for BHC public/private status | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| IBES (Institutional Brokers' Estimate System) | Equity analysts' consensus long-term EPS growth forecasts for newly public banks | [I/B/E/S](/wiki/commercial/ibes/) (licensed) |
| Thomson-Reuters Institutional Holdings (13F) | Institutional investor portfolio turnover; 1990-2012 | [Thomson 13F (s34)](/wiki/commercial/thomson-13f/) (licensed) |
| Riskmetrics / Capital IQ | Employee stock option grant data (B-S value, duration) | no page yet |

Sample: 178,980 bank-quarter observations for 7,166 (3,251) unique banks (BHCs);
IPO identification sample 17,754 bank-quarter observations for 528 unique
commercial banks (276 BHCs), 1990-2012; baseline tests use the 10-year window
1997-2006 (8,237 observations, Table I p. 3231). CAMELS ratings are quarterly.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13502) if you are:
studying bank risk regulation and the role of stock market incentives;
replicating the DD design with confidential supervisory data (the Internet
Appendix has additional robustness tables and placebo tests);
building on the short-termism mechanism in nonbank financial intermediaries;
or extending the analysis to international banking systems or post-crisis periods.
The locators above point to the exact tables.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(6), December 2025, pp. 3223-3261.
DOI: [10.1111/jofi.13502](https://doi.org/10.1111/jofi.13502). Paywalled; Wiley
standard terms-of-use (not CC). This distillation was extracted by an LLM on
2026-06-03 and is **not human-verified or independently reproduced**.

> Falato, Antonio, and David Scharfstein. "The Stock Market and Bank Risk-Taking."
> *The Journal of Finance* 80, no. 6 (December 2025): 3223-3261.
> DOI: 10.1111/jofi.13502. Extract-only: no redistribution of the verbatim article.


==============================================================================
# The Value of Bank Lending: Flanagan (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/flanagan-value-bank-lending-2025/
# Distilled: Using novel realized cash flows for 8,100 syndicated term loans (1992-2014) and a private-equity-style risk-adjustment methodology, Flanagan (2025) finds that banks earn 177 bps annualized gross risk-adjusted returns on loan cash flows, add roughly $75 million of value annually per loan portfolio, and that shareholders receive near-zero net risk-adjusted returns once lending expenses are deducted. J. Finance 2025, CC BY-NC 4.0. Eight core results with source locators, datasets used, the economic framework, the method (risk-adjusted profit adapted from Gupta and Van Nieuwerburgh (2021)), and empirical specifications.
# Tags: paper-summary, banking, credit-supply, financial-intermediation
==============================================================================

**What this is.** The paper's core results, the economic framework it builds on, and the risk-adjustment method it adapts from private equity to bank loans, with the defining equations: enough to know what it found and how, without reading all 45 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13465).

## TL;DR

The paper asks: do banks produce real value through lending, and who captures it? Using a novel dataset of realized cash flows for 8,100+ U.S. syndicated term loans (1992-2014), Flanagan adapts the Gupta and Van Nieuwerburgh (2021) private equity risk-adjustment method to estimate the risk-adjusted profit (RAP) of bank loans relative to public-market benchmarks. Banks earn on average 177 basis points annualized gross risk-adjusted returns, consistent with providing valuable screening and monitoring services. These returns are larger when borrowers face more severe financing frictions and when banks invest more resources in lending services. Bank-level RAP also persists over time, indicating skill heterogeneity. However, once commercial lending expenses (182 bps, primarily staff compensation) are deducted, shareholders receive near-zero net risk-adjusted returns, consistent with competitive equilibrium: the present value of loan cash flows covers lending costs, not shareholder rents.

## Core results

Magnitudes and significance are as reported; `\*` = 10%, `\*\*` = 5%, `\*\*\*` = 1%. Locators point into the source PDF (pp. 2017-2061).

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Baseline gross RAP: banks earn 177 bps annualized** risk-adjusted returns on syndicated loan cash flows | Table II Panel B col (2), p. 2037 | RAP = $0.032 per $1 invested; annualized psi = 1.77%\*\*\* (SE 0.059) |
| R2 | **Higher RAP for financially constrained borrowers**: top vs bottom financial-constraint quartile earns 93% higher annualized RAP | Table III, p. 2041 | Top-quartile psi = 2.42%\*\*\*, bottom = 1.26%\*\*\*; H-L = 1.17%\*\*\* (SE 0.22) |
| R3 | **Shorter-maturity loans earn 53% higher annualized RAP**, consistent with a fixed cost of screening | Table IV, p. 2043 | Low-maturity psi = 2.35%\*\*\*, high-maturity = 1.54%\*\*\*; H-L = -0.81%\*\*\* (SE 0.26) |
| R4 | **Loans to high fixed-cost industries earn 44-47% higher RAP**, consistent with greater irreversibility requiring more screening | Table IV, p. 2043 | High fixed-cost psi = 2.17%\*\*\*, low = 1.51%\*\*\*; H-L = 0.66%\*\*\* (SE 0.18) |
| R5 | **Tighter earnings-based covenants (EBC) earn higher RAP**, consistent with monitoring intensity; higher renegotiation probability earns lower RAP (substitution) | Table IV, p. 2043 | High EBC tightness psi = 2.44%\*\*\*, low = 1.66%\*\*\*; H-L = 0.78%\*\*\* (SE 0.25) |
| R6 | **Bank-level RAP persists 1-4 years ahead**; Dollar Value-Added is a stronger persistence predictor than RAP alone | Table VI, p. 2048 | L4.RAP = 0.101\*\*\* (SE 0.032); L4.DVA = 0.473\*\*\* (SE 0.121); R2 for DVA ~0.30 vs ~0.14 for RAP |
| R7 | **Shareholders receive near-zero net risk-adjusted returns** after deducting commercial lending expenses of 182 bps | Table VII, p. 2050 | Net psi = approx -5 bps, statistically indistinguishable from zero; 113 bps staff comp + 69 bps other expenses |
| R8 | **Corporate bond placebo confirms methodology**: risk-adjusting corporate bond cash flows yields near-zero RAP, validating the loan estimates | Table VIII Panel B, p. 2053 | Corporate bond RAP Ann = 0.015 (SE 0.102), not significant; full-sample risk-free return = 2.77%\*\*\* is eliminated by risk adjustment |

**Overall (paper's conclusion).** Bank loan cash flows have large, economically significant gross risk-adjusted returns that compensate bank employees for their costly screening and monitoring services. The cross-sectional pattern of these returns is consistent with banks providing more valuable intermediation to more financially constrained borrowers. However, shareholders of banks receive approximately zero net risk-adjusted returns from syndicated lending, consistent with competitive provision of lending services and with the model of Philippon (2010). These results help reconcile the empirical literature finding low bank shareholder value (Begenau and Stafford (2019)) with classic theories of banking emphasizing productive lending services (Leland and Pyle (1977), Diamond (1984)).

## Theory / model

The paper has no formal structural model but provides an economic framework (Section I.A, pp. 2025-2027) that motivates the empirical tests and the interpretation of RAP.

**Cost of capital decomposition.** A firm's cost of borrowing via a bond market is

$$
r_{\text{firm,bond}} = r_f + r_{\text{risk}} + \theta_{\text{bond}},
$$

where $$r_f$$ is the risk-free rate, $$r_{\text{risk}}$$ is the risk premium in a frictionless market, and $$\theta_{\text{bond}}$$ is the financing friction specific to the bond market. The cost of a bank loan is

$$
r_{\text{firm,loan}} = r_f + r_{\text{risk}} + \theta_{\text{loan}} + \psi,
$$

where $$\theta_{\text{loan}} \leq \theta_{\text{bond}}$$ (banks mitigate financing frictions through information production and monitoring) and $$\psi \geq 0$$ is the bank's required return for supplying lending services. The incentive compatibility constraint for banks is

$$
\psi - \gamma \geq 0,
$$

where $$\gamma$$ is the cost of compensating loan officers and covering other expenses. In competitive equilibrium $$\psi = \gamma$$, and the borrower takes a loan if the net cost reduction is positive:

$$
(\theta_{\text{bond}} - (\theta_{\text{loan}} + \psi)) \geq 0.
$$

**Total value decomposition.** The total value from mitigated financial frictions decomposes into the present value flowing to the lender and the value flowing to the borrower (pp. 2026-2027). Letting $$X_t$$ be the outstanding principal balance,

$$
\underbrace{\sum_{h=1}^{H} \frac{(\theta_{\text{bond}} - \theta_{\text{loan}}) X_{t-1}}{(1 + r_f + r_{\text{risk}})^h}}_{\text{Total Value}} \approx \underbrace{\sum_{h=1}^{H} \frac{r_{\text{firm,loan}} \cdot X_{h-1} + (X_h - X_{h-1})}{(1 + r_f + r_{\text{risk}} + \theta_{\text{loan}})^h} - X_0}_{\text{Present Value to Lender}} + \underbrace{\sum_{h=1}^{H} \frac{(\theta_{\text{bond}} - (\theta_{\text{loan}} + \psi)) X_{h-1}}{(1 + r_f + r_{\text{risk}})^h}}_{\text{Borrower Value}}.
$$

Empirically, the present value to the lender corresponds directly to the RAP estimated in the paper, and is a lower bound on the total social value generated. The annualized version of this present value is the parameter $$\psi$$ estimated throughout.

**Key hypotheses.** (i) Classic banking theory predicts higher RAP for more financially constrained borrowers, where screening and monitoring provide greater value. Leland and Pyle (1977), Diamond (1984), and Holmstrom and Tirole (1997) all predict banks create value through information production; here, the incentive compatibility condition $$\psi - \gamma \geq 0$$ ties the bank's required spread directly to its lending costs. (ii) Risk-shifting theory predicts the opposite. (iii) Competitive equilibrium predicts near-zero net shareholder returns ($$\psi = \gamma$$). (iv) Persistence in bank-level RAP, in the spirit of Berk and van Binsbergen (2015) for mutual funds, indicates skill heterogeneity in lending services. Indirect evidence from James (1987) and Berger and Udell (1995) shows borrowers benefit from bank lending through higher stock prices and better loan pricing, consistent with the framework.

## Method

The method adapts the Gupta and Van Nieuwerburgh (2021) private equity strip-by-strip risk-adjustment (RAP) methodology to bank loan cash flows, with several extensions. It builds on `panel-regression` and `fama-macbeth` for factor selection and inference.

**Core pricing equation.** Let $$R^k_{t+h}$$ denote the cumulative return on public security $$k$$ from $$t$$ to $$t+h$$. The no-arbitrage pricing equation for the SDF $$M_{t,t+h}$$ is

$$
\text{E}_t[M_{t,t+h} R^k_{t+h}] = 1. \tag{1}
$$

Gupta and Van Nieuwerburgh (2021) estimate the regression

$$
X^i_{t+h} = a_{t+h} + b_h R^k_{t+h} + e^i_{t+h}, \tag{2}
$$

where $$X^i_{t+h}$$ is the cash flow to loan portfolio $$i$$ at horizon $$h$$, normalized to a $1 investment. The key identification assumption is that $$b_h R^k_{t+h}$$ spans all priced risk in the cash flows, so residuals are orthogonal to the SDF.

**Loan benchmark funds.** Because loan principal balances amortize (more than 80% repaid within 4 years, Figure 3, p. 2032), the paper instruments public security returns $$R^k_{t+h}$$ with the loan's outstanding balance $$z^i_t$$, defining instrumented returns $$\tilde{R}^i_{t+h} = R_{t+h} z^i_t$$. Their price satisfies

$$
\text{E}_t[M_{t,t+h} \tilde{R}^i_{t+h}] = \text{E}_t[M_{t,t+h} R_{t+h} z^i_t] = z^i_t. \tag{3}
$$

Two types of benchmark funds implement this (similar in spirit to the benchmark fund construction in Korteweg and Nagel (2016) for venture capital): a *rollover investment benchmark* $$\bar{F}^{i,k}_{t+h}$$ (pays out the change in loan balance and reinvests the rest each period, price = $1 by equation (4)) and a *gain investment benchmark* $$\bar{G}^{i,k}_{t+h}$$ (goes long risky asset and short risk-free bond, accumulates compounded returns, winds down with the loan balance, price = $0 by equation (5)). The main regression is then

$$
X^i_{t+h} = a_{t+h} + \sum_{k=1}^K \left[ b^k \bar{F}^{i,k}_{t+h} + c^k_h \bar{G}^{i,k}_{t+h} \right] + e^i_{t+h}. \tag{6}
$$

**RAP estimation.** The unconditional mean RAP is

$$
\widehat{RAP} = \text{E}[\text{E}_t[\sum_{h=1}^H M_{t,t+h} X^i_{t+h}]] - 1 = \frac{1}{N}\sum_{i=1}^N [\sum_{h=1}^H P^{\$}_{t,h} \hat{a}_{t+h} + \sum_{k=1}^K \hat{b}^k + \sum_{h=1}^H P^{\$}_{t,h} \hat{e}^i_{t+h}] - 1. \tag{8}
$$

The annualized risk-adjusted return is

$$
\hat{\psi} = \frac{\widehat{RAP}}{\text{WAL}}, \tag{9}
$$

where WAL is the weighted-average life of the outstanding loan balance. Inference uses a nonparametric block bootstrap (100 replications) following Driessen, Lin, and Phalippou (2012).

**Risk factors.** The baseline model uses: a risk-free floating-rate rollover benchmark (analogous to a LIBOR-linked bond), a Term factor (10-year Treasury returns), BBB-rated corporate bond returns, HY-rated corporate bond returns, and CRSP value-weighted stock returns plus the bottom-quintile size portfolio. The model's $$R^2 > 0.99$$ (Table II Panel A, p. 2037), indicating benchmark funds span nearly all variation in loan cash flows.

**Noninterest expense hedonic regression.** To estimate net risk-adjusted returns to shareholders, the paper estimates commercial loan expense ratios using the approach of Hanson et al. (2015). For FRY-9C bank holding companies (1994-2014), the pooled regression is

$$
\frac{\text{Expense}_{it}}{\text{Asset}_{it}} = a + \sum_{k=1}^K b^{(k)} \frac{\text{Asset}^{(k)}_{it}}{\text{Asset}_{it}} + \sum_{j=1}^J c^{(j)} \frac{\text{Deposit}^{(j)}_{it}}{\text{Asset}_{it}} + dX_{it} + e_{it}, \tag{10}
$$

estimated with Fama-MacBeth (1973) cross-sectional regressions and Newey-West (1) standard errors. The commercial loan expense ratio is the predicted value for a hypothetical bank investing only in commercial loans with wholesale funding.

## Empirical specifications

All regressions use quarterly loan cash flows normalized to a $1 investment. The main estimating equation is (6) above, estimated by OLS with nonparametric block bootstrap inference.

**Baseline (R1).** Equation (6) is estimated on 259,300 quarterly observations of the full loan sample (8,125 loans, 1992Q3-2014Q1). Factors: Rf (floating-rate rollover), Term, BBB, HY, Stock. Risk-adjusted returns are computed via equation (8), annualized by WAL (equation (9)). The model achieves $$R^2 = 0.991$$ (Table II Panel A), confirming near-complete spanning.

**Financial constraints (R2).** Loans are sorted into four quartile buckets by the first principal component of four financial-constraint indicators (log firm size with negative sign, log firm age with negative sign, Firm Unrated indicator, Firm Issued Bond indicator with negative sign). Equation (6) is estimated separately for each bucket. Differences in bucket-level annualized RAP test whether banks earn more when borrowers face more severe frictions (Table III, p. 2041). The H-L spread is bootstrapped across the two buckets.

**Screening, monitoring, renegotiation (R3-R5).** Loans are sorted by: (i) contractual maturity (above/below bottom and top quartile), (ii) average industry fixed costs (SG&A from Compustat, two-digit SIC), (iii) predicted ex ante renegotiation probability (Prob(Reneg)) using a linear probability model on loan characteristics, and (iv) EBC tightness (Murfin (2012) measure, following Kermani and Ma (2020)). Equation (6) is estimated separately for each half and differences are bootstrapped (Table IV, p. 2043).

**Bank-level persistence (R6).** Bank-level portfolios aggregate individual loan cash flows weighted by lead-lender retention fraction. Equation (7) (the conditional version of equation (6)) yields a time series of bank-level risk-adjusted returns. Panel regressions with quarter-time fixed effects regress these on lags L4, L8, L12, L16 (in quarters), with standard errors clustered by bank (Table VI, p. 2048). Dollar Value-Added is RAP times portfolio size.

**Net shareholder returns (R7).** Commercial lending expense ratios from equation (10) are subtracted from gross annualized RAP to obtain net shareholder risk-adjusted returns (Table VII, p. 2050; Table IA.XXIV in Internet Appendix for bank-level results).

**Placebo validation (R8).** Equation (6) is applied to corporate bond cash flows (fixed-rate, maturities up to 8 years, from Mergent FISD with TRACE transaction prices). Near-zero bond RAP validates the spanning assumption (Table VIII, p. 2053).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Dealscan (Refinitiv Loan Connector) | Loan originations, amortization schedules, interest rate spreads, covenants, lead lender retention; primary source for loan cash flow construction | [DealScan](/wiki/commercial/dealscan/) (licensed) |
| Compustat (via Chava-Roberts 2008 linking file) | Borrower financial characteristics (firm size, age, fixed costs, financial constraints); performance-pricing covenant ratios | [WRDS / Compustat](/wiki/commercial/wrds/) (licensed) |
| CRSP | Stock return factors for benchmark fund construction; value-weighted market return; size quintile portfolio | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| SEC EDGAR (10-K, 10-Q, 8-K filings, web-scraped) | Loan prepayment and refinancing dates not in Dealscan; identifies 94% of prepayment dates accurately | [SEC EDGAR](/wiki/datasets/edgar/) |
| FRY-9C bank holding company data | Bank-level expense decomposition (noninterest expense, compensation, other); bank financial constraint measures | no page yet |
| Corporate bankruptcy databases (UCLA, 8-K) | Default identification and recovery rate estimation (industry-by-year LGD from Moody's Annual Default Report) | no page yet |
| Mergent FISD / TRACE | Corporate bond cash flows and transaction prices for placebo validation test | no page yet |
| Federal Reserve / Treasury (yield curve) | Risk-free ZCB term structure for discounting; 3-month T-bill rate as risk-free proxy | no page yet |
| Moody's Annual Default Report | Industry-by-year loss-given-default (LGD) estimates for loan recovery rates | no page yet |

Sample: 8,125 senior floating-rate term loans to U.S. public borrowers, originated 1992Q3-2014Q1. Security returns data cover 1992Q3-2021Q2. Bank-level tests include 31 banks with at least 40 quarters of data.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13465) if you are: building or testing a risk-adjustment methodology for non-traded private credit instruments; studying the distribution of value added between bank shareholders, employees, and borrowers; extending the cross-sectional tests to other loan types (mortgages, small business); or applying the hedonic expense regression of Hanson et al. (2015) to decompose lending costs. The internet appendix contains extensive robustness tests, the loan-matching algorithm, and the simulation exercise.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(4). This distillation was extracted by an LLM on 2026-06-05 and is **not human-verified or independently reproduced**. Licensed CC BY-NC 4.0; the verbatim PDF is not hosted here.

> **Attribution (CC BY-NC 4.0).** Flanagan, Thomas.
> "The Value of Bank Lending."
> *The Journal of Finance* 80, no. 4 (August 2025): 2017-2061.
> DOI: 10.1111/jofi.13465. © 2025 The Author(s).
> Licensed under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Intermediary Leverage Shocks and Funding Conditions: Fontaine, Garcia & Gungor (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/fontaine-intermediary-leverage-shocks-funding-2025/
# Distilled: Broker-dealer aggregate leverage responds to both demand and supply disturbances with opposite effects on expected returns and funding conditions. Disentangling the two shocks resolves sign puzzles on raw leverage risk across equity, bond, and option markets and confirms intermediary constraints as a priced source of risk. J. Finance 2025, paywalled. Seven core results with source locators, datasets used, the econometric model, and the structural VAR identification procedure.
# Tags: paper-summary, asset-pricing, intermediary-asset-pricing, factor-models
==============================================================================

**What this is.** The paper's core results, the econometric model of leverage demand and supply shocks, and the structural VAR identification strategy with its defining equations: enough to know what it found and how, without reading 43 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13407).

## TL;DR

Broker-dealer aggregate leverage responds to both demand shocks (customers seeking immediacy) and supply shocks (financiers relaxing funding constraints). These two shocks both raise leverage but shift the intermediaries' marginal value of wealth in opposite directions: supply shocks improve funding conditions and carry a positive price of risk; demand shocks tighten funding conditions and carry a negative price of risk. A parsimonious two-shock structural model with a funding-conditions instrument achieves a cross-sectional R-squared of 92% over equities, bonds, and options, versus 8% for raw leverage alone. Disentangling the shocks also resolves why raw leverage carries positive price-of-risk estimates in bond markets but negative (or insignificant) ones in option markets, and why leverage is largely uncorrelated with stock market liquidity in the data.

## Core results

Magnitudes and significance are as reported; `\*` = 5%, `\*\*` = 1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Price of symmetric leverage demand/supply risk is significant with 92% cross-sectional fit | Figure 4, p. 77; text p. 76 | lambda = 2.12 (95% CI [1.38, 3.55]); R-squared = 92.1% (CI [78%, 98%]) |
| R2 | Supply shock price is positive and demand shock price is negative, as predicted, across all test assets | Table I cols (2)-(3), p. 80 | lambda_s = 4.05 (t=2.56); lambda_d = -4.47 (t=-2.38); raw leverage alone (col 1) lambda_l = 4.56 (t=2.45) but R-squared only 7.9% |
| R3 | Decomposing leverage raises cross-sectional R-squared from 8% to 90%-93% | Table I, p. 80 | R-squared: col (1) raw = 7.9%; col (2) supply only = 90.5%; col (3) demand only = 88.6%; col (4) symmetric = 93.1% |
| R4 | Price-of-risk magnitude is consistent across equities, bonds, and options when demand/supply shocks are used | Table II, p. 85 | Equities: lambda = 2.04 (t=2.85); Bonds: 2.04 (t=3.72); Options: 2.22 (t=2.55); raw leverage price negative and insignificant for options |
| R5 | Model confirms AEM's result using AEM's own test assets, reconciling the weaker but positive raw leverage price they find | Table VI, p. 95 | lambda = 1.86 (t=2.20) symmetric; supply price 3.25 (t=2.11); demand price -4.54 (t=-1.68); R-squared = 82% |
| R6 | Leverage supply shocks have a significant negative effect on stock illiquidity (improve liquidity), while demand shocks do not | Table V Panel A, pp. 93-94 | Supply coefficients: -42.8 (t=-1.79) to -0.001 (t=-2.40) monotonically across illiquidity deciles; demand shocks insignificant in every portfolio |
| R7 | Model explains why raw leverage price of risk switches sign across asset classes: supply betas dominate in bonds, demand betas dominate in options | Figure 8, p. 91; Table IV Panel B, p. 90 | Uncentered R-squared = 48% for model vs. OLS estimates of raw leverage price; sign correct in every asset class |

**Overall (paper's conclusion).** Disentangling demand and supply disturbances to broker-dealer leverage substantially strengthens evidence for a central role of intermediaries in asset pricing. Both shocks carry consistent and significant prices of risk across equities, bonds, and options, with opposite signs. The mixing of these two shocks in raw leverage explains previously puzzling sign reversals (e.g., in Adrian, Etula, and Muir (2014) and He, Kelly, and Manela (2017)) and the weak correlation between leverage and market liquidity documented by Brunnermeier and Pedersen (2009). Future work is needed to identify the deeper structural mechanisms driving each type of shock and to explain their varying importance across markets.

## Theory / model

The econometric model (Section I, p. 62) represents broker-dealer log leverage $$\text{LEV}$$ as the sum of two independent disturbances (eq. 1, p. 63):

$$
\text{LEV} = \mu_l + b_d e^d + b_s e^s, \tag{1}
$$

where $$b_d, b_s > 0$$ are positive loadings, $$e^d$$ is the leverage demand shock (customers demanding immediacy), and $$e^s$$ is the leverage supply shock (financiers relaxing funding constraints). Both shocks have zero mean and unit variance.

Intermediaries' marginal value of wealth $$\phi$$ is driven by the same shocks in opposite directions (eq. 2, p. 63):

$$
\phi = \gamma + \alpha_d e^d - \alpha_s e^s, \tag{2}
$$

with $$\alpha_d, \alpha_s, \gamma > 0$$. The demand shock $$e^d$$ raises both leverage and marginal value of wealth (tighter funding); the supply shock $$e^s$$ raises leverage but reduces marginal value of wealth (easier funding). Asset $$i$$ earns excess return $$xR_i$$ given by (eq. 3, p. 63):

$$
xR_i = \mu_i + \beta_{i,d} e^d + \beta_{i,s} e^s + e^i, \tag{3}
$$

where $$e^i$$ is an idiosyncratic shock. Intermediaries price assets such that their marginal value of wealth spans the pricing kernel (eq. 4, p. 64):

$$
\mathbb{E}[xR_i] = -\frac{\text{Cov}[\phi, xR_i]}{\mathbb{E}[\phi]}. \tag{4}
$$

Equations (1)-(4) jointly pin down expected returns (eq. 5, p. 64):

$$
\mu_i = \beta_{i,d} \lambda_d + \beta_{i,s} \lambda_s = \beta_i^\top \lambda, \tag{5}
$$

with prices of risk $$\lambda_d = -\alpha_d \gamma^{-1} < 0$$ and $$\lambda_s = \alpha_s \gamma^{-1} > 0$$.

**Implication 1** (p. 64): Risky assets have $$\beta_{i,d} < 0$$ and $$\beta_{i,s} > 0$$, and the corresponding prices of risk $$\lambda_d < 0$$ and $$\lambda_s > 0$$.

The raw leverage factor $$L = \text{LEV} - \mathbb{E}[\text{LEV}] = b_d e^d + b_s e^s$$ is a mix, so the raw leverage beta is (eq. 6, p. 64):

$$
\beta_{i,l} \equiv \frac{\text{Cov}(L, xR_i)}{\text{Var}(L)} = (\sigma_l^2)^{-1}(b_d \beta_{i,d} + b_s \beta_{i,s}), \tag{6}
$$

which can take either sign because it mixes demand and supply betas. The price of raw leverage risk in a cross-section depends on the dispersion and covariance of demand and supply betas (eq. 7, p. 64):

$$
\lambda_l = c(b_s \omega_s^2 \lambda_s + b_d \omega_d^2 \lambda_d + \omega_{ds}(b_s \lambda_d + b_d \lambda_s)), \tag{7}
$$

where $$c = \sigma_l^2 (b^\top \Omega b)^{-1} > 0$$. **Implication 2** (p. 65): the sign of $$\lambda_l$$ depends on the dispersion of demand and supply betas and their correlation, which can differ across asset classes.

To verify the prediction for market liquidity, illiquidity $$\Lambda_i$$ is modeled as proportional to intermediaries' marginal value of wealth (eq. 20, p. 92):

$$
\Lambda_i = \delta_i (\phi - \gamma), \tag{20}
$$

leading to **Implication 3**: supply shocks have negative population coefficients in illiquidity regressions, demand shocks have positive coefficients, and the sign of the raw leverage coefficient is determined by $$(b_d - b_s) \alpha \delta_i$$ (eqs. 21-22, p. 92).

## Method

**Identification.** The identification strategy (Section II.A, pp. 65-67) uses a funding-conditions instrument $$Z$$ that is correlated with both types of shocks:

$$
Z = \mu_Z + a_d e^d - a_s e^s, \tag{8}
$$

$$
u = \begin{bmatrix} u^z \\ u^l \end{bmatrix} = \begin{bmatrix} Z - \mu^z \\ \text{LEV} - \mu^l \end{bmatrix} = \begin{bmatrix} a_d & -a_s \\ b_d & b_s \end{bmatrix} \begin{bmatrix} e^d \\ e^s \end{bmatrix} = Ae. \tag{9}
$$

The variance of observed innovations is $$\text{Var}(u) = AA^\top$$ (eq. 10, p. 66), which provides three restrictions but leaves four parameters $$(a_d, a_s, b_d, b_s)$$ underdetermined. The paper achieves point identification by imposing the economic symmetry restriction $$\lambda_s = -\lambda_d = \lambda > 0$$ (equivalently $$\kappa = 1$$ in $$\lambda_s = -\kappa\lambda_d$$), which links the ratio of reduced-form price-of-risk coefficients to structural parameters (eq. 13, p. 66):

$$
\frac{c_l}{c_z} = \frac{a_s - a_d \kappa}{b_s + b_d \kappa}. \tag{13}
$$

**Identification 1** (p. 67) states: given independent shocks, positive parameters, and symmetric prices of risk, the structural parameters are identified in closed form:

$$
a_d = \frac{\sigma_z^2 \varphi_a}{2a_s}, \quad a_s = \sigma_z \sqrt{\frac{1 \pm \sqrt{1 - \varphi_a^2}}{2}},
\quad b_d = \frac{\sigma_l^2 \varphi_b}{2b_s}, \quad b_s = \sigma_l \sqrt{\frac{1 \pm \sqrt{1 - \varphi_b^2}}{2}}, \tag{14}
$$

where $$\varphi_a, \varphi_b \leq 1$$ depend on $$\text{Var}(u)$$ and $$C$$ (eq. 14, p. 67).

**Estimation procedure** (Section II.C, pp. 69-70). A VAR(1) model is specified for $$y_t = [\text{LEV}_t, Z_t]^\top$$:

$$
y_{t+1} = a + \Phi y_t + u_{t+1}, \tag{17}
$$

estimated by OLS to recover forecast errors $$\hat{u}_{t+1}$$ and their covariance $$\hat{\Sigma}_u$$. The reduced-form coefficient $$\hat{C}$$ is recovered from a cross-sectional OLS regression of average returns $$\mathbb{E}_T[xR_i]$$ on covariances $$\mathbb{E}_T[xR_i u]$$. The matrix $$\hat{A}$$ follows from eq. (14), structural shocks from $$\hat{e}_t = \hat{A}^{-1}\hat{u}_t$$, and the price of risk from $$\hat{\lambda} = \hat{A}^\top \hat{C}$$. Standard errors are from a block bootstrap to account for serial correlation and heteroskedasticity.

**Instrument construction** (Section II.D, p. 70). The funding-conditions proxy $$\mathit{FUND}$$ is the first principal component of three Treasury-market measures: the TED spread ($$\mathit{TED}$$), the Hu, Pan, and Wang (2013) noise measure ($$\mathit{HPW}$$), and the Fontaine and Garcia (2012) term-structure factor ($$\mathit{FG}$$), monthly data January 1986 to December 2021. The approach builds on Goldberg (2020) and Goldberg and Nozawa (2021), who identify demand and supply shocks in Treasury and corporate bond inventory, and is also consistent with Du, Hebert, and Huber (2022), who show leverage constraints do not always bind. A higher $$\mathit{FUND}$$ value signals tighter funding conditions / higher marginal value of intermediary wealth, so it co-moves positively with demand shocks and negatively with supply shocks, as required.

The estimated $$\hat{A}$$ matrix with 95% bootstrap confidence intervals is (eq. 19, p. 74):

$$
\hat{A} = \begin{bmatrix} 0.58 & -0.34 \\ (0.27, 0.77) & -(0.55, 0.02) \\ 5.16 & 3.87 \\ (2.03, 8.35) & (0.78, 7.46) \end{bmatrix}. \tag{19}
$$

## Empirical specifications

**Baseline asset pricing test (R1, R2, R3).** The two-stage Fama-MacBeth regression (Section III.B, pp. 79-81) is estimated on a balanced panel of ~125 test assets (equities, Treasury bonds, corporate bonds, S&P 500 options). First stage: estimate return betas for each asset by regressing excess returns on the identified shocks or on the raw leverage innovation. Second stage (cross-sectional, without a constant as recommended by Shanken 1996 and Kroencke-Thimme 2021): regress average returns on betas to recover the price-of-risk vector $$\hat{\lambda}$$. $$t$$-statistics use Shanken-corrected standard errors. The uncentered $$\bar{R}^2$$ confidence interval follows Lewellen, Nagel, and Shanken (2010).

For the symmetric model (column (4) of Table I), only the difference $$\beta_s - \beta_d$$ enters the second stage (imposing $$\lambda_s = -\lambda_d$$), yielding a parsimonious one-parameter price-of-risk estimate $$\hat{\lambda} = 2.02$$ (close to the structural model's $$\hat{\lambda} = 2.12$$).

**Asset-class regressions (R4).** The same specification run separately on equities (1986-2021, value-weighted), bonds (1986Q2-2021Q4, equally weighted), and options (1986Q2-2021Q4, equally weighted S&P 500 call and put portfolios from Constantinides, Jackwerth, and Savov 2013). Consistent sign and magnitude of $$\hat{\lambda}$$ across classes (Table II, p. 85).

**AEM replication (R5).** Table VI (p. 95) uses AEM's original test assets: 25 size- and book-to-market-sorted FF portfolios, 10 momentum portfolios, and 12 Treasury bonds. The raw leverage factor produces a positive $$\hat{\lambda}_l = 6.16$$ (t=2.64) in column (1), consistent with AEM; the symmetric demand/supply decomposition delivers $$\hat{\lambda} = 1.86$$ (t=2.20) in column (4) with a higher $$\bar{R}^2$$.

**Liquidity regressions (R6).** Panel time-series regressions (Section III.E, pp. 91-94) of quarterly changes in Amihud illiquidity ratios on the identified leverage supply and demand shocks plus contemporaneous market returns, with Newey-West $$t$$-statistics using three lags (Table V, p. 93):

$$
\Delta\text{Illiq}_{i,t} = a + b_i e_t^s + c_i x R_{m,t} + \varepsilon_t \quad \text{(Panel A)},
$$

$$
\Delta\text{Illiq}_{i,t} = a + b_i e_t^d + c_i x R_{m,t} + \varepsilon_t \quad \text{(Panel B)},
$$

$$
\Delta\text{Illiq}_{i,t} = a + b_i \Delta\text{LEV}_t + c_i x R_{m,t} + \varepsilon_t \quad \text{(Panel C)}.
$$

Estimated for 10 illiquidity-decile portfolios (Row I) and 10 volatility-decile portfolios (Row II). Supply-shock coefficients in Panel A are negative, monotone, and significant; demand-shock coefficients in Panel B are insignificant; raw leverage in Panel C is negative but never significant at 5%.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Federal Reserve Flow of Funds (Table L.129) | Quarterly broker-dealer aggregate leverage LEV (total financial assets / book equity), 1986Q2-2021Q4 | [FRED](/wiki/datasets/fred/) (public; Flow of Funds tables) |
| TED spread (EuroDollar LIBOR minus T-bill) | Component of FUND instrument; funding-conditions proxy | No page yet |
| Hu, Pan & Wang (2013) noise measure | Component of FUND instrument; bond yield noise as funding proxy | No page yet |
| Fontaine & Garcia (2012) liquidity factor | Component of FUND instrument; term-structure-based funding measure | No page yet |
| CRSP (via WRDS) | Stock returns and market data for equity test portfolios; Amihud illiquidity construction | [WRDS](/wiki/commercial/wrds/) (licensed) |
| Compustat (via WRDS) | Book-to-market, size for FF25 equity portfolio sorts | [WRDS](/wiki/commercial/wrds/) (licensed) |
| CRSP bond files | Treasury bond returns for 12-bond test-asset panel (2- to 10-year maturities); security-level data | [WRDS](/wiki/commercial/wrds/) (licensed) |
| Constantinides, Jackwerth & Savov (2013) option portfolios | S&P 500 unlevered call and put portfolios (27 each), 1986Q2-2021Q4 | No page yet |
| Corporate bond data | Corporate bond test portfolios sorted on illiquidity, volatility, and funding betas | No page yet |

Sample: January 1986 to December 2021 (monthly for FUND, quarterly for asset pricing tests).

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13407) if you are: building intermediary asset pricing models and need the structural identification procedure and closed-form parameter expressions (Section I-II and Internet Appendix Section I); revisiting the sign puzzle in AEM or He, Kelly, and Manela (2017) (Section IV); studying the leverage-liquidity nexus and want the panel illiquidity regression results (Section III.E and Table V); or extending the framework to other asset classes or to a model where the leverage constraint does not always bind.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(1), February 2025. This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The article is paywalled (Wiley VOR); only extraction is permitted here.

> Fontaine, Jean-Sebastien, Rene Garcia, and Sermin Gungor.
> "Intermediary Leverage Shocks and Funding Conditions."
> *The Journal of Finance* 80, no. 1 (February 2025): 57-99.
> DOI: [10.1111/jofi.13407](https://doi.org/10.1111/jofi.13407).
> © 2024 the American Finance Association.
> Paywalled; extract-only redistribution.


==============================================================================
# Minority Representation at Mortgage Lenders: Frame, Huang, Jiang, Lee, Liu, Mayer & Sunderam (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/frame-impact-minority-representation-mortgage-2025/
# Distilled: Using new data linking U.S. mortgage applications to individual loan officers via NMLS and confidential HMDA, the paper shows that minority borrowers face lower completion, approval, and origination rates when matched with White loan officers, but these gaps shrink substantially under minority loan officers, and that minority-officer-matched loans also default less, consistent with an informational advantage rather than favoritism. J. Finance 2025, paywalled. Eight core results with source locators, datasets used, the identification strategy, and the estimating equations.
# Tags: paper-summary, household-finance, mortgage-lending, racial-disparities
==============================================================================

**What this is.** The paper's core results, the datasets it links, and the estimating equations: enough to know what it found and how, without reading all 52 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13428).

## TL;DR

Using a new panel that links 5.65 million U.S. home purchase mortgage applications (2018-2019 HMDA) to the individual loan officers who handled them (via NMLS), Frame, Huang, Jiang, Lee, Liu, Mayer, and Sunderam (2025) establish two facts: (1) minorities are significantly underrepresented among loan officers (15% minority share versus 29-39% in comparable white-collar professions), and (2) minority borrowers are about 2 percentage points less likely to have their applications completed, 1.2-3 percentage points less likely to be approved (for high-discretion applications), and 2.5 percentage points less likely to originate a loan when handled by White loan officers. These gaps shrink substantially when the loan officer is also a minority. Critically, default rates on minority loans originated by White officers are 1.7-2.2 percentage points higher than for White borrowers, while minority-officer-matched minority loans default at the same rate as comparable White borrower loans. The pattern is consistent with minority loan officers having an informational advantage in handling minority borrower applications, rather than simple favoritism.

The paper contributes to three strands of the literature. First, it adds to the long tradition beginning with Munnell et al. (1996) on racial disparities in mortgage approval. Bhutta, Hizmo, and Ringo (2024) show the approval gap is largely explained by observed risk factors; this paper shows the residual gap is explained by the absence of minority loan officers who can supply soft information. Second, it extends the cultural-proximity credit result of Fisman, Paravisini, and Vig (2017) from Indian banks to the U.S. mortgage market, where automated underwriting and hard information dominate. Third, it complements Ambrose, Conklin, and Lopez (2021) and Bartlett et al. (2022): the minority-officer effect is strongest at small banks and weakest at FinTech lenders, consistent with FinTech reducing scope for loan-officer soft information.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Minority applicants are **1.9 pp less likely to complete** applications under White loan officers (within-officer comparison); gap is **1.1 pp smaller** under minority officers | Table III, cols. (1)-(2), p. 1224 | Minority: beta = -0.019\*\*\* (0.001); Minority x Minority Officer: beta = +0.011\*\*\* (0.002) |
| R2 | For high-discretion applications, minority applicants are **2.9 pp less likely to be approved** under White loan officers; gap is **1.2 pp smaller** under minority officers | Table III, col. (5)-(6), pp. 1224-1227 | Minority: -0.029\*\*\* (0.002); Minority x Minority Officer: +0.012\*\*\* (0.003) in col. (5) |
| R3 | In all-in origination rates, minority applications are **5 pp less likely to originate** under White loan officers; gap is **2.5 pp smaller** (about 50%) under minority officers | Table III, cols. (7)-(8), p. 1224 | Minority: -0.050\*\*\* (0.002); Minority x Minority Officer: +0.025\*\*\* (0.003) in col. (7) |
| R4 | IV estimates (day-of-week instrument): high-discretion minority approval gap is **3.6 pp** under White officers, and minority-officer interaction **fully offsets the gap** | Table IV, Panel C, col. (3), p. 1233 | Minority: -0.031\*\*\* (0.003); Minority x Minority Officer: +0.036\*\*\* (0.016) |
| R5 | FHA default analysis (OLS): minority borrowers with White officers default **1.8 pp more**; Minority x Minority Officer interaction is **-2.2 pp**, eliminating the excess default rate | Table V, col. (1), p. 1237 | Minority: +0.018\*\*\* (0.001); Minority x Minority Officer: -0.022\*\*\* (0.002) |
| R6 | FHA default IV: minority borrowers with White officers default **1.7 pp more**; IV interaction is **-5.2 pp**, fully eliminating the excess | Table V, col. (4), p. 1237 | Minority: +0.017\*\*\* (0.002); Minority x Minority Officer: -0.052\*\*\* (0.025) |
| R7 | Cross section of mechanism: effect concentrated in **same-race/ethnicity pairings** and in counties with a **high share of non-native English speakers** and **low college share** | Table VII, Panel A cols. (1),(3), pp. 1241-1243; Panel B cols. (2),(3) | Same Race triple interaction: +0.008\*\*\* (0.003); High Non-English Share triple interaction: +0.005\* (0.003) |
| R8 | Minorities are **underrepresented** among loan officers: 15% of loan officers are minorities versus 60.7% White (2019), while minorities are 39.3% of the U.S. labor force; minority share systematically below local population share | Table I, Panel A, p. 1219; Figure 1, p. 1221 | 84.6% White, 8.9% Hispanic, 1.8% Black, 4.7% Asian among N=255,277 loan officers |

**Overall (paper's conclusion).** Minority loan officers have an informational advantage in processing applications from minority borrowers: they achieve higher approval rates and lower default rates simultaneously, a pattern inconsistent with taste-based discrimination. The underrepresentation of minorities among loan officers therefore reduces minority access to credit, and this effect persists even in the hard-information-intensive U.S. mortgage market where automated underwriting systems dominate.

## Theory / model

The paper has no formal model. The paper's identification strategy rests on two testable hypotheses:

**Hypothesis 1 (supply-side discrimination / information advantage).** If minority loan officers have better soft information about minority borrowers, they can help those borrowers complete stronger applications, achieve higher approval rates, and generate loans that perform better (lower defaults). This is the authors' preferred interpretation.

**Hypothesis 2 (taste-based discrimination).** If White loan officers discriminate against minority borrowers based on taste, they would apply stricter standards: minority borrowers handled by White officers would have lower approval rates (as observed) but also lower default rates (the paper's Table V rules this out; default rates are in fact higher for minority loans handled by White officers).

The paper tests the hypotheses by examining approval and default rates jointly. Taste-based discrimination predicts lower approvals AND lower defaults for White-officer-handled minority loans. Information advantage predicts lower approvals AND higher defaults (because excluded minority borrowers are creditworthy, so the approved pool selected by stricter White officers is adversely selected relative to the approved pool selected by informed minority officers). The data match the information-advantage prediction.

**Identification.** Endogenous matching of loan officers to borrowers is a central concern. Two approaches address it:

1. Tight fixed effects: branch-year and branch-year-officer fixed effects isolate within-officer, within-branch variation in how the same officer treats minority versus White applicants.
2. Day-of-the-week instrument (Section II.C, pp. 1228-1234): exogenous variation in whether a minority officer handles a specific application is generated by minority officer work schedules.

## Method

**BIFSG race imputation** (pp. 1215-1216). Loan officer race/ethnicity is not observed in NMLS. The paper applies the Bayesian Improved First Name Surname Geocoding (BIFSG) method of Voicu (2018) to infer each officer's race/ethnicity. For surname $$s$$, first name $$f$$, and ZIP code $$z$$, the posterior probability of belonging to race group $$r$$ is (equation 1, p. 1215):

$$
p(r \mid s, f, z) = \frac{p(r \mid s) \times p(f \mid r) \times p(z \mid r)}{\sum_{r=1}^{6} p(r \mid s) \times p(f \mid r) \times p(z \mid r)} \tag{1}
$$

where $$p(r \mid s)$$ is the probability of belonging to race group $$r$$ given surname (from the 2010 Census surname list), $$p(f \mid r)$$ is the probability of having first name $$f$$ given race $$r$$ (from the Tzioumis 2018 list), and $$p(z \mid r)$$ is the probability of being in ZIP code $$z$$ given race $$r$$ (from the 2010 Census). Officers are assigned to the race group with the highest posterior probability.

**Day-of-the-week instrument** (pp. 1228-1229). For application $$i$$ opened at branch office $$b$$ on day of the week $$d$$ in week $$w$$, the instrument is the share of applications at the same branch on the same day of the week during the prior 12 weeks ($$w - 12$$ to $$w - 1$$) handled by minority officers:

$$
Z_{i,b,d,w} = \frac{\text{\#Minority Officer Applications}_{b,d,w-12 \to w-1}}{\text{\#Applications}_{b,d,w-12 \to w-1}}
$$

The first stage regresses an indicator for the application being handled by a minority officer on the instrument, branch-week fixed effects, day-of-the-week fixed effects, and controls (p. 1228):

$$
\mathbf{1}\{\text{Minority Officer}\}_{i,b,d,w} = \alpha_{b,w} + \beta Z_{i,b,d,w} + \gamma' \mathbf{X}_{i,b,d,w} + \varepsilon_{i,b,d,w} \tag{3}
$$

First-stage F-statistics exceed 15 across all samples (Table IV, Panel A). Covariate balance tests (Table IV, Panel B) show the instrument is uncorrelated with borrower age, income, loan amount, FICO, LTV, DTI, and AUS recommendation code.

## Empirical specifications

The main estimating equation (equation 2, p. 1223) is a linear probability model:

$$
Y_i = \beta_1 \mathbf{1}\{\text{Minority}\}_i + \beta_2 \mathbf{1}\{\text{Minority Officer}\}_i + \beta_3 \mathbf{1}\{\text{Minority}\}_i \times \mathbf{1}\{\text{Minority Officer}\}_i + \gamma' X_i + \varepsilon_i \tag{2}
$$

where $$Y_i$$ is in turn: application completion, approval (conditional on completion), all-in origination, or default (90+ days delinquent). The parameter of interest is $$\beta_3$$, the differential effect of having a minority loan officer on outcomes for minority versus White applicants. Standard errors are two-way clustered by lender and county.

**Application-level specifications (Table III, pp. 1223-1228):**

- Columns (1) and (3), (5), (7): branch-year fixed effects and property-county fixed effects, plus Basic App Controls (loan type indicators, 10-year age bins, income-to-MSA-median centile bins, log(loan amount), jumbo indicator, joint application indicator). This exploits cross-officer variation within the same branch-year.
- Columns (2) and (4), (6), (8): replace branch-year FE with branch-year-officer FE, so identification is within-officer (comparing how the same officer treats minority versus White applicants). The $$\beta_2$$ coefficient on Minority Officer is absorbed; only $$\beta_3$$ is identified.
- For approval regressions (cols. 3-6), the sample is restricted to completed applications and the Extended App Controls add FICO-bin, LTV-bin, and DTI-bin indicators (all interacted with loan type) plus AUS output code fixed effects. The sample is further split into "low-discretion" (AUS approval rate > 90%) and "high-discretion" (AUS approval rate <= 90%) applications.

**FHA default specification (Table V, p. 1237):** The dependent variable is an indicator for the FHA loan ever becoming 90+ days delinquent. Controls (FHA Controls) include log(loan amount), income-to-MSA centile bins, FICO-bin, LTV-bin, DTI-bin (all interacted with loan type), interest rate, first-time buyer indicator, branch-year FE, property-county FE, and origination-month FE. The IV variant uses branch-month FE and day-of-the-week FE with the same instrument $$Z_{i,b,d,w}$$ recalculated on FHA loans at the branch over prior 12 months.

**Mechanism tests (Tables VII-IX, pp. 1240-1251):** Same base specification augmented with triple interactions for: same-race/ethnicity pairing, low-income borrower, small bank, FinTech lender, rural county, high non-English share, low college share (Table VII); loan officer minority application share and experience (Table VIII); and linear hard information variables (credit score, DTI) to test differential reactions to hard information (Table IX).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| NMLS Consumer Access (2012-2019) | Nationwide loan officer panel: name, employer, work address, career history; source for BIFSG race imputation | [NMLS](/wiki/datasets/nmls/) |
| Confidential HMDA (2018-2019, Federal Reserve) | Mortgage applications matched to loan officers via NMLS ID; includes FICO, LTV, DTI, AUS code (added to confidential version from 2018) | [no page yet] |
| FHA insured mortgage data (2000-2018, FHA/Federal Reserve) | Population of FHA single-family originations; used for default analysis (90+ days delinquent through 2019-Q3) | [FHA](/wiki/confidential/fha/) |
| Black Knight McDash (matched to HMDA 2018-2019) | Monthly performance data; 60-day default within 24 months of origination; 36% match of HMDA approved mortgages | [no page yet] |
| U.S. Census Bureau (ZIP code level) | Demographic and economic characteristics (minority population share, PIPC, population density, non-English share, college share) | [no page yet] |

Sample: HMDA analysis covers 5.65 million first-lien 30-year fixed-rate home purchase mortgage applications (owner-occupied single-family properties) in 2018-2019, after filters. FHA default sample: ~3.37 million loans originated 2012-2018, tracked through 2019-Q3.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13428) if you are: studying the supply-side determinants of racial disparities in mortgage lending; evaluating the role of soft information in a setting dominated by hard information and automated underwriting; assessing whether minority officer representation has welfare-improving effects via credit expansion with no increase in default; or working on fair lending policy or the economics of racial diversity in financial services. The Internet Appendix (18+ tables) contains the BIFSG validation, subsample analyses by race/ethnicity pairing, shopping behavior robustness, and additional mechanism tests.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(2), April 2025. Copyright 2025 the American Finance Association; portions contributed by U.S. Government employees are in the public domain in the USA. This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The underlying confidential HMDA and FHA microdata are proprietary and were not accessed here. Extract-only; no PDF hosted.

> Frame, W. Scott, Ruidi Huang, Erica Xuewei Jiang, Yeonjoon Lee, Will Shuo Liu, Erik J. Mayer, and Adi Sunderam.
> "The Impact of Minority Representation at Mortgage Lenders."
> *The Journal of Finance* 80, no. 2 (April 2025): 1209-1260.
> DOI: 10.1111/jofi.13428.


==============================================================================
# Baby Booms and Asset Booms: Francke & Korevaar (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/francke-baby-booms-asset-booms-2025/
# Distilled: Using centuries of data from Amsterdam and Paris, this paper shows that lagged birth rates are a major predictable driver of house prices, with high birth rates 25 to 29 years ago raising rent-price ratios and high birth rates 60 to 64 years ago lowering them; the effect concentrates in house prices rather than rents, consistent with age-dependent entry into and exit from homeownership. J. Finance 2025, CC BY 4.0. Six core results with source locators, datasets used, the estimating equation, and the mechanism analysis.
# Tags: paper-summary, real-estate, demographics, housing, macro, panel-regression
==============================================================================

**What this is.** The paper's core results, the empirical model, and the mechanism
analysis: enough to know what it found and how, without reading all 36 pages.
To replicate or extend it, read the full source at the
[original](https://doi.org/10.1111/jofi.13480).

## TL;DR

Using centuries of housing transaction data from Amsterdam (1550-1884) and rent
data from Paris (1500-1831), Francke and Korevaar show that lagged birth rates
are a major and predictable driver of house prices. A high birth rate 25 to 29
years ago, when a large cohort enters prime home-buying years, raises house
prices relative to rents by about 4% per percentage-point increase in the
five-year birth rate. A high birth rate 60 to 64 years ago, when a large cohort
exits homeownership at death or through a move to senior housing, lowers
rent-price ratios by a similar magnitude. These effects concentrate
in house prices and not in rent prices, pointing to age-concentrated entry into
and exit from homeownership rather than general housing consumption demand. The
two lags together explain about 18% of total variation in rent-price ratios over
250 years. Mechanism analysis finds that sale probabilities respond sluggishly to
demographic demand shocks, and that spatial segmentation between rental and
owner-occupied markets contributes to the large price effects.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%.
Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Lagged birth rates 25-29 and 60-64 years ago predict rent-price ratios**, with the two lags explaining 18.4% of total variation over 250 years | Table II col. 3, p. 3040 | Birth rate 25-29 yrs ago: coeff. -4.251\*\*\* (SE 1.130); birth rate 60-64 yrs ago: coeff. 4.001\*\*\* (SE 0.837); adj. R2 = 0.448 |
| R2 | **Effect concentrates in house prices**, with birth rates 25-29 yrs ago raising house prices and birth rates 60-64 yrs ago lowering them | Table III cols. 1-3, p. 3041 | Coeff. on B_{t-25/29} = 4.457\*\*\* (SE 1.314); coeff. on B_{t-60/64} = -4.033\*\*\* (SE 1.377); adj. R2 = 0.587 |
| R3 | **Rent prices show little to no response** to lagged birth rates in Amsterdam or Paris; rents are not the channel | Table III cols. 4-9, p. 3041 | B_{t-25/29} on Amsterdam rents: -0.562 (SE 0.651), insignificant; B_{t-60/64}: -0.862\* (SE 0.509); Paris rents: similarly weak |
| R4 | **Transaction probability responds to demographic shocks but with a delay**, peaking years after price effects, consistent with slow supply adjustment | Figure 5, p. 3047 | Birth-rate lag ~37 yrs raises sale probability by 0.1 pp (+4%); price effects peak at lag ~25, sale-probability effects at lag ~37 |
| R5 | **Spatial segmentation supports the mechanism**: young cohorts raise prices in low-homeownership (rental) areas; homeownership-age cohorts raise prices in high-homeownership areas | Figure 6, p. 3049 | Young-cohort (teen) lag: ~2% higher house price growth in low- vs high-homeownership areas; early-30s lag: ~1% excess growth in high-homeownership areas |
| R6 | **No significant effect on bond or young-cohort dividend yields**, confirming the effect is housing-specific and not general asset demand | Table V, p. 3053 | B_{t-25/29} on bond yield: 1.555 (SE 1.318), insignificant; on dividend yield: -1.192 (SE 0.792), insignificant; only old-cohort lag shows weak positive dividend yield effect |

**Overall (paper's conclusion).** Demographics have been a major, predictable
driver of house prices relative to rents over multiple centuries. The effect
arises because entry into and exit from homeownership are strongly
age-concentrated, and because other market participants respond slowly to shifts
in ownership demand rather than immediately converting rental units to
owner-occupied ones and vice versa. The effect is specific to housing: there is
no evidence of similar effects on bond yields and only weak evidence for
dividend yields, ruling out a general life-cycle asset-demand explanation.

## Theory / model

The paper has no formal dynamic equilibrium model. Instead, it builds on the
insight that total housing demand in year $$t$$ can be written as
$$D_t = \sum_j \hat{\alpha}_j N_{jt}$$, where $$N_{jt}$$ is the number of individuals
aged $$j$$ and $$\hat{\alpha}_j$$ is a (constant) age-specific demand weight estimated
from cross-sectional data on residents' housing choices. This is the
Mankiw and Weil (1989) framework (p. 3036). The aggregate house price or index
$$h_t$$ then depends on $$D_t$$ via a simple time-series regression:

$$
h_t = \delta_0 + D_t \delta_1 + \varepsilon_t \tag{background}
$$

Since the authors do not observe age-specific population counts $$N_{jt}$$ in the
historical setting, they substitute a linear approximation: the change in demand
$$\Delta D_t$$ is a linear function of lagged birth rates $$B_{t-\text{lag}}$$, so
that changes in house prices (in logs) become:

$$
\Delta h_t \approx \alpha + \sum_{\text{lag}} B_{t-\text{lag}} \beta_{\text{lag}} + \varepsilon_t
$$

**Identification logic.** The paper uses lagged birth rates as the key predictor
rather than current demographic structure, for three reasons stated on p. 3023:
(i) lagged birth rates predate the outcomes by decades and so are less likely to
be jointly driven by current economic conditions; (ii) migration, which is
endogenous to economic opportunity, induces endogeneity in current age structure
but not in birth rates from generations ago; and (iii) using lags allows a test
for predictability, since birth cohorts and their eventual housing demand are
known far in advance. The paper focuses on rent-price ratios (rather than price
levels) to further reduce sensitivity to shared confounders: if past birth rates
correlate with current economic activity, that would affect house prices and
rents similarly, but the ratio difference would remain.

The identification strategy builds on the idea from DellaVigna and Pollet (2007)
that predictable demographic demand shifts generate price effects because
investors do not fully anticipate them. It also updates the findings of
Poterba (2001), who found limited evidence that demographic shifts affect
aggregate asset prices; the present paper shows the housing market is the
exception, not a general effect across all assets.

No causal identification device (instrument, natural experiment, RDD) is
deployed; the strategy is predictive/descriptive using deep historical variation
to exploit many decades of pre-demographic-transition fluctuations in birth
rates.

## Method

The main estimating equation is a time-series predictive regression of five-year
log changes in rent-price ratios, house prices, or rents on lagged five-year
birth rates (equation 1, p. 3037):

$$
y_t - y_{t-5} = \alpha + \sum_{\text{lag} \in L} B_{t - \text{lag}/4 + \text{lag}} \beta_{\text{lag}} + z_t' \gamma + \varepsilon_t \tag{1}
$$

where $$y$$ is the log rent-price ratio $$(r - h)$$, the log house price index
$$h$$, or the log rent index $$r$$; $$B_{t - \text{lag}/4 + \text{lag}}$$ is the total
five-year birth rate summed over lags $$t-\text{lag}$$ through $$t-\text{lag}+4$$;
$$z_t$$ is a vector of contemporaneous control variables (birth rate, mortality,
nuptiality, migration, wage growth, inflation, GDP per capita log change); and
$$L$$ is a set of lag lengths between 15 and 70 years (in five-year steps,
corresponding to cohorts aged 15 to 74 at time $$t$$).

Serial correlation from overlapping five-year differences is addressed by
Newey-West standard errors with a lag length of five (p. 3037). The paper
estimates equation (1) for various sets $$L$$: a single lag, two lags (young and
old cohort), and all lags jointly. The main reported results use two lags
corresponding to birth rates 25-29 years ago (peak homeownership-entry cohort)
and 60-64 years ago (peak exit cohort).

The probability-of-sale mechanism is tested using a discrete-time hazard model
for the probability that property $$i$$ sells at holding duration $$d = t_i - s_i$$
(equations 2a-2b, p. 3046):

$$
\Pr(D_{is} = t_i - s_i) = H_{is}(t_i - s_i) S_i(t_i - s_i - 1) \tag{2a}
$$

$$
\Pr(D_{it} > T - s_i) = S_i(T - s_i) \tag{2b}
$$

where $$H_{is}(d) = (1 + \exp(-x_{isd}' \beta))^{-1}$$ is a logistic hazard function
and $$S_{is}(d) = \prod_{j=1}^d (1 - H_{is}(j))$$ is the survival probability.
This is estimated by logistic regression in property-period format.

## Empirical specifications

**Main specification (Table II, p. 3040; Table III, p. 3041).** The baseline
uses five-year log changes in the Amsterdam rent-price ratio as the dependent
variable, with two birth-rate lags (25-29 yrs and 60-64 yrs). All variables are
in five-year differences to address unit-root concerns (unit-root tests reject
the null in levels; p. 3038). Six columns vary the control set:

- Column (1): no controls (adj. R2 = 0.178)
- Column (2): demographic controls (birth rate, mortality, nuptiality, migration)
- Column (3): demographic + economic controls (wage growth, inflation, GDP per capita); baseline
- Column (4): baseline + lagged controls matching the birth-rate lag windows
- Column (5): baseline + housing-quality growth control
- Column (6): baseline + interest rate

The coefficients on the two main birth-rate lags are stable across all six
specifications: approximately -4 (young cohort) and +4 (old cohort), both
significant at 1%.

**House prices vs. rents (Table III).** The same specification is run separately
for Amsterdam house prices, Amsterdam rents, and Paris rents. House prices show
large significant effects (coeff. ~4.5 and -4.0). Rents show near-zero and
generally insignificant effects, both in Amsterdam and Paris, confirming that
the rent-price ratio result is driven by house prices.

**Robustness: supply constraints (Table IV, p. 3051).** The baseline is
interacted with a supply-constrained dummy (SC = 1 before 1668 and after 1855,
the two Amsterdam expansion episodes). Interaction effects are insignificant,
and the main birth-rate coefficients are unchanged, ruling out that results are
driven by periods of constrained supply.

**Other assets (Table V, p. 3053).** The same two-lag specification is run for
five-year changes in Dutch government bond yields and Dutch East India Company
dividend yields. The young-cohort lag is insignificant for both. The old-cohort
lag shows a weakly significant positive effect only on dividend yields
(coeff. 1.525\*\*, SE 0.751), consistent with estate sales rather than general
life-cycle asset demand.

**Segmentation (Figure 6, p. 3049).** Separate house price indices for streets
with low, medium, and high homeownership rates (from the 1805 Amsterdam rental
census) are estimated via repeat-sales and regressed on the same birth-rate
lags. The spatial pattern of coefficients by lag length matches the predicted
housing life-cycle: young cohorts raise prices in low-HO (rental) areas,
mature cohorts raise prices in high-HO (owner-occupied) areas.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Amsterdam housing transaction data (mandatory aldermen registrations, 1620-1811; 19th-century repeat-sales on Herengracht) | Main house price index via Bayesian repeat-sales (Francke 2010); segmentation analysis | [Amsterdam housing](/wiki/datasets/amsterdam-housing-transactions/) |
| Paris repeat-rent indices (Eichholtz, Korevaar, and Lindenthal 2020) | Rent price series for Paris 1500-1831 (new rental contracts only) | [Paris rents](/wiki/datasets/paris-rents/) |
| Amsterdam repeat-rent index (Eichholtz, Korevaar, and Lindenthal 2020) | Rent price series for Amsterdam 1550-1884 | [Amsterdam housing](/wiki/datasets/amsterdam-housing-transactions/) |
| Amsterdam archival civil registers (Amsterdam City Archives, from 1554) | Annual births, deaths, marriages for demographic rates | No page yet |
| Paris demographic data (historians + official Paris statistics) | Annual births, deaths, marriages for demographic rates 1500-1831 | No page yet |
| Dutch government bond yields (provincial debt pre-1810, national debt post-1810) | Test of general asset-demand mechanism | No page yet |
| Dutch East India Company (VOC) dividend yields (Golez and Koudijs 2018) | Test of general asset-demand mechanism, 1629-1782 | No page yet |
| 1805 Amsterdam rental census (Amsterdam City Archives) | Street-level homeownership rates for segmentation analysis | [Amsterdam housing](/wiki/datasets/amsterdam-housing-transactions/) |
| OECD panel on house prices, rents, demographics (1970-2020) | Modern-context robustness check (Internet Appendix Section VI) | No page yet |

Sample: Amsterdam rent-price ratio 1550-1884 (N=256 five-year observations);
Paris rents 1500-1831 (N=242-331 observations); probability-of-sale sample
covers all Amsterdam repeat-sales pairs 1620-1811. Annual frequency, aggregated
to five-year differences for main regressions.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13480) if you are:
investigating demographic drivers of house prices or rent-price ratios, building
a model of the housing life-cycle and its aggregate price implications, or
studying the historical Amsterdam or Paris housing and rental markets. The
Internet Appendix (available online) contains the full list of historical data
sources (Table IA.V), the age-distribution evidence for the United States
(Section II.A), the OECD modern-context replication (Section VI), and further
robustness checks including all-lag regressions (Table IA.VI) and the
probability-of-sale two-lag results (Table IA.VII).

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(5). This distillation was
extracted by an LLM on 2026-06-05 and is **not human-verified or independently
reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not
hosted in this batch.

> **Attribution (CC BY 4.0).** Francke, Marc, and Matthijs Korevaar.
> "Baby Booms and Asset Booms: Demographic Change and the Housing Market."
> *The Journal of Finance* 80, no. 5 (October 2025): 3021-3056.
> DOI: 10.1111/jofi.13480. (C) 2025 The Author(s).
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Test Assets and Weak Factors: Giglio, Xiu & Zhang (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/giglio-test-assets-weak-factors-2025/
# Distilled: Giglio, Xiu, and Zhang show that weak factors and test asset selection are deeply connected, and introduce Supervised Principal Component Analysis (SPCA), an iterative procedure that screens test assets by correlation with the target factor before applying PCA, enabling consistent risk premium estimation even when some latent factors are weak. J. Finance 2025, paywalled. Eight core results with source locators, datasets used, the model (linear factor model with weak factors), and the method (SPCA algorithm) with its defining equations.
# Tags: paper-summary, asset-pricing, factors, factor-models, weak-factors
==============================================================================

**What this is.** The paper's core results, the model it builds on (the linear factor model with weak factors), and the method it contributes (SPCA) with the defining equations: enough to know what it found and how, without reading all 61 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13415).

## TL;DR

Giglio, Xiu, and Zhang show that weak factors and test asset selection are two faces of the same problem: a factor is weak precisely in the cross section of test assets chosen by the researcher. They propose Supervised Principal Component Analysis (SPCA), a procedure that screens test assets by their correlation with the factor of interest $$g_t$$ at each iteration before extracting a principal component. This ensures that only assets with nontrivial exposure to the relevant factor are used, effectively strengthening the factor within the selected subset and enabling consistent risk premium estimation even when some latent SDF factors are weak or when some priced factors are omitted. Applied to a large cross section of 901 to 1,672 characteristic-sorted equity portfolios (1976 to 2020), SPCA estimates risk premia close to model-free averages for tradable factors, achieves substantially higher out-of-sample hedging $$R^2$$ than PCA for factors with weak exposures, and diagnoses that standard observable factor models (CAPM, FF3, FF5) miss important latent pricing factors.

## Core results

Magnitudes and significance are as reported. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **SPCA gives risk premia consistent with model-free averages for the market factor**, a strong factor, across all tuning parameters | Table IV, p. 300; Figure 5, Panel A, p. 302 | Market RP estimates 68-74 bps/month for p = 3, 5, 7, 11; average excess return 74/62 bps (train/eval); never statistically different at 5% |
| R2 | **SPCA risk premia are consistent with model-free averages for all tradable factors** (with minor exceptions); the intermediary capital factor has a significant positive RP even among nontradables | Table IV, pp. 300-301; Figure 5, pp. 302; Figure 6, p. 303 | Momentum RP ~112 bps (p=3) varying across p=3-11; HML RP 37-50 bps; liquidity factor RP 70-95 bps out-of-sample (p.305) |
| R3 | **Almost all nontradable macro factors (IP growth, uncertainty, consumption, term spread, credit, oil) have risk premia indistinguishable from zero**; equity markets cannot span their variation | Table IV, pp. 300-301; Figure 6, p. 303 | Positive out-of-sample $$R^2$$ is rare for macro nontradables; LN1/LN2 macro factors: $$R^2$$ near zero or negative; Liquidity: $$R^2$$ 0-4% |
| R4 | **SPCA out-of-sample hedging $$R^2$$ is higher or equal to PCA's, often with fewer factors** | Figure 5, pp. 302; Figure 6, p. 303 | Market: $$R^2 > 0.98$$ for all p, q; Momentum: SPCA reaches $$R^2 > 70\%$$ with p=3; PCA needs p >= 6; Intermediary Cap: SPCA $$R^2 \approx 50\%$$, PCA much lower |
| R5 | **SPCA degrades little when informative test assets are removed**; PCA degrades sharply | Figure 8, p. 311 | Momentum: SPCA $$R^2$$ from 86% to 77% without momentum assets; PCA from 76% to 48%. Profitability: SPCA 71% to 60%; PCA 41% to 14% |
| R6 | **In simulations, SPCA has far smaller bias than PCA, rpPCA, Ridge, Lasso, and two-pass estimators** for weak factors | Table I, p. 293 | Weak factor V bias: SPCA -3.4 bps vs PCA -35 bps at T=240; RMSE 14.6 vs 36.5 bps; PLS ranks second among alternatives |
| R7 | **SPCA achieves out-of-sample Sharpe ratios closest to the theoretical optimum** (0.256) among all estimators in SDF recovery simulations | Table III, p. 295 | SPCA SR: 0.193 (T=120), 0.226 (T=240), 0.241 (T=480); PCA: 0.084, 0.110, 0.227; rpPCA: 0.134, 0.192, 0.242 |
| R8 | **SPCA diagnoses that CAPM, FF3, and FF5 all miss important pricing factors**: SPCA Sharpe ratio grows well above each model's own Sharpe ratio as additional latent factors are extracted | Figure 9, p. 315 | Market (CAPM) Sharpe = 0.46; SPCA SR grows from ~0.5 to ~1.4+ with 1-20 factors (CZ data); FF5+MOM+BAB+QMJ best but still misspecified in CZ data |

**Overall (paper's conclusion).** SPCA resolves the weak factor problem in empirical asset pricing by treating factor weakness as a property of the test assets rather than the factor itself. It consistently estimates risk premia and recovers the SDF in the presence of weak, omitted, and mismeasured factors, and provides a diagnostic tool for observable factor models. Empirically, nearly all nontradable macro factors have risk premia statistically indistinguishable from zero in equity markets, while standard observable factor models miss important latent pricing factors.

## Theory / model

The paper studies a standard linear latent factor model. Suppose an $$N \times 1$$ vector of test asset excess returns $$r_t$$ follows (equation 1, p. 266):

$$
r_t = \beta\gamma + \beta v_t + u_t, \qquad \mathbb{E}(v_t) = \mathbb{E}(u_t) = 0 \text{ and } \text{cov}(v_t, u_t) = 0, \tag{1}
$$

where $$\beta$$ is an $$N \times p$$ matrix of factor exposures, $$v_t$$ is a $$p \times 1$$ vector of factor innovations ($$v_t = f_t - \mu_f$$), and $$u_t$$ is an $$N \times 1$$ vector of idiosyncratic errors. Factors $$f_t$$ may be latent or observable. The SDF in terms of latent factor innovations is (equation 2, p. 266):

$$
m_t = 1 - \gamma^\top \Sigma_v^{-1} v_t, \tag{2}
$$

and the tradable SDF representation in terms of test asset excess returns is (equation 3, p. 266):

$$
\tilde{m}_t = 1 - b^\top (r_t - \mathbb{E}(r_t)), \tag{3}
$$

where $$b$$ satisfies $$\mathbb{E}(r_t) = \Sigma b$$ and $$\Sigma = \text{cov}(r_t)$$. The observable factor proxy vector $$g_t$$ is linked to the latent factors by (equation 4, p. 267):

$$
g_t = \xi + \eta v_t + z_t, \tag{4}
$$

where $$\xi = \mathbb{E}(g_t)$$, $$\eta$$ is a $$d \times p$$ matrix of factor loadings of $$g_t$$ on $$v_t$$, and $$z_t$$ is measurement error orthogonal to $$v_t$$. The risk premium of $$g_t$$ is $$\gamma_g = -\text{cov}(m_t, g_t) = \eta\gamma$$.

**Definition of weak factors.** A factor is strong relative to a cross section of $$N$$ test assets when the eigenvalues $$\lambda_i(\beta^\top\beta)$$ grow at rate $$N$$ for all $$i = 1, \ldots, p$$ (the standard "pervasive" assumption of Bai and Ng (2002)). A factor is weak when some $$\lambda_i(\beta^\top\beta)$$ grow at a slower rate than $$N$$. The necessary condition for consistency of PCA-based risk premium estimation in the multi-factor case is (equation 6, p. 275):

$$
N / (\lambda_{\min}(\beta^\top\beta) T) \to 0. \tag{6}
$$

When this fails for even one factor, the PCA estimator of Giglio and Xiu (2021) is inconsistent for risk premia (Proposition 1, p. 270, establishes this in the single-factor case). The same breakdown occurs in the observable-factor context studied by Kan and Zhang (1999), who showed that Fama-MacBeth regressions with useless factors produce invalid inference. Rank deficiency in the beta matrix produces the same failure even when each factor is individually strong (equation 7 example, p. 275).

## Method

SPCA is an iterative selection-and-projection procedure. It adds a supervised screening step (Step S1) to the PCA-based risk premium estimator of Giglio and Xiu (2021). For the single-factor case (Algorithm 2, p. 273):

- **S1 (Selection).** Select a subset $$\hat{I} \subset \langle N \rangle$$ of test assets by correlation with $$g_t$$:
$$
\hat{I} = \left\{ i \mid T^{-1} \| \bar{R}_{[i]} \bar{G}^\top \| \geq c_q \right\},
$$
where $$c_q$$ is the $$(1-q)$$-quantile of the absolute covariances $$\{ T^{-1} | \bar{R}_{[i]} \bar{G}^\top | \}_{i \in \langle N \rangle}$$. Only the top $$qN$$ assets are kept.
- **S2.** Run Steps S1 to S3 of the PCA-based Algorithm 1 on the selected return matrix $$\bar{R}_{[\hat{I}]}$$, $$\bar{G}$$, and $$p = 1$$.

The output is $$\hat{\gamma}_g^{\text{SPCA}} := \hat{\eta}\hat{\gamma}$$.

For the general multifactor case, SPCA iterates selection and projection (Algorithm 3, p. 277). At each step $$k$$:

**(S1.a)** Select $$\hat{I}_k$$ using covariance with residuals of $$G_{(k)}$$:
$$
\hat{I}_k = \left\{ i \mid T^{-1} \| (\bar{R}_{(k)})_{[i]} \bar{G}_{(k)}^\top \|_{\text{MAX}} \geq c_q^{(k)} \right\}. \tag{9}
$$

**(S1.b)** Apply Algorithm 1 to $$(\bar{R}_{(k)})_{[\hat{I}_k]}$$ and $$\bar{G}_{(k)}$$ to extract the $$k$$-th latent factor $$\hat{V}_{(k)}$$.

**(S1.c)** Project $$\bar{R}_{(k)}$$ onto $$\hat{V}_{(k)}$$ to obtain $$\hat{\beta}_{(k)}$$.

**(S1.d)** Update residuals: $$\bar{R}_{(k+1)} = \bar{R}_{(k)} - \hat{\beta}_{(k)}\hat{V}_{(k)}^\top$$, $$\bar{G}_{(k+1)} = \bar{G}_{(k)} - \hat{\eta}_{(k)}\hat{\gamma}_{(k)}$$.

Stop at $$k = \hat{p}$$ when $$c_q^{(k)} < c$$ for threshold $$c$$ (equation 10). The final risk premium estimate is $$\hat{\gamma}_g^{\text{SPCA}} = \sum_{k=1}^{\hat{p}} \hat{\eta}_{(k)}\hat{\gamma}_{(k)}$$.

**Consistency (Theorem 1, p. 279).** Under mild moment conditions, if $$\log(NT)(N_0^{-1} + T^{-1}) \to 0$$ and tuning parameters satisfy:
$$
c \to 0, \quad c^{-1}(\log NT)^{1/2}(q^{-1/2}N^{-1/2} + T^{-1/2}) \to 0, \quad qN/N_0 \to 0, \tag{11}
$$
then $$\hat{\gamma}_g^{\text{SPCA}} \xrightarrow{P} \eta\gamma$$. Theorem 2 (p. 280) additionally gives a CLT: $$\sqrt{T}(\hat{\gamma}_g^{\text{SPCA}} - \eta\gamma) \xrightarrow{d} \mathcal{N}(0, \Phi)$$ under stronger conditions including $$\lambda_{\min}(\eta^\top\eta) \gtrsim 1$$, i.e., $$g_t$$ contains at least as many variables as true factors.

**Comparison with alternatives.** The Ridge estimator (Kozak, Nagel, and Santosh (2020)) converges at rate $$(N+T)/(\lambda_p T)$$, which fails when condition (6) fails (Theorem 4a, p. 286). The Lasso estimator is consistent but at a slower rate $$\|b\|_1 \sqrt{\log N / T}$$ (Theorem 4b). Lettau and Pelger (2020) propose risk-premium PCA (rpPCA) for weak factors, but the paper shows rpPCA is inconsistent for risk premia in the multifactor weak-factor setting studied here (Internet Appendix Section I). SPCA combines factor structure (via PCA) with supervision (via screening), achieving the $$T^{-1/2}$$ rate without requiring the strong sparsity assumption.

Tuning parameters $$p$$ (number of factors) and $$\lfloor qN \rfloor$$ (number of selected assets) are chosen by three-fold cross-validation, maximizing the time-series $$R^2$$ of the hedging portfolio for $$g_t$$ in the training sample.

## Empirical specifications

The empirical analysis uses monthly data from March 1976 to December 2020, split into training (first half) and evaluation (second half) subsamples. Two test asset universes are used:

- **Main (Chen and Zimmermann (2022)):** 901 characteristic-sorted portfolios (as many as each anomaly's original paper used, 2 to 10 sorts) plus 49 industry portfolios from Ken French. The CZ data use the April 2021 release.
- **Robustness (Hou, Xue, and Zhang (2020)):** 1,672 portfolios sorted by characteristics (momentum, value, investment, profitability, intangibles, frictions).

Factors studied include eight tradable factors (market excess return, HML, SMB, RMW, CMA, momentum, BAB, QMJ) and twelve nontradable factors (liquidity, intermediary capital, IP growth, three LN macro principal components, three Jurado-Ludvigson-Ng uncertainty indexes, term spread, credit spread, unemployment, two sentiment indexes, oil, consumption growth). All factor data are at the monthly frequency.

**Risk premium estimation (Table IV, R1-R3).** SPCA is applied factor-by-factor ($$d=1$$, each factor as its own $$g_t$$) and jointly ($$d=p$$, all factors simultaneously) with $$p \in \{3, 5, 7, 11\}$$. For each $$(p, q)$$ pair, SPCA: (i) selects $$\lfloor qN \rfloor$$ assets with the largest absolute covariance with $$g_t$$; (ii) runs PCA on those assets to extract $$p$$ factors; (iii) uses Fama-MacBeth-type time-series regressions to estimate risk premia. The out-of-sample $$R^2$$ of the hedging portfolio for $$g_t$$ is reported in the evaluation half.

**SDF diagnosis (Figure 9, R8).** For each observable factor model $$g_t$$ (CAPM; the Fama and French (1993) three-factor model FF3; FF5; FF5+Momentum; FF5+Momentum+BAB+QMJ), SPCA extracts up to 20 latent factors using $$g_t$$ as supervisor. The out-of-sample Sharpe ratio of the SPCA-based SDF (right-hand side of equation 20, p. 288) is compared to the Sharpe ratio of $$g_t$$. A SPCA Sharpe ratio exceeding the model's own Sharpe indicates missing factors.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Chen and Zimmermann (2022) open-source asset pricing library | Main test asset cross section: 901 characteristic-sorted equity portfolios, April 2021 release, 1976m3-2020m12 | [Open Source Asset Pricing](/wiki/datasets/open-source-asset-pricing/) (no page yet) |
| Hou, Xue, and Zhang (2020) replicating anomalies library | Robustness test asset cross section: 1,672 characteristic-sorted portfolios | No page yet |
| Ken French Data Library | 49 industry portfolios (added to CZ cross section); Fama-French factor returns (FF3, FF5) for benchmarking | [Ken French library](/wiki/datasets/ken-french/) |
| Tradable factor returns (market, SMB, HML, RMW, CMA, BAB, QMJ, momentum) | Observable factor proxies $$g_t$$ for risk premium estimation | No page yet |
| Nontradable factor series (liquidity, intermediary capital, IP growth, LN macro PCs, uncertainty, sentiment, term, credit, unemployment, oil, consumption) | Nontradable observable factor proxies $$g_t$$; sources include Pastor-Stambaugh, AQR, FRED, Ludvigson-Ng, Jurado-Ludvigson-Ng, Baker-Wurgler, national accounts | No page yet |

Sample: monthly, 1976m3-2020m12 (about 537 months). Training: 1976m3-1998m7; evaluation: 1998m8-2020m12 (approximately equal halves).

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13415) if you are:
estimating risk premia for a potentially weak factor and need a consistent estimator robust to omitted factors and measurement error; diagnosing whether an observable factor model (CAPM, FF3, FF5, or richer) spans the SDF; extending the asymptotic inference results (Theorems 1-2, 4-5) to new DGPs; or applying SPCA to asset classes or frequencies beyond the monthly US equity setting studied here. The locators above point to the key tables and figures.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(1), February 2025. Copyright 2024 the American Finance Association. Paywalled; extract-only. This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**.

> Giglio, Stefano, Dacheng Xiu, and Dake Zhang. "Test Assets and Weak Factors."
> *The Journal of Finance* 80, no. 1 (February 2025): 259-319.
> DOI: 10.1111/jofi.13415. Copyright 2024 the American Finance Association.
> This page is an **extract** by the Institute for Automated Research;
> paywalled source, extract-only rights.


==============================================================================
# Allocation of Socially Responsible Capital: Green & Roth (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/green-allocation-socially-responsible-capital-2025/
# Distilled: This paper develops a tractable equilibrium framework in which social and commercial investors compete to finance entrepreneurs with varying profit and social value profiles. It shows that values-aligned ESG strategies are inefficient at creating social impact and identifies alternative impact-aligned strategies that both increase welfare and financial returns. Supported by a laboratory experiment documenting heterogeneous social preferences. J. Finance 2025, paywalled. Five core results with source locators, the model, method, and empirical specifications.
# Tags: paper-summary, esg, sustainable-finance, social-investing, impact-investing
==============================================================================

**What this is.** The paper's core results, the equilibrium model of social investor competition, and the laboratory experiment documenting investor preference heterogeneity: enough to understand what it found and how, without reading all 27 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13425).

## TL;DR

Green and Roth (2025) build a tractable competitive equilibrium model in which social investors (who care about financial returns and social value) and commercial investors compete to finance entrepreneurs. The central insight is that values-aligned ESG strategies, which dominate real-world socially responsible investing, are inefficient at creating social welfare: they displace commercial capital that would have funded the same firms anyway, and competition among social investors transfers rents to entrepreneurs rather than funding new socially valuable projects. Impact-aligned strategies, which prioritize investments where the social investor is pivotal, generate strictly positive social welfare and higher financial returns.

The paper contributes to a growing literature on social preferences in asset markets. Pastor, Stambaugh, and Taylor (2021) and Pedersen, Fitzgibbons, and Pomorski (2021) study equilibrium models in which values-aligned tilts shift capital toward socially preferred firms. Oehmke and Opp (2024) show social investors cannot generate impact because they would rather not invest than fund an improved but still polluting company; the present paper shows competition among social investors to hold valuable investments is the key friction, distinct from that mechanism. Broccardo, Hart, and Zingales (2022) argue that engagement and tilting strategies (exit vs. voice) are more effective than divestment. Landier and Lovo (2023) study how ESG investing can be optimized for impact. On the empirical side, Bonnefon et al. (2023) conduct an experiment distinguishing values alignment from impact alignment; this paper adds explicit preference heterogeneity and shows investors may mis-operationalize preferences. The enterprise-impact concept builds on Brest, Gilson, and Wolfson (2019).

A laboratory experiment with 389 participants confirms significant investor heterogeneity, with about 34% holding impact-aligned preferences, 29% holding values-aligned preferences, and 37% being purely financially motivated. Moreover, roughly 47% of participants who initially chose values-aligned options revised their choices when shown an alternative offering higher impact and higher financial return, suggesting widespread mis-operationalization of impact preferences.

## Core results

Magnitudes and significance are as reported; `\*\*\*`/`\*\*`/`\*` = 1%/5%/10%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Three latent investor classes: commercial (36.8%), values-aligned (28.8%), impact-aligned (34.4%); classes sharply differ in WTP for values vs. impact | Table I, p. 777 | Class 2 (values): WTP Values = $0.724\*\*\*, WTP Impact = -$0.167\*\*\*; Class 3 (impact): WTP Values = $0.060, WTP Impact = $0.903\*\*\* |
| R2 | 30% of participants classified as values-aligned by revealed preferences self-report impact-aligned preferences; preference mis-operationalization is widespread | §IV.B.2, p. 778 | 30% of revealed-values-aligned investors self-reported impact alignment; 95% of respondents have modal class membership probability above 80% |
| R3 | Nearly half of investors who chose values-aligned options revised when shown impact-improving alternative with equal or higher financial return | §IV.B.2, p. 779 | 109 out of 231 (approx. 47%) revised their choice in the follow-up; 59% selected values-aligned option in at least one of six test scenarios |
| R4 | Values-aligned investors create zero net social value in equilibrium: financial concessions bid up prices of firms that commercial capital would have funded, transferring rents to entrepreneurs | Proposition 2, §II.D, p. 768 | Model result (no point estimate): deviation to fund unfinanced firm weakly increases both total welfare and investor's financial return; commercial firms earn zero marginal social value |
| R5 | Impact-aligned investors exhibit negative assortative matching: higher-altruism investors fund lower social-value firms to preserve space for lower-altruism investors at high social-value opportunities | Proposition 1, §II.B, p. 766 | Model result (no point estimate): for firms with equal profit, higher altruism investor matches with lower social-value firm; contrast with values-aligned positive assortative matching (Lemma 1) |

**Overall (paper's conclusion).** Values-aligned investment strategies, which resemble the construction of conventional ESG and emissions-reduction portfolios, have limited impact because they simply displace commercial investors who would have supported some socially valuable firms anyway. The financial concession made by values-aligned investors is wasteful, transferring rents to entrepreneurs rather than funding new social value. Impact-aligned investment strategies, which prioritize firms that could not attract commercial capital, generate greater social impact and higher financial returns. The empirical evidence confirms that a substantial share of real investors have impact-aligned preferences but are incorrectly operationalizing those preferences through values-aligned strategies.

## Theory / model

The model has two types of players: a finite set $$E$$ of entrepreneurs and a finite set $$S$$ of social investors. Each entrepreneur $$i$$ holds a project requiring one unit of capital that generates profit $$\pi_i \in \mathbb{R}^+$$ and social value $$w_i \in \mathbb{R}$$, both publicly observable (pp. 760-761). There is also an elastic commercial capital market supplying financing at required rate $$r^C$$.

A contract specifies a transfer $$r_i$$ (cost of capital) from the entrepreneur to the investor; entrepreneur utility is $$\pi_i - r_i$$. Social welfare is $$W = \sum_{i \in \bar{E}} w_i$$ where $$\bar{E}$$ is the set of financed entrepreneurs (p. 762).

**Values-aligned investors** maximize the sum of their financial return and the social value of the firm they finance (eq. 1, p. 760):

$$
r_i + \theta_i w_i \tag{1}
$$

where $$\theta_i \in \{\theta^1, \ldots, \theta^K\} \subset \mathbb{R}^+$$ is investor $$i$$'s altruism strength.

**Impact-aligned investors** maximize their financial return and the effect of their investment on aggregate social welfare (eq. 2, p. 761):

$$
r_i + \theta_i \sum_{j \in \bar{E}} w_j = (r_i + \theta_i w_i) + \theta_i \sum_{j \in \bar{E} \setminus i} w_j \tag{2}
$$

where $$\bar{E}$$ is the set of entrepreneurs that receive financing. The key difference is that impact-aligned investors internalize consequences for all funded firms, not just the one they own.

**Equilibrium concept.** Pure-strategy subgame perfect equilibrium: in the acceptance stage, each entrepreneur accepts the contract maximizing their share of profits; in the offer stage, each investor chooses the contract maximizing their utility among contracts that will be accepted (p. 762).

**Key lemmas for values-aligned investors.** Lemma 1 (p. 763): investors and entrepreneurs exhibit positive assortative matching (higher altruism $$\theta_i$$ matches with higher social value $$w_i$$). Lemma 2 (p. 763): cost of capital is decreasing in social value $$w_i$$. The incentive compatibility condition is

$$
r_i + \theta_i w_i \geq r_j + \theta_i w_j \tag{3}
$$

meaning no social investor prefers to undercut another social investor.

**Proposition 1** (p. 766): Impact-aligned investors exhibit negative assortative matching. Among firms with fixed profit $$\pi$$, higher altruism investors finance firms with lower social value $$w_i$$. The incentive compatibility condition for impact-aligned investors is

$$
\pi_i + \theta_i w_i \geq r^C \tag{4}
$$

**Enterprise impact** (p. 770): the enterprise impact of firm $$i$$ is $$e_i \equiv w_i - v_i$$, where $$v_i$$ is the social value of the capital employed by the investor who supports firm $$i$$. Proposition 3 states that increasing firm profitability $$\pi_i$$ (holding $$w_i$$ fixed) weakly increases enterprise impact because a more profitable firm can attract commercial capital, freeing scarce socially motivated capital for other uses.

## Method

The paper combines two methods: a complete-information equilibrium model solved analytically, and a laboratory experiment estimated with a latent class logit model.

**Analytical model.** The baseline model is solved for pure-strategy subgame perfect equilibrium using an offer-then-accept timing (§I.C, p. 761). The equilibrium is characterized by lemmas and propositions derived from the incentive compatibility conditions (eqs. 3 and 4 above). Section III extends the baseline to a model of incomplete information in which $$\pi_i$$ and $$w_i$$ are public signals of expected private types $$(\pi^i, w^i) \in \{\pi^L, \pi^H\} \times \{w^L, w^H\}$$. Under Proposition 4, impact-aligned investors in the incomplete-information model finance all firms $$(\pi_i, w_i)$$ satisfying

$$
w_i \geq \bar{u}^S / \theta, \quad \pi_i \leq \bar{\pi}(w_i) \equiv 1 - \left(\bar{u}^S - \pi^L\right) / (\theta w_i) \tag{5}
$$

where $$\bar{u}^S$$ is the threshold utility of the marginal social investor (p. 772).

The model builds on `mechanism-design` in the game-theoretic sense (investor-offer, entrepreneur-accept timing), characterizing equilibrium investment strategies and welfare properties. Internet appendix extensions include endogenous commercial cost of capital (§I.A), endogenous firm responses (§I.C), and atomistic investors in the limit (§II.C.11).

**Laboratory experiment.** Individual utility of investor $$j$$ owning stock $$k$$ is modeled as (p. 776):

$$
u_{j,k} = \beta^r_{c_j} r_k + \beta^v_{c_j} w_k + \beta^i_{c_j} w_k p_k + \epsilon_{j,k}
$$

where $$r_k$$ is the stock's financial return, $$w_k$$ is its social value (charitable donation), $$p_k \in \{0,1\}$$ indicates whether the investor's purchase is pivotal for the donation, $$c_j$$ is the latent class of investor $$j$$, and $$\epsilon_{j,k}$$ follows a Type I extreme value distribution.

With $$C = 3$$ latent classes, the model is estimated as a `latent-class-logit` (following Heckman and Singer (1984) and Greene and Hensher (2003)): class membership probabilities $$\pi_c$$ and class-specific preference parameters $$\beta_c$$ are jointly estimated by maximum likelihood. Willingness to pay for values alignment is $$\beta^v / \beta^r$$ and for impact alignment is $$\beta^i / \beta^r$$.

## Empirical specifications

**Survey design.** In November 2023, 400 US-based stock-market investors were recruited via the Prolific platform; 11 were dropped for failing attention checks, leaving $$n = 389$$. Participants completed $$S = 14$$ pairwise investment scenarios plus 3 attention checks in random order. For each scenario $$s$$, participants chose between two stocks $$k$$ and $$k'$$ with attributes: financial return $$r_k \in \{\$3.00, \$3.50, \$4.00, \$4.50\}$$, charitable donation $$w_k \in \{\$0.50, \$1.00, \$1.50, \$2.00\}$$, and pivotality indicator $$p_k \in \{0, 1\}$$.

The aggregate social value generated by the participant's choice is (p. 775):

$$
w^{agg} = w_k i_k^{p_k} + w_{k'} (1 - i_k)^{p_{k'}}
$$

where $$i_k$$ indicates investment in stock $$k$$. One of 14 decisions was randomly selected for real-life payment, ensuring incentive compatibility.

**Latent class logit estimation.** The utility model with $$C = 3$$ classes is estimated by maximum likelihood on the pairwise choice data. Standard errors are in parentheses; significance at 1%/5%/10% is denoted `\*\*\*`/`\*\*`/`\*`. The model produces class-specific WTP estimates and class membership probabilities.

**Mis-operationalization test.** Six of the 14 scenarios were structured so that one choice is consistent only with values-aligned preferences (requires accepting lower financial return and lower social impact for higher values alignment). Participants who chose the values-aligned option in any of the six scenarios were given follow-up questions asking if they would revise their choice given information that the alternative offers higher social impact at equal or higher financial return (§IV.B.2, p. 778).

**No standard regression specification.** The paper presents no panel or cross-sectional regression with fixed effects; the empirical section relies entirely on the latent class logit and the descriptive revision rates from the follow-up exercise.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Prolific online survey / laboratory experiment (November 2023, n = 389) | Elicits revealed preferences for values alignment vs. impact alignment across 14 pairwise investment scenarios; identifies three latent investor classes and measures revision behavior | no page yet (author-collected primary data) |

Sample: 389 US-based stock-market investors recruited on Prolific in November 2023; 14 pairwise investment scenarios plus 3 attention checks per participant.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13425) if you are: building theoretical models of ESG investing or impact investing, especially ones that endogenize competition between social investors; evaluating whether values-aligned portfolio strategies (ESG tilt, exclusion screens) can generate social welfare; designing experiments or surveys to elicit social investing preferences; or studying negative assortative matching in two-sided markets with altruistic agents. The formal proofs and Internet Appendix extensions (endogenous commercial cost of capital, continuous firm production functions, atomistic investors) are in the supporting information.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(2), April 2025. This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The article is paywalled (Wiley VOR terms); extract-only.

Green, Daniel, and Benjamin N. Roth. "The Allocation of Socially Responsible Capital." *The Journal of Finance* 80, no. 2 (April 2025): 755–781. DOI: 10.1111/jofi.13425. © 2025 the American Finance Association.


==============================================================================
# The Credit Line Channel: Greenwald, Krainer & Paul (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/greenwald-credit-line-channel-2025/
# Distilled: Using confidential U.S. loan-level supervisory data (FR Y-14Q), Greenwald, Krainer, and Paul document that the COVID-19 surge in bank credit was driven by large firms drawing existing credit lines, which crowded out term lending to smaller firms and amplified the decline in aggregate investment. A calibrated structural model quantifies the credit line channel as the transmission mechanism. J. Finance 2025, paywalled. Eight core results with source locators, datasets used, the model, and the method.
# Tags: paper-summary, macro, credit-markets, bank-lending, firm-heterogeneity
==============================================================================

**What this is.** The paper's core results, the structural model, and the
empirical specifications: enough to know what it found and how, without
reading all 47 pages. To replicate or extend, read the full source at the
[original](https://doi.org/10.1111/jofi.13486).

## TL;DR

The paper uses confidential FR Y-14Q loan-level data covering over 200,000
U.S. firms across 2012-2020 to establish that the COVID-19 surge in bank
credit was driven almost entirely by large firms drawing down existing credit
line commitments, a pattern reminiscent of the 2007-09 crisis documented by
Ivashina and Scharfstein (2010). Banks that experienced larger credit line
drawdowns reduced their term lending supply more (crowding out), especially
to smaller firms that rely on term loans. The identification approach adapts
Khwaja and Mian (2008) to the Y14 panel. Smaller firms could not replace this
lost credit and consequently cut investment and cash holdings, echoing the
real-effects mechanism in Chodorow-Reich (2014). A calibrated structural model
with two firm types (constrained/term-loan-only and unconstrained/credit-line
access) shows that the predetermined pricing of credit lines is the key
mechanism: as unconstrained firms borrow heavily at fixed credit line spreads
(an insurance device documented by Sufi (2009) and Acharya and Steffen (2020)),
banks face tighter capital requirements and raise spreads on term loans,
redirecting credit away from constrained firms. The financial accelerator
framework of Bernanke, Gertler, and Gilchrist (1999) motivates the modeling
approach. The model also reveals that the Fed's bond market intervention had
large indirect effects via credit-line repayment, confirming and extending the
cross-sectional evidence of Darmouni and Siani (2025). Contemporaneous evidence
that smaller firms drew less of their unused capacity is provided by
Chodorow-Reich et al. (2022). In aggregate, the decline in investment is more
than 70% larger with credit lines than in a counterfactual economy without
them, even though aggregate bank-firm credit increases.

## Core results

Magnitudes and significance are as reported. `\*\*` = 5%, `\*\*\*` = 1%.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Undrawn credit line commitments are nearly 40% larger than total used bank credit (lines + term loans combined); top 10% of firms hold 71% of undrawn credit | Table I, p. 3144; Figure 3, p. 3146 | Committed credit = $2,231B; used credit = $941B; ratio 2.37x; Lorenz curve for unused credit is far more bowed than for used credit |
| R2 | 96% of the credit increase during 2020:Q1 flowed to the top 10% of firms by size, driven by drawdowns of existing credit lines | Figure 5, p. 3149 | Change in existing credit lines by large firms explains 77% of the total increase; bottom 90% of firms saw only modest increases |
| R3 | Banks with larger credit line drawdowns contracted term lending more; coefficient on delta Credit Line Usage = -1.96 to -2.74 across specifications | Table III col. (1)-(3), p. 3152 | beta_h=0 = -1.96\*\* (SE 0.72) baseline; -2.74\*\*\* (SE 0.93) with extended FE and controls; effects intensify through 2020:Q3 (-3.63\*\*, col. 5) |
| R4 | Banks with lower pre-crisis capital buffers restricted term lending more in response to drawdowns | Table IV, p. 3155 | Coefficient on delta CL Usage = -3.05\*\*\* (SE 1.05); interaction with Cap-Buffer = +1.25\*\*\* (SE 0.36); deposit inflows do not offset the crowding out (col. 7) |
| R5 | Firms exposed to banks with larger drawdowns experienced reductions in total debt; effect is concentrated in small firms | Table V col. (1)-(2), p. 3157 | OLS beta = -2.63\*\*\* (SE 0.69); small firms: -2.61\*\*\* (SE 0.68); large firms: 0.97 (SE 5.72), indistinguishable from zero |
| R6 | Small firms cut capital expenditures and cash holdings when total debt fell; 12 cents and 24 cents less per $1 decline in debt | Table V col. (4), (7), p. 3157 | IV beta_capex = 0.03\*\* (SE 0.01) for small firms; IV beta_cash = 0.07\*\*\* (SE 0.01); large firms' effects not significant |
| R7 | In the calibrated structural model, the decline in investment is 74% larger in the Credit Lines economy than in the Term Loans counterfactual | Figure 7, p. 3175 | Investment at t=1: -3.3% (Credit Lines) vs -1.9% (Term Loans); bank loans +33.0% vs +3.0%; aggregate corporate bonds lower in Credit Lines economy |
| R8 | The Fed's bond market intervention (SMCCF) raised constrained firm investment via indirect credit-line spillovers; the policy effect is more than five times larger in the Credit Lines economy | Internet Appendix Figure IA.6, referenced p. 3176 | Investment decline at t=1 is 3.2 pp more without intervention in Credit Lines economy vs only 0.6 pp more in Term Loans economy; constrained firm debt moves from +4.4% to -6.3% without the policy |

**Overall (paper's conclusion).** Credit lines are central to macroeconomic
shock transmission. Their predetermined pricing insulates large firms from
rising spreads, concentrating credit drawdowns among the least financially
constrained firms. This crowds out term lending to small, bank-dependent
firms and amplifies the decline in aggregate investment despite increasing
total bank-firm credit. Cross-sectional access to pre-committed credit has
quantitatively first-order aggregate consequences.

## Theory / model

The model has three household types: constrained entrepreneurs (type C),
unconstrained entrepreneurs (type U), and savers (type S). Entrepreneur
type $$j$$ has exponential utility over dividends $$C_{j,t}$$ (eq. 6, p. 3159):

$$
U_{j,t} = E_t \sum_{k=0}^{\infty} \beta_j^k \frac{(1 - \exp(-\zeta_D C_{j,t}))}{\zeta_D}. \tag{6}
$$

Firm type $$j$$ produces output with a Cobb-Douglas technology:

$$
Y_{j,t} = Z_t K_{j,t-1}^{\alpha} \bar{N}_j^{1-\alpha},
$$

where $$Z_t$$ is aggregate TFP, $$K_{j,t-1}$$ is capital, and $$\bar{N}_j$$ is fixed
labor. The representative firm of type $$j$$ maximizes firm value (eq. 11, p. 3163):

$$
V_{j,t} = D_{j,t} + \exp(\tilde{a}_t) \eta_{A,j} \frac{A_{j,t}^{1-\zeta_A}}{1-\zeta_A} + E_t\!\left[\Lambda_{j,t+1} V_{j,t+1}\right], \tag{11}
$$

where $$D_{j,t}$$ is dividends, $$A_{j,t}$$ is cash (with precautionary utility
weight $$\eta_{A,j}$$ and curvature $$\zeta_A$$), and the SDF is

$$
\Lambda_{j,t+1} = \beta_j \exp(-\zeta_D (C_{j,t+1} - C_{j,t})). \tag{12}
$$

The firm's budget constraint (eq. 13, p. 3164) is:

$$
D_{j,t} = \underbrace{(1-\tau)(Y_{j,t} - wN_j)}_{\text{after-tax profit}} + \underbrace{(1-(1-\tau)\delta)\tilde{Q}_{j,t} K_{j,t-1}}_{\text{old capital}} + \bar{\pi}^{-1} A_{j,t-1}
- \bar{\pi}^{-1}\!\left[(1-\tau)r_{t-1} + \nu + \kappa_j \Gamma_{\omega,j}(\tilde{\omega}_{j,t})\right] B_{j,t-1} - (1-\tau)S_{j,t-1}
- Q_{j,t} K_{j,t} - A_{j,t} + B_{j,t}^*, \tag{13}
$$

where $$\kappa_j \Gamma_{\omega,j}(\tilde\omega_{j,t})$$ are expected violation costs
(the covenant channel), $$B_{j,t}^*$$ is new debt, and $$S_{j,t}$$ tracks promised
spread payments. The covenant violation threshold is (eq. 10, p. 3163):

$$
\tilde\omega_{j,t} = \frac{\bar\pi^{-1} B_{j,t-1}}{\theta X_{j,t}}, \tag{10}
$$

with smoothed EBITDA $$X_{j,t} = (1-\rho_X)(Y_{j,t} - wN_j) + \rho_X \bar\pi^{-1} X_{j,t-1}$$
(eq. 8, p. 3162).

The bank holds capital $$k_t$$ and faces the capital requirement (eq. 14, p. 3164):

$$
k_t \geq \chi^B \underbrace{(B_{C,t}^{\text{loan}} + B_{U,t}^{\text{loan}})}_{\text{used credit}} + \chi^L \underbrace{(\bar{L} - B_{U,t}^{\text{loan}})}_{\text{undrawn lines}}, \tag{14}
$$

where $$\chi^B = 0.08$$ and $$\chi^L = 0.04$$ match Basel risk weights. The bank
maximizes (eq. 15, p. 3165):

$$
v_t = \underbrace{d_t}_{\text{dividends}} - \left(\frac{\eta_k}{\bar{k}^{\zeta_L}}\right) \frac{k_t^{1+\zeta_L}}{1+\zeta_L} + E_t\!\left[\Lambda_{S,t+1} v_{t+1}\right]. \tag{15}
$$

In equilibrium the short-term spread on loans to constrained firms is
$$s_{C,t}^{\text{loan}} = (1+r)\chi^B \eta_k k_t^{\zeta_L}$$ (eq. A.20, p. 3180), so credit
line drawdowns that raise $$k_t$$ directly raise spreads on term loans to
constrained firms, generating the crowding-out mechanism.

**Identification.** The paper takes the existence and pricing of credit lines
as exogenous, calibrated to the data. The COVID-19 shock is modeled as a
combination of three AR(1) processes: a negative TFP shock ($$\varepsilon_{Z,1}
= -0.1059$$), a positive cash-demand shock ($$\varepsilon_{a,1} = 0.3227$$), and a
positive corporate bond spread shock ($$\varepsilon_{s,1} = 0.1408\%$$), calibrated
to match aggregate data patterns.

## Method

The paper uses two complementary methods: a reduced-form credit-supply
regression (building on `panel-regression` and `instrumental-variables`) and a
structural general equilibrium model estimated via `method-of-simulated-moments`
targeting empirical regression coefficients.

**Descriptive regression (equation 1, p. 3146).** To document the cross-sectional
distribution of undrawn capacity, the paper regresses:

$$
\frac{\text{Unused Credit}_{i,t}}{\text{Committed Credit}_{i,t}} = \alpha_t + \tau_k + \beta \boldsymbol{X}_{i,t-4} + u_{i,t}. \tag{1}
$$

Time FE $$\alpha_t$$, industry FE $$\tau_k$$, lagged firm characteristics $$\boldsymbol{X}_{i,t-4}$$
include size, age, public status, EBITDA, leverage, tangible assets, and
investment grade. This builds on `panel-regression`.

**COVID credit supply specification (equation 2, p. 3150).** The paper adapts
the Khwaja and Mian (2008) firm fixed effect approach to the Y14 data:

$$
\frac{L_{i,t+h}^{j,k} - L_{i,t-1}^{j,k}}{0.5(L_{i,t+h}^{j,k} + L_{i,t-1}^{j,k})} = \alpha_{i,k}^h + \beta^h \frac{\Delta\text{Credit Line Usage}_t^j}{\text{Assets}_{t-1}^j} + \gamma^h X_{i,t-1}^j + u_{i,h}^{j,k} \tag{2}
$$

for $$h = 0, 1, \ldots$$ quarters after 2019:Q4. $$L_{i,t}^{j,k}$$ is term lending from
bank $$j$$ to firm $$i$$ of loan type $$k$$. The symmetric growth rate in the LHS
bounds the variable in $$[-2, 2]$$ and handles zero observations. The firm-credit-type FE
$$\alpha_{i,k}^h$$ absorbs all common demand shifts, so $$\beta^h$$ captures credit
supply. Sample: firms with term loans only (excluding cases where the same firm
has both term loans and credit lines at the same bank), 2019:Q4 to 2020:Q1-Q4.

**Firm outcome specification (equations 3-5, pp. 3156-3158).** Total debt
growth for firm $$i$$ is regressed on credit line exposure:

$$
\frac{D_{i,t+1} - D_{i,t-1}}{0.5(D_{i,t+1} + D_{i,t-1})} = \alpha_m + \beta \,\text{CL Exposure}_{i,t} + \gamma X_{i,t-1} + u_{i,t+1}, \tag{3}
$$

where $$\text{CL Exposure}_{i,t} = \sum_{j=1}^J \omega_{i,t-1}^j \left(\frac{\Delta\text{Credit Line Usage}_t^j}{\text{Assets}_{t-1}^j}\right)$$ (eq. 4, p. 3156) weights bank-level
drawdowns by each firm's borrowing share. The two-stage IV specification (eq. 5,
p. 3158) uses $$\text{CL Exposure}_{i,t}$$ and $$\sum_j \omega_{i,t-1}^j$$ as
instruments to recover the causal effect of debt changes on investment and cash:

$$
y_{i,t+1} = \bar\alpha_m + \bar\beta \frac{D_{i,t+1} - D_{i,t-1}}{0.5(D_{i,t+1}+D_{i,t-1})} + \bar\gamma X_{i,t-1} + \bar{u}_{i,t+1}, \tag{5}
$$

where $$y$$ is either CAPEX/Assets or $$\Delta\text{Cash}/\text{Assets}$$. Industry FE
and firm controls throughout; standard errors clustered by bank.

## Empirical specifications

**Unused capacity distribution (R1, Table II, p. 3147).** Regression (1) on
156,010 observations from 31,209 firms, 2012:Q3-2019:Q4. $$R^2 = 0.27$$. All
coefficients statistically significant at 1%.

**COVID credit supply (R2-R4, Table III, p. 3152).** Regression (2) at $$h=0$$
on the multilender subsample, 2019:Q4-2020:Q1. Baseline: 1,678 observations,
749 firms, 28 banks; FE: Firm $\times$ Rate type. Extended FE (col. 2): Firm
$\times$ Rate $\times$ Remaining Maturity bins; loan purpose FE (col. 3).
The crowding-out coefficient $$\beta^0$$ is -1.96 to -2.74 and significant at
the 5-1% level. Columns (4)-(6) extend to $$h=1,2,3$$: effects intensify through
2020:Q3 (-3.63\*\*) then abate by 2020:Q4 once credit lines are repaid.

**Capital buffer heterogeneity (R4, Table IV, p. 3155).** Regression (2) with
the interaction of delta CL Usage and bank Cap-Buffer (voluntary capital above
the regulatory minimum). Three specifications adding bank controls and their
interactions; main coefficient -3.05 to -4.62, interaction with Cap-Buffer
+1.25 to +3.34, all significant. Deposit inflow coefficient near zero (col. 7),
rejecting liquidity as the mechanism in favor of bank capital requirements.

**Firm outcomes (R5-R6, Table V, p. 3157).** Regression (3) OLS: 3,164 firm-quarter
observations; beta = -2.63\*\*\* for all firms. IV regressions (eq. 5) on 2,717
observations; first-stage F-stat 248 (all firms), 183 (small). Small firm CAPEX
coefficient 0.03\*\* (SE 0.01); large firm coefficient -0.02 (SE 0.14). Small
firm cash coefficient 0.07\*\*\* (SE 0.01). Industry FE (two-digit NAICS) throughout.

**Structural model (R7-R8).** Calibrated to match three regression coefficients
from Table V (columns 2, 4, 7) exactly; key parameters reported in Table VI
(p. 3168). TFP, cash-demand, and bond spread shocks calibrated to match
COVID-19 aggregate data patterns. Counterfactual: Term Loans economy where
unconstrained firms cannot access credit lines.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| FR Y-14Q (H.1 schedule, BHCs) | Primary loan-level data: committed and used credit by type (credit line vs. term loan), firm financials, quarterly 2012:Q3-2020:Q4; 207,505 distinct TINs | [FR Y-14Q](/wiki/confidential/fr-y14q/) (proprietary-confidential) |
| Compustat | Financial statements for public firms; replaces BHC-collected financials for public firms | [WRDS / Compustat](/wiki/commercial/wrds/) (licensed) |
| Orbis (Bureau van Dijk) | Financial statements for private firms; supplements Y14 BHC-collected data for private firms | [Orbis (BvD)](/wiki/commercial/orbis-bvd/) (licensed) |
| Federal Reserve H.8 releases | Aggregate U.S. commercial bank balance sheet series; used for Figure 1 context | No page yet |

Sample: 2012:Q3-2020:Q4, quarterly. Descriptive analysis: 2012:Q3-2019:Q4
(pre-COVID "normal times"). COVID credit supply regressions: 2019:Q4-2020:Q4.
Firm outcome regressions: 2019:Q2-2020:Q2. 207,505 distinct firm TINs total;
3,222 are public; the rest are private, making this substantially broader than
typical corporate finance samples. 28 BHCs in the credit supply sample.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13486) if you are:
studying credit channel transmission and want the full set of robustness
checks (Internet Appendix Tables IA.V-IA.XII); building or calibrating a
structural model of bank-firm credit with heterogeneous borrowers; studying
the Federal Reserve's corporate bond purchase programs and their indirect
effects; or extending the analysis to different crisis episodes or bank
regulatory environments. Table VI (p. 3168) contains the full calibration;
Figure 7 (p. 3175) shows the aggregate impulse responses.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(6), December 2025,
pp. 3137-3183. DOI: [10.1111/jofi.13486](https://doi.org/10.1111/jofi.13486).
This distillation was extracted by an LLM on 2026-06-03 and is **not
human-verified or independently reproduced**. The paper is paywalled (Wiley
VOR terms only; no CC licence); this page contains only extracted findings
and equations, not the verbatim text. To cite:

> Greenwald, Daniel L., John Krainer, and Pascal Paul.
> "The Credit Line Channel."
> *The Journal of Finance* 80, no. 6 (December 2025): 3137-3183.
> DOI: 10.1111/jofi.13486. © 2025 the American Finance Association.


==============================================================================
# The Disappearing Index Effect: Greenwood & Sammon (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/greenwood-disappearing-index-effect-2025/
# Distilled: The abnormal return from being added to or removed from the S&P 500 fell from an average of 7.4% in the 1990s to statistically indistinguishable from zero in the 2010s, driven by index migrations from the S&P MidCap and an overall rise in market liquidity around index events. J. Finance 2025, paywalled. Seven core results with source locators, datasets used, the model (demand-curve price impact), and the empirical decomposition.
# Tags: paper-summary, asset-pricing, equities, passive-investing, index-funds
==============================================================================

**What this is.** The paper's core results, the model (demand-curve price impact decomposition), and the empirical method behind the decline of the S&P 500 index effect: enough to know what changed and why, without reading all 42 pages. To replicate or extend, read the original at [https://doi.org/10.1111/jofi.13410](https://doi.org/10.1111/jofi.13410).

## TL;DR

The abnormal return earned by a stock added to (or removed from) the S&P 500 peaked at roughly 7-16% in the 1990s and has since fallen to essentially zero in the 2010s, despite index-tracking assets growing from near zero to approximately 7% of market capitalization. Greenwood and Sammon (2025) show this is not explained by changing firm composition. The primary drivers are (i) the growing share of "migrations" from the S&P MidCap 400, where simultaneous forced selling by MidCap trackers offsets forced buying by S&P 500 trackers, and (ii) a factor-of-20 decline in the demand-curve multiplier M (the price impact per unit demand shock), reflecting that active managers and non-S12 institutions now step in to absorb the index demand shock. The paper interprets this as the market adapting to a predictable, repeated trading opportunity, consistent with Lo's (2004) adaptive markets hypothesis.

Early studies by Shleifer (1986) and Harris and Gurel (1986) documented the original index inclusion effect. Bennett, Stulz, and Wang (2020) first noted its decline between 1997 and 2017; this paper extends their analysis to the full 1980-2020 period and covers both additions and deletions. Preston and Soe (2021) also document the decline, and Vijh and Wang (2022) document smaller returns for MidCap-to-S&P migrations. McLean and Pontiff (2016) provide the analogy that anomalies decay after academic publication. Chinco and Sammon (2024) estimate the passive-ownership share is larger than conventionally measured, which implies even larger mechanical demand shocks than previously assumed.

## Core results

Magnitudes and significance as reported; `\*\*\*`/`\*\*`/`\*` = 1%/5%/10%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **S&P 500 addition CAR fell from 7.4% in the 1990s to statistically indistinguishable from zero in the 2010s** | Table I, p. 667; Figure 2, p. 667 | 1980s: 3.4%, 1990s: 7.4%\*\*\*, 2000s: 5.2%\*\*\*, 2010s: 0.8% (insignificant). Decline 2000s to 2010s: -4.3pp\*\*\* (SE 0.93) |
| R2 | **S&P 500 deletion CAR (removal effect) likewise collapsed to zero** | Table I, p. 667; Figure 2, p. 667 | 1980s: -4.6%\*\*, 1990s: -16.1%\*\*\*, 2000s: -12.4%\*\*\*, 2010s: -0.6% (insignificant). Decline 2000s to 2010s: +11.8pp\*\*\* (SE 2.56) |
| R3 | **Changing firm composition (size, arbitrage risk, turnover, analyst coverage) explains only a small part of the decline** | Table II, pp. 671-672 | With controls, decade fixed effects fall from 7.4% to 7.4% (1990s) and from 0.8% to 1.8% (2010s); residual 2010s vs 1990s gap remains -5.6pp\*\*\* (p=0.000) |
| R4 | **Index migrations from MidCap explain a large portion of the decline in addition returns** | Table III, p. 676; Figure 6, p. 677 | Direct additions: 10.2% (1990s), 8.8% (2000s), 5.4% (2010s). Migrations: 6.7% (1990s), 2.7% (2000s), -1.8% (2010s). Migration-nonmigration gap widens from 3.6pp to 7.2pp |
| R5 | **The demand-curve multiplier M declined by a factor of roughly 20 for additions from the 1990s to 2010s** | Table V, p. 683 | Additions M: 6.75 (1995-99), 3.58 (2000-09), 0.37 (2010-20); implied elasticity -0.15 to -2.72. Deletions M: 10.76 (1995-99), 4.52 (2000-09), 0.70 (2010-20); implied elasticity -0.09 to -1.44 |
| R6 | **Decline in M is robust to composition controls and to extended windows that capture front-running** | Table VI, pp. 684-685 | With controls (col 2): M falls from 6.6 to 0.7 (additions). Extended window Ann-20 to Eff+1 (col 3): M falls from 9.1 to 1.3. Factor-of-7 decline even under the most generous front-running window |
| R7 | **Index effect decline extends to other index families (Russell 1000/2000, S&P MidCap, SmallCap, Nasdaq 100)** | Table VIII, p. 693 | Pooled additions: 4.2pp decline from 2000s to 2010s (5% significant). Pooled deletions: 10.3pp decline (1% significant). Individual index results weaker statistically |

**Overall (paper's conclusion).** The index effect grew from the 1980s through the 1990s as passive investing expanded, deepening the arbitrage opportunity. The market then adapted: active managers, institutional investors, and coordinated trading desks now provide liquidity around index events, eliminating the abnormal return on average despite continued growth in index fund assets. The primary mechanisms are migrations from the S&P MidCap (which offset demand shocks) and the rise of sophisticated liquidity provision (which lowers the multiplier M by a factor of ~20). Increased predictability of index changes plays only a minor role.

## Theory / model

The paper has no formal theoretical model but uses a simple structural equation as the organizing framework. Price impact is modeled as a constant-elasticity demand curve hit by a demand shock (equation 1, p. 658):

$$
\text{Price Impact}_{it} = M \times D_{it} \tag{1}
$$

where $$\text{Price Impact}_{it}$$ is the percentage change in price, $$D_{it}$$ is the percentage of market capitalization bought upon index addition (or sold upon deletion), and $$M$$ is minus one over the demand elasticity. Given the rise in indexation, a naive application of this model predicts that price impact should have grown since the 1980s, because the demand shock $$D$$ has been growing. The paper's puzzle is that average price impact (CAR) has instead declined, implying $$M$$ must have fallen substantially.

Taking means of equation (1) by decade and separating migrations (which face offsetting demand from MidCap trackers) from direct additions gives the decomposition (equation 4, p. 681):

$$
\overline{\text{CAR}} = M \times \bar{D} = M \times \left( w \cdot \bar{D}_{\text{Migrations}} + (1 - w) \cdot \bar{D}_{\text{NonMigrations}} \right) \tag{4}
$$

where $$w$$ is the fraction of additions or deletions that are migrations and $$\bar{D}_{\text{Migrations}}$$ and $$\bar{D}_{\text{NonMigrations}}$$ are their respective average net demand shocks. This allows the paper to back out $$M$$ as the ratio of average CAR to the average weighted demand shock, separately by decade beginning in 1995 (when MidCap data start).

**Identification.** The paper is descriptive: it documents time-series variation in event-study returns and uses a structural decomposition to separate the demand-shock channel (migrations) from the multiplier channel (liquidity). There is no causal identification design. The variation exploited is the historical widening of passive ownership and the exogenous timing of S&P 500 inclusion decisions.

## Method

The estimator for individual-event abnormal returns is the market-adjusted cumulative abnormal return (CAR), defined as (equation 2, p. 665):

$$
\text{CAR}_{it} = R_{it} - R_{S\&P\,500, t} \tag{2}
$$

where for announcement returns $$R$$ is the cumulative return from the day before to the day after the announcement, and for effective-date returns $$R$$ is the cumulative return from the day before implementation to the day after. The total return (announcement + effective window) spans the last trading day before announcement to the first trading day after the effective date; the average window is 4.8 days for additions and 5.8 days for deletions (p. 665).

To test whether composition shifts explain the trend, the authors run a cross-sectional regression on the pooled event-level sample (equation 3, p. 670):

$$
\text{CAR}_{it} = b_1 \text{Turn}_{i,t-1} + b_2 \text{Size}_{i,t-1} + b_3 \text{WZ}_{i,t-1} + b_4 \text{Cover}_{i,t-1} + \sum_{k=1}^{4} \gamma_k \mathbf{1}_{\text{era}=k} + e_{it} \tag{3}
$$

where $$\text{Turn}_{i,t-1}$$ is abnormal turnover in the month before the index change (volume/shares outstanding minus market average), $$\text{Size}_{i,t-1}$$ is market capitalization relative to total S&P 500 capitalization at announcement, $$\text{WZ}_{i,t-1}$$ is the Wurgler and Zhuravskaya (2002) arbitrage-risk measure (CAPM residual variance over the prior year), $$\text{Cover}_{i,t-1}$$ is analyst coverage from IBES, and $$\mathbf{1}_{\text{era}=k}$$ are decade fixed effects. All characteristics are demeaned. Standard errors use White (1980) heteroskedasticity-consistent variance estimators (Table II, p. 672).

For the multiplier estimation, the regression interacts the average net demand shock $$D_{\text{era}=k}$$ with decade indicators (equation 5, p. 686):

$$
\text{CAR}_{it} = b_1 \text{Turn}_{i,t-1} + b_2 \text{Size}_{i,t-1} + b_3 \text{WZ}_{i,t-1} + b_4 \text{Cover}_{i,t-1} + \sum_{k=1}^{3} \gamma_k \mathbf{1}_{\text{era}=k} \times D_{\text{era}=k} + e_{it} \tag{5}
$$

The $$\gamma_k$$ coefficients estimate the multiplier $$M$$ for each decade after controlling for firm characteristics (Table VI, pp. 684-685). The paper builds on the `event-study` and `panel-regression` primitives.

## Empirical specifications

All specifications are event-study regressions with market-adjusted returns. Key design choices:

- **Sample:** 736 S&P 500 additions and 731 deletions from Siblis Research, 1980-2020, matched to CRSP; pre-1990 announcement dates from Barberis, Shleifer, and Wurgler (2005). Observations with ACPERM within 100 days (acquisitions) and spinoffs excluded. Consistent sample for characteristic regressions requires Thompson S12 matching.
- **Return window (baseline):** Last trading day before announcement to first trading day after effective date (total window, roughly 5 days average). Sensitivity: extended to Ann-20 (20 trading days pre-announcement) to Eff+1 to capture front-running (Table VI cols 3-4, 7-8).
- **Event-study benchmark:** S&P 500 index return over the same window (market-adjusted). The market return is used rather than risk-model betas to avoid contamination from the index change itself (p. 665).
- **Composition controls (Table II):** CAR regressed on demeaned $$\text{Turn}$$, $$\text{Size}$$, $$\text{WZ}$$, $$\text{Cover}$$, and decade dummies; N=610 additions, 237 deletions; robust standard errors (White 1980). R-squared of 0.26-0.35 for additions.
- **Migration analysis (Tables III, V, VI):** MidCap changes from Siblis Research from 1995; MidCap 400 trackers identified by name/return correlation (at least 99.5% correlation to the MidCap index). Net demand shock D measured as Thompson S12 net buying by index trackers (change in split-adjusted shares held quarter before to quarter after, divided by shares outstanding).
- **Multiplier estimation (Tables V-VI):** M backed out as $$\overline{\text{CAR}} / \bar{D}$$ by decade (Table V), then via the interacted regression (equation 5) with characteristic controls (Table VI). Demand elasticity implied as $$\varepsilon = -1/M$$.
- **Institutional ownership (Table VII):** Thompson 13F ownership changes quarter before to quarter after compared to tracker net buying; active/passive distinction from Appel, Gromley, and Keim (2016).
- **Other indices (Table VIII):** Russell 1000/2000 from FTSE Russell 1990-2020; S&P MidCap/SmallCap from Siblis 1995-2020; Nasdaq 100 from Siblis 1995-2020. Returns market-adjusted; events use 10 days before effective date to one day after (no announcement dates for MidCap/SmallCap/Nasdaq); clustered by year.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CRSP daily stock returns and shares outstanding | Cumulative abnormal returns; split-adjustment factors; market-cap computation | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| Thompson S12 mutual fund holdings (quarterly) | Identifying S&P 500 and MidCap 400 tracking funds; measuring net buying/selling around index changes | [WRDS](/wiki/commercial/wrds/) (licensed) |
| Thompson 13F institutional holdings (quarterly) | Changes in total institutional ownership around index changes (Table VII) | [WRDS](/wiki/commercial/wrds/) (licensed) |
| Siblis Research S&P 500 addition/deletion history | Announcement and effective dates for S&P 500 changes, 1980-2020; also MidCap 400 and SmallCap changes | No page yet |
| Barberis, Shleifer, and Wurgler (2005) | Pre-1990 S&P 500 addition announcement dates not in Siblis | No page yet |
| IBES analyst coverage | Analyst coverage count for each firm at earnings announcement before the index change (control variable Cover) | [WRDS](/wiki/commercial/wrds/) (licensed) |
| WRDS Intraday Indicators (TAQ-based) | Value-weighted average effective bid-ask spread (percent effective spread) from TAQ data 1993-2022, using Holden-Jacobsen (2014) method | [WRDS](/wiki/commercial/wrds/) (licensed) |
| Virtu Financial implementation shortfall | Implementation shortfall for midcap stocks 2009-2021 | No page yet |
| FTSE Russell index membership | Russell 1000 and Russell 2000 additions/deletions, 1990-2020 | No page yet |

Sample: S&P 500 changes 1980-2020; MidCap/SmallCap/Nasdaq/Russell changes from 1990 or 1995 depending on data source. Returns at daily frequency; holdings at quarterly frequency.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13410) if you are: replicating the demand-curve decomposition or extending the multiplier analysis to post-2020 data; studying the mechanics of passive investing and price impact around index rebalancing events; testing whether similar patterns hold for factor-index (e.g., ESG or smart-beta) additions/deletions; or evaluating the market efficiency implications of mechanical, predictable institutional demand shocks. Table I (p. 666-667) gives the full year-by-year CAR history; Tables V-VI (pp. 683-685) give the migration-adjusted M estimates; Table VIII (p. 693) gives the other-index results.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(2), April 2025. This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The paper is paywalled (Wiley VOR licence); only extracts are reproduced here.

> Greenwood, Robin, and Marco Sammon. "The Disappearing Index Effect."
> *The Journal of Finance* 80, no. 2 (April 2025): 657–698.
> DOI: 10.1111/jofi.13410. © 2024 the American Finance Association.
> All rights reserved. Extract-only reproduction under fair-use/commentary.


==============================================================================
# Wealth and Insurance Choices: Gropper & Kuhnen (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/gropper-wealth-insurance-choices-evidence-2025/
# Distilled: Using administrative data on 63,000 U.S. households, Gropper and Kuhnen find that wealthier individuals hold more life insurance coverage, contradicting canonical theory that predicts a negative wealth-insurance relationship. The positive correlation persists after controlling for risk preferences, pricing, bequest motives, background risk, financial literacy, employer benefits, and liquidity constraints. J. Finance 2025, paywalled. Seven core results with source locators, datasets used, the theoretical framework, and the empirical specifications.
# Tags: paper-summary, household-finance, insurance, life-insurance, wealth
==============================================================================

**What this is.** The paper's core results, the theoretical framework it tests (Lewis 1989), and the empirical specifications behind each finding: enough to know what it found and how, without reading all 44 pages. To replicate or extend it, read the full source at [https://doi.org/10.1111/jofi.13426](https://doi.org/10.1111/jofi.13426).

## TL;DR

Standard models of insurance demand (Mossin (1968), Lewis (1989), Gollier (2003)) predict that wealthier agents, having greater risk-bearing capacity, purchase less insurance. Using administrative records from a U.S. financial services firm covering 63,000 individuals and 2.5 million person-month observations (September 2015 to March 2019), Gropper and Kuhnen document the opposite: wealthier households hold larger term life insurance coverage limits, are more likely to have insurance, and are less likely to let policies lapse. This positive wealth-insurance correlation survives controls for risk preferences, pricing, bequest motives, background risk, financial literacy, employer-provided benefits, and liquidity constraints. Liquidity constraints and background risk each explain a portion of the pattern, but a significant positive wealth effect remains. The findings support newer theories that view insurance as consumption-smoothing across time (Rampini and Viswanathan (2019), Ericson and Sydnor (2018), Casaburi and Willis (2018)) but call for further empirical work to identify supply-side or other demand-side mechanisms not yet studied. A contemporaneous paper by Armantier, Foncel and Treich (2023) finds a similar positive wealth-insurance correlation for auto insurance using survey data; this paper complements it by using administrative data on term life insurance with direct dollar-coverage measures. Koijen, Van Nieuwerburgh and Yogo (2016) document that households fail to rebalance portfolios away from life insurance as they age; the patterns here suggest persistence in the wealth-coverage relationship as well.

## Core results

Magnitudes and significance are as reported; `\*\*\*`/`\*\*`/`\*` = 1%/5%/10%. Locators point into the source PDF. Wealth variables are in hundreds of thousands of dollars in all OLS tables, so a coefficient of 12.54 on financial wealth means a $100,000 increase in wealth raises the dependent variable by 12.54 percentage points (Tables II-III) or $12,540 (Table IV-V when DV is in dollars).

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Financial wealth predicts higher probability of having life insurance**, opposite to theory | Table II col 3, p. 1138 | Coefficient = 12.54 (t=18.94\*\*\*); $100,000 increase associated with 11-13 pp higher probability of owning life insurance |
| R2 | **Wealthier individuals are less likely to lapse their life insurance policies** | Table III col 3, p. 1139 | Coefficient = -3.46 (t=-6.40\*\*\*); $100,000 increase in financial wealth associated with 3.1 pp decrease in lapsation probability |
| R3 | **Financial wealth predicts larger coverage limits** at both extensive and intensive margins | Table IV col 4, p. 1142; Table IV col 5 (policyholders only) | Full sample: $1 increase in wealth raises coverage by $0.64 (t=14.17\*\*\*; col 3 without housing wealth: 0.73, t=16.17\*\*\*); policyholders only: $1.41 (t=10.52\*\*\*) |
| R4 | **Within-person wealth increases predict higher coverage**, ruling out fixed individual omitted variables | Table V, p. 1145 | $100,000 within-person increase in financial wealth raises probability of having insurance by 2 pp (t=10.18\*\*\*) and coverage limit by $0.05 per $1 (t=4.72\*\*\*); coefficient attenuated relative to cross-section due to inertia/adjustment costs |
| R5 | **Background risk (consumption and income volatility) explains part of the wealth-insurance correlation** but does not eliminate it | Table VIII, p. 1154 | Controlling for consumption volatility: financial wealth coefficient = 0.62 (t=14.01\*\*\*), nearly the same as baseline 0.64; consumption volatility itself is a positive predictor of coverage (coefficient = 1.96, t=5.37\*\*\*) |
| R6 | **Liquidity constraints reduce coverage and increase lapsation** but cannot account for the wealth-coverage correlation | Table X Panel B, p. 1160 | Credit use ratio positively predicts lapsation (coefficient = 0.04, t=5.39\*\*\*) and negatively predicts coverage (coefficient = -119.66, t=-6.06\*\*\*); after controlling for it, financial wealth coefficient remains 0.63 (t=13.76\*\*\*) |
| R7 | **Education and employer-provided insurance do not eliminate the positive wealth effect** on coverage | Table IX, p. 1156 | College degree associated with $44,000 higher coverage (coefficient = 0.44, t=14.62\*\*\*); wealth effect persists in both government (coefficient = 0.40, t=11.11\*\*\*) and nongovernment subsamples (coefficient = 0.72, t=10.52\*\*\*) |

**Overall (paper's conclusion).** Wealthier U.S. households have more life insurance coverage along every dimension examined: ownership, coverage limits, and lapsation behavior. This contradicts the core prediction of canonical insurance demand models. Several frictions, including background risk and liquidity constraints, explain part of the positive correlation. However, a substantial, economically meaningful positive wealth-insurance relationship remains after all examined channels are controlled for simultaneously (Appendix Table A.VI), pointing to supply-side mechanisms or demand-side factors not yet studied empirically.

## Theory / model

The paper does not develop a new model. It uses the Lewis (1989) framework as the theoretical benchmark. The model of Lewis (1989) yields the following formula for optimal term life insurance coverage (p. 1133), under CRRA utility with risk aversion parameter $$\gamma$$:

$$
C_t = \frac{1}{(1 - \lambda\pi_t)} \left[\frac{(1-\lambda\pi_t)}{\lambda(1-\pi_t)}\right]^{\frac{1}{\gamma}} A_t - \frac{1}{(1-\lambda\pi_t)} W_t \tag{1}
$$

where $$C_t$$ is the extent of life insurance coverage purchased, $$A_t$$ is the value of the asset to be insured (the NPV of dependents' future consumption), $$W_t$$ is the financial wealth of the individual (net of the insured asset), $$\pi_t$$ is the per-period probability of death, $$\lambda$$ is the insurance loading factor ($$\lambda = 1$$ implies actuarially fair insurance; $$\lambda > 1$$ means insurance is costly), and $$\gamma$$ is the CRRA coefficient.

The key theoretical prediction is $$\partial C_t / \partial W_t < 0$$: holding the value of the insured asset $$A_t$$ fixed, wealthier individuals purchase less insurance because they have greater risk-bearing capacity (p. 1133). The model also predicts $$\partial C_t / \partial A_t > 0$$: more valuable insured assets generate higher coverage demand. The paper confirms this second prediction empirically but finds that the first prediction fails in the data: $$\partial C_t / \partial W_t > 0$$ in all specifications (p. 1141).

The paper also sketches why newer theoretical frameworks (Rampini and Viswanathan (2019), Ericson and Sydnor (2018), Casaburi and Willis (2018)) can generate a positive wealth-insurance relationship when insurance is viewed as a savings instrument (state-contingent Arrow-Debreu securities) rather than pure risk transfer: poorer households face higher marginal utility of current consumption and therefore choose to save less through insurance. These theories predict that liquidity-constrained individuals buy less insurance, consistent with the authors' evidence that more constrained individuals lapse more (p. 1158).

## Method

The paper applies linear OLS panel regressions with state fixed effects (cross-sectional) and person-plus-state fixed effects (within-person panel). There is no single proposed estimating equation; the paper runs variants of the following cross-sectional specification for each outcome (pp. 1138-1146):

$$
Y_i = \alpha + \beta_1 \, W_i^{\text{financial}} + \beta_2 \, W_i^{\text{housing}} + \mathbf{X}_i' \boldsymbol{\delta} + \text{StateFE}_i + \varepsilon_i \tag{2}
$$

where $$Y_i$$ is either an insurance-ownership indicator, a lapsation indicator, or the dollar coverage limit; $$W_i^{\text{financial}}$$ is average financial wealth (in hundreds of thousands of dollars); $$W_i^{\text{housing}}$$ is estimated housing wealth; and $$\mathbf{X}_i$$ is a vector of controls including the NPV of future dependents' spending (the theoretical value of the insured asset $$A_i$$), dependents-times-mortgage and dependents-times-other-loans interactions, and the probability of death in the next year. Standard errors are clustered by state (pp. 1138-1139).

For the within-person analysis (Table V), the specification adds person fixed effects and is estimated at the person-month level:

$$
Y_{it} = \alpha_i + \alpha_s + \beta_1 \, W_{it}^{\text{financial}} + \beta_2 \, W_{it}^{\text{housing}} + \mathbf{X}_{it}' \boldsymbol{\delta} + \varepsilon_{it} \tag{3}
$$

with $$\alpha_i$$ a person fixed effect, $$\alpha_s$$ a state fixed effect, and standard errors clustered by person (p. 1145). The within-person estimator eliminates fixed unobserved individual heterogeneity (e.g., permanent risk aversion, innate financial literacy) that could confound the cross-sectional estimates.

The paper builds on `panel-regression` throughout. It does not use an instrumental variable or a natural experiment; identification rests on conditional ignorability (controlling for observables). The authors acknowledge that supply-side mechanisms or unobserved demand-side factors may drive the residual correlation (p. 1161).

## Empirical specifications

**Coverage ownership (Table II).** The dependent variable is an indicator equal to 100 if the individual ever had a term life insurance policy. OLS with state FEs and ~59,722 individuals. Specifications progressively add controls; the headline coefficient on financial wealth is 12.54 (t=18.94\*\*\*) in column (3) and 10.98 (t=19.04\*\*\*) when housing wealth is also included (column 4). The NPV of dependents' future spending (the insured asset $$A$$) has the theoretically predicted positive sign (coefficient = 0.47, t=12.08\*\*\*), validating the approach.

**Lapsation (Table III).** Dependent variable: indicator equal to 100 if the individual ever let a term life insurance policy lapse. OLS with state FEs and ~19,875 individuals (those who ever had a policy). Financial wealth coefficient = -3.46 (t=-6.40\*\*\*); housing wealth = -3.07 (t=-5.41\*\*\*) with the housing wealth measure.

**Coverage limits (Table IV).** Dependent variable: dollar coverage limit (zero for non-policyholders). OLS with state FEs and 57,586 observations (56,519 in column 4 with housing wealth). Financial wealth coefficient = 0.73 (t=16.17\*\*\*) in the full sample (column 3) and 0.64 (t=14.17\*\*\*) adding housing wealth (column 4). The paper's headline figure of "$0.64" refers to column 4. Among policyholders only: financial wealth coefficient = 1.41 (t=10.52\*\*\*) (column 5). Housing wealth coefficient = 0.10 (t=7.83\*\*\*).

**Within-person analysis (Table V).** Person-month level, 2,027,159 observations, 55,696 unique individuals. Person plus state FEs. Financial wealth coefficient = 1.90 (t=10.18\*\*\*) for the ownership indicator and 0.05 (t=4.72\*\*\*) for the coverage limit (in dollar terms per dollar of wealth). Smaller magnitudes than cross-section due to inertia in coverage rebalancing.

**Insurance pricing (Table VI).** Dependent variable: annual premium in cents per dollar of coverage (N=8,212 policyholders with observed premiums). Financial wealth has a positive coefficient (0.04, t=4.17\*\*\*), meaning wealthier individuals pay somewhat higher prices per dollar of coverage. This runs counter to the Rampini and Viswanathan (2019) pricing channel but may reflect longer policy terms among the wealthy (p. 1150).

**Background risk (Table VII, Table VIII).** Background risk is measured as the annualized standard deviation of monthly consumption, income, and wealth. Wealthier individuals face more volatile consumption and income (Table VII: financial wealth coefficient on consumption volatility = 0.03, t=12.04\*\*\*). Adding background risk controls barely changes the financial wealth coefficient on coverage (Table VIII: from 0.64 baseline to 0.62-0.66 with different volatility measures).

**Liquidity constraints (Table X).** Credit use ratio (monthly credit card spend divided by credit limit) negatively predicts coverage (-119.66, t=-6.06\*\*\*) and positively predicts lapsation (0.04, t=5.39\*\*\*). After controlling for liquidity constraints, the financial wealth coefficient on coverage remains 0.63 (t=13.76\*\*\*).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Proprietary administrative data from a U.S. financial services firm | Primary dataset: individual-level bank account balances, monthly income/spending flows, term life insurance coverage limits and premiums, demographic information; 63,141 individuals, Sep 2015-Mar 2019 | No page yet |
| Corelogic transaction prices | Housing wealth estimates: zip-code home price percentiles by income quintile; used to assign housing wealth to homeowners | [CoreLogic](/wiki/commercial/corelogic/) (licensed) |
| American Community Survey (5-year tables) | Income quintile boundaries by zip code, used to assign homeowners to income quintiles for housing wealth estimation | No page yet |
| CDC National Vital Statistics System | Age-sex specific probability of death in the next year, merged at the individual level | No page yet |
| Survey of Consumer Finances (2016 SCF) | Representativeness check: income and wealth distribution comparison (Appendix Table A.I) | No page yet |
| Zillow Home Value Index (ZHVI) | Housing wealth volatility: standard deviation of monthly zip-code ZHVI as the measure of housing wealth volatility | No page yet |

Sample: 63,141 individuals, September 2015 to March 2019 (43 months). 2,500,000 person-month observations. Individuals broadly represent the middle 50% of the U.S. income distribution (Table I, p. 1137).

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13426) if you are: studying the determinants of household insurance demand using administrative data; investigating whether canonical insurance theories hold in household-level data; evaluating the role of liquidity constraints in insurance choice; or extending the analysis to supply-side mechanisms, business-cycle variation, or other insurance products. The appendix (Tables A.I-A.VI) contains robustness checks and alternative insured-asset definitions (NPV of future labor income vs. consumption).

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(2). Paywalled; copyright 2025 the American Finance Association. This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. Extract-only; no PDF hosted here.

> Gropper, Michael J., and Camelia M. Kuhnen.
> "Wealth and Insurance Choices: Evidence from U.S. Households."
> *The Journal of Finance* 80, no. 2 (April 2025): 1127–1170.
> DOI: 10.1111/jofi.13426.


==============================================================================
# Sending Out an SMS: Grubb, Kelly, Nieboer, Osborne & Shaw (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/grubb-sending-out-sms-automatic-2025/
# Distilled: At-scale field experiments at major U.K. banks show that automatic enrollment into just-in-time overdraft text alerts reduces unarranged overdraft and unpaid item charges 17% to 19% and arranged overdraft charges 4% to 8%, implying potential annual market-wide savings of GBP 170 million to GBP 240 million. J. Finance 2025, CC BY 4.0. Eight core results with source locators, datasets used, the identification strategy, and the estimating specification.
# Tags: paper-summary, household-finance, consumer-finance, behavioral-finance
==============================================================================

**What this is.** The paper's core results, the identification strategy (two randomized controlled trials at major U.K. banks), the estimating specification (DiD with individual and month fixed effects), and the behavioral mechanism evidence: enough to know what it found and how, without reading all 48 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1111/jofi.13404).

## TL;DR

The paper runs at-scale field experiments at two large U.K. retail banks (Banks A and B) with a combined sample of 1.1 million banking customers in 2017 to 2018. Customers are automatically enrolled into various overdraft text alert treatments; a control group receives only the alerts already required by regulation. All tested stand-alone alerts (both just-in-time and early-warning) significantly reduce the overdraft charges they target by 2% to 19%, with just-in-time alerts reducing unarranged overdraft and unpaid item (UOD and UI) charges by 17% to 19% and arranged overdraft (AOD) charges by 4% to 8%. These findings directly informed FCA regulation expanded to all U.K. banks in 2019. Incremental early-warning alerts layered on top of existing just-in-time alerts show no statistically detectable effect. The primary mechanism is that alerts prompt customers to log into their accounts and transfer funds on the same day, rather than cut spending. Alerts help customers across the income and overdraft-frequency distribution, with absolute benefits highest for heavy overdrafters.

## Core results

Magnitudes are as reported; `\*\*`/`\*\*\*` = 5%/1% significance. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Stand-alone just-in-time alerts reduce total monthly charges by 3-19%** across both banks | Table V Panel A, p. 487 | Treatment 1 (Bank A AOD alert): -0.528 GBP/month total (8.0%)\*\*\*; Treatment 4 (Bank B UOD+UI): -0.399 GBP/month (3.8%)\*\*\*; Treatment 6 (Bank B UOD only): -0.557 GBP/month (19%)\*\*\* |
| R2 | **Just-in-time AOD alerts reduce AOD charges 4-8%**; early-warning AOD alert reduces them 2.4% | Table V Panel A, p. 487 | Treatment 1: -0.529 GBP/month (8.4%)\*\*\*; Treatment 2: -0.302 GBP/month (3.8%)\*\*\*; Treatment 3: -0.194 GBP/month (2.4%)\*\*\* |
| R3 | **Incremental early-warning alerts have statistically insignificant effects** on charges once just-in-time alerts are in place | Table V Panel B, p. 488; Table VI col. 7, p. 489 | Treatments 7-9 postmandate: point estimates near zero, not statistically significant; 95% CI rules out effects larger than 14p/month postmandate |
| R4 | **Alerts reduce days in overdraft** by 4-21% per month across stand-alone treatments | Table VI, p. 489 | Treatment 1: -0.499 days/month (8.7%)\*\*\*; Treatment 4: -0.063 days/month (15%)\*\*\*; Treatment 6: -0.081 days/month (21%)\*\*\* |
| R5 | **Absolute benefit rises with overdraft propensity**; low-income and heavy overdrafters both benefit from alerts | Tables VII-VIII, pp. 492-493 | Frequent overdrafters: -1.30 GBP/month AOD (Treatment 1)\*\*\*; Rare overdrafters: -0.09 GBP/month (39% effect size)\*\*\*; all account-inflow terciles show significant reductions |
| R6 | **Alerts raise same-day account logins by 10-53%** on the day they are received | Table X, p. 498 | Treatment 1 (AOD just-in-time): +0.495 logins on day 0 (53% increase); Treatment 9 (early-warning, premandate): +0.27 logins; Treatment 10 (UI): +0.13 logins |
| R7 | **Primary mechanism is fund transfers**, not spending cuts; each alert triggers up to 0.27 same-day transfers averaging GBP 465 | Table XII, p. 500; Table XI, p. 499 | Treatment 1: +0.272 transfers on day 0 (52% increase, average value GBP 465 per transfer, GBP 126 per alert); debit card reduction economically small (-0.08 transactions) |
| R8 | **Alerts reduce the probability of exceeding overdraft thresholds** at all day counts from 0 to 15 days in a month | Table IX, pp. 496-497 | Treatment 1: -3.71pp (11%) for any overdraft day; -3.03pp (11%) for over 3 days; -2.58pp (10%) for over 5 days; -1.74pp (8.4%) for over 10 days |

**Overall (paper's conclusion).** Just-in-time overdraft alerts provide large consumer benefits without offsetting harm: automatic enrollment is the key because active opt-in rates were below 8% prior to mandates. Alerts eliminate less than half of overdraft charges arising from inattention, so scope remains for further interventions. The FCA expanded the CMA's mandate for just-in-time UOD and UI alerts to cover more banks in 2019 and added a mandate for just-in-time AOD alerts, based directly on these findings.

## Theory / model

The paper has no formal structural model. The tested hypothesis is the **inattention hypothesis**: consumers overdraft not because they value credit above its cost, but because they are unaware of their current account balance at the time of transactions (Stango and Zinman (2009), Armstrong and Vickers (2012)). If inattention is the primary friction, then real-time text alerts that notify customers when their balance approaches or crosses an overdraft threshold should allow customers to take corrective action (transfer funds, cut spending) and avoid overdraft charges.

The only prior study of overdraft alert experiments is the contemporaneous work of Ben-David, Mintz, and Sade (2021), who study e-mail alerts sent by the personal finance application Mint. The present study finds larger effects (17-19% vs. 3-9%), consistent with text messages being opened by 99% of recipients while Mint e-mails are opened by only 28-31% of users (p. 472). In contrast to Stango and Zinman (2014), who find that raising overdraft fee salience via a survey question does not affect account inflows, this paper shows that just-in-time text alerts do affect inflows via same-day transfers.

A secondary hypothesis motivates early-warning alerts specifically: by warning customers before the overdraft threshold is crossed, they have more time to cut spending rather than react after the fact. The paper also considers the "waterbed effect" (Agarwal et al. (2015)): mandated reductions in overdraft charges might cause banks to raise other fees, partially offsetting consumer benefits. No evidence for or against this is available from the experimental data; it is left as a question for future work (p. 508).

The **identification strategy** is randomized assignment. Two large U.K. retail banks randomly assigned eligible customers (those with a valid mobile number who incurred overdraft charges and whose balance dropped below GBP 1,000 in the prior six months) to treatment and control groups. Bank A used stratified (block) randomization on key pretreatment variables; Bank B used simple random sampling. Balance on pretreatment observables is verified in Internet Appendix Table IA.XXII. The FCA served as a pre-registration authority: trial dates and sample sizes were agreed in a "Terms of Reference" document before the experiments began (Section IX of the Internet Appendix). An ethics review board approved the protocol (p. 478).

## Method

The primary estimator is an intent-to-treat (ITT) difference-in-differences specification with individual and month fixed effects. The estimating equation is (equation 1, p. 484):

$$
Y_{i,t} = \beta_1 \text{Treatment}_i \times I(t \geq 7) + \beta_2 \text{Treatment}_i \times I(t = 7) + \eta_i + \mu_t + \epsilon_{i,t} \tag{1}
$$

where $$Y_{i,t}$$ is the outcome variable for individual $$i$$ in month $$t$$, $$\text{Treatment}_i$$ is an indicator equal to one if customer $$i$$ was assigned to the treatment group, $$I(t \geq 7)$$ is an indicator for the treatment period (months 7-11), $$I(t = 7)$$ is an indicator for the first treatment month (included separately because observed overdraft charges lag by approximately 25 days due to billing cycles), $$\eta_i$$ are individual fixed effects, $$\mu_t$$ are month fixed effects, and $$\epsilon_{i,t}$$ is an error term. Standard errors are clustered by individual and month. The parameter of interest is $$\beta_1$$, the ITT effect of being automatically enrolled into the treatment alert.

For treatments 4 and 6, which use a staggered-rollout design (design B2 in Figure 3), months 10 and 11 are excluded because all units are treated in those months, following the approach of Goodman-Bacon (2021) who shows that the standard DiD estimate is hard to interpret when treatment effects vary over time.

The behavioral-response analysis uses a second specification to measure same-day account activity following alerts (equation 2, p. 496):

$$
Y_{i,t} = \sum_{k=-3}^{3} \gamma_k (k\text{-Days.after.predicted.alert}_{i,t}) + \lambda \text{Treatment}_i \times I(t \geq 7)
+ \sum_{k=-3}^{3} \beta_k (k\text{-Days.after.alert}_{i,t}) + \eta_i + \mu_t + \epsilon_{i,t} \tag{2}
$$

where $$k\text{-Days.after.alert}_{i,t}$$ equals one if and only if day $$t$$ falls $$k$$ days after the first treatment alert customer $$i$$ received. The indicator for actual alert arrival is instrumented by the predicted alert arrival interacted with treatment (equation 3, p. 497):

$$
k\text{-Days.after.predicted.alert}_{i,t} \times \text{Treatment}_i \times I(t \geq 7) \tag{3}
$$

This IV is needed because receiving an alert is endogenous to opt-in and opt-out decisions, and predicted-alert indicators have measurement error. Customer ($$\eta_i$$) and day ($$\mu_t$$) fixed effects are included; standard errors are clustered by customer and day.

## Empirical specifications

**Main treatment effects (R1-R2, Table V).** The headline specification is equation (1) estimated separately for each of the 11 treatments, with outcomes: (i) observed AOD charges, (ii) UOD and UI charges (sum), and (iii) total charges. Observed charges are deducted from individual accounts once per billing cycle. Effect sizes are reported as $$-100 \cdot \hat{\beta}_1 / \text{baseline mean}$$ where baseline mean is the control group mean in the treatment period. Sample sizes range from 64,654 customers (Treatment 6) to 274,471 (Treatment 10). Adjusted $$R^2$$ ranges from 0.36 to 0.78, reflecting the absorbing power of individual fixed effects.

**Secondary outcomes (R3-R4, Table VI).** The same specification is applied to inferred charges, number of unpaid items, days in overdraft per month, number of overdraft episodes (by length), and number of zero-day overdraft episodes. For treatment 9, pre- and post-mandate periods are estimated separately (design A) to distinguish the stand-alone from incremental alert effect.

**Heterogeneity (R5, Tables VII-VIII).** The main specification is re-estimated within subsamples split by (i) pretreatment account inflows tercile (proxy for income: GBP 0-1,500, GBP 1,500-3,000, GBP 3,000+/month) and (ii) pretreatment overdraft propensity (rare, occasional, frequent overdrafters based on the pretreatment median of positive charges).

**Distribution of days in overdraft (R8, Table IX).** The specification is re-estimated with the binary outcome that days in overdraft exceed thresholds $$d \in \{0, 3, 5, 10, 15, 20, 25, \text{full month}\}$$. Effects are reported in percentage points. Standard errors cluster by customer and month. Stand-alone treatments only.

**Behavioral response (R6-R7, Tables X-XII).** Equations (2)-(3) estimated at Bank A on daily outcomes: number of account logins, number of debit card transactions, and number of account transfers, in the six-day window around alert arrival. IV first-stage F-statistics exceed 10,300,300. Effects are relative to a baseline day defined as four or more days before the first predicted treatment alert.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| FCA-commissioned bank transaction panel (Banks A and B), 2017-2018 | RCT treatment and control data: monthly overdraft charges (AOD, UOD, UI), daily account logins, debit card transactions, transfers; 1.1 million customers, 11 months | No page yet (proprietary-confidential FCA data) |
| FCA representative 6-bank panel, 2015 | Descriptive statistics on U.K. current account market: 1,366,355 accounts, overdraft charges by bank (Table II) | No page yet (proprietary-confidential FCA data) |
| U.K. Competition and Markets Authority (CMA) 2016 data | Context on overdraft usage: over two-thirds of accounts have an overdraft facility; 50% to 60% of days in overdraft avoidable via savings or credit cards | No page yet |
| Follow-up consumer survey (linked to Banks A and B experiments) | Survey of enrolled customers on alert attitudes, actions taken, and reasons for opting out (Internet Appendix Section I) | No page yet |

Sample: treatment period is November 2017 to March 2018 (months 7-11); pretreatment period is May to October 2017 (months 1-6). The FCA-commissioned data include all current accounts held by sampled customers and all their transactions during the sample period.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13404) if you are: designing or evaluating overdraft alert mandates in any jurisdiction; studying the role of inattention in consumer borrowing behavior; building on randomized experiments at U.K. banks under FCA auspices (the Internet Appendix contains the full treatment table, survey details, and all additional robustness tables); or interested in how at-scale field experiments can directly inform regulation. The behavioral-response tables (Tables X-XII) are the most detailed evidence on how alerts mechanically reduce charges.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(1), February 2025. This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Grubb, Michael D., Darragh Kelly, Jeroen Nieboer, Matthew Osborne, and Jonathan Shaw.
> "Sending Out an SMS: Automatic Enrollment Experiments for Overdraft Alerts."
> *The Journal of Finance* 80, no. 1 (February 2025): 467–514.
> DOI: 10.1111/jofi.13404. © 2024 The Author(s).
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Excess Capacity, Marginal q, and Corporate Investment: Grullon & Ikenberry (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/grullon-excess-capacity-marginal-q-2025/
# Distilled: When managers anticipate excess capacity, average q becomes a biased proxy for marginal q; augmenting Tobin's q model with asset utilization (sales scaled by total capital including intangibles) substantially improves explanatory power in time-series and cross-sectional investment regressions, eliminates the paradoxical negative q-investment relation, and explains why investment rates have declined for decades despite rising average q. J. Finance 2025, paywalled. Eight core results with source locators, the theoretical model, and the estimating specifications.
# Tags: paper-summary, corporate-investment, tobin-q, excess-capacity, intangible-capital
==============================================================================

**What this is.** The paper's core results, the theoretical model (a demand-constrained investment problem showing average q becomes biased for marginal q), and the estimating specifications that resolve the q-investment paradox: enough to understand what was found and how, without reading all 60 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1111/jofi.13439).

## TL;DR

Tobin's q predicts firms should invest more when market value exceeds replacement cost, yet average q has been rising for decades while corporate investment has been declining. This paper shows the paradox arises because average q is a biased proxy for marginal q whenever managers anticipate excess capacity: the prospect of underutilizing new capital lowers its marginal benefit, creating a wedge between average and marginal q. Augmenting investment regressions with asset utilization (sales scaled by total capital including intangibles, following Peters and Taylor (2017)) restores a positive q-investment relation, raises time-series R-squared from 0.29 to over 0.95, lowers out-of-sample MSE by 89%, and explains why the investment decline has been most severe in industries with the greatest drop in asset utilization. The findings hold across all 10 Fama-French industries, all firm-size deciles, and all G7 countries. Economic rigidities in labor and product markets, not financial constraints or automation, appear to be the primary cause of the secular capacity buildup.

## Core results

Magnitudes and significance are as reported in the source. `*`/`**`/`***` = 10%/5%/1%.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **q-investment relation flips from negative to positive** when asset utilization is included as a control | Table I Panel A (p. 1545), Table III Panel A (p. 1555) | Traditional model: ln(qtot) coeff = -0.35 (t = -2.82); augmented model (CurrAU proxy): coeff = +0.21 (t = 6.31) |
| R2 | **Augmented model R-squared triples** in time-series investment regressions | Table III Panel A (p. 1555) | Traditional R2 = 0.29; augmented model R2 = 0.95-0.98 depending on asset utilization proxy; asset utilization coeff approx 1.0-1.2 (t = 23-39) |
| R3 | **Firm-level panel confirms results**: augmented model R2 rises significantly after controlling for asset utilization | Table VI Panel A (p. 1561) | Traditional within-firm R2 = 0.28 (N = 140,655); augmented R2 = 0.35-0.45; asset utilization coeff = 0.43-0.63 (t = 14-29) |
| R4 | **No structural break in q-investment relation post-1996** once excess capacity is controlled | Table III Panels B-C (pp. 1555-1556) | Pre-1996: augmented q coeff = 0.24-0.32 (all positive and significant); post-1996: augmented q coeff = 0.09-0.23 (all positive and significant); break disappears |
| R5 | **Augmented model outperforms out-of-sample**: MSE 89% below traditional q model | Figure 8 (p. 1568) | Augmented model MSE = 0.0098%; traditional q model MSE = 0.0881%; naive model MSE = 0.0902%; traditional q model barely beats naive |
| R6 | **Decline in investment rates across industries fully explained by decline in asset utilization** | Table IX (p. 1574) | ln(CurrAU) coeff = 0.83 (t = 5.80); ln(qtot) coeff = -0.10 (t = -0.85, insignificant); long-run cross-industry R2 rises from 0.51 to 0.79 |
| R7 | **Effect is strongest in largest firms** (consistent with excess capacity, not financial constraints): augmented model improvement increases monotonically with firm size | Table VII Panels B-D (pp. 1563-1564) | Asset utilization coeff = 0.55-1.27 across all 10 size deciles (all significant); q coeff flips from -0.36 (largest decile, traditional) to +0.17 (augmented); largest improvement in decile 10 (R2 rises from 0.31 to 0.94-0.98) |
| R8 | **International evidence (G7) confirms results**: asset utilization resolves negative q-investment relation in all G7 countries | Table XII (p. 1586) | US augmented R2 = 0.93-0.96; other G7 R2 = 0.65-0.93; asset utilization elasticity positive and significant in all countries; US unique in showing negative traditional q relation |

**Overall (paper's conclusion).** After correcting for the measurement error induced by anticipated excess capacity, Tobin's q model of investment works as theory predicts: average q is positively related to investment throughout 1974-2021, with no structural break post-1996 (contra Andrei, Mann, and Moyen (2019)). The secular decline in corporate investment rates (documented by Jones and Philippon (2016)) is explained by a secular buildup of excess capacity, driven by economic rigidities that prevent firms from fully adjusting output prices and production costs during negative demand shocks. Tobin's marginal q theory has always worked; measurement error from ignoring excess capacity has obscured the relation.

## Theory / model

The paper builds a continuous-time model of a price-taking firm that maximizes the present value of cash flows, subject to a demand constraint and convex capital adjustment costs (Section III.A, p. 1546). The model follows the intuition in Precious (1985) and the analytical framework of Licandro (1992a) and Licandro (1992b), who show that excess capacity drives a wedge between average and marginal q. The maximization problem is:

$$
\max V(t) = \int_t^\infty e^{-r(s-t)} \{ p(s) F[K(s), L(s)] - w(s)L(s) - I(s) - C[I(s), K(s)] \} \, ds \tag{3}
$$

subject to the law of motion for capital:

$$
I(t) = \dot{K}(t) \tag{3a}
$$

and a demand constraint (the key addition relative to Hayashi (1982)):

$$
F[K(t), L(t)] \leq \bar{Q}(t) \tag{3b}
$$

where $$r$$ is the discount rate, $$p$$ is the output price, $$F[K,L]$$ is the production function, $$w$$ is the wage rate, $$K$$ is capital, $$L$$ is labor, $$I$$ is investment, $$C[I,K]$$ is the capital installation cost, and $$\bar{Q}$$ is the demand ceiling.

Using a Lagrangian with costate variable $$q$$ (average/market q) and Lagrange multiplier $$\lambda_1$$ on the demand constraint (p. 1546):

$$
\mathcal{L} = pF[K,L] - wL - I - C[I,K] + qI + \lambda_1(\bar{Q} - F[K,L]) \tag{4}
$$

The first-order condition for capital (Euler equation) is:

$$
\dot{q} - rq = -pF_K[K,L] + C_K[I,K] + \lambda_1 F_K[K,L] \tag{7}
$$

Solving forward and imposing the transversality condition yields:

$$
q(t) = \int_t^\infty e^{-r(s-t)} \{ (p(s) - \lambda_1(s)) F_K[K(s),L(s)] - C_K[I(s),K(s)] \} \, ds \tag{12}
$$

This shows that when the firm is demand-constrained ($$\lambda_1 > 0$$), the shadow value of capital (marginal $$q$$) declines, because some of the productivity of additional capital cannot be monetized. Marginal $$q$$ is a decreasing function of expected excess capacity.

Under Hayashi's (1982) homogeneity assumption (production function $$F$$ and installation cost $$C$$ both homogeneous of degree one), integrating the Euler equation forward yields the key result (equation 23, p. 1548):

$$
q(t) = \frac{V(t)}{K(t)} - \int_t^\infty e^{-r(s-t)} \frac{\lambda_1(s) F[K(s),L(s)]}{K(t)} \, ds \tag{23}
$$

When there is no demand constraint ($$\lambda_1 = 0$$ for all $$s$$), marginal $$q$$ equals average $$q$$ as in Hayashi (1982). As $$\lambda_1$$ increases (demand more binding), the wedge between marginal and average $$q$$ grows, making average $$q$$ a progressively worse proxy for marginal $$q$$. The wedge equals the present value of future output that cannot be sold due to demand constraints, scaled by current capital.

## Method

The paper applies two complementary empirical strategies. The first (aggregate time-series) closely follows the standard investment literature; the second (firm-level panel) uses fixed effects to control for unobserved heterogeneity. Both build on `panel-regression` and `fama-macbeth` as technique primitives. Prior explanations for weak q-investment relations include measurement error in q (Erickson and Whited (2000)), intangible assets, financial frictions, and declining competition; this paper adds anticipated excess capacity as a previously unexplored source. The finding that the augmented model works best for the largest firms corroborates Grullon, Hund, and Weston (2018), who document that investment-q sensitivity is negative specifically for large firms.

**Measuring total investment and total q.** Following Peters and Taylor (2017), the paper adjusts both investment and average $$q$$ for intangible capital. Total investment is (equation 1, p. 1539):

$$
itot_t = \frac{\sum_{i=1}^N (I_{i,t} + 0.3 \times SG\&A_{i,t} + R\&D_{i,t})}{\sum_{i=1}^N (K_{i,t-1} + KINT_{i,t-1})} \tag{1}
$$

where $$I$$ is capital expenditure, $$SG\&A$$ is selling, general, and administrative expenses (net of R&D), $$R\&D$$ is research and development, $$K$$ is net PP&E, and $$KINT$$ is the replacement cost of intangible capital from Peters and Taylor (2017). Total average $$q$$ is (equation 2, p. 1539):

$$
qtot_t = \frac{\sum_{i=1}^N V_{i,t}}{\sum_{i=1}^N (K_{i,t} + KINT_{i,t})} \tag{2}
$$

where $$V$$ is adjusted market value (market equity plus long-term debt minus net working capital).

**Asset utilization proxy.** Aggregate asset utilization is defined as (equation 24, p. 1549):

$$
au_t = \frac{\sum_{i=1}^N S_{i,t}}{\sum_{i=1}^N (K_{i,t} + KINT_{i,t})} \tag{24}
$$

where $$S$$ denotes total sales. Three proxies for expected asset utilization are used: $$CurrAU_{t-1}$$ (sales-to-total-capital ratio at $$t-1$$, available to managers at the start of year $$t$$), $$ExpAU_{t-1}$$ (realized sales at $$t$$ scaled by total capital at $$t-1$$, a proxy for managers' one-year-ahead expectation), and $$AvgAU_{t-1}$$ (average of the previous two, smoothing temporary shocks).

The predictive validity of these proxies is established via Fama and MacBeth (1973) cross-sectional regressions relating future $$AU_{t+n}$$ to current $$AU_{t-1}$$ across horizons $$n = 1, 3, 5$$ years (Table II, p. 1552): elasticities range from 0.76 to 0.91 with R-squared of 0.65-0.85, confirming current utilization reliably predicts future utilization.

## Empirical specifications

**Aggregate time-series specification (Table III, p. 1555).** The augmented investment equation, estimated in log-log form on annual aggregate Compustat data 1974-2021 with Newey-West (1987) standard errors:

$$
\ln(itot_t) = \alpha + \beta_1 \ln(qtot_{t-1}) + \beta_2 \ln(AU_{t-1}) + \varepsilon_t \tag{}
$$

where $$AU_{t-1}$$ is one of the three proxies. No fixed effects (aggregate time-series). Newey-West with one lag corrects for autocorrelation. This specification produces the headline results (R1, R2): $$\beta_1$$ flips from -0.35 (t = -2.82, traditional) to +0.21 (t = 6.31, augmented with CurrAU), and R-squared jumps from 0.29 to 0.95.

**Industry-level time-series specification (Table IV, p. 1557).** The same log-log form estimated separately for each of the 10 Fama and French (1997) industries on annual aggregated firm-level data:

$$
\ln(itot_{j,t}) = \alpha_j + \beta_1 \ln(qtot_{j,t-1}) + \beta_2 \ln(AU_{j,t-1}) + \varepsilon_{j,t} \tag{}
$$

where $$j$$ denotes industry. No pooled fixed effects; separate intercepts. Asset utilization is uniformly positive and significant across all 10 industries.

**Firm-level panel specification (Table VI, p. 1561).** Annual firm-level data 1974-2021 (N = 140,655 firm-years) with firm and year fixed effects:

$$
\ln(itot_{i,t}) = \alpha_i + \gamma_t + \beta_1 \ln(qtot_{i,t-1}) + \beta_2 \ln(AU_{i,t-1}) + \varepsilon_{i,t} \tag{}
$$

Standard errors are Newey-West with one lag. All ratios winsorized at the 1st and 99th percentiles to mitigate outliers. The within-firm R-squared rises from 0.28 to 0.35-0.45 (result R3).

**Size-decile analysis (Table VII, p. 1563).** The time-series specification re-estimated separately by total-capital decile (tangible plus intangible capital). Identifies the excess-capacity-versus-financial-constraints distinction (result R7): augmented model improvement is largest for the largest firms, where financial constraints are least binding.

**Long-run cross-industry specification (Table IX, p. 1574).** Cross-sectional OLS of long-run investment change on long-run q change and long-run asset utilization change, where long-run change is defined as (equation 25, p. 1572):

$$
\Delta \ln(X) = \ln(\bar{X}_{Post}) - \ln(\bar{X}_{Pre}) \tag{25}
$$

with $$\bar{X}_{Post}$$ the 2019-2021 average and $$\bar{X}_{Pre}$$ the 1974-1976 average. No constant (first differences). Establishes that declining asset utilization, not rising q, explains the secular decline in investment rates (result R6).

**Out-of-sample forecasting (Figure 8, p. 1568).** Rolling 20-year windows forecast $$itot_t$$ using information available at $$t-1$$; the augmented model MSE (0.0098%) is compared to the traditional q model (0.0881%) and a naive model (0.0902%) using rolling and expanding windows.

**International analysis (Table XII, p. 1586).** The aggregate time-series specification replicated country by country on Worldscope data for G7 countries, 1981-2021. Market value defined as equity plus total assets minus book equity minus net working capital (Worldscope lacks the long-term debt decomposition available in Compustat).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Compustat (WRDS) annual fundamentals | Primary dataset: PP&E, sales, R&D, SG&A, market value, long-term debt, working capital; intangible capital (KINT) via Peters and Taylor (2017) algorithm; firm sample 1974-2021 | [WRDS / Compustat](/wiki/commercial/wrds/) (licensed) |
| Federal Reserve capacity utilization (FRED) | Survey-based capacity utilization at industry level (manufacturing, mining, electric/gas utilities); used to validate asset utilization as a proxy for capacity utilization (Figures 6, 9, 13) | [FRED](/wiki/datasets/fred/) |
| Worldscope | International accounting data for G7 countries (Canada, France, Germany, Italy, Japan, UK, US), 1981-2021; used for Section VII international robustness | No page yet |
| USPTO patent data | Patents granted 1974-2021; used as supplementary evidence for rising intangible capital (Figure 1, Panel B) | No page yet |
| BEA National Income and Product Accounts (NIPA) | Aggregate capital and value-added data (Table S.5.a); used in Section V.C to document government underreporting of intangibles | No page yet |

Sample: Compustat non-financial, non-utility firms (SIC 4900-4999 and 6000-6999 excluded) with at least $5 million in PP&E or sales, annual frequency, 1974-2021. N = 140,655 firm-years.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13439) if you are: (i) building investment models that control for capacity utilization or intangible capital, (ii) testing or extending Tobin's q theory, (iii) investigating why corporate investment rates have declined since the 1980s, (iv) examining the role of economic rigidities (price and wage stickiness) in generating persistent excess capacity, or (v) using Peters and Taylor (2017) intangible capital measures. The locators above point to the exact tables and figures for each result.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(3), June 2025. This distillation was extracted by an LLM (claude-sonnet-4-6) on 2026-06-06 and is **not human-verified or independently reproduced**. The paper is paywalled under Wiley standard VOR terms; extract-only.

> Grullon, Gustavo, and David L. Ikenberry. "Excess Capacity, Marginal *q*, and Corporate Investment." *The Journal of Finance* 80, no. 3 (June 2025): 1533-1592. DOI: 10.1111/jofi.13439. © 2025 the American Finance Association.


==============================================================================
# In Too Deep: Guenzel (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/guenzel-too-deep-effect-sunk-2025/
# Distilled: Quasi-random cost shocks in fixed-exchange-ratio stock mergers show that higher acquisition costs reduce subsequent divestiture rates by 8% to 9%, providing the first cleanly identified field evidence that sunk costs distort corporate investment decisions. J. Finance 2025, paywalled. Seven core results with source locators, the conceptual framework, the identification design, and the estimating equations.
# Tags: paper-summary, behavioral-corporate-finance, mergers-acquisitions
==============================================================================

**What this is.** The paper's core results, the conceptual framework formalizing sunk cost effects on divestiture decisions, and the identification strategy: enough to know what it found and how, without reading all 54 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13430).

## TL;DR

This paper provides the first cleanly identified field evidence that sunk costs distort corporate investment decisions. In fixed-exchange-ratio (Fixed Shares) stock mergers, the final dollar acquisition cost is unknown at agreement signing. Aggregate market fluctuations between merger agreement and completion create plausibly exogenous variation in acquisition costs. Higher quasi-random acquisition costs strongly predict that acquiring firms hold on to the acquired business rather than divesting it: an interquartile cost increase reduces annual divestiture rates by 8% to 9%. Placebo tests using post-completion market fluctuations find no effect, supporting the sunk cost interpretation. The effect is concentrated in firm-years when the acquiring CEO (who personally incurred the cost) is still in office, and in financially unconstrained firms. Mechanism tests rule out learning, investment budget constraints, and CEO entrenchment as primary explanations; the evidence is most consistent with managerial behavioral frictions generated by sunk cost thinking.

## Core results

Magnitudes and significance are as reported; `\*\*`/`\*\*\*` = 5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Quasi-random acquisition cost increases reduce divestiture rates by ~8%**: the main sunk cost effect, robust to controls | Table III, col. (1), p. 1617 | Cox coefficient on delta-C = -0.065 (z=-2.77`\*\*\*`); interquartile cost increase (1.28 pp of market cap) reduces divestiture rate 8% |
| R2 | **Effect robust to time-varying controls** (acquirer prior-year return, industry distress indicator): coefficient barely changes | Table III, col. (2), p. 1617 | Coefficient -0.068 (z=-2.89`\*\*\*`); with time-varying interactions up to -0.077 (col. 4); interquartile effect 8-9.4% |
| R3 | **Placebo: post-completion market fluctuations do NOT predict divestiture rates**: only pre-completion (sunk) cost variation matters | Table IV (Panel A), p. 1621 | All five placebo coefficients between 0.009 and 0.011, z-stats 0.30-0.40, all insignificant |
| R4 | **Within-divestiture sample (divested-only) replicates**: sunk cost effect holds without case-control matching | Table V, p. 1622 | Coefficient -0.070 to -0.065 (z=-2.29 to -2.50`\*\*`); interquartile effect 8.0-8.1% |
| R5 | **Financial constraints dampen the effect**: distortions ~13% for unconstrained firms, ~6% for constrained | Table VII (Panel A), p. 1626 | Unconstrained: -0.113 to -0.106 (z=-3.42 to -3.94`\*\*\*`); Constrained: -0.047 to -0.046 (insignificant) |
| R6 | **Effect is CEO-specific**: concentrated in firm-years when the acquiring CEO is still at the helm; 30-50% lower once that CEO departs | Table VIII, col. (2), p. 1628 | Acquiring CEO: -0.105 (z=-2.51`\*\*`); New CEO: -0.060 (z=-1.65, insignificant); New CEO base effect +0.575`\*\*\*` |
| R7 | **Concentrated in diversifying acquisitions** (a proxy for inferior deal quality): no significant effect for same-industry deals | Table X, p. 1633 | Diversifying: -0.068 to -0.079 (z=-2.89 to -3.20`\*\*\*`); Same-industry: -0.011 to -0.035 (all insignificant) |

**Overall (paper's conclusion).** Quasi-random increases in acquisition costs cause firms to hold on to acquired businesses longer, consistent with managers taking sunk costs into account when deciding whether to divest. The mechanism appears to be an intrapersonal, CEO-specific behavioral friction: the effect disappears when the CEO who made the costly acquisition is replaced, extending the observation by Weisbach (1995) that divestitures are more likely after CEO changes (here shown to reflect a sunk cost channel rather than general strategy shifts). Financial constraints partially counteract the distortion, consistent with constraints limiting Malmendier and Tate (2005)-style behavioral investment distortions. Same-industry deals are unaffected, whereas diversifying acquisitions (a proxy for inferior deal quality flagged by Kaplan and Weisbach (1992)) concentrate the distortions. These patterns are most consistent with sunk costs generating psychological frictions in managerial decision making (Staw and Hoang (1995)), and are difficult to reconcile with CEO learning, investment budgets, or CEO entrenchment as the primary driver. Contemporaneous work by Cronqvist and Pely (2024) concludes that many divestitures are "corrections of failure," consistent with the efficiency costs of delay documented here.

## Theory / model

The paper has no structural model estimated by moments. It begins with a simple three-period reduced-form conceptual framework (Section I.A, pp. 1599-1601, Figure 1) that formalizes the prediction being tested.

**Setup.** At $$t=0$$, a manager buys an asset at total cost $$\bar{C} = C + \Delta C$$, where $$C$$ is known at the investment decision and $$\Delta C$$ is a mean-zero random variable realized between $$t=0$$ and $$t=1$$. At $$t=1$$, the manager decides to keep or divest the asset. The asset's market price (divestiture value) is $$P$$, the firm-specific interim cash flow (synergy) is $$X$$, and the long-run payoff if kept is $$Z$$. A divestiture yields $$P = Z$$ (competitive buyer market). The manager has potentially nonstandard preferences.

**Standard manager.** For $$\kappa = 0$$, the manager at $$t=1$$ solves:

$$
\max_{d_1 \in \{0,1\}} \; (1-d_1)(X + Z) + d_1 P \tag{Result 1, p. 1601}
$$

where $$d_1 = 1$$ denotes divestiture. The standard manager divests if and only if $$X + Z < P$$. With $$P = Z$$, this reduces to $$X < 0$$: the divestiture decision is independent of the realized cost shock $$\Delta C$$.

**Sunk cost manager.** For $$\kappa > 0$$, the manager incurs a disutility from divesting that is increasing in the total cost $$\bar{C}$$. The manager solves (p. 1600):

$$
\max_{d_1 \in \{0,1\}} \; (1-d_1)(X + Z) + d_1 \!\!\underbrace{\!\!\left(P - \kappa\bar{C}\right)}_{\text{net of sunk cost disutility}} \tag{Result 2, p. 1601}
$$

The sunk cost manager divests if and only if $$X < -\kappa(\bar{C}) = -\kappa(C + \Delta C)$$. A larger realized cost shock $$\Delta C$$ raises the threshold and makes divestiture less likely. This is the testable prediction: the probability of divestiture at $$t=1$$ is decreasing in $$\Delta C$$ for a sunk cost manager ($$\kappa > 0$$) and independent of $$\Delta C$$ for a standard manager ($$\kappa = 0$$).

The Internet Appendix extends the framework to a prospect-theory setting (Kahneman and Tversky (1979), Thaler (1980)) showing that diminishing sensitivity to losses generates the same prediction: a higher cost shock codes as a larger loss domain for the manager, making continued holding relatively more attractive.

## Method

The paper applies three estimators to a hand-collected dataset of Fixed Shares M&A deals.

**Cox (1972) proportional hazards model (main estimator).** The primary specification models the hazard of divestiture as (Section III.E, p. 1616, equation 4):

$$
h(t \mid \mathbf{X}_i) = h_0(t) \exp(\boldsymbol{\delta}' \mathbf{X}_i) \tag{4}
$$

where $$t$$ is survival time (years since acquisition), $$h_0(t)$$ is the unspecified baseline hazard, and $$\mathbf{X}_i$$ includes the main variable of interest $$\Delta C_i$$, deal- and firm-level controls, acquirer and target industry fixed effects, and acquisition year fixed effects. The model treats nondivested acquisitions as right-censored (censoring date: December 15, 2018, or acquirer takeover date). Time-varying covariates are accommodated by reshaping data into one-year-long sub-spells. Proportional hazards assumption is tested via Schoenfeld (1982) residuals; some control variables require time-interaction corrections.

**Logit and stratified hazard models.** Robustness to using a logit model instead of the hazard model (Efron (1988), Jenter and Kanaan (2015)), and to stratified Cox (1972) models. These produce qualitatively identical results (Internet Appendix, Section IV.A).

**Two-stage control function approach.** To address concerns about endogenous acquisition cost changes, the paper also implements a two-stage approach (Wooldridge (2015)): in the first stage, regress the actual cost change on the market-induced component and controls; in the second stage, include the residual from the first stage in the hazard model to control for endogeneity. Results are stable (Internet Appendix Table IA.VII).

This paper builds on `survival-analysis` (the Cox proportional hazards model) and `panel-regression` primitives, and uses a `natural-experiment` identification design: market fluctuations during the binding merger agreement period shift acquisition costs quasi-randomly across deals in the same year, analogous to `instrumental-variables` logic.

## Empirical specifications

**Acquisition cost change construction (pp. 1610-1611, equations 1 and 1').** The endogenous change in acquisition cost induced by the acquirer's own stock price movements is (equation 1):

$$
\Delta C_i^{Acq} = \Delta R_i^{Acq} \times \%\text{stock}_i \times \frac{\text{Deal Value}_i}{\text{Market Cap}_i^{Acq}} \tag{1}
$$

where $$\%\text{stock}_i \in (0,1]$$ is the fraction of merger consideration paid in stock, relative deal value is the deal value at agreement relative to the acquirer's pre-announcement market capitalization, and $$\Delta R_i^{Acq} = \sum_{t=\tau_1+2}^{\tau_2} R_{i,t}^{Acq}$$ is the cumulative daily acquirer return during the transaction period (merger agreement date $$\tau_1$$ to completion date $$\tau_2$$).

To isolate exogenous variation, the acquirer's daily return is replaced by the daily market return, adjusted for expected market appreciation and the acquirer's industry beta (equation 2'):

$$
\Delta R_i = \sum_{t=\tau_1+2}^{\tau_2} \hat{\beta}_{i,\tau_1}\!\left(R_t^{Mkt} - E_{\tau_1}\!\left[R_t^{Mkt}\right]\right) \tag{2'}
$$

The market-driven cost change used as the main variable of interest is then (equation 1'):

$$
\Delta C_i = \Delta R_i \times \%\text{stock}_i \times \frac{\text{Deal Value}_i}{\text{Market Cap}_i^{Acq}} \tag{1'}
$$

**Main estimating equation (p. 1611, equation 3):**

$$
\Pr(\text{Divestiture}_{i,t}) = \alpha + \kappa\,\Delta C_i + \boldsymbol{\delta}'\mathbf{X}_{i,t} + \nu_{j(\text{Acq})} + \nu_{j(\text{Tar})} + \mu_{t_0} + \varepsilon_{i,t} \tag{3}
$$

where $$i$$ denotes an acquisition, $$t$$ is years elapsed since acquisition, $$t_0$$ is the acquisition (calendar) year, $$\text{Divestiture}_{i,t}$$ is an indicator for the year of divestiture, $$\Delta C_i$$ is the market-induced cost change (equation 1'), $$\nu_{j(\text{Acq})}$$ and $$\nu_{j(\text{Tar})}$$ are acquirer and target industry fixed effects, and $$\mu_{t_0}$$ are acquisition year fixed effects. The null hypothesis (no sunk cost effects) is $$\kappa = 0$$. Standard errors are clustered by quarter of acquisition (treatment is assigned by market fluctuations between merger agreement and completion, which cluster within calendar quarters).

**Identifying variation.** Market fluctuations between merger agreement and completion shift acquisition costs across deals in the same year (same $$\mu_{t_0}$$) but in different calendar quarters: Table II confirms that the market return during the transaction period (i) strongly predicts firm returns (Panel A, F-statistic above 70), and (ii) is unpredictable from deal and firm characteristics (Panel B, joint F-statistic for 8 covariates = 0.56, p-value = 0.81). This validates the "as good as randomly assigned" assumption conditional on acquisition year.

**Key robustness tests.** Placebo using post-completion market fluctuations (Table IV, pp. 1621-1622): coefficients insignificant across all 10 specifications. Fixed Dollar placebo (Table VI, p. 1624): the hypothetical cost change for Fixed Dollar deals (which do not have exchange-ratio-induced cost variation) is insignificant while the actual Fixed Shares effect remains. Results are robust to financial constraint controls, alternative clustering, alternative time specifications, and gradual removal of high-withdrawal-probability observations.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| SDC Platinum M&A database | Starting universe of domestic acquisitions by US public acquirers 1980-2016; deal characteristics and divestiture flags | [SDC Platinum](/wiki/commercial/sdc-platinum/) (licensed) |
| Nexis (LexisNexis) news search | Divestiture identification: systematic search for newspaper articles and news wires for acquisitions not flagged by SDC | No page yet |
| SEC EDGAR filings (10-K, 10-Q, 8-K, S-4, Exhibit 21) | Hand-collected merger agreement terms (exchange ratio type, deal terms); divestiture verification | [SEC EDGAR](/wiki/datasets/edgar/) |
| CRSP monthly and daily returns | Acquirer stock returns during transaction period; acquirer market capitalization; beta estimation; control variables | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| Compustat | Acquirer financial characteristics; industry market-to-book; leverage; financial constraint construction | [WRDS / Compustat](/wiki/commercial/wrds/) (licensed) |
| Execucomp | CEO tenure and compensation data for ~50% of acquirers; CEO change dates | [WRDS / Execucomp](/wiki/commercial/wrds/) (licensed) |
| SEC filings, BoardEx, Bloomberg, Capital IQ, Who's Who | Hand-collected CEO education and biographical data for CEO sophistication tests | No page yet |

Sample: US public acquirers, acquisitions 1980-2016, divestitures tracked through December 2018. Main sample: 558 Fixed Shares acquisitions (279 divested), 4,461 firm-year observations.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13430) if you are: replicating the sunk cost hazard model (the Internet Appendix contains the full identification validity tests, Schoenfeld residual analysis, and the two-stage control function approach); designing an identification strategy for sunk cost effects in other investment contexts (R&D, VC, financial intermediation); studying CEO-level behavioral mechanisms in M&A decision making; or examining efficiency costs of delayed divestitures (Section V.D, counterfactual analysis). The locators above point to the exact tables.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(3), June 2025. This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The article is paywalled (Wiley/American Finance Association). Extract-only; no PDF hosted.

> Guenzel, Marius. "In Too Deep: The Effect of Sunk Costs on Corporate Investment."
> *The Journal of Finance* 80, no. 3 (June 2025): 1593-1646.
> DOI: [10.1111/jofi.13430](https://doi.org/10.1111/jofi.13430).
> © 2025 the American Finance Association.
> Distilled by the Institute for Automated Research (extract-only, not reproduced).


==============================================================================
# Thirty Years of Change: Guernsey, Guo, Liu & Serfling (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/guernsey-thirty-years-change-evolution-2025/
# Distilled: Using a new machine-learning-constructed dataset covering nearly all U.S. public firms from 1991 to 2020, this paper documents that classified (staggered) board usage has not declined overall; rather, its life-cycle dynamics have changed substantially by decade and IPO cohort, driven by falling collective-action costs and rising innovation-related investment. J. Finance 2025, paywalled. Seven core results with source locators, datasets used, the empirical design, and the ML data-construction method.
# Tags: paper-summary, corporate-governance, takeover-defenses, classified-boards
==============================================================================

**What this is.** The paper's core findings, the ML data-construction method, and the empirical design: enough to know what it found and how, without reading all 50 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1111/jofi.13485).

## TL;DR

The paper constructs a novel classified board dataset covering nearly all U.S. public firms from 1991 to 2020 using a Random Forest (RF) Classifier applied to DEF 14A proxy filings from SEC EDGAR, yielding roughly three times more firm-year observations than commercial databases like ISS. Using this dataset, the paper documents three decades of change in how firms use classified (staggered) boards over their life cycles. In the 1990s, classified board usage was sticky across the life cycle, consistent with high collective-action costs preventing optimal adjustment. Since 2001, and especially in the 2010s, firms have become increasingly likely to declassify their boards as they mature. The value-reversal pattern linking classified boards to firm value (positive for young firms, negative for mature ones) that was strong in the 1990s and 2000s has largely disappeared in the 2010s, suggesting classified board adjustments have become more optimal. Mechanisms investigated include rising R&D intensity among young firms, growing passive institutional ownership, falling bid-ask spreads, and increased hedge fund activism.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Classified board usage has NOT declined overall: non-S&P 1500 firms increased from 42% to 51.8% (1991-2020), even as S&P 500 firms fell from 60% to 12.2% | Figures 1A2, p. 2985; text p. 2986 | Non-S&P 1500: +9.8pp over 30 years; S&P 500: -47.8pp; divergence masked in ISS-only samples |
| R2 | Life-cycle slope of classified board usage steepened across decades: in the 1990s it was essentially flat; in the 2011-2020 decade it dropped from 72.8% to 33.0% across age groups | Table II (col. 1,4,8), pp. 2989-2990; Figure 2, p. 2988 | Full-sample age >=16 coefficient: -8.6pp\*\*\*; 2011-2020 IPO cohort age >=16 coefficient: -5.8pp\*\*\* |
| R3 | Board declassification rates have increased substantially over time: mature firms (age >= 16) declassify at 3.5% per year in the 2010s vs 0.6-1.1% in the 1990s | Figure 3, p. 2992 | Approximately 3-5x higher annual declassification rates for mature firms in the 2010s vs 1990s |
| R4 | Being added to the S&P 1500 caused declassification only in the 2011-2020 decade: propensity-matched DiD shows significant negative post-treatment CB coefficients for this cohort but not the prior two | Figure 8, p. 3006 | Statistically significant negative coefficients in t+1 to t+3 for 2011-2020 cohort (90% CI excludes zero); no effect in 1991-2000 or 2001-2010 |
| R5 | The classified-board value-reversal pattern (positive for young firms, negative for mature ones) was present in the 1990s and 2000s but largely disappeared in the 2010s | Table V, pp. 3009-3010 | Age 0: CB = +0.161\*\*\* (full sample); age >= 2: CB negative and significant (up to t=-3.87). In 2010s, only age 3-4 remains significant (-0.210\*\*); all older age groups insignificant |
| R6 | Market response to declassification announcements was significantly positive in 1999-2000 (CAR = 5.9%-6.1%) but has been statistically insignificant across the rest of the sample period | Figure 9, p. 3013 | Full-sample CAR[-2,+2]: positive but p-value = 0.588; 1999-2000 CARs significantly positive and largest in any two-year window |
| R7 | Declassifying the board improved operating ROA in the 1990s (especially for mature firms) but not in the 2000s or 2010s | Figure 10, p. 3014 | 1991-2000 cohort: positive post-treatment ROA shifts visible in matched DiD; 2001-2010 and 2011-2020: coefficients cluster near zero with wide confidence intervals |

**Overall (paper's conclusion).** Contrary to conventional wisdom, classified boards are not going extinct across the U.S. corporate sector; the decline is concentrated in S&P 1500 firms. The life-cycle dynamics of classified board usage have changed profoundly over three decades: from high stickiness in the 1990s to more dynamic adjustment in the 2010s. Evidence on mechanisms suggests that rising R&D intensity among young firms, growing passive institutional ownership that reduced collective-action costs, falling bid-ask spreads, and increased shareholder activism all contributed to this shift. The disappearance of the value-reversal pattern in the 2010s implies that classified board adjustment has become more optimal, reducing the friction that caused firms to maintain this defense too long in earlier decades.

## Theory / model

The paper has no formal theoretical model; instead it organizes its analysis around the cost-benefit framework of Johnson, Karpoff, and Yi (2022) and the related literature on takeover defenses. The tested theoretical propositions are:

**Hypothesis 1 (life-cycle costs and benefits).** A classified board can be value-enhancing for young firms engaged in long-term, information-intensive investments (R&D, intangibles) because it protects managers from short-termist takeover pressure. As firms mature and reduce such investments, the costs of a classified board (managerial entrenchment, reduced monitoring) outweigh its benefits, implying the optimal strategy is to declassify (pp. 2974, 2993).

**Hypothesis 2 (collective-action costs).** High collective-action costs among dispersed shareholders (free-rider problems, heterogeneous time horizons, information asymmetry) prevent optimal declassification even when the defense is value-destroying. Falling collective-action costs over time, driven by rising passive institutional ownership, reduced bid-ask spreads, and increased activist pressure, should enable more dynamic adjustment (pp. 2974-2975, 2993-3001).

**Hypothesis 3 (value reversal).** Because collective-action frictions prevented optimal adjustment in the 1990s and 2000s, the relation between classified boards and firm value should be positive for young firms and negative for mature ones (the value-reversal pattern). As frictions fell in the 2010s, enabling more optimal adjustment, this reversal should diminish (p. 2976).

The paper explicitly states that it cannot establish causality between trends in the mechanism variables and trends in classified board usage, because no valid instrument or natural experiment cleanly identifies the channel (p. 2975).

This paper extends and relates to several prior studies. Cremers, Litov, and Sepe (2017) documented declining classified board usage in S&P 1500 firms; the present paper shows that trend does not hold for the broader population. Field and Lowry (2022) documented increasing classified board adoption at IPO; the present paper adds decade and cohort decomposition of that trend. The RF classifier builds on and formalizes the approach introduced in Guernsey, Sepe, and Serfling (2022).

## Method

**RF Classifier for classified board status.** The core methodological contribution is a five-step machine-learning pipeline that uses text from DEF 14A proxy filings in SEC EDGAR to classify the classified board status of all U.S. public firms from 1991 to 2020, extending the ISS database (which covers only S&P 1500 firms) to the near-universe of public firms (pp. 2978-2982).

The steps are:

1. Obtain all DEF 14A filings from SEC EDGAR with mentions of "elect" or "stagger" (179,942 unique CIK-FDATE pairs; reduced to 110,511 firm-year observations after cleaning).
2. Apply two parallel text-extraction strategies: (i) locate 150-word windows following "Proposal 1. Election of Directors" headings (covers 85.3% of filings), and (ii) keyword search for "class" and "term" in context, finding classified board language in 66.0% of the sample.
3. Build a training sample by merging with ISS data (39,998 DEF 14A filings that match ISS; 80% training, 20% test).
4. Convert text to unigrams and bigrams, apply Porter stemming, retain phrases appearing in at least 1,000 filings (corpus of 2,287 phrases), and run the RF Classifier (tuning: number of trees, max depth, min samples per split/leaf, max features per tree, cross-validated in training sample).
5. Apply the best RF model to predict classified board status for all filings; extend to microfiche-era filings (pre-EDGAR, 1991-1995) via hand collection.

Out-of-sample prediction accuracy is 97.3% (1.5% false negatives, 1.2% false positives). Compared to a refined keyword search, the RF Classifier reduces total error by an 81% improvement (error rate 1.6% vs 8.7% for the refined keyword approach). The 25 most important predictor phrases (accounting for 37.6% of variable importance) include "three year," "three class," and "divid" (from the Porter stemmer, stemming "divided" as in "directors are divided into three classes").

**Empirical designs.** The paper uses several complementary designs:

- OLS panel regressions with firm-age dummies, firm-level controls, industry fixed effects, and/or firm fixed effects (Figures 1-4, Tables II-IV, following Johnson, Karpoff, and Yi (2022)).
- A stacked propensity-score-matched DiD design around S&P 1500 index additions (Figure 8, p. 3006), where matched treatment/control pairs are required to have a classified board in the three years before treatment.
- An OLS firm-value (Tobin's Q) regression with firm controls and industry-year fixed effects, estimated separately by age group and decade (Table V).
- A short-run event study of announcement CARs[-2,+2] around shareholder meeting votes to approve declassification, using the CRSP equal-weighted index as the benchmark (Figure 9, p. 3013).
- A stacked propensity-score-matched DiD for post-declassification operating ROA, using firm-treatment cohort and cohort-year fixed effects (Figure 10).

## Empirical specifications

**Life-cycle OLS (main design, Figures 1-4, Table II).** For each firm $$i$$ in year $$t$$, the primary estimating equation is (p. 2985):

$$
CB_{it} = \sum_{t=1991}^{2020} \beta_t \text{Year}_t \left(\text{or } \sum_{a=0}^{\geq 16} \omega_a \text{Age}_{it}\right) + \Gamma X_{it} + \eta_k + \gamma_i + \varepsilon_{it} \tag{1}
$$

where $$CB_{it}$$ equals one if firm $$i$$ has a classified board in year $$t$$; $$\text{Year}_t$$ is a year-group indicator; $$\text{Age}_{it}$$ is an age-group indicator (years since IPO, grouped 0-1, 2-3, 4-5, 6-7, 8-10, 11-15, $$\geq 16$$); $$X_{it}$$ includes $$\text{Ln}(\text{MVE}_{t-1})$$, $$\text{IO}_t$$, $$\text{Delaware}_t$$, and $$\text{S\&P 1500}_t$$; $$\eta_k$$ are two-digit SIC industry fixed effects; and $$\gamma_i$$ are firm fixed effects. Regressions for Figures 1-4 are estimated separately for each decade or IPO cohort. Standard errors are clustered by firm throughout. The dependent variable for declassification regressions (Figure 3) is $$\text{Declass}_{it}$$, which equals one if the firm had a classified board in $$t-1$$ but not in $$t$$.

**Mechanism regressions (Tables III-IV, Figures 6-7).** Same specification as equation (1) but with the dependent variable replaced by passive institutional ownership (Big Three IO, Quasi IO), bid-ask spread ($$\text{Ln}(\text{Bid-Ask Spread})_{it}$$), hedge fund activism, or the likelihood of receiving a shareholder proposal to declassify. Table III adds interactions of age groups with five-year period dummies:

$$
CB_{it} = \sum_{t=1991}^{2020} \left[\beta_t (\text{IndexIO}_{it} \times \text{Year}_t) + v_t (\text{NonIndexIO}_{it} \times \text{Year}_t)\right] + \Gamma X_{it} + \gamma_i + \tau_t + \varepsilon_{it} \tag{2}
$$

where $$\text{IndexIO}_{it}$$ is Big Three or quasi-indexer ownership (normalized by one standard deviation).

**S&P 1500 DiD (Figure 8).** A stacked DiD with propensity-score matching (logit on $$\text{Ln}(\text{MVE})$$, $$\text{IO}$$, $$\text{Delaware}$$; caliper 0.025; matched within two-digit SIC and fiscal year). For treatment cohort $$j$$ and event time $$t$$:

$$
CB_{ijt} = \sum_{t=-3}^{3} \beta_t (\text{Treat}_{ij} \times \text{Time}_{jt}) + \gamma_{ij} + \delta_{jt} + \Gamma X_{ij} \times \text{Post}_{jt} + \varepsilon_{ijt} \tag{3}
$$

where $$\text{Time}_{jt-1}$$ is the excluded base year, $$\gamma_{ij}$$ are firm-treatment-cohort fixed effects, and $$\delta_{jt}$$ are treatment-cohort-year fixed effects. The matched sample has 624 treatment and 1,159 control firms.

**Tobin's Q regressions (Table V).** OLS with the dependent variable $$\text{Tobin's Q}_{it} = (\text{prcc\_f} \times \text{csho} + \text{at} - \text{ceq}) / \text{at}$$, estimated separately for each of seven age groups (0, 1, 2, 3-4, 5-6, 7-9, $$\geq 10$$) and four sample periods (full, 1991-2000, 2001-2010, 2011-2020). Controls and industry-year fixed effects are included; standard errors are clustered by firm.

**Event study CARs (Figure 9).** Market model estimated over the [-210,-11] trading-day window before each declassification meeting date (857 meetings, 1995-2020), using the CRSP equal-weighted index. CARs computed over [-2,+2]. Coefficients regress CAR on year dummies with and without firm-level controls and SIC2 fixed effects; 90% confidence intervals based on firm-clustered standard errors.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CRSP monthly stock returns | Market value of equity (MVE), stock return data, event study benchmark index | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| Compustat annual fundamentals | Book value of assets, Tobin's Q construction, firm characteristics | [WRDS / Compustat](/wiki/commercial/wrds/) (licensed) |
| SEC EDGAR DEF 14A filings | Text input to the RF Classifier; classified board status for all public firms | [SEC EDGAR](/wiki/datasets/edgar/) |
| ISS (RiskMetrics / IRRC) governance database | Training labels for the RF Classifier; classified board status for S&P 1500 firms | No page yet |
| Thomson Reuters 13-F institutional holdings | Institutional ownership (IO), Big Three IO, quasi-indexer IO | [Thomson 13F (s34)](/wiki/commercial/thomson-13f/) (licensed) |
| Voting Analytics (ISS / Diligent) | Shareholder proposals to declassify boards, 2003-2020 | No page yet |
| Hedge fund activism data (Brav, Jiang & Kim) | Hedge fund targeting by firm-year, 1994-2018 | No page yet |
| Wall Street Journal articles | Media attention to corporate governance; keyword search, 1998-2020 | No page yet |

Sample: 137,032 firm-year observations, 1991-2020 (annual frequency). Classified board sample: 66,262 firm-years. Unitary board sample: 70,770 firm-years. The dataset introduced by the paper (classified board status for nearly all public firms via RF Classifier + hand-collection) covers roughly 3x the firm-year observations of commercial databases.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13485) if you are:
studying the evolution of corporate governance defenses over long horizons across the full population of public firms (not just index firms); interested in the RF Classifier methodology for extending ISS-type governance data to the full public-firm universe; testing the value-reversal hypothesis of Johnson, Karpoff, and Yi (2022) with a more comprehensive sample; or investigating the role of passive institutional ownership and shareholder activism in governance change. Table II (p. 2990) contains the primary life-cycle regressions; Table V (pp. 3009-3010) contains the firm-value results; Figures 1-3 (pp. 2985-2992) display the decade-by-decade and cohort-by-cohort patterns.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(5). The licence is Wiley publisher terms (not CC); this page is extract-only. This distillation was extracted by an LLM on 2026-06-05 and is **not human-verified or independently reproduced**.

> Guernsey, Scott, Feng Guo, Tingting Liu, and Matthew Serfling. "Thirty Years of Change: The Evolution of Classified Boards." *The Journal of Finance* 80, no. 5 (October 2025): 2971–3020. DOI: 10.1111/jofi.13485.


==============================================================================
# Collusion in Brokered Markets: Hatfield, Kominers & Lowery (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/hatfield-collusion-brokered-markets-2025/
# Distilled: Models collusion in brokered markets (e.g., US residential real estate) as a repeated extensive-form game, showing that brokers can sustain prices substantially above marginal cost even with many independent agents and easy entry, by refusing to work with price deviators within-period. J. Finance 2025, paywalled. Six core results with source locators, the model, and the equilibrium construction.
# Tags: paper-summary, market-microstructure, real-estate, collusion, market-design
==============================================================================

**What this is.** The paper's core model, theorems, and policy implications: enough to know what was proved and how, without reading the full 46 pages. To replicate or extend it, read the original at [doi.org/10.1111/jofi.13432](https://doi.org/10.1111/jofi.13432).

## TL;DR

The paper explains how the U.S. residential real estate brokerage industry can maintain commissions roughly at 6% of the transaction value, far above competitive cost, even though the industry has many independent agents and easy entry. This is the "enormous puzzle" noted by Hsieh and Moretti (2003). Modeled as a repeated extensive-form game, brokers can sustain collusion by refusing to work with any agent who cuts prices within the current period. Because buyers and sellers expect a price deviator to be excluded from the network of other agents, they demand a large discount to work with that agent, making a small price cut unprofitable. This mechanism works even as the number of agents grows arbitrarily large, unlike standard repeated-game models where collusion breaks down as the number of firms increases. The paper also shows that rebate bans and agent specialization supported by agency fees expand the scope for collusion, while eliminating agency fees ("decoupling") can reduce it.

## Core results

Magnitudes are as reported in the theorems and figures; all results hold for discount factor $$\delta \geq \frac{1}{2}$$.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Highest sustainable prices remain bounded away from marginal cost as market concentration approaches zero; seller price equals full seller surplus | Theorem 1, Figure 1, pp. 1427-1428 | $$p^*_S = v_S$$ for all $$\alpha$$; $$\lim_{\alpha \to 0}(p^*_B + p^*_S) = (v_B + v_S)(1 - \kappa_S) > 0$$ |
| R2 | Simple exclusion equilibrium (prices never change after deviations) achieves same prices as optimal equilibrium as concentration goes to zero | Theorem 2, Corollary 1, Figure 3, pp. 1440-1441 | $$q^*_B = v_B - \kappa_S(v_B + v_S)$$, $$q^*_S = v_S$$; $$\lim_{\alpha \to 0}(p^*_B, p^*_S) = (q^*_B, q^*_S)$$ |
| R3 | Rebate bans raise sustainable collusive prices whenever market concentration is sufficiently low | Theorem 3, Figure 4, p. 1446 | With rebate ban: $$\lim_{\alpha \to 0}(p^*_B + p^*_S) = v_B(1-\kappa_S) + v_S$$; strictly higher than without ban when $$\hat\alpha < \kappa_S v_S / (v_S + (1-\kappa_S)v_B)$$ |
| R4 | Agent specialization with agency fees raises industry profits above the symmetric baseline | Theorem 4, Figure 5, p. 1449 | $$\lim_{\sigma \to 0}(p^*_B + p^*_S) = (v_B + v_S)(1-\kappa_S) + \kappa_S c$$, strictly above baseline |
| R5 | Eliminating agency fees when seller-proficient agents represent buyers weakly reduces sustainable profits | Theorem 5, Figure 6, p. 1451 | Industry revenue weakly below the agency-fee equilibrium; limit revenue same only as $$\sigma \to 0$$ |
| R6 | When buyer valuations are sufficiently low relative to sellers, eliminating agency fees reduces both buyer and seller prices; decoupling weakens collusion | Theorem 6, Corollary 2, pp. 1452-1453 | When $$v_B (1-\kappa_S)/\kappa_S < c < v_S$$: buyer price = 0, seller price below $$v_S$$; total revenue strictly below agency-fee equilibrium |

**Overall (paper's conclusion).** Brokered markets are structurally prone to collusion because the brokerage requirement means that each broker must cooperate with other brokers to complete transactions. This gives incumbents the power to punish price cutters immediately and in-period, so that the gains from deviating vanish even as the number of agents grows large. The result persists under simple exclusion strategies requiring minimal coordination, making it robust to the practical difficulty of fine-tuning punishment across many agents.

## Theory / model

The paper models a brokered market as a repeated extensive-form game played over discrete infinite time with a common discount factor $$\delta \in (0, 1)$$ (p. 1422). There is a finite set of agents $$A$$ with market concentration $$\alpha \equiv \frac{1}{|A|}$$. In each period, a continuum of short-lived buyers $$B_t$$ and sellers $$S_t$$ arrive. Each agent has a buyer capacity $$\kappa_B$$ and a seller capacity $$\kappa_S$$, with $$\kappa_S \leq \kappa_B < \frac{1}{2}$$.

The stage game has four steps (pp. 1422-1424):

1. Each agent $$a$$ posts a buyer price $$p^a_{B,t} \in \mathbb{R}$$ and a seller price $$p^a_{S,t} \in \mathbb{R}$$, publicly observed.
2. Buyers and sellers rank agents and are assigned via random rationing, so no agent represents more than $$\kappa_B$$ buyers or $$\kappa_S$$ sellers.
3. Each agent $$a$$ invites other agents $$\bar{a}$$, including a contingent agency fee $$f^{a \leftarrow \bar{a}}_t \in \mathbb{R}$$ per transaction paid from $$\bar{a}$$'s seller-side commission to $$a$$.
4. Each agent accepts or rejects invitations; the resulting directed network of accepted invitations determines which buyer-seller pairs can be matched.

The agent payoff per period is (pp. 1424-1425):

$$
|\mathbf{B}_t(a)| \sum_{\bar{a} \in A^{a\Rightarrow}_t} \left( |\mathbf{S}_t(\bar{a})| (p^a_{B,t} + f^{a \leftarrow \bar{a}}_t) \right) + |\mathbf{S}_t(a)| \sum_{\bar{a} \in A^{a\leftarrow}_t} \left( |\mathbf{B}_t(\bar{a})| (p^a_{S,t} - f^{\bar{a} \leftarrow a}_t) \right)
$$

which simplifies, as agents in the cooperation phase split profits evenly, to (p. 1425):

$$
|\mathbf{B}_t(a)| \sum_{\bar{a} \in A^{a\Rightarrow}} \left( |\mathbf{S}_t(\bar{a})| (p^a_{B,t} + p^a_{S,t}) \right) + |\mathbf{S}_t(a)| \sum_{\bar{a} \in A^{a\leftarrow}} \left( |\mathbf{B}_t(\bar{a})| (p^a_{S,t} - f^{\bar{a}\leftarrow a}_t) \right).
$$

A key equilibrium refinement is (buyer-and-seller) coordination-proofness: no positive-measure subset of buyers and sellers can jointly deviate to improve all their payoffs (p. 1426). This rules out coordination-failure equilibria in which buyers and sellers refuse to sign up with any agent.

**Identification of the key friction.** The model's main departure from standard Bertrand models is two-sided intermediation: to facilitate a transaction, both the buyer's agent and the seller's agent must agree to work together. A price deviator who cuts prices attracts buyers or sellers away from other agents, but those other agents can refuse network links to the deviator, reducing the probability of a transaction for any buyer or seller who signed up with the deviator. As a result, buyers and sellers demand a large discount from a price deviator, not just a small epsilon discount (p. 1429).

**Theorem 1 (Optimal Collusion, p. 1427).** For $$\delta \geq \frac{1}{2}$$, the highest sustainable industry profits are achieved with prices:

$$
p^*_B = \begin{cases} v_B & \alpha \geq (1-\delta)\kappa_B\kappa_S \\ \dfrac{(1-\delta)\kappa_B(v_B - \kappa_S(v_B+v_S)) + \alpha v_S}{(1-\delta)\kappa_B - \alpha} & \alpha \leq (1-\delta)\kappa_B\kappa_S \end{cases} \tag{1}
$$

$$
p^*_S = v_S. \tag{2}
$$

Moreover, $$\lim_{\alpha \to 0}(p^*_B + p^*_S) = (v_B + v_S)(1 - \kappa_S) > 0$$.

The seller price always equals the full seller surplus $$v_S$$; buyers receive a price below $$v_B$$ because, given that each buyer accesses only $$\kappa_S$$ sellers through a deviating agent, cutting buyer prices is more effective at deterring entry than cutting seller prices.

## Method

The equilibrium is constructed in three phases: a cooperation phase, a $$\hat{a}$$-collusive punishment phase, and a $$\{\hat{a}, a\}$$-collusive punishment phase (p. 1432, Figure 2, p. 1434). This multi-phase construction is necessary because, as Mailath, Nocke, and White (2017) show, simple penal codes sufficient for Abreu (1988)-style infinitely repeated normal-form games are not sufficient in repeated extensive-form games: agents who comply with punishing a deviator must themselves be rewarded in the punishment phase. The paper builds directly on Hatfield et al. (2020), who analyze collusion with syndication using a similar extensive-form repeated-game framework.

**Key quantity: buyer and seller deviation prices.** In the cooperation phase, buyers and sellers are willing to work with a price deviator $$\hat{a}$$ only if the prices $$(\hat{p}^{\hat{a}}_B, \hat{p}^{\hat{a}}_S)$$ satisfy both (p. 1431, eq. 4):

$$
(v_B - \hat{p}^{\hat{a}}_B)\kappa_S \geq v_B - p^*_B \qquad \text{and} \qquad (v_S - \hat{p}^{\hat{a}}_S)\kappa_B \geq v_S - p^*_S.
$$

The highest prices at which buyers and sellers will work with a deviator are therefore (eq. 6, p. 1431):

$$
p^\circ_B = v_B - \frac{1}{\kappa_S}(v_B - p^*_B) \qquad \text{and} \qquad p^\circ_S = v_S - \frac{1}{\kappa_B}(v_S - p^*_S). \tag{6}
$$

Because $$\kappa_S < 1$$, these deviation prices are substantially below $$p^*_B$$ and $$p^*_S$$: the agent must cut prices far enough that buyers and sellers prefer working with a lower-quality network.

**Incentive constraint for non-deviators.** An agent $$a \neq \hat{a}$$ is willing to exclude a price deviator if the discounted future profits from adhering exceed the current gain from working with the deviator (eq. 10-12, pp. 1437-1438):

$$
\frac{\delta}{1-\delta}\frac{\alpha}{1-\alpha}(q^*_B + q^*_S) \geq \frac{\alpha}{1-\alpha}\left[(1-\kappa_S)\kappa_B(p^{\hat{a}}_B + p^*_S) + (1-\kappa_B)\kappa_S(p^*_B + p^{\hat{a}}_S)\right].
$$

This simplifies to the condition $$\delta/(1-\delta) \geq \kappa_B + \kappa_S$$, which holds as long as $$\delta \geq \frac{1}{2}$$. The key observation is that both current profits from working with $$\hat{a}$$ and future profits from adherence are proportional to $$\alpha$$, so the inequality is independent of market concentration.

**Theorem 2 (Exclusion Equilibrium, p. 1440).** For $$\delta \geq \frac{1}{2}$$, there exists an exclusion equilibrium (prices never change after deviations) with prices:

$$
q^*_B = v_B - \kappa_S(v_B + v_S), \qquad q^*_S = v_S. \tag{22, 23}
$$

**Theorem 3 (Rebate Ban, p. 1446).** With a rebate ban (constraint $$p^a_{B,t} \geq 0$$ for all $$a, t$$), for $$\delta \geq \frac{1}{2}$$:

$$
p^*_B = \begin{cases} v_B & \hat\alpha \geq \kappa_S \\ \dfrac{v_B - \kappa_S(v_B + v_S) + \hat\alpha v_S}{1 - \hat\alpha} & \hat\alpha \in \left[\kappa_S \frac{v_S}{v_S + (1-\kappa_S)v_B},\, \kappa_S\right] \\ v_B(1-\kappa_S) & \hat\alpha \leq \kappa_S \frac{v_S}{v_S + (1-\kappa_S)v_B} \end{cases}, \quad p^*_S = v_S, \tag{24, 25}
$$

where $$\hat\alpha = \frac{\alpha}{(1-\delta)\kappa_B}$$. Profits are strictly higher than without a rebate ban whenever $$\hat\alpha < \kappa_S v_S/(v_S + (1-\kappa_S)v_B)$$.

## Empirical specifications

This is a pure-theory paper. There are no regression specifications, data, or empirical tests. The paper provides calibrated numerical examples for Figures 1-7 using parameter values such as $$\delta = 3/4$$, $$v_B = 3$$, $$v_S = 5$$, $$\kappa_B = 1/5$$, $$\kappa_S = 1/6$$, to illustrate how prices vary with market concentration. These figures confirm that buyer prices can be negative for sufficiently low market concentration and that seller prices remain at $$v_S$$ throughout.

The paper applies the model to several policy-relevant cases:

- **Rebate bans** (§III.A): Modeled as the constraint $$p^a_{B,t} \geq 0$$; shown to raise collusive profits for low concentration (Theorem 3). Han and Hong (2011) had argued rebate bans are anticompetitive; the model confirms and formalizes this. Christie and Schultz (1994) documented analogous in-period punishments by NASDAQ market makers via odd-eighth avoidance.
- **Agent specialization and agency fees** (§III.B): A model with buyer-exclusive agents $$A_B$$ and seller-proficient agents $$A_S$$, where seller-proficient agents incur cost $$c \leq v_B$$ to represent buyers. Theorems 4-6 characterize optimal prices. Barwick (2018) proposed eliminating agency fees; Corollary 2 formalizes when this lowers collusive prices.
- **For-sale-by-owner and buyer self-representation** (§III.C): Shown not to resolve collusion because self-representing sellers/buyers can also be excluded by incumbent agents.
- **iBuyers** (§III.D): Predicted to offer the same agency fees as traditional brokers to avoid ostracism (consistent with observed iBuyer behavior).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| U.S. residential real estate industry (stylized facts: 6% commission, observed steering behavior) | Motivating application and empirical context for the model | No page yet |

This is primarily a theory paper. No quantitative datasets are used; all results are derived from the formal model. Motivating facts are sourced from prior empirical work (Hsieh and Moretti (2003), Federal Trade Commission (1983)) and DOJ/FTC reports.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13432) if you are: (i) working on the theory of collusion in intermediated markets and need the proofs in Appendices B-C and the Internet Appendix; (ii) analyzing policy proposals for real estate brokerage (rebate bans, decoupling, MLS reform) and want the formal comparative statics; (iii) extending the model to finite buyers/sellers or to settings where price observability is imperfect (§II.C.4); or (iv) studying other two-sided intermediated markets (municipal bonds, Nasdaq dealer markets, venture capital) where analogous exclusion mechanisms may apply.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(3), June 2025. DOI: [10.1111/jofi.13432](https://doi.org/10.1111/jofi.13432). This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The article is paywalled; this page reproduces only short extracts for scholarly commentary under fair use. No PDF is hosted.

> Hatfield, John William, Scott Duke Kominers, and Richard Lowery.
> "Collusion in Brokered Markets."
> *The Journal of Finance* 80, no. 3 (June 2025): 1417-1462.
> DOI: 10.1111/jofi.13432. © 2025 the American Finance Association.
> Extract-only; all rights reserved by the publisher.


==============================================================================
# War Discourse and the Cross Section: Hirshleifer, Mai & Pukthuanthong (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/hirshleifer-war-discourse-cross-section-2025/
# Distilled: a war-discourse factor (WarFac) derived from 7 million New York Times articles via semisupervised topic modelling predicts the cross section of stock returns with a significant, negative return premium across six broad sets of test assets. J. Finance 2025, paywalled. Eight core results with source locators, datasets used, the model, and the method with defining equations.
# Tags: paper-summary, asset-pricing, anomalies, text-as-data, factors, disaster-risk, return-predictability, fama-macbeth, portfolio-sort, peer-reviewed, unreplicated, data:nyt-news, data:wrds, data:ken-french, data:open-source-asset-pricing
==============================================================================

**What this is.** The paper's core results, datasets, and theory: enough
to know what it found without reading all 49 pages. To replicate or extend
it, consult the [original](https://doi.org/10.1111/jofi.13482) (paywalled)
or request data and code from the authors.

## TL;DR

Using 7,000,000 *New York Times* articles spanning 1871 to 2019, the paper
builds a monthly war-discourse index (*War*) via a semisupervised topic model
(sLDA, one seed word: *war*), then defines the war factor WarFac as the
AR(1) innovation in *War*. Loadings on WarFac significantly and negatively
predict expected returns across six sets of test assets covering up to 4,964
portfolios (138 HXZ long-short anomalies, 1,372 HXZ single-sorted, 904 CZ
single-sorted, 360 ML-based nonlinear, and 128 and 2,190 own-constructed
portfolios), with a monthly return premium ranging from about -0.66% to
-3.32% per month. WarFac is incremental to the Fama-French six-factor model
and to news-based uncertainty indexes (NVIX, GPR). A mimicking portfolio
(WMP) earns an annualised Sharpe ratio of 1.73 and passes the Pukthuanthong
et al. (2019) factor-identification protocol and the Giglio-Xiu three-pass test.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%.
Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | WarFac commands a **significant negative return premium** on 138 HXZ long-short anomaly portfolios | Table I Panel A, Table II Panel A, p. 3607–3611 | λ = -1.33%/mo (t = -2.87\*\*\*) standalone; remains -0.47%\*\* with all FF6+M4+DHS+Q5 factors; R² = 48% as single factor vs 51–77% for multifactor benchmarks (FF6 59%, M4 65%, DHS 51%, Q5 77%) |
| R2 | WarFac return premium is **negative and significant for all six sets** of test assets; consistently ranks in the top three among 11 nontraded factors | Table I (all panels), p. 3607–3610 | λ ranges from -0.66%\*\* (HXZ single-sorted, t = -2.25) to -3.32%\*\*\* (ML portfolios, t = -3.42); no other nontraded factor achieves this across all six sets |
| R3 | WarFac **explains 62% of cross-sectional variance** in ML-based nonlinear portfolio returns as a single factor, outperforming FF6 (41%), M4 (40%), DHS (35%), Q5 (58%) | Table II Panel D, p. 3617–3618 | Single-factor R² = 62%; adding WarFac to FF6 raises R² by 34%; common pricing error falls from 3.3% to near zero |
| R4 | WarFac return premium is **incremental to traded factors** (WMP, MKT, SMB, HML, RMW, CMA, MOM and mispricing factors); CMA and WarFac are the only factors significant across all six test-asset sets | Table III (all panels), p. 3619–3622 | WarFac: -1.33%\*\*\* to -3.32%\*\*\* depending on test assets; CMA also consistently significant; WMP -2.19%\*\* to -3.32%\*\*\* |
| R5 | WarFac is **incremental to NVIX and GPR** uncertainty indexes; NVIX and GPR do not command significant return premia across all test assets | Table IV (all panels), p. 3623–3624 | With all three factors, WarFac: -1.04%\*\* (HXZ long-short), -2.76%\* (ML portfolios); NVIX_War2Fac: insignificant for HXZ and ML; GPRFac: insignificant across all panels |
| R6 | WarFac **prices industry portfolios** with a negative premium, incremental to the CrisisFac (crisis event counts) of Berkman et al. (2011) | Table V, p. 3626 | 30-industry portfolios: λ(WarFac) = -0.24%\* (t = -1.89) alone; -0.32%\*\* (t = -2.39) with CrisisFac and CWarFac jointly; 49-industry: -0.28%\*\* (t = -2.16) in joint specification |
| R7 | WMP (the traded mimicking portfolio for WarFac) has a **Sharpe ratio of 1.73**, the highest among all factors in the sample, and generates significant alphas against all benchmark factor models | §V.A and Internet Appendix Table IA.III, p. 3627–3628 | Monthly average return = -3.32%, monthly SD = 6.64%; annualised Sharpe = 1.73; monthly alpha vs all factors ≈ 3.10%\*\*\* (t-stat in Internet Appendix); WMP passes three-pass test and factor-identification protocol |
| R8 | WarFac **captures a distinct tail risk**: return premium survives after controlling for CAPM beta, bear beta, downside beta, VIX beta, volatility beta, jump beta, coskewness, skewness beta, tail beta, and idiosyncratic volatility | §VII.A and Internet Appendix Table IA.VII, p. 3632 | WarFac premium remains significant with all tail-risk mimicking portfolios included; it is the only nontraded factor with a significant beta return premium on HXZ single-sorted portfolios in this horse race |

**Overall (paper's conclusion, p. 3634).** Loadings on the war-discourse factor
strongly predict the cross section of stock returns with a negative premium,
consistent with rational rare-disaster hedging (good hedges earn low premia)
or with behavioral overweighting of war prospects (war-sensitive stocks are
overpriced). The war premium is incremental to all standard factor models
and to other news-based uncertainty measures, and is driven by factual war
news rather than opinion articles.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| *New York Times* full text, Jan 1871–Oct 2019 (~7M articles) | Source corpus for sLDA topic modelling; constructs the *War* index | no page yet; proprietary/licensed archive |
| CRSP monthly stock returns and characteristics | Returns for all six test-asset sets; underlying portfolio construction data | [WRDS / CRSP / Compustat](/wiki/commercial/wrds/) (licensed) |
| Compustat | Firm fundamentals for anomaly characteristic construction | [WRDS / CRSP / Compustat](/wiki/commercial/wrds/) (licensed) |
| Hou, Xue & Zhang (2020) HXZ anomaly portfolios (138 long-short, 1,372 single-sorted) | Primary test-asset sets; Jul 1972–Dec 2016 | no page yet; available from HXZ replication files |
| Chen & Zimmermann (2022) single-sorted portfolios (904) | Third test-asset set; available from open-source-asset-pricing project | [Open Source Asset Pricing](/wiki/datasets/open-source-asset-pricing/) |
| Bryzgalova, Huang & Julliard (2023) ML-based nonlinear portfolios (360) | Fourth test-asset set; tree-based nonlinear portfolios | no page yet |
| Ken French Data Library (size, B/M, momentum portfolios) | Basis assets for WMP time-series mimicking; Fama-French factor benchmarks | [Ken French Data Library](/wiki/datasets/ken-french/) |
| Berkman, Jacobsen & Lee (2011) crisis event counts | CrisisFac and CWarFac benchmarks for §IV.B industry tests; data from sites.duke.edu/icbdata | no page yet |
| NVIX (Manela & Moreira 2017) | Benchmark news-based uncertainty index; horse-race test in §IV.A | no page yet |
| GPR index (Caldara & Iacoviello 2022) | Benchmark geopolitical risk index; horse-race test in §IV.A | no page yet |

Sample period for asset pricing tests: Jul 1972–Dec 2016 (532 months). *War* index: Jan 1871–Oct 2019.

## Theory / model

This paper has no original structural model. It tests two competing,
observationally equivalent theoretical frameworks:

1. **Rational rare-disaster risk** (Barro 2006, 2009; Gourio 2008; Gabaix
   2012, p. 3601): investors demand a risk premium for bearing war-related
   disaster risk. Assets that pay off when war risk is high are good hedges
   and therefore command lower expected returns. A negative cross-sectional
   return premium on WarFac betas is the central prediction.

2. **Behavioral overweighting** (Daniel, Hirshleifer & Subrahmanyam 2001;
   Tversky & Kahneman 1992 cumulative prospect theory, p. 3601): investors
   overweight the probability of rare salient disasters such as war,
   overvaluing stocks that do well under high war risk, so those stocks
   subsequently earn lower returns. The same negative premium arises for
   behavioral reasons.

The paper tests both frameworks by constructing a text-based proxy for
investor attention to war risk rather than relying on realized war events,
which have small sample sizes. Identification rests on the rolling-forward
estimation of the sLDA model and the AR(1) residual (WarFac), which ensures
only past information is used at each point in time, avoiding look-ahead
bias (p. 3591).

## Method

The method has two components: (1) constructing the *War* index via
semisupervised topic modelling (sLDA), and (2) building WarFac as the
innovation in *War* via a rolling AR(1). It builds on `slda-topic-model`
and `ar1-innovation`. The *War* index used here is the same one that the
companion aggregate-return study of Hirshleifer, Mai, and Pukthuanthong
(2025, Review of Financial Studies) uses; this paper extends it to
cross-sectional pricing (p. 3590).

**sLDA topic model (pp. 3596-3598).** Each month $$t$$, the model is estimated
on all *New York Times* articles in the preceding 120 months (the rolling
window $$[t-119, t]$$). Using Gibbs sampling, the model infers, for each
document $$d$$, the document-topic distribution $$\theta_d$$ (a vector of topic
probabilities) and, for each topic $$k$$, the topic-word distribution
$$\phi_k$$ (a vector of word probabilities). The seed word for the War topic
is *war* (a single word, for parsimony and to avoid researcher discretion
in seed-word selection). The global monthly weight of topic $$k$$ in month
$$t$$ is the length-weighted average across all articles $$d$$ in month $$t$$:

$$
\text{War}_t = \frac{1}{\sum_d \text{len}_d} \sum_d \text{len}_d \cdot \theta_{d,k=\text{War}}
$$

- $$\text{len}_d$$ is article length in n-gram count.

The rolling window allows
topic-word distributions $$\phi_k$$ to shift with language over time, which
is essential for a corpus spanning 1871 to 2019 (p. 3597).

**AR(1) innovation (p. 3603, equations 3 and 4).** Following Berkman,
Jacobsen & Lee (2011), Liu & Matthies (2022), and Giglio & Xiu (2021),
WarFac is defined as the residual from a rolling AR(1) fit to *War*,
estimated at each month $$t$$ using data from 1926 to $$t$$ to avoid
look-ahead bias:

$$
\text{War}_t = \rho_0 + \rho \cdot \text{War}_{t-1} + u_t \tag{3}
$$

$$
\text{WarFac}_t = u_t \tag{4}
$$

The AR(1) coefficients $$(\rho_0, \rho)$$ are re-estimated each month on the
growing window of available data. Results are robust to using an ARMA(1,1)
residual or a rolling-regression residual (pp. 3591, 3603 fn. 13).

**War-mimicking portfolio (WMP).** The traded version of WarFac is
constructed using the cross-sectional approach of Lehmann & Modest (1988):
the slope from the monthly second-pass cross-sectional regression of
asset returns on WarFac betas is the monthly WMP return (p. 3627). As a
robustness check, the time-series approach projects WarFac onto the space
of excess returns of 360 tree-based portfolios plus basis assets:

$$
\text{WarFac}_t = \alpha + \beta' R^e_t + \epsilon_t \tag{8}
$$

$$
\text{WMP}_t = \hat{\beta}' R^e_t \tag{9}
$$

- $$R^e$$ is the vector of excess returns on basis assets.
- $$\hat{\beta}$$ is estimated by OLS on the full sample.

## Empirical specifications

All asset pricing tests use monthly data, July 1972 to December 2016
(T = 532 months for WarFac; T = 522 for tests including NVIX_War and GPR).

**First pass: factor loadings (eq. 1, p. 3602).** For each test asset
$$i = 1, \ldots, N$$, excess returns are regressed on a vector of factors $$F_t$$
in a multivariate time-series regression:

$$
R^e_{it} = \alpha_i + \beta_{iF}' F_t + \epsilon_{it}, \quad i = 1, \ldots, N \tag{1}
$$

The paper reports $$\text{avg}(|t|)$$ (average absolute beta t-statistic) and the
number of assets with $$|t| \geq 1.65$$ (the 5% one-sided threshold). This
first pass is run for each of the six test-asset sets separately.

**Second pass: cross-sectional return premium (eq. 2, p. 3602).** Time-series
average excess returns are regressed cross-sectionally on the estimated factor
loadings:

$$
\bar{R}^e_{i} = \lambda_0 + \beta_{iF}' \lambda_F + e_i \tag{2}
$$

- $$\lambda_F$$ is the vector of return premium slopes.
- Standard errors are Shanken (1992) corrected.
- The paper reports $$\lambda$$ and its $$t$$-statistic, cross-sectional $$R^2 = 1 - \sigma^2_e / \sigma^2_{\mu}$$, and mean absolute pricing error $$\text{MAPE} = |\bar{e}|$$.
- Under rational pricing, $$\lambda_0 = 0$$.

**Industry portfolios: rolling Fama-MacBeth with betas (eqs. 5-6, p. 3625).**
For the industry pricing tests (§IV.B), betas are estimated over a rolling
60-month window for excess returns on factor $$X$$ (WarFac, CrisisFac, or
CWarFac) plus market, size, and value controls:

$$
R^e_{it} = \alpha_i + \beta_{it} X_t + \beta^{\text{MKT}}_{it} \text{MKT}_t + \beta^{\text{SMB}}_{it} \text{SMB}_t + \beta^{\text{HML}}_{it} \text{HML}_t + \epsilon_{it}, \quad \text{window: } t\text{-}59 \text{ to } t \tag{5}
$$

Cross-sectional betas are ranked into quintiles each month $$t$$ and rescaled
to $$[0, 1]$$. The monthly return premium is estimated by:

$$
R^e_{it} = \lambda_{0t} + \lambda_t \beta_{i,t-1} + \lambda^{\text{MKT}}_t \beta^{\text{MKT}}_{i,t-1} + \lambda^{\text{SMB}}_t \beta^{\text{SMB}}_{i,t-1} + \lambda^{\text{HML}}_t \beta^{\text{HML}}_{i,t-1} + e_{it} \tag{6}
$$

- Time-series averages of $$\lambda_t$$ are reported; statistical significance uses Newey-West (1987) standard errors.
- Sample period for industry tests: July 1926 to December 2018 (T = 1,110 months for Panels A and B of Table V, p. 3626).

**WMP spanning test (eq. 7, p. 3628).**

$$
\text{WMP}_t = \alpha + \beta' F_t + \epsilon_t \tag{7}
$$

- $$F_t$$ is the vector of benchmark traded factors.
- $$\alpha$$ measures whether WMP expands the mean-variance frontier.
- Monthly alpha of WMP against all factors combined is approximately 3.10%, significant at the 1% level (Internet Appendix Table IA.III, p. 3628).

## When to read the full paper

Use the original at [doi.org/10.1111/jofi.13482](https://doi.org/10.1111/jofi.13482)
(paywalled) if you are: replicating (data and code available from authors on
request per fn. 16); extending the sLDA war-discourse methodology to other
corpora or time periods; doing a literature review where the full robustness
battery (seed-word variants, ARMA(1,1), sLDA vs LDA comparisons, tail-risk
horse races) matters; or auditing a specific coefficient. The locators above
point to the exact table.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(6), December 2025,
pp. 3589–3637. © 2025 the American Finance Association. This distillation
was extracted by an LLM on 2026-05-31 and is **not human-verified or
independently reproduced**. The underlying article is paywalled; no verbatim
PDF is hosted here.

> Hirshleifer, David, Dat Mai, and Kuntara Pukthuanthong. "War Discourse and
> the Cross Section of Expected Stock Returns." *The Journal of Finance* 80,
> no. 6 (December 2025): 3589–3637. DOI: 10.1111/jofi.13482.
> © 2025 the American Finance Association. Extract-only; redistribution
> of the original article is subject to Wiley/AFA terms.


==============================================================================
# Scope, Scale, and Concentration: Hoberg & Phillips (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/hoberg-scope-scale-concentration-21st-2025/
# Distilled: Using doc2vec text analysis of firm 10-Ks, Hoberg and Phillips document that U.S. firms expanded their product market scope by 50-70% from 1989 to 2017, primarily through acquisitions and R&D rather than capital expenditures, with scope expansion raising firm valuations by 29.5% of the interquartile range while leaving traditional Herfindahl-Hirschman Index concentration measures flat once scope is accounted for. J. Finance 2025, paywalled. Eight core results with source locators, datasets used, the method (D2V-Scope), and the empirical specifications with equations.
# Tags: paper-summary, corporate-finance, industrial-organization, firm-dynamics
==============================================================================

**What this is.** The paper's core results, the new D2V-Scope measure it constructs, and the two-stage instrumental variable framework used to identify the effect of scope on corporate policies and performance: enough to know what it found and how, without reading the full 52 pages. To replicate or extend it, read the full source at [doi.org/10.1111/jofi.13400](https://doi.org/10.1111/jofi.13400).

## TL;DR

Hoberg and Phillips use doc2vec text analysis of firm 10-Ks to measure firm product market scope across 300 D2V industries. The average U.S. firm's scope increased by roughly 50 to 70% from 1989 to 2017, driven primarily by acquisitions and R&D rather than capital expenditures, while Compustat segment counts stayed flat. Scope expansion is positively related to firm valuation (a 25th-to-75th percentile scope increase raises market-to-book by 0.31, or 29.5% of the interquartile range), contradicting the conglomerate discount in the prior literature. Traditional industry concentration (HHI) appears to rise over this period, but scope-adjusted HHI is flat since 1997, suggesting that firms increasingly compete across more overlapping markets.

## Core results

Magnitudes and significance are as reported; `\*` = 5%, `\*\*` = 1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Average D2V-Scope increased roughly 50-70%** from 1989 to 2017 while Compustat segment counts remained flat | Figure 1 (middle and upper panels), p. 434 | D2V-Scope rises from ~6 to ~11; NAICS-Scope also rises; Compustat segments stable at ~1.4-1.5 |
| R2 | **Modern multi-industry firms operate in closely related industries**: 53% of all operating industry pairs are in the highest horizontal TNIC similarity decile | Table VII, Panel A, p. 439 | 52.8% of pairs (all firms) in most-similar decile; 15.6% in next decile; least-similar decile only 1.7% |
| R3 | **Scope incentives raise acquisitions and R&D but not CAPX**: 2SLS using peer-based redeployability and opportunity set as instruments | Table XII, pp. 447-448 | Acquisition probability +6.55 pp (t=3.5); Divestiture -4.15 pp (t=-2.9); R&D/Assets +0.82 pp (t=3.4); CAPX coefficient = 0.000, t=-0.06 |
| R4 | **Scope expansion increases firm valuation** by 0.31 M/B points from 25th to 75th percentile, representing 29.5% of the interquartile range | Table XIII row 1, p. 451 | 2SLS coefficient on D2V-Scope = 0.100 (t=5.22); contradicts the conglomerate discount of Lang and Stulz (1994) and Berger and Ofek (1995) |
| R5 | **Scope raises sales and asset growth but not ROA**, consistent with expansion into still-profitable but lower-profitability industries | Table XIII rows 2-4, p. 451 | Sales growth coefficient = 0.034 (t=5.96); Asset growth = 0.046 (t=8.55); OI/Assets = -0.001 (t=-0.44, n.s.) |
| R6 | **D2V segments are more informative than Compustat segments** for predicting profitability in-sample and out-of-sample | Table IX, p. 443 | D2V adj R2 = 31.6% vs Compustat 23.7% (OI/Assets, in-sample); out-of-sample D2V adj R2 = 3.8% vs Compustat 2.9% |
| R7 | **Scope-adjusted HHI is flat since 1997**; the apparent rise in traditional HHI is explained by firms operating in more overlapping markets | Figures 4, 5, pp. 458-459 | Scope-adjusted HHI stable at ~0.11-0.14 since 1997; traditional SIC HHI rose from ~0.19 to ~0.28 over 1989-2016 |
| R8 | **Scope expansion is financed by equity issuance, not debt**, consistent with intangible and redeployable assets lacking collateral value | Table XV, p. 453 | Equity issuance +3.1 pp (65.3% of mean) from 25th to 75th pct scope (t=7.23); Debt issuance coefficient = 0.002 (t=0.73, n.s.) |

**Overall (paper's conclusion).** The 21st-century firm is a high-scope firm that serves multiple related product markets through flexible production, innovation, and acquisition, all without increasing formal Compustat segments. This scope expansion creates value, is consistent with economies of scope (not agency-driven conglomeration), and, once accounted for, explains most of the apparent rise in industry concentration. Traditional Compustat-based measures substantially understate the scope of modern U.S. firms.

## Theory / model

The paper has no formal structural model. Its theoretical motivation draws on the economies-of-scope literature (Panzar and Willig (1977, 1981); Teece (1980); Maksimovic and Phillips (2002)) and the agency-cost view of diversification (Jensen (1986); Lang and Stulz (1994); Berger and Ofek (1995)). The paper empirically distinguishes between two views:

1. **Scope-as-synergy**: Firms with redeployable assets and related product-market opportunities expand scope at low cost, generating positive NPV. This predicts higher valuations for high-scope firms.
2. **Scope-as-agency**: Managers empire-build into unrelated markets, generating a diversification discount. This predicts lower valuations.

The paper's findings support the synergy view. The paper additionally models Compustat segment underreporting: managers evaluating related industries holistically do not report them as separate segments, so segment counts do not rise with scope. This is confirmed by a regression (Table VI, p. 436) showing that Compustat segments are negatively associated with the number of highly related D2V industry pairs a firm operates in (average coefficient -0.019, t=-4.32), while positively associated with weakly related (average coefficient 0.166, t=6.41) and unrelated pairs.

**Identification.** The paper uses a two-stage instrumental variable (2SLS) strategy (described in Section V.A-B, pp. 446-450). The two instruments for scope are:

- **Sectoral Redeployment Potential**: the average cosine similarity of the asset utilization vectors between the NAICS industry of the focal firm's close peers and the NAICS industries of the firm's more distant peers, following the BEA capital flows methodology of Kim and Kung (2017). High values mean close-peer assets can be cheaply redeployed to distant-peer markets (eq. 3, p. 423). First-stage coefficient = 1.156 (t=3.74; Table XI row 1, p. 447).
- **Sectoral Opportunity Set Potential**: one minus the HHI of the NAICS industries served by the focal firm's distant peers (eq. 4, p. 424). High values mean distant peers span a wide set of markets, indicating a thick scope-expansion opportunity set. First-stage coefficient = 2.124 (t=11.94; Table XI row 1, p. 447).

Both instruments are constructed from characteristics of peers who are not the focal firm itself, reducing first-degree endogeneity concerns.

## Method

**D2V-Scope construction** (the paper's headline methodological contribution; Appendix B, pp. 463-464). The measure uses a doc2vec embedding model trained on 10-K Item 1 business descriptions from all Compustat firms in the base year 1997. The 300-dimensional vector space represents each firm's product offering. Five steps:

1. Run k-means clustering on single-segment firm vectors to identify 450 candidate industries; add word2vec dialects (vocabulary) for each cluster.
2. Prune 150 boilerplate or redundant clusters to obtain 300 D2V industries.
3. Compute term-specific weights $$w_{k,n}$$ for each word $$n$$ in industry $$k$$ as:

$$
w_{k,n} = (\text{cosine similarity of word } n \text{ to centroid } k) \times (\text{word-specific HHI across industries}) \tag{B.Step 3}
$$

4. Compute each firm $$i$$'s exposure to industry $$k$$ in year $$t$$ (equation B1, p. 464):

$$
E_{i,k,t} = \frac{\sum_{n=1}^{N} B_{i,k,n,t} \cdot w_{k,n}}{\sum_{n=1}^{N} w_{k,n}} \tag{B1}
$$

where $$B_{i,k,n,t} = 1$$ if the firm uses word $$n$$ of industry $$k$$'s dialect in year $$t$.

5. Tag firm $$i$$ as operating in industry $$k$$ if $$E_{i,k,t} \geq \bar{E}$$, where the threshold $$\bar{E}$$ is fixed at the 2% granularity level from the 1997 base year. D2V-Scope is the count of industries exceeding this threshold.

**Alternative: NAICS-Scope** (Section II.A.2, pp. 421-422) uses 311 four-digit NAICS industry descriptions from the 2017 NAICS manual. The overlap ratio for firm $$i$$ in industry $$j$$ in year $$t$$ is (equation 1, p. 421):

$$
Q_{i,j,t,\text{NAICS}} = \frac{\#\text{words overlapping in } D_{\text{NAICS},j} \text{ and } V_{i,t}}{\#\text{words in } D_{\text{NAICS},j}} \tag{1}
$$

NAICS-Scope is then the count of industries above a fixed threshold $$\bar{Q}_{\text{NAICS}}$$ (eq. 2, p. 422):

$$
\text{NAICS-Scope}_{i,t} = \sum_{j=1}^{K} \text{Indicator}\{Q_{i,j,t,\text{NAICS}} > \bar{Q}_{\text{NAICS}}\} \tag{2}
$$

**Redeployability instruments.** Local asset redeployability is the weighted average cosine similarity of asset utilization vectors across industries spanned by close and distant peers (equation 3, p. 423):

$$
\text{LocalAssetRedep}_{i,t} = \sum_{j,k: s,t, j \neq k \in \text{NAICS-4}} F_{i,t,j,\text{near}} \cdot F_{i,t,k,\text{distant}} \left\langle \frac{A_j}{A_j \cdot \mathbf{1}} \cdot \frac{A_k}{A_k \cdot \mathbf{1}} \right\rangle \tag{3}
$$

The opportunity set instrument is one minus the HHI of the distribution of distant-peer NAICS-4 industries (equation 4, p. 424):

$$
\text{LocalScopeExpOppSet}_{i,t} = 1 - \sum_{j \in \text{NAICS-4}} F_{i,t,j,\text{distant}}^2 \tag{4}
$$

## Empirical specifications

All regressions include firm and year fixed effects; standard errors are clustered by firm. The headline two-stage specification (Sections V.B and V.C, pp. 446-453) is:

**First stage** (Table XI, p. 447): For firm $$i$$ and year $$t$$,

$$
\text{Scope}_{i,t} = \alpha + \beta_1 \cdot \text{SectRedepPot}_{i,t} + \beta_2 \cdot \text{SectOppSetPot}_{i,t} + \gamma X_{i,t-1} + \mu_i + \mu_t + \varepsilon_{i,t} \tag{FS}
$$

where $$X_{i,t-1}$$ includes log assets, log age, market-to-book (in some specs), and TNIC HHI. With D2V-Scope as the dependent variable: $$\hat{\beta}_1 = 1.156$$ (t=3.74), $$\hat{\beta}_2 = 2.124$$ (t=11.94). The instruments are strong (Kleibergen-Paap r-k statistic significant at 1%) and Hansen J-tests are not rejected at 5% for most outcome variables.

**Second stage** (Tables XII-XV, pp. 448-453): The instrumented scope variable is used as the regressor in the outcome equation:

$$
y_{i,t} = \alpha + \delta \cdot \widehat{\text{Scope}}_{i,t} + \gamma X_{i,t-1} + \mu_i + \mu_t + \varepsilon_{i,t} \tag{SS}
$$

Key second-stage results with D2V-Scope instrumented:

- **Investment** (Table XII): Acquirer dummy $$\hat{\delta}$$ = 0.019 (t=3.45); Target dummy = -0.012 (t=-2.89); R&D/Assets = 0.002 (t=3.40); CAPX/Assets = 0.000 (t=-0.06); Vertical integration = 0.002 (t=12.88).
- **Outcomes** (Table XIII): Valuation = 0.100 (t=5.22); Sales growth = 0.034 (t=5.96); Asset growth = 0.046 (t=8.55); OI/Assets = -0.001 (t=-0.44).
- **Financing** (Table XV): Equity issuance = 0.009 (t=7.23); Debt issuance = 0.002 (t=0.73); Dividends/Assets = -0.001 (t=-2.22).

**Profitability validation** (Table IX, p. 443): Generalized fixed effects model estimating industry-level profitability parameters $$\mu_{k,t}$$:

$$
\text{OIassets}_{i,t} = \sum_{k=1}^{300} \omega_{i,k,t} \cdot \mu_{k,t} + \varepsilon_{i,t} \tag{5}
$$

where $$\omega_{i,k,t}$$ is the textual exposure weight for D2V (or the fraction of sales for Compustat). D2V segments yield in-sample adjusted R2 of 31.6% vs. 23.7% for Compustat on single-segment firms (OI/Assets).

**Scope-concentration link** (Section VI, pp. 455-459): Scope-adjusted HHIs computed using D2V industry assignments and textual-intensity weights; scope-adjusted HHI is flat since 1997 (Figures 4 and 5), while traditional SIC HHI rises. A secondary granularity-based approach loads firm product descriptions onto SIC-2 (broad) vs. SIC-3 (narrow) vocabulary over time, confirming firms operate at coarser granularity in later years.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Compustat annual fundamentals (WRDS) | Assets, R&D, CAPX, profitability, segments, financing; main panel 1989-2017 | [WRDS](/wiki/commercial/wrds/) (licensed) |
| SEC EDGAR 10-K filings (Item 1) | Source text for D2V-Scope and NAICS-Scope construction; all Compustat firm-years with 10-K | [EDGAR](/wiki/datasets/edgar/) |
| TNIC / Hoberg-Phillips (2016) | Industry classification for peer identification; pairwise similarity used for instruments and scope validation | [TNIC](/wiki/datasets/tnic/) |
| SDC Platinum | Acquisition and divestiture events (acquirer dummy, target dummy) | [no page yet] |
| BEA capital flows tables (1997) | 180-asset utilization vectors for 123 BEA industries, used to compute local asset redeployability instrument | [no page yet] |
| Venture Expert (VentureXpert) | VC funding similarity measure: startup business descriptions for computing VC funding similarity | [no page yet] |
| NAICS Manual (2017) | 963-page NAICS manual with 311 four-digit industry descriptions used to construct NAICS-Scope | [no page yet] |

Sample: 101,535 firm-year observations, 1989-2017 annual. SIC 6000-6999 (financials) and 4900-4949 (utilities) excluded. Firms required to have sales and assets of at least $1 million in both current and prior year.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13400) if you are: constructing D2V-Scope or NAICS-Scope measures (Appendix B gives the full five-step algorithm); studying the relationship between firm scope and industry concentration; testing whether scope-adjusted HHI changes the interpretation of the rising-concentration literature (Grullon, Larkin, and Michaely (2019)); or using the peer-based redeployability and opportunity set instruments for scope endogeneity. The locators above point to the exact tables and figures.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(1), February 2025. Pages 415-466. DOI: 10.1111/jofi.13400. Published by Wiley on behalf of the American Finance Association. This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The article is paywalled; only extracts are reproduced here under fair-use commentary.

> Hoberg, Gerard, and Gordon M. Phillips. "Scope, Scale, and Concentration: The 21st-Century Firm."
> *The Journal of Finance* 80, no. 1 (February 2025): 415-466.
> DOI: 10.1111/jofi.13400. © 2024 the American Finance Association.
> Paywalled. This page is an extract-only summary by the Institute for Automated Research.


==============================================================================
# Auctions versus Negotiations: Hoffmann & Vladimirov (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/hoffmann-auctions-versus-negotiations-role-2025/
# Distilled: When payments can have a contingent component (equity, royalties, performance bonuses), a seller facing fewer bidders in optimally structured negotiations can earn strictly higher revenue than an auction with one more competing bidder. The key driver is bargaining power over the payment structure, not reserve-price setting. J. Finance 2025, CC BY-NC 4.0. Six core results with source locators, the model, and the formal propositions.
# Tags: paper-summary, market-design, auctions, mergers-acquisitions, contract-theory
==============================================================================

**What this is.** The paper's core propositions, the model it builds on (a
seller choosing between depth and breadth of bidder competition when payments
can be contingent), and the theoretical mechanism it contributes: enough to
know what it found and how, without reading all 45 pages. To replicate or
extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13446).

## TL;DR

The paper develops a theory of auctions versus negotiations that allows for
general (state-contingent) payment structures. A seller choosing between
optimal negotiations with a small group of bidders and an ascending-bid
auction with one more bidder can strictly prefer negotiations - even against
the benchmark result of Bulow and Klemperer (1996) that auctions dominate. The key
driver is not the reserve price but bargaining power over the payment
structure: when the asset is complementary to bidder productivity (synergies
increase in types), negotiating for contingent payments (equity, royalties,
performance bonuses) extracts more rent than cash competition. The paper
builds on the rent-extraction efficiency trade-off studied by Inderst and
Vladimirov (2019), extends the full-surplus extraction result of Liu and
Bernhardt (2021) to general securities, and uses the security-bid auction
framework of DeMarzo, Kremer, and Skrzypacz (2005). The motivating empirical
fact that negotiations with few bidders are as common as auctions without
lower premia comes from Boone and Mulherin (2007). The auction revenue
benchmark draws on Myerson (1981)'s revenue equivalence theorem. Negotiations
dominate if the type distribution is sufficiently dispersed, absolute
valuations are high, and the complementarity condition holds.

## Core results

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | If the seller can extract the full surplus in bilateral negotiations, her expected revenue is **strictly higher** than from an auction with two cash-bidding competitors | Proposition 1(i), p. 1780 | With uniform $$\theta$$, bilateral negotiations yield $$\Pi^{fb}_{neg} = (\alpha_Y - \alpha_N + \frac{1}{2}(\beta_Y - \beta_N))\Delta x$$ versus auction $$\Pi_{comp} = (\alpha_Y - \alpha_N + \frac{1}{3}(\beta_Y - \beta_N))\Delta x$$; negotiations up to 50% higher in Example 1 |
| R2 | Full-surplus extraction requires the seller to negotiate for **contingent payments**: the first-best contract has $$\Delta w^{fb} > 0$$ and is feasible iff valuations are increasing in productivity (complements case, $$\beta_Y/\beta_N \in (1, \alpha_Y/\alpha_N]$$) | Proposition 1(ii), pp. 1779-1780 | The first-best contract is $$\{w^{fb}, \Delta w^{fb}\}$$ with $$\Delta w^{fb} = (1-\beta_N/\beta_Y)\Delta x > 0$$ and $$w^{fb} = \alpha_Y(\beta_N/\beta_Y - \alpha_N/\alpha_Y)\Delta x$$ |
| R3 | In the **substitutes case** ($$v'(\theta) < 0$$), an auction with one more bidder always yields higher expected revenue than optimal negotiations | Proposition 2, p. 1786 | In substitutes, seller optimally demands pure cash in negotiations (Lemma 2) and bidders choose cash in auctions (Lemma 3), so payment structure is irrelevant; Bulow-Klemperer result applies directly |
| R4 | In the complements case ($$v'(\theta) > 0$$) when full surplus extraction is infeasible, **efficient bilateral negotiations dominate competition** if and only if the Gini coefficient $$G$$ of the productivity-type distribution satisfies $$G \geq \frac{\alpha_Y/\alpha_N - 1}{\beta_Y/\beta_N - 1}$$ | Proposition 3, equation (20), p. 1788 | Condition is always satisfied if $$\alpha_Y/\alpha_N \geq 1/G$$; always holds for any $$\beta_Y/\beta_N > \alpha_Y/\alpha_N$$ if $$\alpha_Y/\alpha_N \geq 1/G$$ |
| R5 | **Setting the payment structure takes precedence** over setting a reserve price in the seller's pecking order: bargaining power over the payment structure has equilibrium value on its own, but bargaining power over the reserve price is valuable only when combined with payment-structure power | Proposition 5, p. 1791 | Formally, $$\Pi^r_{neg}$$ is always smaller than $$\Pi_{comp}$$, while $$\Pi^{s+r}_{neg}$$ can be larger than $$\Pi_{comp}$$, and $$\Pi^s_{neg}$$ can also be larger than $$\Pi_{comp}$$ |
| R6 | The optimal selling mechanism with $$n \geq 2$$ bidders is a **two-stage mechanism**: Stage 1 is a standard English auction to identify the highest valuation bidder; Stage 2 is a take-it-or-leave-it offer with the seller's preferred payment structure to the last remaining bidder | Proposition 6, p. 1793 | The Stage 2 offer is: complements case with full surplus extractable, demands $$\omega^{FB}$$; otherwise demands $$w = 0$$, $$\Delta w = (1 - p_{N}(\tilde\theta)/p_Y(\tilde\theta))\Delta x$$ at the optimal reserve |

**Overall (paper's conclusion).** Negotiations over payments are valuable in
many corporate finance settings. The paper resolves the theoretical puzzle
that negotiations are widely observed even when their revenue advantage over
auctions is difficult to explain with reserve prices alone. The key value
driver in negotiations is bargaining power over the mix of cash and contingent
pay, not the reserve price. Negotiations are more likely to dominate when the
asset creates higher synergies at more productive types, valuations are
dispersed, and the type-independent component of valuations is high relative
to the total upside.

## Theory / model

The paper studies a single seller (she) selling an indivisible asset to $$n$$
risk-neutral bidders (they/he) indexed $$i = 1, \ldots, n$$. The asset can be a
takeover target, patent, or employee's human capital. All parties are
risk-neutral and there is no discounting.

**Project cash flows and bidder types.** Each bidder $$i$$ has a productivity
(quality) type $$\theta_i$$ drawn independently from distribution $$F$$ on
$$[0, 1]$$. The project either fails (cash flow $$x \geq 0$$) or succeeds (cash
flow $$x + \Delta x > x$$). The probability of success depends on bidder type
and whether the bidder acquires the asset ($$a_i = Y$$) or not ($$a_i = N$$),
with the linear specification (equation (2), p. 1776):

$$
p_{a}(\theta) = \alpha_a + \beta_a \theta, \quad a \in \{Y, N\}, \quad \theta \in [0,1], \tag{2}
$$

where $$\alpha_a, \beta_a > 0$$ and $$\alpha_a + \beta_a \leq 1$$. A bidder's
expected cash flow under allocation $$a$$ is $$X_a(\theta) = x + p_a(\theta)\Delta x$$,
strictly increasing in $$\theta$$. A bidder's valuation is his willingness to
pay for the asset (equation (1), p. 1776):

$$
v(\theta) := X_Y(\theta) - X_N(\theta) = (p_Y(\theta) - p_N(\theta))\Delta x. \tag{1}
$$

The **complements case** arises if $$\beta_Y/\beta_N > 1$$ so that $$v'(\theta) > 0$$
(more productive types have higher willingness to pay). The **substitutes
case** arises if $$\beta_Y/\beta_N < 1$$ so that $$v'(\theta) < 0$$.

**Payment contracts.** Payments can be in general securities. If bidder $$i$$
acquires the asset, he pays the seller $$w_i$$ in the low-cash-flow state and
$$w_i + \Delta w_i$$ in the high-cash-flow state. Here $$w_i$$ is the cash
payment, $$\Delta w_i \geq 0$$ is the contingent payment, and the payment
structure is captured by the ratio $$\omega_i = \Delta w_i/(w_i + \Delta w_i)$$.
Examples of contingent payments include royalties, stock options, and
performance bonuses.

The seller's expected payment from a contract $$\omega = \{w, \Delta w\}$$ when
the buyer's type is $$\theta$$ is (p. 1779):

$$
\pi(\theta, \omega) := w + p_Y(\theta)\Delta w.
$$

Full surplus extraction requires $$\pi(\theta, \omega) = v(\theta)$$ for all
$$\theta$$, which from equation (3) (p. 1779) gives:

$$
w + p_Y(\theta)\Delta w = (p_Y(\theta) - p_N(\theta))\Delta x \quad \text{for all } \theta. \tag{3}
$$

The first-best contract solving (3) is (equations (4)-(5), p. 1779):

$$
\Delta w^{fb} = \left(1 - \frac{\beta_N}{\beta_Y}\right)\Delta x, \tag{4}
$$

$$
w^{fb} = \alpha_Y\left(\frac{\beta_N}{\beta_Y} - \frac{\alpha_N}{\alpha_Y}\right)\Delta x. \tag{5}
$$

This contract is feasible ($$w^{fb}, \Delta w^{fb} \geq 0$$) if and only if
$$\beta_Y/\beta_N \in (1, \alpha_Y/\alpha_N]$$, that is, in the complements case but
not too steeply.

**Game structure.** At $$t = 0$$ the seller decides between (i) negotiations:
choosing the optimal mechanism for the $$n$$ bidders already present, including
setting the payment structure; or (ii) competition: attracting one more bidder
so that $$n + 1$$ bidders compete in a standard ascending-bid (English) auction
where bidders choose their own payment structure. Cash flows are realized at
$$t = 1$$ and the winning bidder pays according to the agreed contract (pp.
1777-1778).

## Method

The paper's solution method is mechanism design with state-contingent payments
and bilateral-contract analysis under asymmetric information, building on
`principal-agent` contracting and elements of `bayesian-persuasion` (the
seller's mechanism design with general securities).

**Optimal negotiations (seller designs the mechanism).** The seller maximizes
expected revenue over a menu of contracts $$W$$, with the set of accepting types
$$\Theta_W \subseteq [0,1]$$. The seller's problem (equation (6), p. 1781) is:

$$
\max_W \int_{\Theta_W} \pi(\theta, \omega_\theta)\,dF(\theta) + \int_{[0,1]\setminus\Theta_W} \underline{w}\,dF(\theta), \tag{6}
$$

subject to feasibility ($$w_\theta, \Delta w_\theta \geq 0$$), individual
rationality, and incentive compatibility. Participation requires (equations
(7)-(8), p. 1781):

$$
v(\theta) - \pi(\theta, \omega_\theta) = \max_{\omega \in W} v(\theta) - \pi(\theta, \omega) \geq 0 \quad \text{for all } \theta \in \Theta_W, \tag{7}
$$

$$
\max_{\omega \in W} v(\theta) - \pi(\theta, \omega) < 0 \quad \text{for all } \theta \notin \Theta_W. \tag{8}
$$

The cutoff type $$\tilde\theta(\omega)$$ indifferent between acquiring and not is
(equation (9), p. 1782):

$$
\tilde\theta(\omega) := \frac{w + \alpha_N \Delta x - \alpha_Y(\Delta x - \Delta w)}{\beta_Y(\Delta x - \Delta w) - \beta_N \Delta x}. \tag{9}
$$

The seller's information rent for type $$\theta > \tilde\theta$$ under the
contingent-only contract ($$w = 0$$) is (equation (10), p. 1782):

$$
v(\theta) - \pi(\theta, \omega) = (\theta - \tilde\theta)\beta_Y \left(\underbrace{\left(1 - \frac{\beta_N}{\beta_Y}\right)\Delta x}_{= \Delta w^{fb}} - \Delta w\right). \tag{10}
$$

**Competition (bidders choose payment structure).** The English auction
establishes a reservation price that each active bidder must match. Remaining
active bidders choose whether to compete in cash or other securities, subject
to the seller's acceptance constraint (equation (12), p. 1785):

$$
\int_0^1 \pi(\theta, \omega)\,d\widetilde{F}(\theta|\omega) \geq \underline{w}. \tag{12}
$$

The equilibrium outcome is given by Lemma 3: with $$n \geq 2$$ bidders, they
optimally offer pure cash payments ($$w > 0$$, $$\Delta w = 0$$), and the winner
pays the second-highest valuation (p. 1785-1786). The seller's expected revenue
is the expected valuation of the bidder with the second-highest valuation
(equation (16), p. 1787):

$$
\Pi_{comp} = \int_0^1 v(\theta)\, 2(1 - F(\theta))\,dF(\theta). \tag{16}
$$

**Revenue difference decomposition.** Taking the difference between
$$\Pi_{neg}(0)$$ (efficient negotiations without a reserve price) and
$$\Pi_{comp}$$ (equations (17)-(19), p. 1787):

$$
\Pi_{neg}(0) - \Pi_{comp} = \int_0^1 v'(\theta)F(\theta)(1-F(\theta))\,d\theta - \int_0^1 \varrho(\theta, 0)\,dF(\theta), \tag{19}
$$

where the first term is positive in the complements case ($$v'(\theta) > 0$$)
and equals $$\Pi^{fb}_{neg} - \Pi_{comp}$$, while the second term is the
bidder's expected information rent in negotiations. Negotiations dominate iff
the rent is sufficiently small, which is governed by condition (20).

## Empirical specifications

This is a pure-theory paper. There are no regressions, datasets, or empirical
specifications. The paper's propositions are established by analytical proofs
in the Appendix (pp. 1801-1813). The paper does derive comparative statics and
testable implications for M&A, patent licensing, and employee compensation.

**Key theoretical comparative statics (Proposition 4, p. 1790).** The revenue
advantage of efficient bilateral negotiations $$\Pi_{neg}(0) - \Pi_{comp}$$ is
higher if:

(i) The bidders' productivity type distribution becomes more dispersed in the
sense of a mean-preserving spread (higher Gini coefficient $$G$$), because more
dispersed valuations lower the auction revenue $$\Pi_{comp}$$ (the second-highest
valuation falls in expectation) while leaving $$\Pi_{neg}(0)$$ unchanged.

(ii) The type-independent component of bidders' valuations, captured by
$$\alpha_Y/\alpha_N$$, is higher, because higher $$\alpha_Y/\alpha_N$$ allows the
seller to demand a larger contingent payment $$\Delta w$$ acceptable to all
types, moving closer to the full-rent extraction contract $$\Delta w^{fb}$$.

**Necessary and sufficient condition for negotiations to dominate (equation (20),
p. 1788):**

$$
G := \frac{\int_0^1 F(\theta)(1-F(\theta))\,d\theta}{\int_0^1 \theta\,dF(\theta)} \geq \frac{\beta_Y/\beta_N\,/\,\alpha_Y/\alpha_N - 1}{\beta_Y/\beta_N - 1}, \tag{20}
$$

where $$G \in (0,1)$$ is the Gini coefficient of the productivity-type
distribution $$F$$. This condition is always satisfied if $$\alpha_Y/\alpha_N \geq 1/G$$.

**Multi-bidder extension (Proposition 7, p. 1794-1795).** The condition for
efficient negotiations with $$n$$ bidders to dominate competition with $$n+1$$
bidders in the complements case (equation (21), p. 1795):

$$
\frac{\int_0^1 F(y)^n (1-F(y))\,dy}{\int_0^1 \left(\int_y^1 \frac{\alpha_Y(q-y)}{\alpha_Y + \beta_Y y} dF(q)\right)(n-1)F(y)^{n-2}\,dF(y)} \geq \frac{\beta_Y/\beta_N\,/\,\alpha_Y/\alpha_N - 1}{\beta_Y/\beta_N - 1}. \tag{21}
$$

This condition holds for any $$\beta_Y/\beta_N > \alpha_Y/\alpha_N$$ if
$$\alpha_Y/\alpha_N \geq 1/G_n$$, where $$G_n$$ is explicitly defined in the
Appendix.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| None (pure theory) | All results are derived analytically | N/A |

No empirical data were used. The paper's claims are theoretical propositions
derived from the formal model. Testable implications for M&A, patent licensing,
and compensation are discussed in Section VII (pp. 1796-1800).

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13446) if you are:
studying optimal mechanism design with general security payments; designing
M&A sale processes and evaluating when negotiations versus auctions maximize
revenue; modeling patent licensing or employee compensation negotiation;
extending the framework to seller private information, common values, or
risk aversion (Internet Appendix); or checking the full formal proofs
(Appendix, pp. 1801-1813).

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(3), June 2025. This
distillation was extracted by an LLM on 2026-06-06 and is **not
human-verified or independently reproduced**. The CC BY-NC 4.0 licence
permits reproduction for non-commercial purposes; the verbatim PDF is not
hosted in this batch.

Citation: Hoffmann, Florian, and Vladimir Vladimirov.
"Auctions versus Negotiations: The Role of the Payment Structure."
*The Journal of Finance* 80, no. 3 (June 2025): 1769-1813.
DOI: 10.1111/jofi.13446. © 2025 The Author(s).
Licensed under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/).
This page is an **adaptation** by the Institute for Automated Research:
core results extracted and re-expressed; **changes were made**.


==============================================================================
# Worker Runs: Hoffmann & Vladimirov (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/hoffmann-worker-runs-2025/
# Distilled: Hoffmann and Vladimirov model how firms design compensation contracts to prevent contagious collective worker departures ("worker runs"), showing that dilutable output-dependent pay and asymmetric compensation structures resolve the coordination problem at no extra cost. J. Finance 2025, CC BY-NC 4.0. Six core results with source locators, the model equations, and the key propositions.
# Tags: paper-summary, corporate-finance, compensation, labor, theory, peer-reviewed
==============================================================================

**What this is.** The paper's core propositions, the model it builds, and the
compensation design results: enough to understand what it found and how, without
reading all 43 pages. To replicate or extend it, read the full source at the
[original](https://doi.org/10.1111/jofi.13424).

## TL;DR

Hoffmann and Vladimirov develop a theory of "worker runs": because workers
privately observe a firm-wide productivity shock, and because a departing
skilled worker reduces the value of remaining workers' output-dependent pay,
an initial departure can trigger further departures even from an otherwise
healthy firm. The modeling of collective turnover as a coordination failure is
similar to the bank-run literature of Diamond and Dybvig (1983) and Goldstein
and Pauzner (2005). Unlike bank runs, however, standard remedies such as
deposit insurance have no direct labor-market analog, so the paper focuses on
compensation design instead. The paper shows that firms can design compensation
contracts to prevent such contagious collective turnover at no extra cost
relative to a frictionless benchmark. The key instruments are (i) making
compensation "dilutable" (promising workers more when others leave) to smooth
workers' expected pay across retention scenarios, and (ii) offering ex-ante
identical workers differently structured contracts (asymmetric compensation) to
ensure that a critical subset always stays, removing the strategic
complementarity for the rest. The asymmetric contracting results extend Winter
(2004) and Halac, Kremer, and Winter (2020) by studying differences in
compensation structure rather than compensation level. The dilution results
build on Oyer (2004)'s insight that firms optimally match workers'
on-the-job pay to their outside options. The paper characterizes optimal
symmetric and asymmetric contracts and derives empirically testable implications
for dilution, compensation structure, and worker targeting.

## Core results

Propositions and results are as stated in the paper. Locators point into the
source PDF.

| # | Result | Locator | Content |
|---|---|---|---|
| R1 | Dilutable or fixed-wage contracts always resolve the coordination problem at zero extra cost versus the frictionless benchmark | Propositions 2-3, pp. 950-953 | Optimal contract sets expected compensation constant in retention level n for each shock realization (W(eps,n) = W(eps,N) for all n), achieved by dilution whenever the output-dependent component is positive |
| R2 | Optimal degree of dilution increases in the equilibrium share of output-dependent pay and in the sensitivity of firm success to retention | Corollaries 1-2, pp. 954-955; Figure 2, p. 955 | Dilution through decreasing output-dependent pay (complements case) or decreasing output-independent pay (substitutes case); both w(n) and delta-w(n) move to achieve smoothing |
| R3 | Symmetric contracts conditioning on overall retention n are weakly cheaper than asymmetric contracts that cannot condition on n | Lemma 3, p. 962 | Conditioning on n allows off-equilibrium promises to differ from equilibrium pay, fully absorbing coordination costs; asymmetric-only contracts must resolve coordination on-equilibrium, which is costlier |
| R4 | When symmetric contracting entails positive coordination costs, combining asymmetric structure with dilution strictly reduces expected costs | Proposition 7, p. 962 | Offering the optimal symmetric contract to N-1 workers and the relaxed-problem contract to worker N halves per-worker coordination cost; generally yields strictly lower total cost |
| R5 | Under optimal asymmetric contracts, higher-ranked workers receive a higher share of output-independent pay and do not necessarily earn more rent | Proposition 5, pp. 959-960 | Higher-ranked workers' decision to stay must be independent of more coworkers, so their output-dependent pay is more mispriced by the coordination friction; firm compensates with safer output-independent pay |
| R6 | Workers retained with higher probability in asymmetric retention policies optimally receive higher output-independent pay when shocks are idiosyncratic; higher output-dependent pay when shocks are systematic with outside options more sensitive than firm output | Internet Appendix Propositions IA.2, IA.5, p. 964 | Compensation-type allocation responds to which friction dominates: coordination (pushes toward fixed pay) vs. noncontractibility of the shock (pushes toward output-dependent pay for high-retention workers) |

**Overall (paper's conclusion).** Mitigating collective contagious turnover
poses additional challenges relative to individual turnover and requires
different compensation solutions. The tools the paper proposes, making
compensation dilutable or offering asymmetric contracts, can be easily
implemented with equity-based pay, profit-sharing bonus pools, retention
bonuses, and title-linked pay differences that are already common in practice
(pp. 970).

## Theory / model

The baseline model (Section I, pp. 943-945) has a firm that hires $$N \geq 2$$
risk-neutral workers at $$t = 0$$. The firm's only asset is a project that
generates cash flows at $$t = 2$$: $$x > 0$$ if the project fails and $$x + \Delta x$$
with $$\Delta x > 0$$ if it succeeds. The probability of success at interim date
$$t = 1$$ depends on an exogenous shock $$\varepsilon$$ (drawn from distribution $$G$$
with support $$[\underline{\varepsilon}, \bar{\varepsilon}]$$) and on the number of workers $$n$$
retained until $$t = 2$$ (p. 943, eq. 1):

$$
p(\varepsilon, n) = \alpha(n) + \beta(\varepsilon)\gamma(n), \tag{1}
$$

with $$\beta(\varepsilon), \beta'(\varepsilon) > 0$$ for all $$\varepsilon$$ and $$\gamma(n) > 0$$ for all
$$n$$, so both the shock and retention increase the probability of success.
Workers observe the shock $$\varepsilon$$ privately at $$t = 1$$ and decide
simultaneously whether to stay or take an outside option of value
$$\underline{w}(\varepsilon) \geq 0$$.

Each worker $$i$$ is offered a compensation contract
$$C_i = (w_i(n), \Delta w_i(n))_{n=1}^{N}$$, which specifies output-independent pay
$$w_i(n)$$ and output-dependent pay $$\Delta w_i(n)$$ conditional on the number $$n$$
of workers retained. The firm's objective is to minimize workers' expected
equilibrium rent (expected compensation minus outside option) subject to
achieving full retention as the unique equilibrium at $$t = 1$$.

The expected surplus of retaining $$n$$ workers is (p. 944, eq. 2):

$$
\Omega(\varepsilon, n) := x + p(\varepsilon, n)\Delta x - n\underline{w}(\varepsilon), \tag{2}
$$

which the paper assumes is positive and nondecreasing in $$n$$ for all $$\varepsilon$$
(retaining more workers is always efficient).

**Worker runs as coordination failure.** For symmetric contracts
$$C_i = C$$, workers play a coordination game at $$t = 1$$. Workers' expected
on-the-job compensation is $$W(\varepsilon, n) := w(n) + p(\varepsilon, n)\Delta w(n)$$,
which increases in $$n$$ whenever output-dependent pay is positive. This creates
strategic complementarities: a worker is more likely to stay if others stay,
since the firm's success probability and thus the value of her equity/bonus
increases with retention. Proposition 1 (p. 947) characterizes the resulting
equilibrium structure. A worker-run equilibrium, in which all workers leave,
exists whenever

$$
W(\varepsilon, N) \geq \underline{w}(\varepsilon) > W(\varepsilon, 1), \tag{3}
$$

that is, staying is attractive only when all others stay but not when alone.
Under condition (3) a full-retention equilibrium also exists, and the
worker-run equilibrium is Pareto-dominated by it.

**Monotonicity-in-retention constraint** (Assumption 1, p. 952): To rule out
unrealistic contracts that reward workers for inducing coworkers to leave, the
paper imposes that workers' expected on-the-job pay must be nondecreasing in
the number of retained workers: $$W_i(\varepsilon, n) \geq W_i(\varepsilon, n-1)$$ for
all $$(\varepsilon, n)$$ and $$i$$. This constraint limits the degree of dilution.

**Dilutability** (Definition 1, p. 948): Compensation of worker $$i$$ is
dilutable at $$(\varepsilon, n)$$ if, holding the success probability constant at any
$$\hat{p} \in [p(\varepsilon, n-1), p(\varepsilon, n)]$$, expected compensation
$$\hat{W}_i(\hat{p}, n) := w_i(n) + \hat{p}\Delta w_i(n)$$ decreases in
retention (eq. 4):

$$
\hat{W}_i(\hat{p}, n) - \hat{W}_i(\hat{p}, n-1)
= w_i(n) - w_i(n-1) + \hat{p}[\Delta w_i(n) - \Delta w_i(n-1)] < 0. \tag{4}
$$

Equity-based pay exhibits this dilutability because the firm's equity "pie" is
shared among fewer workers when some depart, so each remaining worker's
percentage stake rises even as the total pie shrinks.

## Method

This is a contract-theory paper with no econometric estimation. The solution
method is backward induction in a three-period ($$t = 0, 1, 2$$) game with
incomplete information (the shock $$\varepsilon$$ is private), building on
`principal-agent`, `repeated-game`, and `mechanism-design` techniques.

The firm's optimization at $$t = 0$$ is to design contracts that implement full
retention as the unique equilibrium at $$t = 1$$ at minimum expected
compensation cost. For symmetric contracts, this reduces to Problem 1 (p. 950,
eqs. 5-6):

$$
\min_{C \in \mathbf{C}^f} \int_{\underline{\varepsilon}}^{\bar{\varepsilon}}
[w(N) + p(\varepsilon, N)\Delta w(N) - \underline{w}(\varepsilon)] \, dG(\varepsilon) \tag{5}
$$

$$
\text{subject to} \quad W(\varepsilon, n) = w(n) + p(\varepsilon, n)\Delta w(n)
\geq \underline{w}(\varepsilon) \quad \forall (\varepsilon, n). \tag{6}
$$

Constraint (6) is the full-retention participation constraint requiring staying
to be dominant for all shock realizations and all retention levels. Together
with the monotonicity-in-retention constraint (Assumption 1), this implies
that optimal contracts satisfy (p. 952, eq. 7):

$$
W(\varepsilon, N) \geq \ldots \geq W(\varepsilon, 1) \geq \underline{w}(\varepsilon)
\quad \forall (\varepsilon, n). \tag{7}
$$

Workers' equilibrium rent can be decomposed as the sum of incremental rents
(p. 952, eq. 8):

$$
R(\varepsilon, N) := [W(\varepsilon,N) - W(\varepsilon,N-1)] + \ldots + [W(\varepsilon,1) - \underline{w}(\varepsilon)] \geq 0. \tag{8}
$$

The key insight is that minimizing total expected rent pushes as many
incremental rents to zero as possible, making expected compensation constant in
$$n$$: $$W(\varepsilon, n) = W(\varepsilon, N)$$ for all $$(\varepsilon, n)$$. This requires
choosing the output-dependent component $$\Delta w(n)$$ to offset the retention
sensitivity of $$p(\varepsilon, n)$$ (pp. 952-953, eqs. 9-10):

$$
\Delta w(n-1) = \frac{\gamma(n)}{\gamma(n-1)} \Delta w(n), \tag{9}
$$

$$
w(n-1) = w(n) + \left[\alpha(n) - \alpha(n-1)\frac{\gamma(n)}{\gamma(n-1)}\right] \Delta w(n). \tag{10}
$$

For asymmetric contracts (Section IV), each worker $$i$$ (indexed by rank in
the iterative dominance ordering) faces Problem 2 (p. 959, eq. 12):

$$
\min_{\{w_i, \Delta w_i\}_{i=1}^{N} \in \mathbf{C}^f} \sum_{i=1}^{N}
\int_{\underline{\varepsilon}}^{\bar{\varepsilon}} [W_i(\varepsilon, N) - \underline{w}(\varepsilon)] \, dG(\varepsilon)
\quad \text{s.t.} \quad (11) \; \forall i, \tag{12}
$$

where the participation constraint for worker $$i$$ at rank $$i$$ is (eq. 11):

$$
W_i(\varepsilon, i) = w_i + p(\varepsilon, i)\Delta w_i \geq \underline{w}(\varepsilon) \quad \forall \varepsilon. \tag{11}
$$

Worker $$i$$'s expected rent under idiosyncratic risk is (p. 959, eq. 13):

$$
\int_{\underline{\varepsilon}}^{\bar{\varepsilon}} [W_i(\varepsilon, N) - \underline{w}] \, dG(\varepsilon)
= \int_{\underline{\varepsilon}}^{\bar{\varepsilon}} (\underline{w} - w_i)
\left(\frac{p(\varepsilon, N)}{p(\underline{\varepsilon}, N)} \cdot \frac{p(\underline{\varepsilon}, N)}{p(\underline{\varepsilon}, i)} - 1\right) dG(\varepsilon). \tag{13}
$$

The ratio $$p(\underline{\varepsilon}, N)/p(\underline{\varepsilon}, i)$$ is decreasing in rank $$i$$, so
higher-ranked workers (lower $$i$$) misprice output-dependent pay more,
making output-independent pay relatively cheaper for the firm.

## Empirical specifications

This is a theory paper. It has no econometric estimation. Section V
(pp. 966-969) translates the propositions into testable cross-sectional
implications.

**Implication 1** (p. 966): When firms relying on hard-to-replace skilled
workers pay with output-dependent compensation, they should make that
compensation dilutable. The degree of dilution is higher when (i) firm
performance is more sensitive to retention and (ii) the share of
output-dependent pay is higher.

**Implication 2** (p. 967): Dilution provisions are more likely to be tied to
output-dependent pay ($$\Delta w(n)$$ decreasing in $$n$$) if technological shocks
increase firm productivity more at higher retention levels (complements), and
to output-independent pay ($$w(n)$$ decreasing in $$n$$) if shocks increase
productivity more at lower retention levels (substitutes).

**Implication 3** (p. 969): (i) Firms can lower the cost of preventing worker
runs by offering identical workers different compensation types, with
higher-ranked workers receiving a higher share of output-independent pay and
lower-ranked workers a higher share of output-dependent pay. Resource-constrained
firms especially benefit. (ii) If firms seek to retain some workers with higher
probability, they should optimally pay those workers with a higher share of
output-independent pay when shocks are idiosyncratic; with a higher share of
dilutable performance/equity-based pay when outside options are more sensitive
than firm output to common shocks.

Tests of these implications require identifying sensitivity of workers' outside
options to systematic shocks (proxied by comovement with closely related peer
firms, using Hoberg and Phillips (2016) network similarity scores), and the
share of output-dependent pay (proxied by proportion of equity-based pay per
Bergman and Jenter (2007)). Time-vesting equity implements the dilution pattern
of Panel A.3 of Figure 2; equity buyback agreements implement Panel A.1;
retention bonuses implement Panels B and C (p. 968).

## Datasets used

This is a theoretical paper. No datasets are used for estimation. The
paper references management and industry survey evidence to motivate the
setting (turnover contagion studies, industry reports on quit rates and
replacement costs) but does not analyze microdata.

| Dataset / Source | Role | Wiki page |
|---|---|---|
| Industry surveys and management studies (cited in motivation) | Motivational evidence on quit rates, contagion, and replacement costs | No page yet |

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13424) if you are: building
a model of collective turnover or contagious labor-market dynamics; designing
compensation contracts for teams of skilled workers in startups or professional
services; studying the retention properties of equity-based versus fixed pay or
bonus pools; or extending the model to richer information structures (global
games refinements) and renegotiation environments, which are analyzed in the
Internet Appendix.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(2), April 2025, pp. 937-979.
DOI: [10.1111/jofi.13424](https://doi.org/10.1111/jofi.13424).

This distillation was extracted by an LLM on 2026-06-06 and is **not
human-verified or independently reproduced**. The CC BY-NC 4.0 licence permits
non-commercial reproduction with attribution; the verbatim PDF is not hosted
in this batch.

> **Attribution (CC BY-NC 4.0).** Hoffmann, Florian, and Vladimir Vladimirov.
> "Worker Runs." *The Journal of Finance* 80, no. 2 (April 2025): 937-979.
> DOI: 10.1111/jofi.13424. (c) 2025 The Author(s).
> Licensed under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Persuading Investors: Hu & Ma (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/hu-persuading-investors-video-based-2025/
# Distilled: Using machine learning to process 1,139 startup pitch videos across visual, vocal, and verbal dimensions, this paper finds that more positive pitch delivery (the Pitch Factor) raises accelerator funding probability by 3 pp (35.2% from baseline), yet funded startups with higher positivity underperform on every long-run measure, consistent with inaccurate investor beliefs (80%) and preference-based taste (20%) as the mechanism. J. Finance 2025, paywalled. Seven core results with source locators, datasets used, the model, and the method.
# Tags: paper-summary, entrepreneurial-finance, venture-capital, persuasion, text-as-data
==============================================================================

**What this is.** The paper's core results, the method it contributes (a three-V video-processing pipeline with Pitch Factor construction), and the experiment identifying the mechanism: enough to know what it found and how, without reading all 50 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13471).

## TL;DR

The paper asks whether how entrepreneurs deliver a pitch, not just what they say, influences venture investment decisions. Using machine learning to process 1,139 startup pitch videos from five major US accelerators (2010-2019), the authors measure persuasion delivery across three dimensions (visual facial expressions, vocal tone, and verbal word choice) and combine them into a single *Pitch Factor*. A one-standard-deviation increase in Pitch Factor raises accelerator funding probability by 3 percentage points (35.2% from baseline). Yet funded startups with higher positivity systematically underperform on employment, follow-on VC, and IPO/acquisition outcomes. An experiment with 102 MBA students shows that approximately 80% of the persuasion effect runs through inaccurate belief formation (investors mistakenly think passionate pitchers are more likely to succeed) and 20% through taste-based preferences, broadly following the mechanisms framework of DellaVigna and Gentzkow (2010). Gender matters: investors penalize women 9 times more than men for being one standard deviation below average in pitch positivity, building on the gender-bias evidence of Ewens and Townsend (2020), but women are essentially overlooked when co-presenting with men.

## Core results

Magnitudes and significance are as reported; `\*\*\*`/`\*\*`/`\*` = 1%/5%/10%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Pitch Factor strongly predicts funding**: 1-SD increase raises funding probability by 3 pp (35.2% from 8.52% baseline) | Table III col 1, p. 2661 | Marginal effect 0.030 (SE 0.007)\*\*\*; stable across content controls in cols 2-5 |
| R2 | **Full-video Pitch Factor dominates thin-sliced and single-channel measures**: contributes 66.55% of R2 in Shapley-Owen decomposition, versus 16.78% (first-word slice) and 16.66% (random slice) | Table VI Panel A col 6, p. 2667 | Full-video Pitch Factor R2 contribution is 4x that of thin-sliced alternatives |
| R3 | **Gender asymmetry**: the penalty for being 1-SD below average in pitch positivity is 9x larger for women than for men in single-gender teams (coefficients 0.218 vs 0.016); women are statistically irrelevant in mixed-gender teams | Table IX cols 1-4, p. 2675 | Women coeff 0.218 (SE 0.061)\*\*\*; men 0.016 (SE 0.009)\*; pooled difference p=0.058; mixed-gender women coeff -0.002 (SE 0.025) |
| R4 | **High-positivity funded startups grow more slowly**: Pitch Factor negatively predicts employment among all startups | Table VIII Panel A col 1, p. 2673 | Coefficient -0.071 (SE 0.018)\*\*\* (inverse-hyperbolic-sine employment) |
| R5 | **Funded high-positivity startups raise less follow-on VC**: Pitch Factor negatively predicts VC fundraising | Table VIII Panels A-B col 2, p. 2673 | Panel A (full sample): -0.009 (SE 0.003)\*\*\*; Panel B (invested subsample N=270): -0.033 (SE 0.014)\*\* |
| R6 | **Investors form inaccurate beliefs**: in experiment, subjects overestimate survival of high-Pitch-Factor startups by 12.2 pp vs realized outcomes | Table X cols 1, 5, p. 2680 | Predicted mu coefficient 0.023\*\*; realized coefficient -0.099\*; miscalibration = 0.122 |
| R7 | **Beliefs vs taste decomposition**: inaccurate beliefs account for 79.2% and taste/preference for 20.8% of total persuasion effect | Table XI col 4, p. 2681; p. 2682 | Taste kappa=0.061; beliefs-mediated bias=0.232; total bias=0.293 |

**Overall (paper's conclusion).** Noncontent delivery features of persuasive communication have statistically and economically large effects on early-stage venture investment decisions. These features do not help investors make better decisions; the evidence from both archival data and the experiment suggests a bias, particularly leading investors to form inaccurate beliefs. Passionate pitches lower the investment bar in a way that reduces the portfolio's true average success probability. The result complements Bernstein, Korteweg, and Laws (2017), who show hard information matters in early-stage investment, by adding that soft delivery is an independent channel; it parallels findings in Kaplan and Sorensen (2021) that boards overweight interpersonal communication skills. Survey evidence in Gompers et al. (2020) that passion is a top VC selection criterion provides a direct motivation. The approach differs from Huang et al. (2023) by using full dynamic videos with three-V dimensions rather than static photos.

## Theory / model

The paper has no formal economic model for asset pricing, but it formalizes the mechanism via a threshold investment model with belief and taste channels (pp. 2677-2678, equations 5-6).

Investor $$j$$ makes a binary investment decision about startup $$i$$ using pitch delivery features $$\theta_i$$, beliefs about success probability $$\mu_{ij}$$, and confidence $$\sigma_{ij}$$. The investor's utility index from investing is (equation 5, p. 2677):

$$
U(\mu_{ij}, \sigma_{ij}, \theta_i) \equiv \gamma_\mu \mu_{ij} + \gamma_\sigma \sigma_{ij} + \kappa \theta_i \tag{5}
$$

and the investment rule is $$I_{ij} = \mathbf{1}[U_{ij} \geq \bar{U}]$$, with $$\gamma_\mu > 0$$ (investors prefer believing startups will succeed) and $$\gamma_\sigma < 0$$ for a risk-averse agent. The beliefs channel enters through $$\mu$$ and $$\sigma$$, which depend on hard information $$Q_i$$ and pitch delivery $$\theta_i$$:

$$
\mu_{ij} = \lambda_\mu Q_i + \psi_\mu \theta_i \tag{6a}
$$

$$
\sigma_{ij} = \lambda_\sigma Q_i + \psi_\sigma \theta_i \tag{6b}
$$

Under this framework $$\theta_i$$ enters the investment decision both via beliefs (coefficient $$\psi_\mu \gamma_\mu + \psi_\sigma \gamma_\sigma$$) and via a direct preference/taste effect (coefficient $$\kappa$$). The overall empirical coefficient $$\beta$$ in the investment regression is $$\kappa + \psi_\mu \gamma_\mu + \psi_\sigma \gamma_\sigma$$, which the experiment estimates separately by eliciting $$\mu_{ij}$$ and $$\sigma_{ij}$$ directly.

Beliefs are inaccurate when the sign of $$\psi_\mu$$ in equation (6a) is the opposite of the sign of the relation between $$\theta_i$$ and realized startup outcomes: investors who form higher expectations for high-$$\theta_i$$ startups are systematically wrong.

## Method

The core methodological contribution is a three-step video-processing pipeline that simultaneously extracts visual, vocal, and verbal (three-V) information from full pitch videos.

**Step 1: Information representation.** Videos are decomposed into an image stream (sampled at 10 frames per second using Face++ facial landmark detection) and an audio stream (48 kHz, analyzed by the `pyAudioAnalysis` Python package). Verbal content is extracted via Google Cloud Speech-to-Text, producing a time-stamped transcript (pp. 2651-2654).

**Step 2: Measure construction with ML algorithms.** Three sets of measures are computed per speaker per video frame or sentence segment (Table II, p. 2656; Appendix, p. 2683):

- *Visual*: Face++ categorizes facial emotions into six dimensions (happiness, sadness, anger, fear, disgust, neutral), aggregated to `Visual-Positive` and `Visual-Negative`.
- *Vocal*: `pyAudioAnalysis` extracts 34 audio features; LSTM models (speechemotionrecognition) give `Vocal-Positive`/`Vocal-Negative`; SVM models give `Vocal-Arousal` and `Vocal-Valence`.
- *Verbal*: The Loughran-McDonald Master Dictionary gives `Verbal-Positive`/`Verbal-Negative`; the Nicolas-Bai-Fiske NBF social-psychology dictionary gives `Verbal-Warmth` and `Verbal-Ability`.

**Step 3: Aggregation into Pitch Factor.** All per-frame and per-sentence measures are averaged to the video level, then subjected to principal-components factor analysis. The single factor with the highest eigenvalue is the *Pitch Factor* (p. 2658). Factor loadings confirm it captures overall positivity: `Vocal-Arousal` (+0.91), `Vocal-Valence` (+0.88), `Visual-Positive` (+0.08), `Verbal-Warmth` (+0.06), and negative loadings on all negativity measures. The ML algorithm is cross-validated against 115 Amazon Mechanical Turk raters; the algorithm matches human rankings with 89.5% consistency (p. 2659).

The method builds on `text-classification` (Loughran and McDonald (2011), Nicolas, Bai, and Fiske (2019)) and `panel-regression` for the econometric analysis, and proposes a new multi-modal video-processing technique (`pitch-factor-video-ml`) as its primary methodological contribution.

## Empirical specifications

**Baseline funding regression (Table III).** The headline specification is a logit regression with marginal effects (equation 1, p. 2659):

$$
I(\textit{Invested})_{ijt} = \alpha + \beta \cdot \textit{Pitch Factor}_i + \gamma \cdot \textit{Controls}_i + \delta_j + \varepsilon_{ijt} \tag{1}
$$

where $$I(\textit{Invested})_{ijt} = 1$$ if startup $$i$$ was funded by accelerator $$j$$ in year $$t$$, Pitch Factor is standardized (mean 0, SD 1), and Controls include three sets of textual content variables (idea novelty via BERT similarity to PitchBook and 10-K filings; dictionary-based content indicators for cash flow, employment, technology, etc.; LIWC linguistic features). Accelerator fixed effects $$\delta_j$$ account for cross-accelerator heterogeneity. Standard errors are clustered at the accelerator-year level. Sample: N=1,139 pitch videos (Table III, p. 2661).

**Long-run performance regression (Table VIII).** For performance outcomes the specification is OLS (or logit for binary outcomes), equation (4):

$$
\textit{Performance}_i = \alpha + \beta \cdot \textit{Pitch Factor}_i + \gamma \cdot \textit{Controls}_i + \delta_{\text{FE}} + \varepsilon_i \tag{4}
$$

Performance variables include inverse-hyperbolic-sine of employment, binary VC raised, inverse-hyperbolic-sine of VC amount raised, IPO/acquisition indicator, and website update frequency (Wayback Machine). Controls add firm age, squared firm age, industry FEs, and region FEs. SE clustered at the industry level.

**Gender heterogeneity (Table IX).** Same logit specification as equation (1), run separately for men-only (N=559) and women-only teams (N=310), then pooled with Pitch-Factor-times-gender interaction. For mixed-gender teams (N=270), male and female Pitch Factors are computed separately and entered jointly.

**Experiment investment regression (Table XI).** Logit on experimental investment decisions $$I_{ij}$$ (equation 7, p. 2681):

$$
I_{ij} = \underbrace{\kappa \cdot \theta_i}_{\text{Taste}} + \underbrace{\gamma_\mu \cdot \mu_{ij} + \gamma_\sigma \cdot \sigma_{ij}}_{\text{Beliefs}} + \delta_j + \varepsilon_{ij} \tag{7}
$$

where $$\mu_{ij}$$ and $$\sigma_{ij}$$ are directly elicited subject beliefs about P(alive|invested) and confidence. Subject fixed effects and startup/team and content controls included. SE two-way clustered at startup and subject levels. N=952 experimental investment rounds (102 subjects times 10 videos minus incomplete observations, p. 2679).

**Omitted-variable test (Table VII, Oster 2019).** To test whether omitted founder quality drives results, the authors apply the Oster (2019) delta/R-max test. For the baseline parameterization ($$R^2_{\max} = \min(2.2 R^2_c, 1)$$, $$\delta=1$$), the identified set is [0.021, 0.023], excluding zero. Even at $$\delta=2$$ and $$R^2_{\max}=1$$ the identified set is [0.006, 0.023], rejecting the null. The $$\delta$$ required to make $$\beta_{adj}=0$$ is 8.06 under the baseline parameterization.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| 1,139 startup pitch videos (YouTube, Vimeo, hand-collected) | Primary source of visual, vocal, verbal delivery measures via ML processing | no page yet |
| Crunchbase and PitchBook | Startup characteristics: founding year, industry, location, funding rounds and amounts, investor count | [PitchBook](/wiki/commercial/pitchbook/) (licensed) |
| LinkedIn (via API) | Founder backgrounds: education, work experience, prior entrepreneurship; used for startup/team quality controls | no page yet |
| Wayback Machine (Internet Archive) | Startup survival measure: website update frequency over three years post-application | [Wayback Machine](/wiki/datasets/wayback-machine/) |
| Amazon Mechanical Turk survey | Human validation of Pitch Factor ratings (N=115 raters, 89.5% algorithmic-human consistency) | no page yet |
| Yale SOM MBA experiment (N=102 subjects) | Elicited investor beliefs and investment decisions for mechanism decomposition | no page yet |

Sample: 1,139 accelerator applications to Y Combinator, MassChallenge, 500 Startups, Techstars, and AngelPad, spanning 2010-2019. Startup performance tracked as of July/August 2023.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13471) if you are: (i) building a video-processing pipeline for economic research and need the detailed ML algorithm and hyperparameter choices (Internet Appendix II); (ii) studying gender bias in VC and need the full heterogeneity analysis across single-gender and mixed-gender teams; (iii) extending the belief-vs-taste decomposition framework to other persuasion settings; or (iv) replicating the sample selection tests (Table V) or the university-incubator robustness sample. The locators above point to the exact tables and figures.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(5). This distillation was extracted by an LLM on 2026-06-05 and is **not human-verified or independently reproduced**. The article is paywalled (Wiley terms and conditions for VOR; not CC-licensed). Extract-only.

> Hu, Allen, and Song Ma. "Persuading Investors: A Video-Based Study."
> *The Journal of Finance* 80, no. 5 (October 2025): 2639-2688.
> DOI: 10.1111/jofi.13471. (c) 2025 the American Finance Association.


==============================================================================
# The Global Credit Spread Puzzle: Huang, Nozawa & Shi (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/huang-global-credit-spread-puzzle-2025/
# Distilled: Structural credit risk models systematically underpredict investment-grade corporate bond spreads over government bonds and swap rates across eight developed economies, constituting a global credit spread puzzle. Incorporating endogenous bond market illiquidity via a He-Milbradt search model substantially mitigates the puzzle and raises individual-bond cross-sectional fit in every country. J. Finance 2025, CC BY-NC-ND 4.0. Eight core results with source locators, datasets used, the models (BC, CDG, HM), and the estimating specifications.
# Tags: paper-summary, asset-pricing, credit-risk, fixed-income, corporate-bonds
==============================================================================

**What this is.** The paper's core results, the models it tests (Black-Cox, Collin-Dufresne-Goldstein, He-Milbradt), and the specifications behind each result: enough to know what it found and how, without reading all 62 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13409).

## TL;DR

The paper asks whether the U.S. credit spread puzzle (CSP) extends globally. Huang and Huang (2012) documented that structural models underpredict U.S. IG corporate-Treasury spreads; this paper tests whether the same pattern holds outside the United States. Using security-level pricing data on IG corporate bonds from eight developed economies and global default data, the authors implement two standard pure-default structural models (Black and Cox (1976) and Collin-Dufresne and Goldstein (2001)) and a reduced-form variant of the He and Milbradt (2014) model with endogenous bond market illiquidity, building on the OTC search-and-bargaining framework of Duffie, Garleanu, and Pedersen (2005). The default boundary is estimated via three methods including those of Feldhutter and Schaefer (2018) and Bai, Goldstein, and Yang (2020). The paper finds robust evidence that pure default-risk structural models tend to underpredict IG credit spreads over swap rates and, even more so, over government bond yields in all countries except Japan, establishing a "global credit spread puzzle" (GCSP). The CDG model improves overall performance but does not resolve the puzzle. However, incorporating endogenous OTC search-and-bargaining frictions into the BC model substantially mitigates the GCSP and raises the model's cross-sectional R-squared for individual IG bond spreads in every country from 19-35% (BC model) to 34-79% (HM model).

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **BC model significantly underpredicts IG credit spreads over swap rates** across most IG/country groups; constitutes a global credit spread puzzle | Table IV, pp. 122-123; Table V, p. 126; Table IX Panel A1, p. 151 | Significantly negative MPE for 53 of 72 IG/country/d bins; at country level, underpredicts for 17 of 24 IGctry/d bins |
| R2 | **Underprediction is substantial in AUS, CAN, GBR, ITA, and USA**; BC model overpredicts for most JPN groups | Table IV Panel A, pp. 122-123 | AUS MPEs for AA+, A, BBB under BC(dFS): (-98, -125, -151) bps; GBR: (-27, -39, -92) bps; JPN AA+: +12.6 bps |
| R3 | **GCSP is robust to the CDG model with stationary leverage ratios**; CDG improves mean pricing but still underpredicts for 16 of 24 IG/country bins | Table IV (CDG rows), pp. 122-124; Table IX Panel A1, p. 151 | CDG underpredicts for 16 IG/country bins vs 17 for BC(dFS); underpredicts for 6 of 8 IGctry groups vs 6 for BC(dFS) |
| R4 | **HM model with endogenous illiquidity substantially mitigates the GCSP**: bins with significantly negative MPE drop from 17 (BC) to 6 (HM) | Table IV (HM rows) pp. 122-123; Table V p. 126; Table IX Panel A1, p. 151 | GBR MPEs narrow from (-27,-39,-92) to (1,13,-26) bps; AUS from (-98,-125,-151) to (6,-15,-10) bps; puzzle disappears in FRA, DEU, ITA, and USA for industrial issuers excluding negative spreads |
| R5 | **HM model substantially raises cross-sectional R2** for individual IG bond spreads in every country | Table IX Panel A2, p. 152 | HM R2: 34% (FRA) to 79% (AUS); BC R2: 19% (AUS) to 35% (CAN); slope coefficient closer to 1 under HM in every country except AUS |
| R6 | **HM model also captures time-series variation in IG spreads better** than BC or CDG in every country | Table IX Panel A3, p. 152 | Time-series correlation rho^IG under HM(dFS): 0.65 (FRA) to 0.94 (GBR) vs 0.55 (FRA) to 0.90 (GBR) under BC(dFS) |
| R7 | **Government bond yield-based GCSP is stronger** than the swap rate-based version and poses a tougher challenge to the HM model | Table IX Panel B1, p. 153 | BC underpredicts for 67 of 72 IG/country/d bins (govt) vs 53 (swap); HM overcomes underprediction for only 9 of 21 IG/country bins (govt) vs 13 of 19 bins (swap, excl. neg. spreads) |
| R8 | **HM model raises R2 for government yield-based spreads** in every country except JPN; gains are largest in AUS and ITA | Table IX Panel B2, p. 153 | R2 under HM(dFS, Govt): 0.08 (CAN) to 0.75 (AUS); R2 under BC(dFS, Govt): 0.02 (CAN) to 0.30 (ITA) |

**Overall (paper's conclusion).** Pure default-risk structural models underestimate IG corporate bond credit spreads in global developed bond markets, especially over government bond yields. Incorporating mean-reverting leverage (CDG) provides limited relief. Incorporating OTC search-and-bargaining frictions (HM) substantially mitigates the GCSP for both swap rate-based and government bond yield-based spreads, and improves the cross-sectional fit of individual bond spreads in every country, suggesting that corporate bond illiquidity is a central missing ingredient in standard structural pricing models.

## Theory / model

The paper tests two classes of structural models.

**Baseline: Black-Cox (BC) model** (p. 110-112). Consider a corporate bond with fixed maturity $$T$$, face value $$K$$, and continuous coupon rate $$c$$. Default occurs when firm value falls to a flat default boundary for the first time before or at $$T$$. Under constant risk-free rate $$r$$, the bond price at time $$t$$ is (equation (1), p. 112):

$$
D^{BC}(t, T) = \frac{cK}{r} + e^{-r(T-t)} K \left(1 - \frac{c}{r}\right)(1 - \pi^Q(t,T)) + K\!\left(R - \frac{c}{r}\right) G(t,T), \tag{1}
$$

where $$\pi^Q(t,T)$$ is the risk-neutral default probability over $$(t, T]$$, $$G(t,T)$$ is the time-$$t$$ price of the Arrow-Debreu default claim, and $$R$$ is the state-dependent recovery rate. The model-implied yield $$y(t,T)$$ solves:

$$
D^{BC}(t,T) = \frac{cK}{y}\left(1 - e^{-y(T-t)}\right) + Ke^{-y(T-t)}, \tag{2}
$$

and the model-implied credit spread is $$s^{BC}(t,T) = y(t,T) - r$$.

**CDG model** (p. 136-137). Collin-Dufresne and Goldstein (2001) augment the BC model to allow stationary leverage ratios. The total debt level $$K_t$$ follows the process (equation (4), p. 137):

$$
d\ln K_t = \kappa\left[-\nu - \ln(K_t/A_t)\right]dt, \tag{4}
$$

where $$\kappa$$ controls the speed at which log-leverage reverts to the target ratio under the risk-neutral measure: $$\ln(\bar{L}^Q) \equiv \frac{-r + \delta + (\sigma^A)^2/2}{\kappa} - \nu$$. Parameters $$\kappa$$, $$\nu$$, and $$\bar{L}^Q$$ are estimated via GMM on single-name CDS spreads for each country.

**HM model with endogenous illiquidity** (Section IV, pp. 138-148). The paper adapts the He and Milbradt (2014) framework to study the incremental effect of OTC search-and-bargaining frictions. L-type investors (hit by liquidity shocks, holding-cost $$\chi$$) and H-type investors (not hit) trade the bond at Poisson intensities $$\lambda$$ (customer-to-dealer) and $$\beta$$ (backward L-to-H transition). The H-type and L-type bond valuation functions $$D_H(t,T)$$ and $$D_L(t,T)$$ satisfy (equation (5), p. 139):

$$
\begin{bmatrix} D_H(t,T) \\ D_L(t,T) \end{bmatrix} = Z^{-1} \begin{bmatrix} c \\ c - \chi \end{bmatrix} K + e^{-Z(T-t)}\!\left(\begin{bmatrix} K \\ K \end{bmatrix} - Z^{-1}\begin{bmatrix} c \\ c-\chi \end{bmatrix}\! K\right)(1 - \pi^Q(t,T)) + U G(t,T) U^{-1}\!\left(\begin{bmatrix} R_H K \\ R_L K \end{bmatrix} - Z^{-1}\begin{bmatrix} c \\ c-\chi \end{bmatrix}\! K\right), \tag{5}
$$

where $$Z$$ is the $$2\times 2$$ matrix of liquidity-adjusted discount factors, $$U$$ diagonalizes $$Z$$, $$\{R_H, R_L\}$$ are type-dependent recovery rates, and $$\pi^Q(t,T)$$ and $$G(t,T)$$ are the same risk-neutral default objects as in the BC model. Specifically:

$$
Z = \begin{bmatrix} r + \xi & -\xi \\ -\lambda\beta & r + \lambda\beta \end{bmatrix} = U \cdot \begin{bmatrix} \widetilde{r}_1 & 0 \\ 0 & \widetilde{r}_2 \end{bmatrix} \cdot U^{-1}, \tag{6}
$$

where $$\xi$$ is the liquidity shock intensity, $$\lambda$$ is the investor-to-dealer meeting intensity, and $$\beta$$ is the bargaining power of investors vis-a-vis dealers. The bid price is $$D_B^{HM}(t,T) = \beta D_H(t,T) + (1-\beta) D_L(t,T)$$ (equation (10), p. 139). The model shares the same $$\mathbb{P}$$-measure default probabilities as BC, so the HM model-implied yield spread decomposes into a BC-implied credit component and a liquidity component from search frictions.

**Identification.** The paper is a descriptive empirical study: no causal claim is made. The BC/CDG models are calibrated to historical default and equity data (no estimation of pricing errors from causal variation). The HM search parameters $$\theta^S = \{\xi, \lambda, \beta, \chi_k, \chi_c\}$$ are estimated country-by-country by minimizing the sum of squared fitting errors to observed BGN proportional bid-ask spreads (equation (14), p. 141), keeping firm-level BC fundamentals fixed. The BC and HM models share the same $$\mathbb{P}$$-measure default probability, so their yield differential isolates the incremental contribution of search frictions.

## Method

The paper evaluates three nested structural models:

1. **BC model** (pure default risk): firm-level parameters $$K/A_t$$, $$\sigma^A$$, $$\delta$$ estimated at the bond-level from Compustat/equity data; country-level Sharpe ratio SR estimated as median across Compustat firms; default boundary $$d$$ estimated via three methods (FS, BGY, HNS) matching physical default probabilities to historical data.

2. **CDG model** (stationary leverage): augments BC with GMM estimation of $$\kappa$$, $$\nu$$, $$\bar{L}^Q$$ using single-name CDS spreads at 1, 2, 3, 5, 7, and 10 years by country and rating (Table VII, p. 137).

3. **HM model** (endogenous illiquidity): fixes BC firm fundamentals and estimates the five search-friction parameters $$\theta^S$$ at the country level by least squares on proportional BGN bid-ask spreads (equation (14), p. 141):

$$
\widehat{\theta}^S = \arg\min_{\theta^S} \sum_t \sum_i \left(\phi(t, T_i; \theta^S) - \phi_{i,T_i}^{obs}\right)^2, \tag{14}
$$

where the model-implied proportional bid-ask spread is $$\phi(t,T;\theta^S) = \frac{D_A^{HM}(t,T) - D_B^{HM}(t,T)}{(D_A^{HM}(t,T) + D_B^{HM}(t,T))/2}$$ (equation (13), p. 141).

Asset volatility $$\sigma^A$$ is estimated following Schaefer and Strebulaev (2008) (equation (3), p. 113):

$$
\sigma^A_{i,t} = \sqrt{(1 - L_{i,t})^2 (\sigma^E_{i,t})^2 + L_{i,t}^2 (\sigma^D_{i,t})^2 + 2(1-L_{i,t}) L_{i,t} \sigma^E_{i,t} \sigma^D_{i,t} \rho^{ED}_{i,t}}, \tag{3}
$$

where $$L_{i,t}$$ is quasi-market leverage, $$\sigma^E_{i,t}$$ is annualized equity volatility from daily stock returns, $$\sigma^D_{i,t}$$ is debt volatility, and $$\rho^{ED}_{i,t}$$ is the stock-bond return correlation.

## Empirical specifications

**Mean pricing error (MPE) test (R1-R4, R7).** The headline test computes the mean pricing error for each credit rating/country/$$d$$-estimate bin:

$$
\text{MPE} = \overline{s^M(t,T)} - \overline{s^{obs}(t,T)},
$$

where $$s^M$$ is model-implied spread and $$s^{obs}$$ is observed spread (over swap rates or government bond yields). The MPE is computed for 72 IG/country/$$d$$ bins (24 country-rating bins times three $$d$$ estimates) and 24 IGctry/$$d$$ bins (country-level aggregates). Standard errors are clustered by bond issue; significance reported at 1%, 5%, 10% two-tailed. The test is run on: (i) the full sample including negative credit spreads; (ii) excluding observations with negative spreads over swap rates; (iii) industrial issuers only. Main results in Tables IV (bond-level, pp. 122-124) and V (country-level, pp. 126-127). Summary in Table IX (pp. 151-153).

**Bond-level panel regression (R5, R8).** To assess the cross-sectional fit of individual IG bond spreads, the paper runs panel regressions of monthly observed spreads on model-implied counterparts for each country, recovering the slope coefficient and $$R^2$$:

$$
s^{obs}_{i,t} = a + b \cdot s^M_{i,t} + \varepsilon_{i,t},
$$

with standard errors clustered by bond issue (Table IX Panels A2 and B2, p. 152-153). The slope under BC($$d^{FS}_{swap}$$) ranges from 0.28 (FRA) to 0.87 (AUS), with $$R^2$$ from 0.19 to 0.35. Under HM, the slope is closer to one in every country except AUS, and $$R^2$$ rises to 0.34-0.79.

**Time-series correlation test (R6).** The time-series correlation $$\rho^{IG}$$ between monthly mean observed and predicted IG spreads is computed for each country-model pair. Under BC($$d^{FS}_{swap}$$), $$\rho^{IG}$$ ranges from 0.55 (FRA) to 0.90 (GBR). Under HM($$d^{FS}_{swap}$$), $$\rho^{IG}$$ ranges from 0.65 (FRA) to 0.94 (GBR) (Table IX Panel A3, p. 152).

**Search-friction parameter validation.** The estimated HM search parameters $$\widehat{\theta}^S$$ are validated against independent empirical proxies (Figure 6, pp. 144-145): $$\widehat{\xi}$$ (liquidity shock intensity) correlates positively with the mutual fund share of corporate bond ownership; $$\widehat{\chi}_k + \bar{c}\widehat{\chi}_c$$ (holding cost) correlates with forced-selling costs at rating downgrades and index exclusions; $$\widehat{\lambda}$$ (dealer-meeting intensity) correlates with the scaled number of dealers quoting each bond from the Markit Bond Pricing Database; $$\widehat{\beta}$$ (bargaining power) is negatively correlated with downgrade frequency.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| ICE BofAML Global Corporate Index and High Yield Index (via Mercury/Bank of America ML) | Monthly bond prices, credit ratings, maturity for IG and HY bonds in 8 countries | No page yet |
| Compustat Global / Compustat NA | Firm balance sheet data (book debt, market equity, book-to-market); merged by issuer name to bond data | [WRDS / Compustat](/wiki/commercial/wrds/) (licensed) |
| CRSP (U.S.) | Stock prices and returns for U.S. firms | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| Bloomberg (BGN bid-ask prices, bond characteristics, shareholder data) | Identification of callability, seniority, security; screening of state-owned firms; proportional bid-ask spreads for HM estimation | [Bloomberg](/wiki/commercial/bloomberg/) (licensed) |
| Markit Bond Pricing Database | Daily trader quotes for individual bonds; number of dealers quoting each bond | [Markit bond pricing](/wiki/commercial/markit/) (licensed) |
| Markit single-name CDS spreads | Five-year CDS spreads for CDG GMM estimation; CDS-implied LGD proxies | [Markit CDS](/wiki/commercial/markit-cds/) (licensed) |
| Moody's Default and Recovery Database (DRD) | Historical issuer-weighted default and recovery rates by rating category and region (1970-2017) | No page yet |
| IHS Markit Bond Pricing Database | Number of distinct quotes and contributing dealers (for lambda proxy) | [Markit bond pricing](/wiki/commercial/markit/) (licensed) |
| Global Financial Data (stock market indexes by country) | TOPIX (JPN), FTSE100 (GBR), DAX (DEU), CAC40 (FRA), FTSE MIB (ITA), TSX Composite (CAN), S&P/ASX200 (AUS) for SR estimation | No page yet |
| OECD macroeconomic data | Macroeconomic covariates for pricing error regressions | [data:fred](/wiki/datasets/fred/) (partial; OECD is a separate source) |
| FRED (Federal Reserve Economic Data) | Additional macroeconomic variables | [FRED](/wiki/datasets/fred/) |
| Lehman Brothers Fixed Income Database + ML U.S. Corporate Bond Database | U.S. corporate bond prices 1987-2015 for U.S. subsample | No page yet |
| Barclays Live (swap rates) | IRS LIBOR swap rates for default-free benchmark construction | No page yet |

Sample: January 1997 to December 2017 for non-U.S. countries (except Italy from 2003 and Australia from 2007); 1987-2015 for U.S. bonds. Monthly frequency.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13409) if you are: replicating the global CSP evidence across multiple countries and rating categories (Tables IV, V for country-by-country MPEs); implementing the HM bond pricing model outside the U.S. (Section IV and Tables VIII, IX for parameter estimates and model validation); building structural credit models for non-U.S. markets; decomposing corporate bond yield spreads into default and liquidity components across eight economies; or examining cross-sectional determinants of IG bond spreads using panel regressions by country.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(1). This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The CC BY-NC-ND 4.0 licence permits sharing with attribution for non-commercial purposes; no derivatives; the verbatim PDF is not hosted in this batch.

> **Citation.** Huang, Jing-Zhi, Yoshio Nozawa, and Zhan Shi.
> "The Global Credit Spread Puzzle."
> *The Journal of Finance* 80, no. 1 (February 2025): 101-162.
> DOI: 10.1111/jofi.13409. © 2024 The Author(s).
> Licensed under [CC BY-NC-ND 4.0](https://creativecommons.org/licenses/by-nc-nd/4.0/).
> This page is a distillation by the Institute for Automated Research: core results extracted and re-expressed; extract-only, not reproduced.


==============================================================================
# Private Equity and Financial Stability: Johnston-Ross, Ma & Puri (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/johnston-ross-private-equity-financial-stability-2025/
# Distilled: Using proprietary FDIC failed-bank bidding data and a quasi-random close-bid design, Johnston-Ross, Ma, and Puri show that PE investors filled the capital gap in the 2008 crisis by acquiring riskier failed banks that incumbent banks avoided, and that PE-acquired banks outperformed on branch preservation, deposit growth, small business lending, and regional employment recovery. J. Finance 2025, paywalled. Seven core results with source locators, datasets used, the empirical design, and the estimating equations.
# Tags: paper-summary, banking, private-equity, financial-stability, bank-failures
==============================================================================

**What this is.** The paper's core results, the empirical design (quasi-random close-bid identification), and the estimating equations: enough to know what was found and how, without reading the full 48 pages. To replicate or extend, read the original at [doi.org/10.1111/jofi.13399](https://doi.org/10.1111/jofi.13399).

## TL;DR

Using proprietary FDIC failed-bank bidding records for 482 bank failures resolved between 2009 and 2014, Johnston-Ross, Ma, and Puri document that PE investors systematically acquired the riskier, more undercapitalized failed banks in regions where neighboring banks were also distressed, filling a funding gap that incumbent banks could not. A close-bid quasi-random design (48 auctions where the PE and bank bids differed by less than 5% of assets) shows that PE-acquired banks subsequently closed fewer branches, grew deposits 35 percentage points faster over three years, expanded small business lending by 32%, lowered SBA loan rates by about 32 basis points, and supported faster regional employment and income growth, all without incurring higher FDIC loss-share claims. PE investors held failed banks for 6.5 years on average before selling, mostly to local banks, earning roughly a 12% IRR.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | PE-acquired bank branches are **14.8 pp less likely to close** within three years post-acquisition | Table VI, col (1), p. 190; quasi-random sample | coef = -0.148\*\*\* (SE = 0.037); base rate = 20.1%; 73.6% reduction |
| R2 | PE-acquired banks are **7.0 pp less likely to close and exit a county entirely** | Table VI, col (2), p. 190 | coef = -0.070\*\* (SE = 0.029) |
| R3 | **Deposit growth 35.6 pp higher** in PE-acquired banks over three years post-acquisition | Table VII, col (2), p. 193; quasi-random sample, two-failed-bank county pairs | coef = 0.356\*\*\* (SE = 0.111) |
| R4 | **SBA loan count grows 32% more** in counties where PE acquires the failed bank | Table VIII, Panel A, col (1), p. 196 | coef = 0.320\*\* (SE = 0.112); R2 = 0.797 |
| R5 | **SBA loan interest rates 32 bp lower** in PE-acquired counties | Table VIII, Panel A, col (3), p. 196 | coef = -0.323\* (SE = 0.147); noncond. mean = 6.42% |
| R6 | **Employment 6.5 pp higher and total personal income 1.5 pp higher** in PE-acquired counties | Table IX, Panel A, cols (2) and (3), p. 197 | employment coef = 0.065\*\* (SE = 0.029); income coef = 0.015\*\*\* (SE = 0.006) |
| R7 | **PE-acquired banks do not claim more FDIC loss-share losses** than bank acquirers | Table X, col (1), p. 199 | coef = -0.015 (t = -0.544), indistinguishable from zero |

**Overall (paper's conclusion).** PE investors complemented incumbent banks in the failed-bank market by acquiring the riskier, harder-to-sell banks that local banks were unwilling or unable to acquire. They stabilized those banks, preserved branch networks, grew deposits, increased small business lending, and contributed to regional economic recovery, without generating excess costs for the FDIC. PE participation in failed-bank resolution helped reduce the FDIC's cost of crisis resolution by an estimated $3.63 billion (roughly 5% of the DIF's total crisis losses).

## Theory / model

The paper has no formal structural model. It tests two complementary hypotheses motivated by the PE literature and the institutional setting of FDIC failed-bank auctions.

**Hypothesis 1 (Complementary selection).** PE investors have higher risk tolerance and more stable funding than distressed local banks, so they should bid on and acquire the riskier segment of the failed-bank market that incumbent banks cannot or will not purchase. The prediction is that PE-acquired banks are more undercapitalized, have riskier loan portfolios (higher C&D and OREO shares), lower profitability, and are located in regions where neighboring banks are also in distress (pp. 163-165).

**Hypothesis 2 (Positive turnaround).** PE investors' expertise in distressed firms (Hotchkiss, Smith, and Stromberg (2021)) and their more stable funding (Bernstein, Lerner, and Mezzanotti (2019)) allow them to turn around failed banks, so PE-acquired banks should outperform bank-acquired failed banks on branch stability, deposit growth, lending, and real outcomes. The paper also draws on the bank value decomposition framework of Egan, Lewellen, and Sunderam (2022) to attribute deposit gains across pricing, branch network, and productivity channels (pp. 194-195). The confounding threat is selection: PE acquires worse banks, so raw comparisons would understate or reverse the effect (p. 186).

**Identification strategy.** The paper resolves the selection problem via a close-bid quasi-random design developed on the foundation of Granja, Matvos, and Seru (2017), who show that local bank capitalization is the key determinant of failed-bank acquirer identity. For failed banks bid on by both PE investors and banks, the paper restricts to auctions where the winner's margin over the cover bid is less than 5% of total bank assets. Within this subsample, whether a PE investor or a bank wins the auction is treated as quasi-random, and Table V shows that the 25 PE-won and 23 bank-won banks in the sample are statistically identical across dozens of pre-auction characteristics (Table V, pp. 188-189).

## Method

**PE-acquisition probability (Section III).** Equation (1) (p. 177) is a logit regression:

$$
\Pr(PE = 1) = \Phi\!\left(\alpha + \beta \cdot X_i + \gamma \cdot \text{Control}_i + \theta_t + \varepsilon_i\right) \tag{1}
$$

where $$X_i$$ is a bank characteristic (tier 1 capital ratio, core deposits, net interest margin, C&D loans, OREO), $$\text{Control}_i$$ includes log total assets, and $$\theta_t$$ are year-quarter fixed effects. The sample is the cross-section of 456 failed banks resolved 2009-2014. Marginal effects at the sample mean are reported (Table III). The builds from `logit-regression` for the selection analysis and from `panel-regression` for the performance analysis.

**Post-acquisition performance (Section IV).** Equation (2) (p. 187) is a local linear panel regression:

$$
\text{Performance}_{b,i,t,z} = \alpha + \beta \cdot PE_i + \gamma \cdot \text{Control}_i + \theta_{t \times z} + \varepsilon_i \tag{2}
$$

where $$\text{Performance}$$ is the outcome (branch closure indicator, deposit growth, SBA loan count/amount/rate, employment growth, income growth) for branch $$b$$ of failed bank $$i$$ that failed in year $$t$$ in region $$z$$, $$PE_i$$ equals 1 if a PE investor won the auction, and $$\theta_{t \times z}$$ are state-by-failed-year fixed effects absorbing local time trends. Standard errors are double-clustered at the state and failed-year levels. The estimation uses the quasi-random subsample (48 auctions) as the preferred specification.

## Empirical specifications

**Selection analysis (Table III).** Cross-sectional logit on 456 failed banks with failed year-quarter fixed effects and robust standard errors. Panel A: dependent variable = PE acquired (1/0); regressors = bank-level health measures (tier 1 ratio, core deposits, net interest margin, C&D loans, OREO). Panel B: same but regressors = neighboring bank health measures (neighboring tier 1 ratio, noncurrent loans, OREO, number of large local banks, number of failed banks in state). Ties results to R1-R2 (Table III, pp. 180-182).

**Branch closures (Table VI).** Panel regression at the branch level. The preferred specification (columns 1-2) uses the quasi-random sample. Dependent variable: indicator = branch closed within three years, or branch closed and county exited. Controls include tier 1 capital, core deposits, C&D loans, OREO, and log assets. Fixed effects: state x failed-year. Standard errors double-clustered at state and failed-year levels (Table VI, p. 190).

**Deposit growth (Table VII).** Panel regression at the bank-county level. Dependent variable: one-year or three-year change in branch-level deposits. For PE acquirers using shelf charters, all branches in the local region are counted; for inflatable charters or multiple acquisitions, combined PE and acquirer branches are included to address consolidation effects. An alternative specification restricts to counties where the acquirer has no pre-existing branch (columns 5-6, confirming results hold without overlap) (Table VII, p. 193).

**Small business lending (Table VIII).** County-level regression on three-year growth in SBA 7(a) loan number, amount, interest rate, and average loan size. Quasi-random sample: 276 county observations. Full sample: 2,181 observations. State x failed-year fixed effects, double-clustered standard errors. Specification matches equation (2) (Table VIII, pp. 195-196).

**Regional economic recovery (Table IX).** County-level panel on three-year growth in startup employment (Census QWI, firm age 0-1), total employment, total personal income (IRS SOI), and per capita income. Same fixed effects and SE clustering as Table VIII. Both quasi-random (276 obs) and full sample (2,181 obs) reported (Table IX, p. 197).

**Loss share claims (Table X).** Bank-level regression on aggregate claimed loss rate (total losses / covered assets) and incurred loss rate (losses net of FDIC reimbursements). Uses proprietary FDIC loss-share portfolio data. Controls match equation (1). State and failed-year fixed effects. 304 observations (full); 38 observations in quasi-random subsample (Table X, p. 199).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| FDIC proprietary failed-bank bidding data (P&A transaction records, bid values, acquirer identities, FDIC least-cost estimates) | Core identification: close-bid quasi-random design; failed-bank selection analysis | no page yet |
| FDIC Call Reports (Consolidated Reports of Condition and Income) | Failed-bank and neighboring-bank financial characteristics (tier 1 capital, loan composition, deposits) | [Call Reports](/wiki/datasets/call-reports/) |
| FDIC Summary of Deposits (SOD) | Branch-level deposit balances and locations for closure and deposit growth analysis | [FDIC Summary of Deposits](/wiki/datasets/fdic-summary-of-deposits/) |
| FDIC Reports of Structure Change | Branch openings/closings; county exit analysis | no page yet |
| FDIC proprietary loss-share data | Loss-share claims by acquirer type; Table X | no page yet |
| Preqin (PE fund data) | PE fund size, vintage, first-time fund indicator; consortium-level PE ownership | [Preqin](/wiki/commercial/preqin/) (licensed) |
| RateWatch | Branch-level deposit interest rates for deposit rate analysis (Panel B, Table VII) | no page yet |
| U.S. Census Quarterly Workforce Indicators (QWI) | County-level startup employment and total employment for regional recovery analysis | [QWI](/wiki/datasets/qwi-census/) |
| IRS Statistics of Income (SOI) | County-level personal income and per capita income | no page yet |
| Small Business Administration (SBA) 7(a) loan data | Number, amount, interest rate, and average size of small business loans by county | [SBA loans](/wiki/datasets/sba-loans/) |
| S&P Global Market Intelligence | PE exit deal values and IPO data; IRR calculation | no page yet |

Sample: 456 failed banks resolved via Purchase and Assumption transactions 2009-2014; 62 PE-acquired, 393 bank-acquired (27 excluded: no least-cost bid or bridge-bank status). Quasi-random sample: 48 banks from close-bid auctions (bid difference below 5% of total bank assets), of which 25 PE-acquired and 23 bank-acquired.

## When to read the full paper

Read the original (link above) if you are: (i) building a related quasi-random design for financial-institution interventions and want the balance test methodology (Table V, pp. 187-189); (ii) studying how PE ownership structures (shelf vs. inflatable charters, BHC formation) interact with bank regulation; (iii) quantifying the FDIC cost savings from PE participation (the $3.63 billion back-of-envelope, pp. 202-203); or (iv) studying the management channel via hand-collected CEO characteristics (Table XII, p. 202). The Internet Appendix (referenced at p. 210) contains additional robustness tables (IA.I through IA.XIII) and the replication code.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(1), February 2025. Published by Wiley on behalf of the American Finance Association. Copyright 2024 the American Finance Association. This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The article is paywalled; no CC licence is present in the Crossref metadata. This page contains extracted summary information only (extract-only).

> Johnston-Ross, Emily, Song Ma, and Manju Puri. "Private Equity and Financial Stability: Evidence from Failed-Bank Resolution in the Crisis." *The Journal of Finance* 80, no. 1 (February 2025): 163-210. DOI: 10.1111/jofi.13399.


==============================================================================
# Going for Broke: de Jong, Kooijmans & Koudijs (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/jong-going-broke-bank-reputation-2025/
# Distilled: Using 18th-century Dutch plantation mortgage-backed securities, this paper shows high-reputation banks originated better mortgages and issued securities retaining 17.5 percentage points more value during market collapse, with the effect attenuated when bankers were shielded from downside risk or had short-run profit focus. J. Finance 2025, CC BY-NC-ND 4.0. Eight core results with source locators, datasets used, the model (banker reputation and MBS quality), and the method (mediation analysis, OLS with MBS fixed effects).
# Tags: paper-summary, financial-intermediation, securitization, bank-reputation
==============================================================================

**What this is.** The paper's core results, the stylized model motivating the empirical analysis, and the estimating specifications with their equations: enough to know what was found and how, without reading all 50 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1111/jofi.13503).

## TL;DR

The paper asks whether bank reputation can improve security quality in opaque markets. It studies 37 Amsterdam merchant banks that securitized West Indian plantation mortgages between 1753 and 1772, the first large-scale mortgage-backed securities market on record. High-reputation banks (measured by the rental value of their office premises, capturing outside business at stake) originated better-quality mortgages and issued securities that retained on average 17.5 percentage points more value during the market collapse. Virtually all of this premium can be traced to better mortgage characteristics at origination, not ex-post behavior. The reputation effect is significantly attenuated for bankers who were married into wealth (shielded from downside risk) or who had a short-run profit focus (partner died with a minor heir). These findings are consistent with a partial-equilibrium model predicting that reputation disciplines behavior only when bankers are personally exposed to long-run reputational losses.

The results contrast sharply with evidence from modern securitization markets. Griffin, Lowery, and Saretto (2014) show that high-reputation banks continued issuing large volumes of poorly performing MBS in the 2000s. Piskorski, Seru, and Witkin (2015) find that similar misrepresentation rates prevailed regardless of bank reputation in modern RMBS. The paper argues these modern failures reflect structural differences: limited liability, bailout expectations, and short-term incentives removed the personal downside exposure that makes reputation effective. On the theoretical side, Winton and Yerramilli (2021) model reputation as a disciplining device in originate-to-distribute lending, providing the framework this paper empirically supports. Hartman-Glaser (2017) shows how reputation can instead lead to pooling equilibria where opportunistic types mimic good types, a channel this paper finds limited evidence for in the historical setting. The companion data paper, de Jong, Kooijmans, and Koudijs (2023), documents the plantation MBS data set and provides evidence on the full intermediation chain. Flandreau and Flores (2009) document a related reputation-quality link in 19th-century sovereign bond markets.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | High-reputation banks originated mortgages with lower LTVs during the boom | Table II, p. 3293 | LTV boom: high-rep ~0.56 vs low-rep 0.62 (SD 0.09); difference -0.06\*\*\* (t = -2.67) |
| R2 | High-reputation banks used elite agents and borrowers more during the boom | Table II, p. 3293 | Nonelite agent boom diff: -0.22\*\* (t = -2.08); nonelite borrower boom diff: -0.35\*\*\* (t = -3.58) |
| R3 | MBS issued by high-reputation banks retained 17.5 percentage points more value in the bust | Table III col. (1), p. 3295 | High-reputation dummy coeff: 17.46\*\*\* (t = 5.10); R² = 0.72; N = 4,605 transactions, 46 MBS |
| R4 | The reputation-price gap is robust to all alternative reputation measures | Table III cols. (2)-(9), p. 3295 | Continuous office value: 24.25\*\*\* (t = 5.14); city government: 14.17\*\*\* (t = 5.85); log ABE vol: 1.63\* (t = 0.85); log N deeds: 2.10 (t = 2.03) |
| R5 | Mortgage characteristics at origination mediate 73% of the reputation-price gap | Table IV Panel B, p. 3300 | Joint ACME: 10.29 (p = 0.10); remaining direct effect after all mediators becomes statistically insignificant |
| R6 | The reputation-price gap is small (about 6 pp) during the boom and large (about 35 pp) during the bust | Figure 5 Panel A, p. 3297 | Boom difference approx. 6 pp (yield diff approx. 34 bps); bust 1778 difference approx. 35 pp (yield diff approx. 895 bps) |
| R7 | Reputational effects are attenuated for bankers with short-run focus or married into wealth | Table V Panel B, p. 3301 | Short-run focus: direct coeff -9.34\*\*\* (t = 3.09); interaction short-run focus x office value: -28.22\* (t = 14.84); married-into-wealth x office value: -24.80\* (t = 12.38) |
| R8 | Banks with poor MBS performance suffered a 59% post-1770 decline in Amsterdam Bank of Exchange trading volume | Figure 6, p. 3303; Section VI.B Internet Appendix | Controlling for time and bank FE, post-1770 ABE trading volume 59% lower for below-median MBS performers |

**Overall (paper's conclusion).** Bank reputation can improve security quality in opaque markets, but only when bankers have substantial personal exposure to long-run reputational losses and are not focused on short-run profits. The evidence from the 18th-century Dutch plantation MBS market shows that high-reputation banks consistently originated better-quality mortgages and that investors suffered far smaller losses on their securities. The contrast with the 2000s RMBS literature is consistent with modern-era limited liability, bailout expectations, and short-term incentives undermining the same mechanism.

## Theory / model

The paper presents a stylized partial-equilibrium model (Section III, pp. 3286-3290) with three types of agents: planters (borrowers), bankers, and investors. Each period a banker originates a loan and sells it to investors.

**Planters.** Plantation $$p$$ has fundamental value $$F_p \in \{\underline{F}, \bar{F}\}$$ with $$\text{Prob}[F_p = \bar{F}] = \pi$$. A planter takes a one-period loan $$\lambda_{p,b,t}$$ from banker $$b$$, paying interest $$R_{b,t}$$ and origination fee $$\varphi$$. The planter's participation constraint pins the fee (eq. 1, p. 3286):

$$
(1 - \varphi) \lambda_{p,b,t} \Omega \geq \bar{R} \lambda_{p,b,t}; \qquad \varphi \equiv \frac{\Omega - \bar{R}}{\Omega}. \tag{1}
$$

Moral hazard: a planter can repay or defraud, taking $$(1-\beta)F_p$$ and leaving nothing. The incentive-compatibility constraint is (eq. 2, p. 3286):

$$
F_p - R_{p,b,t} \lambda_p \geq (1 - \beta) F_p. \tag{2}
$$

This defines maximum loan sizes $$\bar{\lambda} \equiv \frac{\beta}{\bar{R}}\bar{F}$$ (high type) and $$\underline{\lambda} \equiv \frac{\beta}{\bar{R}}\underline{F}$$ (low type) (eq. 3, p. 3286):

$$
\bar{\lambda} \equiv \frac{\beta}{\bar{R}}\bar{F}; \qquad \underline{\lambda} \equiv \frac{\beta}{\bar{R}}\underline{F}. \tag{3}
$$

**Bankers.** Each banker $$b$$ has discount factor $$\delta_b$$, outside activities yielding $$\rho_b$$ per period, and independent wealth $$\omega_b$$. A banker who misrepresents a low-type planter as high-type earns extra fees but loses fraction $$\gamma = 1$$ of outside activities permanently if discovered. The utility loss from falling below the threshold $$\rho_b$$ is quadratic (eq. 4, p. 3287):

$$
U(y, \rho_b) = \begin{cases} y & \text{if } y \geq \rho_b \\ y - A(\rho_b - y)^2 & \text{if } y < \rho_b \end{cases}. \tag{4}
$$

**Market.** The market continues each period with probability $$\zeta_t$$ (unknown; updated by Bayesian learning to $$\mu_{t+1}$$). Bankers learn about an impending end one period before investors.

**Equilibrium types.** Proposition 1 (pp. 3288-3289) shows that in equilibrium two types participate: Good (G) bankers who always provide $$\underline{\lambda}$$ honestly, and Mediocre (M) bankers who provide $$\underline{\lambda}$$ while the market is set to continue but misrepresent if they learn it is ending. The conditions are (eqs. 9-10, p. 3288):

$$
G: \quad \varphi(\bar{\lambda} - \underline{\lambda}) \leq \frac{\delta_b}{1 - \delta_b} \left[\rho_b + A(\rho_b - \omega_b)^2\right] \tag{9}
$$

$$
M: \quad \varphi(\bar{\lambda} - \underline{\lambda}) - \frac{\delta_b}{1-\delta_b} \frac{(1-\delta_b \zeta_t)}{(1-\delta_b \zeta_t (1-\pi))} \varphi \bar{\lambda} \leq \frac{\delta_b}{1-\delta_b}\left[\rho_b + A(\rho_b - \omega_b)^2\right] \leq \varphi(\bar{\lambda}-\underline{\lambda}). \tag{10}
$$

**Key testable predictions.** Prediction 2 (from Proposition 1, Lemma 1): high-reputation banks provide better loans; the gap is especially pronounced when the market is about to end. Prediction 5 (Lemma 1): the effect is attenuated for bankers with high $$\omega_b$$ (married into wealth) or low $$\delta_b$$ (short-run focus, i.e., discounting the future more heavily). Prediction 6 (by design): banks that sold worse MBS see a decline in other activities.

## Method

The paper applies two main estimators: OLS with MBS fixed effects for the price regressions, and average causal mediation effect (ACME) analysis (Baron and Kenny (1986), Imai et al. (2011)) for the mediation analysis.

**Reputation measure.** Bank reputation $$\rho_b$$ is proxied by the rental value of a bank's Amsterdam office block from the 1742 census (Oldewelt (1945)), reflecting outside business activities unrelated to plantation MBS. High-reputation: office value above the median. The measure is verified to correlate with city-government positions, ABE trading volume, and notarial deed counts (Table I, pp. 3274-3275).

**MBS price regression (equation 11, p. 3294).** For each auction transaction between 1768 and 1796:

$$
P_{i,b,t} = \beta \rho_b + \eta_t + \varepsilon_{i,b,t}, \tag{11}
$$

where $$P_{i,b,t}$$ is the price of MBS $$i$$ issued by bank $$b$$ in year $$t$$ as a percentage of par, $$\rho_b$$ is the reputation measure, and $$\eta_t$$ are auction-year fixed effects. Standard errors are clustered at the MBS level (46 clusters). Each transaction is weighted by the inverse of the number of auction transactions for that MBS times its market share, so $$\beta$$ captures the loss a hypothetical investor equally split between high- and low-reputation banks would have experienced.

**Time-path specification (equation 12, p. 3296).** To test Prediction 3, the regression is estimated separately for high- and low-reputation MBS:

$$
P_{i,t} = \eta_t + \varepsilon_{i,t}, \tag{12}
$$

where $$\eta_t$$ are annual averages for each group (same weights as eq. 11). This traces the price divergence over time, showing a small 6 pp gap during the boom (1769-1770) widening to approximately 35 pp in 1778.

**Mediation analysis.** The mediators (mortgage characteristics) are introduced sequentially into equation (11) to estimate the ACME (Imai et al. (2011)). Mediators are: timing of mortgage origination, fraction via nonelite agents, fraction to nonelite borrowers, and average LTF within an MBS. The joint ACME uses block-bootstrapping (10,000 resamples) to compute p-values. Results in Table IV, p. 3299-3300.

## Empirical specifications

**Mortgage quality regressions.** The sample is all mortgages extended 1750-1770 by Amsterdam merchant banks issuing MBS (N = 315, 26 banks). Each observation is weighted by mortgage sum. The specification compares means for high- vs low-reputation banks before (1750-1768) and during (1769-1770) the boom, testing differences-in-differences (Table II, p. 3293). No instrument; identification relies on the timing of low-reputation banks entering the market and the differential change in quality across bank types during the boom period.

Key mortgage quality outcomes:
- LTV (mortgage amount / appraised value): high-rep boom ~0.56 vs. low-rep boom 0.62, diff -0.06\*\*\* (t = -2.67)
- Fraction nonelite borrowers: boom diff -0.35\*\*\*, indicating high-rep banks maintained elite borrower screens during the boom
- LTF (mortgage / fundamental value, a debt-to-income analog): high-rep LTF increased 0.29\*\*\* (t = 2.66) during boom vs low-rep increase 0.50\*\*\* (t = 4.56), diff -0.21 (insignificant but economically large)

**MBS price regression.** Main sample: all MBS by Amsterdam merchant banks up to 1772 for which auction-price data are available (N = 46 MBS, 23 merchant banks; 4,605 auction transactions 1768-1796). Weighted OLS with auction-year FE and MBS-level clustered SE. The baseline result of 17.46 percentage points (t = 5.10) is robust to nine alternative reputation measures (Table III, cols. 1-9, p. 3295), block-bootstrapped SE, different regression weights, and annual-level aggregation.

**Heterogeneity tests (Table V, pp. 3301-3302).** The continuous office value interacted with short-run focus and married-into-wealth dummies:
- Short-run focus (partner died with minor heir) reduces the reputation effect to near zero (interaction coeff -28.22\*, t = 14.84)
- Married into wealth reduces the reputation effect by more than 60% in the price regression (interaction coeff -24.80\*, t = 12.38) and makes it statistically insignificant in the LTF regression

Standard errors are clustered at the bank level (23 or 24 clusters depending on spec).

**Reputational losses (Figure 6, p. 3303).** Post-1770 ABE trading volume is regressed on bank FE and time FE; above-/below-median MBS performers are separated on the estimated bank FE. The 59% post-1770 gap is the difference in log ABE volume between good- and poor-MBS-performance banks, controlling for time trends.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Hand-collected plantation mortgage records (Amsterdam notarial archives, Suriname colonial archives) | Mortgage characteristics: LTV, borrower type, agent type, mortgage sum, appraisal values, loan-to-fundamental ratios; 1750-1770 | No page yet |
| 1742 Amsterdam property census (Oldewelt 1945) | Bank reputation measure: rental value of bank office block | No page yet |
| Amsterdam Bank of Exchange (Wisselbank) account books, 1765-1795 | Bank trading volumes for reputational-loss test (half-yearly) | No page yet |
| MBS secondary-market auction prices (Amsterdam notarial archives) | MBS price outcomes: transaction prices 1768-1796, N = 4,605 transactions over 46 MBS | No page yet |
| Amsterdam estate tax records | Investor wealth distribution and portfolio holdings (N = 889 estates 1768-1796) | No page yet |

Sample: Suriname plantation MBS market, 1753-1796. Core price regressions: 1768-1796, 46 MBS, 23 merchant banks, 4,605 auction transactions. Mortgage analysis: 1750-1770, 315 mortgages, 26 merchant banks.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13503) if you are: studying the conditions under which bank reputation disciplines security quality in opaque markets; researching the history of financial innovation and the first MBS market; examining how banker incentive structures (limited liability, short-termism, wealth shielding) affect security issuance; or building on the mediation analysis framework to separate reputation effects from observable skill. Table III (p. 3295) contains the main price regressions; Tables IV and V (pp. 3298-3302) contain mediation and heterogeneity results; Figure 5 (p. 3297) traces the price gap over time.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(6). This distillation was extracted by an LLM on 2026-06-03 and is **not human-verified or independently reproduced**. The CC BY-NC-ND 4.0 licence permits sharing with attribution for non-commercial purposes but no derivatives; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY-NC-ND 4.0).** de Jong, Abe, Tim Kooijmans, and Peter Koudijs.
> "Going for Broke: Bank Reputation and the Performance of Opaque Securities."
> *The Journal of Finance* 80, no. 6 (December 2025): 3263-3312.
> DOI: 10.1111/jofi.13503. © 2025 The Author(s).
> Licensed under [CC BY-NC-ND 4.0](https://creativecommons.org/licenses/by-nc-nd/4.0/).
> This page is a distilled summary by the Institute for Automated Research;
> it is not a reproduction of or derivative from the original article.


==============================================================================
# Regulatory Fragmentation: Kalmenovitz, Lowry & Volkova (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/kalmenovitz-regulatory-fragmentation-2025/
# Distilled: Using the full text of the Federal Register (1994-2019), the paper constructs a firm-specific measure of regulatory fragmentation and documents that fragmentation increases firm costs (SG&A +4.3% SD), reduces productivity (TFP -3.6% SD) and profitability (ROA -5.3% to -5.9% SD), slows growth, deters entry, and pushes out small firms, with inconsistency across agencies driving more harm than mere duplication. J. Finance 2025, paywalled. Seven core results with source locators, datasets used, the measurement framework, and the estimating specifications.
# Tags: paper-summary, regulation, regulatory-burden, text-as-data, firm-real-outcomes
==============================================================================

**What this is.** The paper's core results, the measurement framework (LDA-based regulatory fragmentation measure), and the regression specifications: enough to know what it found and how, without reading all 46 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1111/jofi.13423). Firm-year data are publicly available at [evolkova.info](http://www.evolkova.info/data/fragmentation/).

## TL;DR

The paper introduces the concept of regulatory fragmentation: the regulation of a single topic by multiple federal agencies. Using the full text of the Federal Register (783,950 documents, 1994-2019), it applies Latent Dirichlet Allocation (LDA) to identify 100 regulatory topics and measure which agencies regulate each. It then matches these topics to firm-level 10-K filings to build a firm-specific, time-varying exposure measure. Across 60,573 firm-year observations (CRSP/Compustat, 1995-2019), higher regulatory fragmentation is associated with significantly higher SG&A costs, lower total factor productivity, lower profitability, slower sales and asset growth, less industry entry, and greater exit of small firms. Inconsistency across agencies (not just redundancy) drives the harm. Agency incentives, proxied by unexplained promotion activity, predict more fragmentation.

## Core results

Magnitudes and significance are as reported; `\*\*\*`/`\*\*`/`\*` = 1%/5%/10%. All variables are normalized so coefficients reflect a one-standard-deviation change. Independent variables are lagged one period. Standard errors clustered at the Fama-French 48-industry level.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Regulatory fragmentation raises SG&A expenses**: higher fragmentation is associated with significantly higher overhead costs in both year+company and industry x year+company FE specifications | Table IV, cols (1)/(4), p. 1106 | Coefficient on Regulatory Fragmentation: 0.043\*\*\* (SE 0.015) [year+co FE]; 0.055\*\*\* (SE 0.014) [ind x yr+co FE] |
| R2 | **Regulatory fragmentation reduces TFP**: a one-SD increase in fragmentation is followed by a 3.2-3.6% SD decrease in total factor productivity | Table IV, cols (2)/(5), p. 1106 | -0.036\*\* (SE 0.016) [year+co FE]; -0.032\* (SE 0.018) [ind x yr+co FE] |
| R3 | **Regulatory fragmentation lowers profitability (ROA)**: the effect ranges from -5.3% to -5.9% of a SD | Table IV, cols (3)/(6), p. 1106 | -0.059\*\*\* (SE 0.013) [year+co FE]; -0.053\*\*\* (SE 0.015) [ind x yr+co FE] |
| R4 | **Regulatory fragmentation slows sales growth**: a one-SD increase is followed by a 9.9-11.0% SD decrease in sales growth | Table V, cols (1)/(4), p. 1109 | -0.099\*\*\* (SE 0.017) [year+co FE]; -0.110\*\*\* (SE 0.022) [ind x yr+co FE] |
| R5 | **Regulatory fragmentation slows asset growth**: a one-SD increase is followed by a 13.9-14.3% SD decrease in asset growth | Table V, cols (2)/(5), p. 1109 | -0.143\*\*\* (SE 0.024) [year+co FE]; -0.139\*\*\* (SE 0.027) [ind x yr+co FE] |
| R6 | **Regulatory fragmentation deters entry and increases small-firm exit**: fragmentation reduces new IPOs, raises small-firm exit rates, and has a net negative effect on industry size; large firms benefit from the resulting barriers | Table VIII, p. 1115 | IPOs: -0.050\*\*\* (SE 0.014); small-firm exit: +0.036\*\* (SE 0.016); large-firm exit: -0.046\*\* (SE 0.018); total peers: -0.035\*\*\* (SE 0.008); all industry x year+company FE |
| R7 | **Regulatory fragmentation reduces lobbying**: firms reduce lobbying as fragmentation rises, consistent with lower returns to regulatory capture when oversight is dispersed across many agencies | Table X, cols (1)/(3), p. 1118 | log(Lobbying $): -0.106\*\*\* (SE 0.032) [year+co FE]; -0.102\*\* (SE 0.038) [ind x yr FE]; n = ~14,983 |

**Overall (paper's conclusion).** Regulatory fragmentation is a costly but previously undocumented dimension of regulatory burden. The negative effects arise primarily from inconsistency, not mere duplication: effects are weaker when agencies co-author documents (coordinated regulation), and stronger when agencies independently regulate the same topic. Agency promotion incentives are positively linked to rulemaking activity outside core expertise, suggesting empire-building as one mechanism driving fragmentation. These findings extend the regulatory-intensity measurement of Kalmenovitz (2023) by adding the multi-agency dimension, are consistent with the theory of economic regulation of Stigler (1971), draw on the evidence of inconsistency across banking regulators in Agarwal et al. (2014), and apply the LDA methodology of Lowry, Michaely and Volkova (2020).

## Theory / model

The paper has no formal model. The empirical strategy is built on two competing hypotheses about the sign of the fragmentation effect, drawn from the industrial organization literature on regulation.

**Hypothesis 1 (fragmentation is beneficial).** Multiple agencies create regulatory competition, let firms choose the least restrictive regulator, and may enable more efficient regulation via a race to the top. Firms can also focus lobbying on the agency most susceptible to capture. Under this view, higher fragmentation should lower costs and raise productivity and growth.

**Hypothesis 2 (fragmentation is costly).** Multiple agencies create duplicative compliance requirements, and more critically, inconsistent requirements that raise uncertainty. Firms cannot fully anticipate how to resolve discrepancies across agencies. Fragmentation also reduces firms' ability to direct lobbying effort. Under this view, higher fragmentation should raise costs and reduce productivity and growth.

The empirical results uniformly support Hypothesis 2. The paper provides no general equilibrium or structural model; its contribution is the measurement framework and the empirical evidence.

The **identification strategy** relies on within-firm, within-industry variation in regulatory fragmentation over time, with company fixed effects absorbing time-invariant firm heterogeneity and industry x year fixed effects removing industry-wide trends. The main endogeneity concern (firm operational changes causing apparent fragmentation changes) is addressed by: (i) excluding firm-years with industry switches, segment count changes, or asset changes exceeding 20% (Table VII); (ii) 1,000-iteration placebo tests that randomly reassign topics to firms (Figure 7); and (iii) coefficient stability analysis following Oster (2019). The authors acknowledge they lack an exogenous shock that randomly shifts fragmentation across firms (p. 1085).

## Method

**Step 1: Topic identification in the Federal Register.** LDA (Latent Dirichlet Allocation) is applied to the full text of 783,950 FR documents (1994-2019) from the Rules, Proposed Rules, and Notices sections. LDA generates 100 topics; each document receives a topic distribution. Topic labels are assigned using CFR subject classifications. For each topic $$i$$ and year $$t$$, the fraction of words written by agency $$a$$ is:

$$
\omega^2_{i,a,t} = \frac{\text{words by agency } a \text{ on topic } i \text{ in year } t}{\text{total words on topic } i \text{ in year } t} \tag{1 (within AgencyHHI)}
$$

The fragmentation of topic $$i$$ in year $$t$$ across agencies (equation (2), p. 1091) equals one minus the HHI:

$$
\text{Regulatory Fragmentation}_{i,t} = 1 - \text{AgencyHHI}_{i,t} = 1 - \sum_{\text{Agency}} \omega^2_{i,a,t} \tag{2}
$$

Values near zero indicate a single dominant agency; values near one (maximum 0.992 with 121 agencies) indicate dispersion across many agencies.

**Step 2: Measuring firm-level topic exposure.** The same LDA model (trained on FR documents) is applied to each firm's annual 10-K. The fraction of the 10-K devoted to topic $$i$$ in year $$t$$ is $$P_{f,i,t}$$. The dispersion of topics within a firm (equation (3), p. 1091) is:

$$
\text{Dispersion of Topics within Firm}_{f,t} = 1 - \sum_{\text{Topic}} P^2_{f,i,t} \tag{3}
$$

**Step 3: Firm-level regulatory fragmentation.** The main measure (equation (4), p. 1093) is the weighted average of topic-level fragmentation, where weights are each topic's share in the firm's 10-K:

$$
\text{Regulatory Fragmentation}_{f,t} = \sum_{i} P_{f,i,t} \cdot \text{Regulatory Fragmentation}_{i,t} \tag{4}
$$

**Step 4: Regulation quantity control.** To separate fragmentation from sheer regulatory volume, the paper constructs a control for the quantity of regulation (equation (5), p. 1095):

$$
\text{RegulationQuantity}_{f,t} = \sum_{\text{Topic}} P_{f,i,t} \cdot \log(\text{Words}_{i,t}) \tag{5}
$$

This is the weighted average of log(FR words) across topics, weighted by firm relevance.

**Validation.** Two alternative dispersion measures are constructed from FR agency mentions of companies (equation (6), p. 1096) and from 10-K agency mentions by companies (equation (7), p. 1096). Panel D of Table II (p. 1098) shows that the primary measure is significantly positively correlated with both alternatives, with a 1-SD increase in fragmentation associated with a 5.0% increase in agency-mention dispersion in the FR.

## Empirical specifications

The main regression specification (equation (8), p. 1104) is:

$$
y_{f,t+1} = \alpha + \beta \cdot \text{Regulatory Fragmentation}_{f,t} + \vec{X}_{f,t}' \gamma + \tau_t + \theta_k + \mu_f + \varepsilon \tag{8}
$$

where $$y_{f,t+1}$$ is a firm-level outcome (SG&A/AT, TFP, ROA, sales growth, asset growth, Emp/AT), $$\tau_t$$ are year fixed effects, $$\theta_k$$ are industry (Fama-French 48) fixed effects, and $$\mu_f$$ are company fixed effects. The tighter specifications replace $$\tau_t + \theta_k$$ with industry x year fixed effects. All continuous independent variables are winsorized at the 0.5% and 99.5% levels and normalized; standard errors are clustered at the Fama-French 48-industry level. Independent variables are lagged one period.

Firm-level controls $$\vec{X}_{f,t}$$ include: log(Words, 10-K), PPE/AT, EBITDA/AT, log(Sales), Tobin's Q, Dispersion of Topics within Firm, and Regulation Quantity.

**Core outcome regressions (Tables IV and V, pp. 1106, 1109):** The main results are estimated for six outcomes: SG&A/AT (costs), TFP (productivity), ROA (profitability), sales growth, asset growth, and Emp/AT (employment). The baseline coefficient on Regulatory Fragmentation ranges from 0.043 (SG&A, positive) to -0.143 (asset growth, negative) in standardized units (R1-R5).

**Industry composition regressions (Table VIII, p. 1115):** For entry (R6), the dependent variable is the count of new peers (IPOs or industry joiners) in year t+1; for exit, it is the rate of peers in year t no longer present in year t+1, separately for small and large firms (defined by above/below-average assets within Fama-French 48 industry). Product-market peers are defined using Hoberg and Phillips (2016) TNIC-3 similarity scores.

**Channel test: co-authorship (Table IX, p. 1117):** The indicator $$\text{CoAuthored}_{f,t}$$ (above-median exposure to topics where FR documents are co-authored by multiple agencies) is interacted with Regulatory Fragmentation in a fully interacted model. The interaction coefficient is negative (e.g., SG&A: Regulatory Fragmentation x CoAuthored = -0.014**, SE 0.006), confirming that inconsistency (solo-authored, potentially conflicting rules) drives more harm than coordination.

**Lobbying regressions (Table X, p. 1118):** The dependent variable is log(1 + lobbying expenditures in USD millions) and raw lobbying expenditures, from the LobbyView database (Kim (2018)), for 14,983-14,986 company-year observations. Regressions include company and year fixed effects, or industry x year fixed effects.

**Agency incentives (Table XI, p. 1121):** The dependent variable is Unexpected Promotions at agency a in year t+1, constructed as residuals from an employee-level promotion model (controlling for tenure, agency x rank, and occupation fixed effects). The key independent variable is the number of FR words written by the agency in notices (or rules) in year t, and its interaction with the share of words in the agency's top 10 topics. The positive coefficient on Words (e.g., Words in notices: 0.615**, SE 0.234) and the negative coefficient on the interaction with core topics support the view that agency employees are rewarded for expanding rulemaking, particularly outside core areas.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Federal Register (full text, 1994-2019) | Main input; LDA topic model trained on 783,950 documents to measure regulatory fragmentation and quantity | [Federal Register](/wiki/datasets/federal-register/) |
| SEC EDGAR (10-K filings) | Applied LDA model to firm 10-Ks to measure firm-topic exposure $$P_{f,i,t}$$; log(Words, 10-K) control | [EDGAR](/wiki/datasets/edgar/) |
| CRSP / Compustat (via WRDS) | Firm-year outcome variables (SG&A, TFP, ROA, sales, assets, Tobin's Q), sample construction; 60,573 firm-years 1995-2019 | [WRDS](/wiki/commercial/wrds/) (licensed) |
| LobbyView lobbying database | Lobbying expenditures for ~14,983 firm-year observations; from Kim (2018) | No page yet |
| Hoberg-Phillips TNIC-3 | Product-market peer similarity scores for industry composition analysis | No page yet |
| OPM/FOIA employee compensation data | Individual-level compensation and rank for 75 federal agencies; used in agency incentive analysis | No page yet |
| Fragmentation firm-year data (evolkova.info) | Authors' public release of the main regulatory fragmentation measure | No page yet |

Sample: 60,573 company-year observations, 1995-2019 (CRSP/Compustat firms with nonmissing 10-Ks in SEC EDGAR). Frequency: annual.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13423) if you are: building a measure of multi-agency regulatory exposure for a specific sector or firm set; extending the fragmentation analysis to enforcement actions or international regulation; studying the industrial organization of the federal government and agency incentives; or replicating the TFP or sales-growth results, which have the largest magnitudes and require careful construction of the LDA model and Imrohoroglu-Tuzel TFP measure. The Internet Appendix (at the publisher's site) contains additional robustness tables and the LDA methodology details.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(2). This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The article is paywalled (Wiley VOR terms; no Creative Commons licence). Extract-only.

> Kalmenovitz, Joseph, Michelle Lowry, and Ekaterina Volkova.
> "Regulatory Fragmentation."
> *The Journal of Finance* 80, no. 2 (April 2025): 1081-1126.
> DOI: 10.1111/jofi.13423. © 2025 the American Finance Association.


==============================================================================
# Intrahousehold Disagreement about Macroeconomic Expectations: Ke (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/ke-intrahousehold-disagreement-macroeconomic-expectations-2025/
# Distilled: Using the Health and Retirement Study and a preregistered randomized survey experiment, Da Ke documents that five in six U.S. married couples disagree about macroeconomic expectations (inflation, recessions, stock returns), and that intrahousehold belief disagreement causally reduces household stock market participation on both the extensive and intensive margins. J. Finance 2025, CC BY 4.0. Eight core results with source locators, datasets used, the empirical model, and the experimental specifications.
# Tags: paper-summary, household-finance, macroeconomic-expectations, beliefs
==============================================================================

**What this is.** The paper's core results, the model it tests (intrahousehold belief aggregation), and the empirical and experimental specifications: enough to know what it found and how, without reading all 43 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1111/jofi.13437).

## TL;DR

Da Ke documents that five in six U.S. married couples hold divergent macroeconomic beliefs, and that this disagreement is persistent within households, largely unexplained by observable characteristics, and causally reduces household stock market investment. Using the Health and Retirement Study (HRS), Ke shows that the mean belief of both spouses predicts portfolio choice 76 to 83% better than either spouse's belief alone, and that the level of spousal disagreement independently depresses participation and equity share. A preregistered MTurk randomized survey experiment confirms the causal mechanism: exogenous pessimism from a spouse reduces stock market participation by 9 percentage points, while exogenous optimism raises it by 4 percentage points, with an asymmetry that is specific to the marital context rather than generic social influence.

## Core results

Magnitudes and significance are as reported; `\*` = 10%, `\*\*` = 5%, `\*\*\*` = 1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Five in six couples disagree** about future macroeconomic developments across all three domains; disagreement is only slightly less than among randomly matched strangers | Figure 1 (p. 1657); text p. 1656 | 83-85% of couples disagree; pseudo couples disagree 87-89%; couples agree only 15-17% of the time |
| R2 | **Intrahousehold belief differences are persistent**: household fixed effects absorb 36-44% of the panel variation; time fixed effects less than 1% | Table III (p. 1663) | Household FE R2: inflation 44.0%, recessions 43.9%, stock returns 36.3%; time FE R2 all below 0.3% |
| R3 | **Mean couple belief predicts portfolios 76-83% better** than either spouse's belief alone | Table V (p. 1667) | Participation coefficient: mean belief 0.303\*\*\* vs. husband 0.170\*\*\* or wife 0.170\*\*\*; equity share: 0.243\*\*\* vs. 0.139\*\*\* or 0.133\*\*\* |
| R4 | **Disagreement reduces stock market participation**: 1-SD higher disagreement reduces participation by 1.1pp (2.4%) | Table VI col. 1 (p. 1668) | Coefficient: -0.054\*\*\* (SE 0.013), controlling for mean belief and demographics |
| R5 | **Disagreement reduces equity share**: 1-SD higher disagreement reduces equity share by 0.8pp (2.9%) | Table VI col. 2 (p. 1668) | Coefficient: -0.037\*\*\* (SE 0.009), controlling for mean belief and demographics |
| R6 | **Causal impact confirmed by experiment**: pessimistic spouse causes -8.8pp participation decline; optimistic spouse causes +4.4pp increase | Table VIII col. 1 (p. 1678) | Pessimistic spouse: -0.088\*\*\* (SE 0.016); optimistic spouse: +0.044\*\*\* (SE 0.015) vs. control arm (same belief) |
| R7 | **Gender asymmetry in responsiveness**: women respond symmetrically to pessimistic and optimistic husbands; men respond mainly to pessimistic wives | Table IX (p. 1679) | Women: optimistic husband +5.4pp\*\*\*, pessimistic husband -9.7pp\*\*\*; men: optimistic wife +3.1pp (insig.), pessimistic wife -6.3pp\*\* |
| R8 | **Marital-context specificity**: subjects are substantially more responsive to a spouse's pessimism than a stranger's; optimism effects are indistinguishable | Table XI (p. 1682) | Spouse vs. stranger pessimism differential: -6pp additional participation decline (significant at 10%); spouse vs. stranger optimism differential: +2pp (insignificant, p > 0.5) |

**Overall (paper's conclusion).** Accounting for interactions between household members would be a potentially important extension of the representative-agent framework in the macrofinance literature. Intrahousehold belief heterogeneity is substantial, persistent, and a genuine driver of portfolio under-investment, consistent with the analogy to group disagreement and underinvestment documented by Garlappi, Giammarino, and Lazrak (2017) in corporate settings. The paper builds on D'Acunto, Malmendier, and Weber (2021), who document a gender gap in inflation expectations within households; extends the stylized fact of low portfolio sensitivity to beliefs from Giglio, Maggiori, Stroebel, and Utkus (2021) and Ameriks, Kezdi, Lee, and Shapiro (2020); and uses the hypothetical portfolio allocation design of Lian, Ma, and Wang (2019).

## Theory / model

The paper has no formal structural model. It tests the following empirical hypotheses motivated by the household macrofinance literature:

**H1 (Belief heterogeneity).** If households consist of members with divergent beliefs, standard models that treat the household as a single agent with a representative belief will mischaracterize how beliefs map to portfolio choices.

**H2 (Mean belief sufficiency).** If the intrahousehold financial decision-making process is truly collective (not dictatorial), the couple's mean belief should predict portfolio outcomes better than either spouse's belief alone.

**H3 (Disagreement effect).** Holding mean belief fixed, a higher level of intrahousehold disagreement should independently reduce stock market participation and equity share, analogous to the group-disagreement underinvestment result of Garlappi, Giammarino, and Lazrak (2017).

**Variance decomposition.** To characterize the nature of the heterogeneity, the paper decomposes the overall cross-sectional variance of macroeconomic expectations into between- and within-household components (pp. 1661-1663). For household $$h$$ with members $$i \in h$$ and belief $$y_i$$ in wave $$t$$:

$$
\sigma^2_Y = \sigma^2_{\text{Between}} + \sigma^2_{\text{Within}} \tag{1}
$$

$$
\sigma^2_{\text{Between}} = \frac{1}{n}\sum_h n_h (\bar{y}_h - \bar{y})^2, \qquad \sigma^2_{\text{Within}} = \frac{1}{n}\sum_h\sum_{i \in h}(y_i - \bar{y}_h)^2 \tag{2-4}
$$

The within-household component accounts for 40-45% of the total cross-sectional variance in stock market expectations (Figure 2, p. 1662), and the result is almost identical after demeaning beliefs by gender, ruling out the gender gap as the sole driver.

To assess persistence, the paper estimates regressions of the intrahousehold belief difference $$y_{h,t}$$ on time fixed effects $$\chi_t$$, household fixed effects $$\phi_h$$, and their combination (pp. 1663):

$$
y_{h,t} = \chi_t + \varepsilon_{1,h,t} \tag{5}
$$

$$
y_{h,t} = \phi_h + \varepsilon_{2,h,t} \tag{6}
$$

$$
y_{h,t} = \phi_{3,h} + \chi_{3,t} + \varepsilon_{3,h,t} \tag{7}
$$

Table III (p. 1663) shows that household fixed effects absorb 36-44% of the panel variation across domains, while time fixed effects absorb less than 1%, confirming that disagreement is a persistent household-level characteristic.

## Method

The paper uses two complementary empirical approaches.

**Observational approach (HRS panel).** Ordinary least squares with household-clustered standard errors, region and survey-wave fixed effects, and a rich set of demographic and economic controls (equation 8, p. 1666):

$$
y_i = \alpha + \beta \cdot \text{Belief}_i + \gamma' \mathbf{X}_i + \varepsilon_i \tag{8}
$$

where $$y_i$$ is stock market participation or equity share for household $$i$$, $$\text{Belief}_i$$ is the husband's belief, the wife's belief, or the couple's mean belief, and $$\mathbf{X}_i$$ contains race, age, education, family income, homeownership, and household wealth. To test the disagreement effect, the specification is extended to (equation 9, p. 1668):

$$
y_i = \alpha + \beta_1 \cdot \text{Mean belief}_i + \beta_2 \cdot \text{Level of belief disagreement}_i + \gamma'\mathbf{X}_i + \varepsilon_i \tag{9}
$$

where level of belief disagreement is the absolute value of the spousal belief difference. Standard errors are clustered at the household level; Probit and Tobit robustness results are reported in the Internet Appendix.

**Experimental approach (MTurk RCT).** The paper uses a two-stage randomized survey experiment preregistered in the AEA RCT Registry (#0007897). Stage one elicits subjects' own one-year stock market return expectations and provides an information treatment (S&P 500 price chart and textual description of recent performance) to a random subsample. Stage two randomly assigns subjects to one of three arms: spouse more optimistic (+10pp vs. subject's belief), spouse more pessimistic (-10pp), or control (same belief). The portfolio allocation task asks subjects to divide $10,000 between a risk-free savings account and the S&P 500 for one year. The main regression is (equation 10, p. 1676):

$$
y_i = \alpha + \beta_1 E_i[R_{1y}] + \beta_2 \cdot \text{Spousemoreoptimistic}_i + \beta_3 \cdot \text{Spousemorepessimistic}_i + \gamma'\mathbf{X}_i + \varepsilon_i \tag{10}
$$

Heteroskedasticity-robust standard errors are reported. The second wave (2023, N=2,000) replaces the spouse's exogenous belief with that of a stranger for a random subsample, enabling the spouse-versus-stranger comparison in equation (11) (p. 1681):

$$
y_i = \alpha + \beta_1 \cdot \text{Spouse}_i + \beta_2 \cdot \text{More optimistic}_i + \beta_3 \cdot \text{More pessimistic}_i + \beta_4 \cdot \text{More optimistic}_i \times \text{Spouse}_i + \beta_5 \cdot \text{More pessimistic}_i \times \text{Spouse}_i + \eta \cdot E_i[R_{1y}] + \gamma'\mathbf{X}_i + \varepsilon_i \tag{11}
$$

where $$\beta_1 + \beta_5$$ captures the differential impact of a pessimistic spouse versus a pessimistic stranger.

## Empirical specifications

**Specification 1: Beliefs and portfolio choice (Table V, p. 1667).** OLS with stock market participation (indicator) or equity share as the dependent variable. Regressors: husband's belief only (col. 1/4), wife's belief only (col. 2/5), or couple mean belief (col. 3/6), plus the full control vector. N=29,549 household-year observations. Household-clustered SEs. The key comparison is whether the mean belief coefficient is more than additive relative to individual spouse coefficients.

**Specification 2: Disagreement and portfolio choice (Table VI, p. 1668).** OLS with participation or equity share as dependent variable. Regressors: level of belief disagreement (absolute spousal belief difference) and mean belief jointly, plus controls. N=29,549. Identifies the independent effect of disagreement holding the average level of optimism constant.

**Specification 3: Sources of intrahousehold disagreement (Table IV, p. 1665).** OLS regressing the belief difference between spouses on spousal differences and averages in age, education, income, cognition, self-efficacy, neuroticism, and information acquisition. N=19,655. Identifies correlates of disagreement (not causal).

**Specification 4: Causal impact of heterogeneity (Table VIII, p. 1678).** OLS (equation 10) on the main experimental sample (N=3,980), with wave fixed effects and controls from Table VII. The key coefficients $$\beta_2$$ and $$\beta_3$$ on the treatment indicators identify the causal impact of a 10pp spousal belief shock on household portfolio choice.

**Specification 5: Heterogeneous effects by gender (Table IX, p. 1679).** Equation 10 estimated separately for female (N=2,481) and male (N=1,499) subjects. Documents the gender asymmetry in responsiveness to spousal beliefs.

**Specification 6: Causal impact of information (Table X, p. 1680).** Adds interactions between the information treatment dummy and the treatment arm indicators, identifying whether information alters the weighting of spousal beliefs at the aggregation stage.

**Specification 7: Spouse versus stranger (Table XI, p. 1682).** Equation 11 on the second-wave subsample (N=2,000). The expressions $$\beta_1 + \beta_4$$ (optimistic spouse vs. optimistic stranger) and $$\beta_1 + \beta_5$$ (pessimistic spouse vs. pessimistic stranger) identify whether the causal effect is specific to the marital context.

**Specification 8: Experimental disagreement effect (Table XII, p. 1683).** Replicates the observational disagreement regressions (equation 9) on the experimental sample, replacing the treatment arm indicators with the mean belief and level of disagreement implied by the experimental assignment.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Health and Retirement Study (HRS) | Primary observational data source: biannual panel of married-couple households over age 50; macroeconomic expectations of both spouses (inflation 1992-2000, recessions 1992-2008, stock returns 2004-2018), portfolio outcomes (stock market participation, equity share), and individual demographic and cognitive characteristics | [HRS](/wiki/datasets/hrs/) |
| Amazon Mechanical Turk (MTurk) survey experiment | Randomized survey experiment; two waves (2021 N=3,000, 2023 N=2,000) of married U.S. residents age 25-64; exogenous variation in spousal beliefs about one-year stock market returns; hypothetical portfolio allocation task | no page yet |

Sample: HRS main sample 29,549 household-year observations (stock returns domain, 2004-2018); inflation sample 18,396 observations (1992-2000); recession sample 24,167 observations (1992-2008). Experiment: 5,000 total subjects across two waves.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13437) if you are: (i) studying collective household decision-making and belief aggregation (exact HRS variable construction in the Appendix, pp. 1684-1686); (ii) designing information experiments in household finance (the two-stage experimental design and preregistration details are in Section II and the Internet Appendix); (iii) extending macrofinance models to allow for intrahousehold heterogeneity (the paper's conclusion identifies this as the key unresolved question); (iv) examining the gender gap in macroeconomic expectations and how it differs from intrahousehold disagreement; or (v) replicating the robustness checks on measurement error, nonlinearity, risk aversion heterogeneity, and the placebo analysis on matched single individuals (Tables IA.III-IA.IX in the Internet Appendix).

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(3). This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Ke, Da. "Intrahousehold Disagreement about Macroeconomic Expectations."
> *The Journal of Finance* 80, no. 3 (June 2025): 1647–1689.
> DOI: 10.1111/jofi.13437. © 2025 The Author(s).
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Propagation of Cyberattacks through the Financial System: Kotidis & Schreft (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/kotidis-propagation-cyberattacks-financial-system-2025/
# Distilled: Using confidential Federal Reserve data on a multiday cyberattack on a technology service provider, Kotidis and Schreft (2025) quantify first-, second-, and third-round propagation effects through the Fedwire payment system, finding that business continuity plans and Federal Reserve support materially mitigated the disruption. J. Finance 2025, U.S. Government work, public domain in the USA. Eight core results with source locators, datasets used, the empirical framework, and the estimating equations.
# Tags: paper-summary, cybersecurity, financial-stability, payment-systems, banking
==============================================================================

**What this is.** The paper's core results, the empirical design (DiD with confidential Federal Reserve data), and the estimating equations: enough to know what was found and how, without reading all 46 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13475).

## TL;DR

This paper is the first to quantify propagation of an actual multiday cyberattack on a major technology service provider (TSP) to the banking sector through the Fedwire payment system. The attack impaired user banks' ability to send payments (first-round effect): on the worst day, users sent 33% fewer payments by number (45% by value) relative to nonusers. Business continuity plans (BCPs) and Fedwire extended hours reduced this by more than half. Exposed nonuser banks (receiver-banks) experienced a liquidity shortfall from reduced incoming payments (second-round effect), which they addressed by borrowing in the discount window or federal funds market, drawing down reserves, and sending payments during extended Fedwire hours. These actions averted broader contagion (third-round effect). Large banks responded more effectively than small ones throughout.

## Core results

Magnitudes and significance are as reported; `*`/`**`/`***` = 10%/5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Users sent significantly fewer payments than nonusers on every day of the attack (first-round effect) | Table II col. (3) and (7), p. 3332 | First day: -33% in number (t-stat implied by -0.395\*\*\*); -45% in value. Mid-period: -13% (number), -19% (value). Last day: -10% (number), -16% (value). All significant at 1%. |
| R2 | Without BCPs (no switching to alternative Fedwire access methods), the first-day drop would have been 100% for users and system-wide disruption would have been twice as large | §V.B, pp. 3333-3334 | Share of all Fedwire payments disrupted: 0.3% observed (0.45%\*0.6%); counterfactual 0.6% (0.6%\*1.0%). |
| R3 | Large users sent substantially fewer payments than small users did relative to nonusers; large users sent 26% fewer payments by value on the first day, versus 63% fewer for small users | Table IV col. (4), pp. 3337-3338 | Small users: -63% value, first day (col. (4) row 1). Large users: -26% (sum of rows 1 and 3, col. (4)). Difference statistically significant. |
| R4 | A 1-percentage-point increase in a receiver-bank's exposure to users is associated with a 0.7% decrease in incoming payments on the first day of the attack (second-round effect) | Table V col. (1), p. 3342 | Coefficient -0.689\*\*\* (SE 0.109). Effect halves in the mid-period (-0.403\*\*\*) and is insignificant on the last day (-0.059). |
| R5 | The contagion was larger for small receiver-banks than large ones; exposure to large users protected large receivers because large users' BCPs were more effective | Table V col. (2)-(3), p. 3342; Table VI, pp. 3344-3345 | Large receivers: -0.389\*\* (SE 0.183) on day 1. Small receivers: -0.667\*\*\* (SE 0.112). Receivers primarily exposed to large users experienced smaller or reversed drops (Table VI). |
| R6 | Exposed small receiver-banks with low reserves were more likely to borrow at the discount window; a 1-pp increase in exposure raised the probability of discount-window borrowing by 0.03% on day 1 | Table VII col. (6), p. 3347 | Coefficient 0.025\*\*\* (SE 0.006). For small banks with high reserves: 0.014\*\* (SE 0.005). Large banks: -0.165\*\* (SE 0.066), i.e., less likely. |
| R7 | Large exposed receivers with high reserves drew them down; a 1-pp increase in exposure is associated with an 18% decrease in reserves on the first day | Table IX col. (1), p. 3350 | Coefficient -18.095\*\*\* (SE 3.735) on log(Reserves). Mid-period -6.749 (insignificant). |
| R8 | Exposed receiver-banks sufficiently compensated for the liquidity shortfall and sent normal payments; no significant third-round effect on payment outflows was detected | Table X col. (1) and (3), p. 3351 | Exposure \* First Day coefficient -0.267\* (SE 0.141); mid-period 0.050 (insignificant); last day -0.020 (insignificant). Large banks sent significantly more payments on the last day (coefficient 1.692\*\*, col. (2)). |

**Overall (paper's conclusion).** The cyberattack had a material impact on individual financial institutions directly and indirectly connected to the TSP, but did not impair the overall financial system. BCPs by users and the TSP, combined with Federal Reserve operational support (Fedwire extended hours, discount window lending), were the primary mitigants. Large banks were more agile in implementing BCPs throughout. The TSP lost approximately 11% of its customers in the year after the attack and another 29% the following year, though other developments may have contributed to these departures.

## Theory / model

The paper has no formal economic model. The analysis tests three hypotheses about propagation through the payment network:

1. The cyberattack disrupts user banks' ability to send Fedwire payments (the common-shock, first-round effect), and BCPs partially offset this disruption.
2. Nonuser banks that normally receive payments from users experience a liquidity shortfall (the contagion, second-round effect), because banks rely on incoming payments to fund outgoing payments (building on Afonso, Kovner, and Schoar (2011)).
3. Unless receiver-banks can obtain alternative funding or rely on reserves, their inability to send their own payments creates further contagion to yet other banks (the third-round effect).

The identification strategy is a difference-in-differences design exploiting the common shock to users: the TSP taking its systems offline is a plausibly exogenous shock to user banks only. Nonusers faced no similar operational disruption. The analysis is in the spirit of Eisenbach, Kovner, and Lee (2022), who develop an ex-ante framework for measuring cyber risk in the US financial system; the present paper provides the first ex-post quantification using an actual event. Acemoglu, Ozdaglar, and Tahbaz-Salehi (2015) provide the theoretical backdrop for systemic risk and stability in financial networks.

The main threat to identification is selection: low-quality customers may match to TSPs with poor cybersecurity and also be less adept at implementing BCPs. To mitigate this, robustness tests restrict the control group to banks that use a competing TSP of similar revenue size, and match on log(assets) via propensity-score matching (pp. 3331, 3317). The paper states that full causal identification is not established (p. 3317). The event is also more severe than the hypothetical cyber run analyzed by Duffie and Younger (2019) because it is multiday and involves operational disruption, not only funding pressure. Crosignani, Macchiavelli, and Silva (2023) study propagation of the NotPetya cyberattack through firm supply chains; the present paper differs by focusing on a financial-sector TSP and Fedwire payment contagion.

The payment-network structure matters for the second-round analysis: large receiver-banks were primarily exposed to large users, while small receiver-banks were more uniformly exposed. Because large users implemented BCPs more effectively, large receivers fared better (pp. 3341, 3343).

## Method

The paper applies a panel difference-in-differences estimator to confidential daily transaction-level Fedwire data. There is no proposed novel method; the contribution is the first quantification of actual (not hypothetical) multiday cyberattack propagation using regulatory micro-data. The design builds on `difference-in-differences`, `panel-regression`, and `event-study` techniques.

**First-round estimator (Equation 1, p. 3330).** The dependent variable is the log change in the number or dollar value of Fedwire payments sent by sender-bank $$s$$ to receiver-bank $$r$$ on day $$t$$ relative to the same day one week earlier:

$$
\Delta \log(\text{Payments})_{srt} = \beta_1 \times \text{Users}_s \times \text{FirstDay}_t + \beta_2 \times \text{Users}_s \times \text{MidPeriod}_t + \beta_3 \times \text{Users}_s \times \text{LastDay}_t + FE + \varepsilon_{srt} \tag{1}
$$

where $$\text{Users}_s$$ is one if sender $$s$$ was a user of the TSP, $$\text{FirstDay}_t$$ ($$\text{LastDay}_t$$) is one on the first (last) day of the event, and $$\text{MidPeriod}_t$$ is one between first and last days. Fixed effects are added progressively. The preferred specification includes receiver-bank-by-day and sender-bank-by-receiver-bank fixed effects. Standard errors are two-way clustered at the sender-bank and day level (Bertrand, Duflo, and Mullainathan (2004)).

The log difference uses the day-before-the-same-weekday convention to account for weekly seasonality in Fedwire flows (Treasury settlement days on Thursdays, mid-month, and end-of-month); the upper 99th percentile of transactions is winsorized (p. 3330).

**Second-round estimator (Equation 2, p. 3342).** For exposed receiver-banks, the dependent variable is the log change in the value of incoming Fedwire payments:

$$
\Delta \log(\text{Payments})_{rt} = \sum_{\text{days}} \left( \beta_{\text{day}} \times \text{Exposure}_r \times \text{DayDummy}_t \right) + FE + \varepsilon_{rt} \tag{2}
$$

where $$\text{Exposure}_r$$ is the share of receiver-bank $$r$$'s total incoming payments (over a look-back window before the attack) originating from user banks. The model includes receiver-bank and day fixed effects. Standard errors are two-way clustered at the receiver-bank and day level.

**Third-round estimator (p. 3351).** Table X regresses the log change in the value of payments sent by exposed receiver-banks (now acting as senders) on the same exposure measure interacted with day dummies, controlling for receiver-bank-by-day FE and sender-bank-by-receiver-bank FE. The large-bank triple interaction isolates size heterogeneity.

**Size heterogeneity (Equation 1 extended, pp. 3335-3336).** Equation (1) is augmented with a $$\text{LargeBank}_s$$ dummy and its interactions with the day dummies and the $$\text{Users}$$ dummy to form a triple difference-in-differences. This captures the additional payment drop for small versus large users (Table IV).

## Empirical specifications

**First-round result (R1, Table II).** Specification: Equation (1). Preferred columns (3) and (7) include receiver-bank-by-day FE and sender-bank-by-receiver-bank FE. Three-month window. U.S. G-SIBs excluded from the sender group (included in robustness, Internet Appendix Table IA.I). Sample: all Fedwire sender-receiver pairs with at least one payment in the window; 546,631 observations.

**Mitigant analysis (R2, Table III).** Identical to Table II but payments after 6:30 p.m. are excluded to isolate normal business-hour effects; the extended-hours contribution is inferred by comparison. Table II column (3) vs. Table III column (3): -0.395 vs. -0.393 first-day effect in number; -0.590 vs. -0.695 in value (Table II col. (7) vs. Table III col. (7)), showing extended hours raised value sent by about 5 percentage points.

**Size heterogeneity (R3, Table IV).** Triple DiD version of Equation (1). Dependent variable: log change in value (columns (4)-(6)) or number (columns (1)-(3)) of payments. Large bank is above-sample-average log(assets). Key interaction terms: $$\text{Users} \times \text{LargeBank} \times \text{FirstDay}$$; $$\text{Users} \times \text{FirstDay}$$. Columns (5)-(6) use extended Fedwire hours; (2)-(3) and (5)-(6) include individual user and TSP adaptation controls.

**Second-round: contagion (R4-R5, Table V).** Specification: Equation (2). Dependent variable: $$\Delta \log(\text{Value of Payments received})$$. Three columns: all receivers, large receivers only, small receivers only. Receiver-bank FE and day FE. Two-way clustered at receiver-bank and day. Observations: 58,357 (all), 5,673 (large), 52,684 (small).

**Contagion by user size (R5, Table VI).** Augments Equation (2) with a second exposure layer: the share of the receiver's user-originated payments coming from large versus small users, at threshold levels of 80%, 50%, 20%. This decomposes the asymmetric contagion finding (pp. 3343-3346).

**Discount window borrowing (R6, Table VII).** LPM regression: $$P(\text{DW}_{rt} > 0 | \text{DW}_{t-1} = 0) = \text{Exposure}_r \times \text{DayDummies}_t + \text{ReceiverBank FE} + \text{FedReservDistrict} \times \text{Day FE} + \varepsilon_{rt}$$. Dependent variable is the dummy for first-time discount-window borrowing at time $$t$$ conditional on no prior use at $$t-1$$. Split by bank size and reserve-to-asset ratio (columns (1)-(6)).

**Reserve drawdowns (R7, Table IX).** Regression of $$\log(\text{Reserves}_{rt})$$ on $$\text{Exposure}_r \times \text{DayDummies}_t$$ for the subset of large banks with relatively high reserves (those from Table VIII column (4)). Receiver-bank FE and day FE; 82 observations.

**Third-round: payments sent by exposed receivers (R8, Table X).** Regression of $$\Delta \log(\text{Value of Payments sent})_{rt}$$ on $$\text{Exposure}_r \times \text{DayDummies}_t$$, with receiver-bank-by-day FE and sender-bank-by-receiver-bank FE, 304,663 observations. Extended and normal business hours compared across columns.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Fedwire Funds Service transaction-level data (confidential) | Daily sender-receiver-pair payment flows; treatment and outcome for first- and third-round DiD | No page yet |
| List of users of the TSP (confidential) | Treatment-group indicator (user vs. nonuser); identifies which banks could not access normal TSP services | No page yet |
| Federal funds loan-level data (Furfine algorithm, confidential) | Interbank borrowing outcome for large exposed receiver-banks; cross-checked against FR 2420 and FHLB 10-Ks | No page yet |
| Discount window daily borrowing records (confidential) | Discount-window borrowing outcome for small exposed receiver-banks | No page yet |
| Federal Reserve confidential reserve accounting records | End-of-day reserves held at the Federal Reserve; used for reserve-drawdown analysis (Table IX) | No page yet |
| Call Reports (FFIEC / FDIC) | Balance sheet data (total assets) for size classification | [Call Reports](/wiki/datasets/call-reports/) (public) |

Sample: event window is a confidential multiday period (not disclosed to protect anonymity of the TSP and the event). The three-month analysis window spans the month before, the event days, and the month after. Frequency: daily payment flows; balance sheet data matched at quarterly frequency.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13475) if you are: designing business continuity or third-party risk management requirements for banks or TSPs; studying payment-system contagion and the role of central-bank operational support; extending the empirical design to other cyber events (Internet Appendix Tables IA.I-IA.IV have robustness results and summary statistics); or evaluating whether larger or smaller banks should hold larger liquidity buffers as a first line of defence against operational disruptions. The locators above point to the exact tables.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(6), December 2025. This distillation was extracted by an LLM on 2026-06-03 and is **not human-verified or independently reproduced**. The underlying data are confidential Federal Reserve records and cannot be reproduced outside the Federal Reserve. Redistribution is extract-only (Wiley VOR terms outside the USA).

> Kotidis, Antonis, and Stacey L. Schreft. "The Propagation of Cyberattacks through the
> Financial System: Evidence from an Actual Event." *The Journal of Finance* 80, no. 6
> (December 2025): 3313-3358. DOI: 10.1111/jofi.13475.
> This article is a U.S. Government work and is in the public domain in the USA.
> Extract-only outside the USA (Wiley VOR terms and conditions).


==============================================================================
# Long-Horizon Exchange Rate Expectations: Kremens, Martin & Varela (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/kremens-long-horizon-exchange-rate-2025/
# Distilled: Two-year-ahead survey forecasts of financial professionals successfully predict currency appreciation both in and out of sample, with estimated slope coefficients close to one. Three macro-finance variables (quanto-implied risk premium, real exchange rate, current account-to-GDP) explain most of the variation in survey expectations, with no residual "secret sauce." J. Finance 2025, paywalled. Eight core results with source locators, datasets used, the theoretical framework (SDF-based UIP identity), and the main empirical specifications.
# Tags: paper-summary, exchange-rates, currency-risk-premia, uip, survey-expectations
==============================================================================

**What this is.** The paper's core results, the SDF-based UIP identity it uses, and the main regression specifications: enough to know what it found and how, without reading all 30 pages. To replicate or extend it, read the full source at [doi:10.1111/jofi.13504](https://doi.org/10.1111/jofi.13504).

## TL;DR

Using monthly Consensus Economics surveys of financial professionals (six high-income currencies against the dollar, December 1994 to March 2019), the paper shows that two-year-ahead exchange rate expectations successfully predict realized currency appreciation both in and out of sample, with slope coefficients statistically close to one and R-squared values around 16-19%. Survey forecasts beat the random walk benchmark of Meese and Rogoff (1983) out of sample, and are the strongest univariate predictor, outperforming the quanto-implied risk premium (QRP) of Kremens and Martin (2019), the real exchange rate predictor emphasized by Dahlquist and Penasse (2022), VIX, and capital flow proxies including the current account variable linked to Gabaix and Maggiori (2015). Three macro-finance variables (QRP, RER, current account-to-GDP) together explain most of the cross-currency and time-series variation in survey expectations, with no residual "secret sauce" from forecasters beyond these observables. The paper also confirms the finding of Nagel and Xu (2023) that short-horizon (one- and three-month) survey forecasts have near-zero predictive power, but documents that long-horizon forecasts predict short-run realizations while short-horizon forecasts do not. The factor loadings on the Dollar and Carry factors of Lustig, Roussanov and Verdelhan (2011, 2014) are included as alternative predictor variables but explain substantially less variation than survey expectations.

## Core results

Magnitudes and significance are as reported; `**`/`***` = 5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Survey expectations are the best univariate predictor** of 24-month realized currency excess returns, with R-squared exceeding all competing variables in the post-GFC sample | Table II, p. 3708 | Survey R-squared = 15.7%; runner-up QRP = 11.6%; RER = 10.4%; VIX = 8.5%; beta-HML = 7.2%; IRD = 1.7%; beta-dollar = 0.9%; CA/GDP = 0.0% |
| R2 | **Survey coefficients are close to one and statistically significant** in panel regressions of realized currency appreciation on survey excess return expectations, controlling for interest rate differentials | Table I Panel A cols (2)-(3), p. 3704 | SXR coefficient = 0.726-0.837 (standard errors 0.212-0.251); R-squared rises from 3.1% (IRD only) to 16.9-19.2% |
| R3 | **Survey forecasts beat the random walk out of sample** at the 24-month horizon, both dollar-based and dollar-neutral, with bootstrapped p-values below 5% | Table III cols (1)-(2), p. 3709 | Dollar-based R-squared-OOS = 19.15%; dollar-neutral = 14.99%; bootstrapped p-value = 8.81% / 3.37% respectively |
| R4 | **QRP, RER, and CA-to-GDP together explain over half the variation in survey expectations** in the post-GFC sample; the full multivariate specification raises R-squared to 53.6% | Table IV col (7), p. 3711 | R-squared = 53.6% (col 7, all variables); trivariate (QRP+RER+CA/GDP only, col 8) R-squared = 52.8%; QRP coefficient = 3.056\*\*\* (0.239), RER = -1.763\*\* (0.678), CA/GDP = -1.274\*\* (0.386) in col (7) |
| R5 | **No "secret sauce"**: the residual component of survey expectations not explained by QRP, RER, CA/GDP has no predictive power for realized returns | Table V col (3), p. 3713 | Coefficient on epsilon(SXR) = 0.177 (SE 0.232), statistically indistinguishable from zero; fitted value coefficient = 1.414 (0.832) |
| R6 | **Short-horizon survey forecasts (1- and 3-month) have near-zero predictive power** for short-run realized returns; the pattern reverses at 12- and 24-month horizons | Table VIII Panel A cols (1)-(4), p. 3718 | SXR\* coefficients at 1M = 0.088 (0.067), 3M = 0.093 (0.102), 12M = 0.237 (0.215), 24M = 0.726 (0.212); R-squared rises from 0.011 to 0.157 |
| R7 | **Long-horizon forecasts predict short-run realizations, but short-horizon forecasts do not predict long-run realizations** | Table VIII Panel A cols (5)-(10), p. 3718 | 24M forecasts predicting 1M RXR\*: coefficient = 1.548 (0.857), R-squared = 0.018; 1M forecasts predicting 24M RXR\*: coefficient = 0.007 (0.009), R-squared = 0.019 |
| R8 | **Forward expectations (3-to-24 month) predict short-run realizations**; spot short-horizon forecasts do not | Table IX col (5), p. 3719 | Coefficient on 3-to-24-month forward expectation predicting 3-month RCA = 0.188 (0.087), significant; coefficient on 3-month spot forecast = -0.062 (0.086), near zero |

**Overall (paper's conclusion).** Long-horizon survey expectations of financial professionals are broadly rational at the two-year horizon: their slope coefficient is close to one and they outperform the random walk. Three macro-finance variables (QRP, RER, CA/GDP) explain most of their variation, consistent with both risk-based and intermediary-constraints views of exchange rate determination. There is no residual forecaster information ("secret sauce") beyond these observables. The finding that long-horizon models outperform short-horizon models at forecasting short-run outcomes presents a puzzle the paper leaves open.

## Theory / model

The paper has no structural model. It grounds its empirical analysis in the standard no-arbitrage SDF identity and uses it to motivate the predictor variables. The fundamental asset pricing equation for any h-period gross dollar return $$R_{t+h}$$ is (p. 3700, eq. 1-2):

$$
\mathbb{E}_t(M_{t+h} R_{t+h}) = 1 \tag{1}
$$

$$
\mathbb{E}_t R_{t+h} - R^{\$}_{f,t,h} = R^{\$}_{f,t,h} \, \text{cov}_t(-M_{t+h}, R_{t+h}) \tag{2}
$$

where $$M_{t+h}$$ is the h-period SDF and $$R^{\$}_{f,t,h}$$ is the U.S. riskless rate. For a currency trade (convert USD to currency $$i$$, invest at the foreign riskless rate $$R^i_{f,t,h}$$, convert back), the return is $$R_{t+h} = R^i_{f,t,h} e_{i,t+h}/e_{i,t}$$ where $$e_{i,t}$$ is the nominal exchange rate. Substituting into equation (2) gives the UIP identity (p. 3700, eq. 3):

$$
\mathbb{E}_t \frac{e_{i,t+h}}{e_{i,t}} - 1 = \underbrace{\frac{R^{\$}_{f,t,h}}{R^i_{f,t,h}} - 1}_{\text{UIP}} + \underbrace{R^{\$}_{f,t,h} \, \text{cov}_t\!\left(-M_{t+h}, \frac{e_{i,t+h}}{e_{i,t}}\right)}_{\text{residual / risk premium}} \tag{3}
$$

The residual in (3) is the currency risk premium. When the marginal investor has log utility and holds the S&P 500 (so $$M_{t+h} = 1/R_{t+h}$$), the residual vanishes and expected appreciation equals the risk-neutral covariance term QRP (eq. 13, p. 3706):

$$
\mathbb{E}_t \frac{e_{i,t+h}}{e_{i,t}} - \frac{R^{\$}_{f,t,h}}{R^i_{f,t,h}} = \frac{1}{R^{\$}_{f,t,h}} \, \text{cov}^{\mathbb{Q}}_t\!\!\left(\frac{e_{i,t+h}}{e_{i,t}}, R_{t+h}\right) \tag{13}
$$

This motivates QRP as an observable proxy for the currency risk premium (and for survey expectations under rational expectations). RER and CA/GDP enter as additional predictors through their empirical association with either risk factors or intermediary balance-sheet constraints.

**Key definitions.** The paper defines the interest rate differential (IRD), realized currency appreciation (RCA), and survey-based currency appreciation (SCA) at horizon h as (pp. 3700-3701, eqs. 4-7):

$$
\text{IRD}_{i,t,h} = \frac{R^{\$}_{f,t,h}}{R^i_{f,t,h}} - 1, \qquad \text{RCA}_{i,t,h} = \frac{e_{i,t+h}}{e_{i,t}} - 1, \qquad \text{SCA}_{i,t,h} = \tilde{\mathbb{E}}_t \frac{e_{i,t+h}}{e_{i,t}} - 1 \tag{4-7}
$$

where $$\tilde{\mathbb{E}}$$ denotes the survey consensus (simple average across respondents). Currency excess returns are $$\text{RXR}_{i,t,h} = \text{RCA}_{i,t,h} - \text{IRD}_{i,t,h}$$ and survey excess return expectations are $$\text{SXR}_{i,t,h} = \text{SCA}_{i,t,h} - \text{IRD}_{i,t,h}$$ (eqs. 6-8, p. 3701).

The quanto-implied risk premium is constructed from quotes on conventional and quanto forwards on the S&P 500 (eq. 12, p. 3706):

$$
\text{QRP}_{i,t,h} = \frac{Q_{i,t,h} - F_t}{R^i_{f,t,h} P_t} = \frac{1}{R^{\$}_{f,t,h}} \, \text{cov}^{\mathbb{Q}}_t\!\!\left(\frac{e_{i,t+h}}{e_{i,t}}, R_{t+h}\right) \tag{12}
$$

where $$Q_{i,t,h}$$ is the quanto forward price and $$F_t$$ is the conventional S&P 500 forward price.

## Method

The paper applies standard panel regression methods to test predictability of exchange rates and to decompose what drives survey expectations. There is no new estimator. The method builds on `panel-regression` and `time-series-forecasting`.

**Baseline regression.** The in-sample predictability test adds survey excess return expectations to the standard UIP regression (eqs. 9-10, p. 3701):

$$
\text{RCA}_{i,t,h} = \alpha_h + \gamma_1 \text{SXR}_{i,t,h} + \gamma_2 \text{IRD}_{i,t,h} + \varepsilon_{i,t,h} \tag{9}
$$

$$
\text{RXR}_{i,t,h} = \alpha_h + \gamma_1 \text{SXR}_{i,t,h} + \gamma_2 \text{IRD}_{i,t,h} + \varepsilon_{i,t,h} \tag{10}
$$

Under UIP, $$\gamma_1 = 0$$ and $$\gamma_2 = 1$$. The paper assesses success by whether $$\hat{\gamma}_1$$ is positive, economically close to one, and statistically significant. Specifications with currency and time fixed effects are also estimated. Standard errors use a nonparametric block-bootstrap (Footnote 7, p. 3703) building on the approach of Hansen and Hodrick (1980), with blocks of length equal to the forecasting horizon and randomized cross-sectional width to account for overlapping observations and cross-sectional correlation; bootstrapped standard errors are typically 10th/90th percentiles of 10,000 resamples.

**Out-of-sample test.** The out-of-sample R-squared following Goyal and Welch (2008) is (eq. 14, p. 3708):

$$
R^2_{\text{OS}} = 1 - \frac{\sum_i \sum_t (\varepsilon^S_{i,t,t+h})^2}{\sum_i \sum_t (\varepsilon^C_{i,t,t+h})^2} \tag{14}
$$

where the numerator is the survey forecast error and the denominator is the competitor model error (random walk or QRP). A dollar-neutral variant $$\bar{R}^2_{\text{OS}}$$ computes errors relative to currency j for each pair (i, j) to net out dollar appreciation effects (eq. 15, p. 3709).

**What informs expectations.** To test what drives survey expectations, the paper regresses survey excess returns (SXR) on contemporaneous macro-finance variables (eq. 16, p. 3710):

$$
\text{SXR}_{i,t,h} = \alpha_h + \gamma_1 X_{i,t} + \gamma_2 \text{IRD}_{i,t,h} + \varepsilon_{i,t,h} \tag{16}
$$

where $$X_{i,t}$$ includes QRP, RER, VIX, CA/GDP, $$\beta^{\$}$$, and $$\beta^{HML}$$. Standard errors are clustered by time and currency. All predictor variables are standardized to unit standard deviation.

**Horizon analysis.** The paper annualizes variables by $$\text{SCA}^*_{i,t,h} = (12/h)\text{SCA}_{i,t,h}$$ to compare across horizons $$h \in \{1, 3, 12, 24\}$$ months (eq. 18, p. 3717). Forward expectations between horizons h and H are defined as (eq. 19, p. 3718):

$$
\text{sca}^{h,H}_{i,t} = \log\!\left(\frac{1 + \text{SCA}_{i,t,H}}{1 + \text{SCA}_{i,t,h}}\right) = \text{sca}_{i,t,H} - \text{sca}_{i,t,h} \tag{19}
$$

## Empirical specifications

**In-sample predictability (R1-R2).** Panel regressions using equations (9) and (10) at the 24-month horizon, post-GFC sample (December 2009 to March 2019, realizations until March 2021) with N = 672 observations (six currencies times approximately 112 months). The key specification is column (2) of Table I (p. 3704): SXR coefficient = 0.726 (SE 0.212), R-squared = 16.9% for RCA; similar results hold with currency and time fixed effects. Panel B extends to the full sample from December 1994, yielding N = 1,340 and similar coefficients.

**Alternative predictors (R1).** Univariate regressions of realized RXR on each alternative predictor variable separately (Table II, p. 3708). QRP data available only from December 2009 (Markit). Survey expectations achieve R-squared = 15.7% in the post-GFC sample, more than one-third higher than QRP (11.6%) and more than 50% higher than RER (10.4%).

**Out-of-sample performance (R3).** Surveys do not require estimated parameters so the out-of-sample test simply compares survey forecast errors to those of the random walk (RCA = 0) and QRP benchmark. Bootstrapped p-values are computed from the same block-bootstrap procedure as in-sample (Table III, p. 3709).

**Decomposing survey expectations (R4-R5).** Table IV (p. 3711) regresses SXR on QRP, RER, VIX, CA/GDP, $$\beta^{\$}$$, and $$\beta^{HML}$$, with standard errors clustered by time and currency. The trivariate specification (QRP, RER, CA/GDP) explains R-squared = 52.8% of variation in survey expectations. Table V (p. 3713) then tests whether the residual $$\varepsilon(\text{SXR})$$ from this regression predicts realized RCA: the coefficient is 0.177 with SE = 0.232, confirming no secret sauce.

**Horizon comparisons (R6-R8).** Table VIII (p. 3718) runs equation (10) separately for $$h \in \{1, 3, 12, 24\}$$ months and also uses long-horizon forecasts ($$\text{SXR}^*_{24}$$) to predict short-run realizations and vice versa. Table IX (p. 3719) decomposes 24-month expectations into three-month spot forecasts plus forward expectations from month 3 to month 12 and 12 to 24, regressing 24-month log realizations on these components.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Consensus Economics surveys | Monthly consensus forecasts of exchange rates at 1-, 3-, 12-, 24-month horizons; the primary SCA/SXR series | no page yet |
| Markit quanto forwards on S&P 500 | Construction of QRP (risk-neutral covariance between FX and equity); 24-month quotes, December 2009 onward | no page yet |
| Reuters forward exchange rates | Forward discounts / interest rate differentials (IRD) by horizon; used interchangeably with IRD under CIP | no page yet |
| IMF International Financial Statistics (IFS) | Current account balance and capital inflows, both scaled by GDP | [IMF IFS](/wiki/datasets/imf-ifs/) |
| BIS real exchange rate | RER predictor variable | [BIS EER](/wiki/datasets/bis-rer/) |
| FRED (VIX) | 30-day S&P implied volatility index as global risk perception proxy | [FRED](/wiki/datasets/fred/) |
| Lane-Milesi-Ferretti (2018) | Net foreign asset positions-to-GDP for robustness | no page yet |

Sample: six high-income currencies (AUD, CAD, EUR, GBP, JPY, KRW) against USD. Baseline sample: December 2009 to March 2019 (realizations until March 2021), N = 672. Full sample: December 1994 to March 2019, N = 1,340.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13504) if you are: testing the rationality of professional FX forecasters at different horizons; evaluating survey expectations as predictors against model-based alternatives; studying what macro-finance variables drive FX risk premia and expectations; or extending the QRP framework of Kremens and Martin (2019) to a broader set of predictors. The Internet Appendix (Appendix S1) contains data-source details, pre-GFC subsample results, robustness to additional currencies and specifications, and currency-specific slope estimates.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(6). This distillation was extracted by an LLM on 2026-06-03 and is **not human-verified or independently reproduced**. The paper is paywalled; extract-only redistribution applies.

> Kremens, Lukas, Ian W. R. Martin, and Liliana Varela.
> "Long-Horizon Exchange Rate Expectations."
> *The Journal of Finance* 80, no. 6 (December 2025): 3695-3724.
> DOI: 10.1111/jofi.13504. Copyright 2025 the American Finance Association.
> Paywalled. This page is an extract by the Institute for Automated Research:
> core results summarized; not a substitute for the original.


==============================================================================
# How Credit Cycles across a Financial Crisis: Krishnamurthy & Muir (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/krishnamurthy-credit-cycles-financial-crisis-2025/
# Distilled: Using credit spreads and credit growth across 17 countries from 1869 to 2022, this paper shows that spread spikes at crisis onset predict worse output losses, especially when precrisis credit growth was high, and that frothy credit markets (low spreads + high credit growth) predict future crises. J. Finance 2025, paywalled. Seven core results with source locators, datasets used, the FZ model of crises, and the estimating specifications.
# Tags: paper-summary, credit-cycles, financial-crises, macro, credit-risk
==============================================================================

**What this is.** The paper's core results, the theoretical framework it tests (the FZ model of financial crises), and the estimating specifications with exact source locators: enough to know what it found and how, without reading all 40 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13431).

## TL;DR

This paper documents how credit spreads and credit growth co-evolve across financial crises in 17 countries from 1869 to 2022. Four stylized facts emerge: (i) crises begin with a sharp spike in spreads, signaling a sudden shift in expectations; (ii) the severity of the recession depends on the size of the spread spike (losses to financial intermediaries) interacted with precrisis credit growth (financial sector fragility); (iii) precrisis spreads are unusually low ("frothy"), not high, despite high credit growth; and (iv) the combination of low spreads and high credit growth is a meaningful predictor of future crises. These facts support fragility-amplification (FZ) theories of financial crises where credit supply expansions precede collapses, but the low-precrisis-spread finding challenges standard FZ models that predict rising spreads as fragility builds, pointing instead toward belief-formation models.

## Core results

Magnitudes and significance are as reported; locators point into the source PDF. Standard errors (Driscoll-Kraay with 8 lags) in parentheses.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Spread spikes at crisis onset predict worse output**: a 1-sigma spread increase in the crisis year reduces three-year cumulative GDP growth by ~2.2% beyond the average crisis path | Table IV col 2, p. 1354 | $$\Delta\hat{s}_{i,t} \times \mathbf{1}_{\text{crisis}}$$ coeff: -2.21 (s.e. 0.74) at 3-yr; -1.48 (s.e. 0.60) at 5-yr |
| R2 | **Fragility amplifies spread shocks**: one-sigma credit growth increases the output loss from a spread spike by ~2.1pp (3-yr) and ~2.6pp (5-yr) | Table IV col 3, p. 1354 | $$\Delta\text{cred}_{i,t} \times \mathbf{1}_{\text{crisis}}$$ coeff: -2.06 (s.e. 0.95) at 3-yr; -2.65 (s.e. 1.10) at 5-yr |
| R3 | **Spread changes fatten the left tail of output**: predictive power is concentrated in bad outcomes; coefficient nearly triples from 75th to 25th quantile | Table V, p. 1357 | Quantile regression coefficient on $$\Delta\hat{s}$$: -0.36 at 75th pctile; -1.12 at 25th pctile |
| R4 | **Precrisis spreads are ~0.28-0.45 sigma too low**: in the three to six years before a crisis, spreads are significantly below the country average; not observed before ordinary recessions | Table VII col 3-4, p. 1361 | $$\mathbf{1}_{t-6,t-4}$$ coeff: -0.26 (s.e. 0.14) without year FE; -0.44 (s.e. 0.15) with year FE |
| R5 | **Low spreads + high credit growth predict crises**: the HighFroth x HighCredit interaction raises the five-year crisis probability by 27pp in-sample, but weakens to 14pp out-of-sample | Table VIII col 3, p. 1363; Table X col 3, p. 1367 | HighFroth x HighCredit OLS: 0.27 (s.e. 0.09) full sample; 0.14 (s.e. 0.07) OOS; froth does not predict nonfinancial recessions |
| R6 | **Results hold in postwar data**: postwar-only (post-1950) estimates are economically similar to full-sample estimates, albeit with larger standard errors | Table XI col 2, p. 1368 | $$\Delta\hat{s} \times \mathbf{1}_{\text{crisis}}$$ postwar: -1.43 (s.e. 0.53) vs. full-sample -2.21 (s.e. 0.74) at 3-yr |
| R7 | **Results robust to crisis dating chronology**: coefficients on spread-change x crisis are economically and statistically similar across JST, Reinhart-Rogoff (2009), and Baron-Verner-Xiong (2021) crisis dates | Table XV, p. 1371 | JST: -2.21 (s.e. 0.74); BVX: -2.47 (s.e. 0.71); RR: -2.36 (s.e. 0.48) |

**Overall (paper's conclusion).** The recessions surrounding financial crises are longer and deeper than nonfinancial recessions. The severity depends on the product of the spread spike (losses) and precrisis credit growth (fragility). Crises are preceded by frothy credit markets, indicating credit supply expansions, not demand. Standard FZ amplification models match the crisis-period facts but miss the low-precrisis-spread evidence. Models with time-varying beliefs (Moreira and Savov (2017); Krishnamurthy and Li (2025)) are better positioned to simultaneously fit both sets of stylized facts. The two prior-work pillars for this paper are Schularick and Taylor (2012), who established that credit booms predict crises and worse recoveries, and Baron and Xiong (2017), who showed that credit expansion coupled with an indicator of investor neglected crash risk predicts crises; this paper adds the credit-spread dimension to both. The crisis-dating framework relies on Jorda, Schularick, and Taylor (2011) (JST), and the risk-premium interpretation of spreads builds on Muir (2017), who documented that risk premia - not default probabilities - drive spread spikes in financial crises.

## Theory / model

The paper does not develop a new formal model. It instead tests a class of theoretical models of financial crises, which the authors label the "FZ model" (fragility-amplification), encompassing Kiyotaki and Moore (1997), Gertler and Kiyotaki (2010), He and Krishnamurthy (2013), Brunnermeier and Sannikov (2014), and Moreira and Savov (2017).

**The FZ structure.** Denote credit losses (p. 1341) as $$z_{i,t}$$, with $$\mathbb{E}_t[z_{i,t}] = 0$$, for country $$i$$ at time $$t$$. Denote the financial sector fragility as $$\mathcal{F}_{i,t}$$. The severity of the crisis depends on $$\mathcal{F}_{i,t} \times z_{i,t}$$: a large shock to a fragile sector triggers bank runs, credit contraction, and deep recession. Credit spreads proxy for both losses and financial sector assets:

$$
s_{i,t-1} = \gamma_{i,0} + \gamma_1 \text{Prob}^{\mathcal{Q}}(z_{i,t} > \underline{z}) \times \mathbb{E}_t^{\mathcal{Q}}[\text{Loss}_{i,t} | \text{crisis}] + \gamma_2 \text{Prob}^{\mathcal{Q}}(z_{i,t} \leq \underline{z}) \times \mathbb{E}_t^{\mathcal{Q}}[\text{Loss}_{i,t} | \text{no-crisis}] \tag{4}
$$

(p. 1359), where $$\text{Loss}_{i,t}$$ is increasing in $$\mathcal{F}_{i,t}$$. Precrisis, as $$\mathcal{F}_{i,t}$$ rises, spreads should rise - but the data show they fall. The reconciliation offered is that $$\text{Prob}^{\mathcal{Q}}(z_{i,t} > \underline{z})$$, the risk-neutral probability of a large loss, falls as credit growth rises, more than offsetting the fragility effect on spreads. This is consistent with models in which agents underestimate crisis likelihood during credit booms (Moreira and Savov (2017); Gennaioli, Shleifer, and Vishny (2013); Bordalo, Gennaioli, and Shleifer (2018)).

**The spread decomposition in a crisis.** In a crisis, illiquidity/fire-sale effects cause $$l_{i,t}$$ (an illiquidity component of spreads) to spike, leading to unexpected losses:

$$
s_{i,t} = \bar{\gamma}_i + \gamma_1 \mathbb{E}_t[\text{Loss}_{i,t}] + l_{i,t} \tag{\text{crisis}}
$$

Outside crises, spreads are better represented without the liquidity spike component and are passive forecasters of output, consistent with existing findings in the literature (Friedman and Kuttner (1992); Gertler and Lown (1999); Gilchrist and Zakrajsek (2012)).

## Method

The paper is a descriptive empirical study. The primary estimator is panel regression with Driscoll-Kraay standard errors allowing arbitrary serial correlation and cross-sectional dependence (8 lags), applied to an annual country-year panel. The paper also uses quantile regressions (Parente and Silva (2016)) and Logit specifications for crisis prediction.

**Spread normalization.** Raw spreads differ in level across countries (junk vs. investment grade). The paper normalizes by dividing by the country's unconditional mean spread (p. 1349, equation 2):

$$
\hat{s}_{i,t} \equiv \text{Spread}_{i,t} / \overline{\text{Spread}}^i \tag{2}
$$

This mean normalization exploits the homogeneity assumption that the sensitivity of spreads to the cycle is proportional to the average spread. The normalization is validated by showing that the coefficient on spread/mean is similar across pre-1940 (IMM bond data, mean spread 5.21%) and post-1940 (multiple sources, mean spread 1.13%) subsamples (Table III, cols 6-7).

**Fragility indicator.** The high-credit-growth variable $$\text{HighCredit}_{i,t}$$ counts the number of years in the past three years in which annual credit growth exceeded the full-sample median (divided by 3, so range 0 to 1); when $$\text{HighCredit}_{i,t} = 1$$ credit has been above median in each of the last three years. The froth indicator $$\text{HighFroth}_{i,t}$$ takes the residual of spread normalized on lagged GDP growth and five-year lagged spreads, averages the below-median dummy over the past five years (p. 1362-1363).

## Empirical specifications

**Main crisis-interaction regression (produces R1, R2, R6, R7).** Run on a panel of country-year observations with crisis and non-crisis dates (p. 1352, equation 3):

$$
\ln\!\left(\frac{y_{i,t+k}}{y_{i,t}}\right) = a_i + a_t + \mathbf{1}_{\text{crisis},i,t} \times b_{\text{crisis}}' Z_{i,t} + \mathbf{1}_{\text{no-crisis},i,t} \times b_{\text{no-crisis}}' Z_{i,t} + c' x_t + \varepsilon_{i,t+k} \tag{3}
$$

where $$y_{i,t}$$ is real per-capita GDP, $$k \in \{3, 5\}$$ years, $$Z_{i,t}$$ includes the normalized spread $$\hat{s}_{i,t}$$, the spread change $$\Delta\hat{s}_{i,t}$$, and credit growth $$\Delta\text{cred}_{i,t}$$, and $$x_t$$ includes two lags of GDP growth. Fixed effects: country ($$a_i$$) and year ($$a_t$$). Standard errors: Driscoll-Kraay with 8 lags. The key coefficient of interest is $$b_{\text{crisis}}$$ on $$\Delta\hat{s}_{i,t}$$ and its interaction with $$\Delta\text{cred}_{i,t}$$ (Table IV). Sample: 826 observations, 15 country groups.

**Quantile regression (produces R3).** Quantile regression of one-year output growth on $$\Delta\hat{s}_{i,t}$$ and $$\Delta\text{Credit}_{i,t}$$ at quantiles 90th, 75th, 50th, 25th, 10th (Table V, p. 1357). Controls: two lags of GDP growth, early-bond-data dummy, country and year fixed effects. Standard errors clustered by year. Sample: 826 observations, 826 at each quantile.

**Precrisis-spread regression (produces R4).** Regression of normalized spreads on crisis-time dummies from $$t - 6$$ to $$t + 5$$ (Table VII, p. 1361):

$$
\hat{s}_{i,t} = a_i + \sum_{\tau=-6}^{5} \delta_\tau \mathbf{1}_{t=\tau} + \text{controls} + \varepsilon_{i,t}
$$

with country fixed effects, early-bond-data dummy, and five-year lagged spread (controls for slow level changes). Standard errors: Driscoll-Kraay with 8 lags. The precrisis window $$t-6$$ to $$t-4$$ is combined into a single dummy in columns (3)-(4) to summarize the "too low" result.

**Crisis prediction (produces R5).** OLS and Logit specifications predicting the cumulative crisis indicator over a five-year horizon (whether any JST-mod crisis occurs in the next five years), on $$\text{HighFroth}_{i,t}$$, $$\text{HighCredit}_{i,t}$$, and their interaction (Tables VIII-IX, pp. 1363-1365). Country fixed effects, no time fixed effects (crisis prediction uses only pre-date information). Standard errors: Driscoll-Kraay with 8 lags (OLS), double-clustered by country and year (Logit). Out-of-sample version (Table X) constructs froth and credit variables in a rolling manner from 20 years after sample start.

All regressions exclude war periods (both world wars) because government intervention distorts bond prices and spread information.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Investors Monthly Manual (IMM), London Stock Exchange bond data, 1869-1929 | Corporate and sovereign bond yields for 17 countries; primary source for pre-1930 credit spreads | No page yet |
| Moody's Baa-Aaa spread (US, 1920-2014) | US credit spread in the modern period | No page yet |
| Global Financial Data (GFD) | Corporate and government bond yields for Australia, Belgium, Canada, Germany, Norway, Sweden, UK, Korea | No page yet |
| Datastream | Bond yields for Ireland, Portugal, Greece (2000-2014); European spreads | No page yet |
| Banque de France nonfinancial corporate spreads (Germany, France, Italy, Spain, 1999-2022) | European corporate credit spreads relative to German Bunds | No page yet |
| Swiss National Bank (SNB) data | Switzerland spreads from 2001 | No page yet |
| JST Macrohistory Database (Jorda, Schularick, Taylor (2017)) | Crisis dates, credit-to-GDP, real GDP per capita for 17 advanced economies | [JST Macrohistory](/wiki/datasets/jst-macrohistory/) (no page yet) |
| Barro-Ursua Macroeconomic Database | Real per-capita GDP; long historical series for advanced economies | [Barro-Ursua](/wiki/datasets/barro-ursua/) |

Sample: 17 countries; spread data 1869-2022 (by country); 1,006 country-year observations total for spreads; 40 financial crisis episodes with spread, credit, and output data available (the "JST mod." crises). Annual frequency.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13431) if you are: building a quantitative model of financial crises that needs empirical targets for spread dynamics and output losses; calibrating a fragility or amplification model to the interaction of losses x fragility; studying the credit-cycle predictor literature and want the international evidence on low-spread precrisis conditions; or interested in why standard FZ models with forward-looking spreads cannot match the frothy precrisis stylized fact. The tables are comprehensive: Table IV for the crisis-interaction results, Table VII for precrisis spread dynamics, Tables VIII-X for crisis prediction.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(3), June 2025. Published under Wiley terms and conditions (paywalled; not open access). This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. Redistribution: extract-only.

> Krishnamurthy, Arvind, and Tyler Muir. "How Credit Cycles across a Financial Crisis."
> *The Journal of Finance* 80, no. 3 (June 2025): 1339–1378.
> DOI: 10.1111/jofi.13431. © 2025 the American Finance Association.


==============================================================================
# Pricing Poseidon: Kruttli, Roth Tran & Watugala (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/kruttli-pricing-poseidon-extreme-weather-2025/
# Distilled: Firms in hurricane landfall regions experience implied volatility increases of up to 18%, reflecting persistent impact uncertainty that takes months to resolve; investors systematically underreact to this uncertainty until Hurricane Sandy (2012) served as a salient learning event. J. Finance 2025, CC BY-NC 4.0. Seven core results with source locators, datasets used, the theoretical model, and the empirical specifications.
# Tags: paper-summary, climate-finance, asset-pricing, options, volatility
==============================================================================

**What this is.** The paper's core results, the theoretical model it builds, and the empirical specifications: enough to know what it found and how, without reading all 50 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13416).

## TL;DR

Extreme weather events generate substantial firm-level uncertainty. Using single-stock options on firms with establishments in hurricane landfall regions (1996-2019, 37 hurricanes, 3,254 firms), the paper documents that implied volatilities rise up to 18% after a hurricane hit and remain elevated for several months, reflecting slow resolution of impact uncertainty. Despite these large increases in expected volatility, investors systematically underreact: the volatility risk premium (VRP) is significantly negative for hit firms, meaning ex ante implied volatility understates subsequent realized volatility (following the VRP definition of Lochstoer and Muir (2022)). This underreaction diminishes after Hurricane Sandy (2012), a salient event that struck the US financial center. Post-Sandy, hurricane uncertainty is priced efficiently and generates a positive expected return premium, consistent with personal experience effects documented by Malmendier and Nagel (2016). Textual analysis of analyst calls (using transcript methodology following Sautner, van Lent, Vilkov, and Zhang (2023a)) identifies five channels of uncertainty: business interruption, physical damages, insurance, supply, and demand.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Implied volatility increases significantly after hurricane landfall** for exposed firms; effect is largest close to the hurricane eye | Table III, p. 806 | 200-mile: lambda = 7.676 (t=3.178) at one month; 100-mile: 9.408 (t=2.801); 50-mile: 17.728\* (t=1.883) at one month post-landfall |
| R2 | **Impact uncertainty persists for about three months** after landfall; IV peaks at ~1 month and declines but remains significant | Figure 5 Panel A, p. 808 | IV effect significant for ~60 trading days (three months) post-landfall; peaks at ~8% around 30 trading days; discussion of hurricanes in analyst calls remains elevated for ~3 months (Panel B) |
| R3 | **Investors underreact to hurricane uncertainty before Sandy**: VRP of hit firms is significantly negative, implying implied volatility understates realized volatility | Table IV, p. 810 | 200-mile: VRP lambda = -6.035\*\*\* (t=-4.414) at one week; -5.315\*\*\* (t=-3.043) at one month; 50-mile: -21.463\*\*\* (t=-2.139) at one week |
| R4 | **After Hurricane Sandy, underreaction diminishes**: VRP effect reverses; post-Sandy hit firms price the uncertainty correctly or at a premium | Table V, p. 812; Figure 6, p. 813 | Post-Sandy interaction: at one week, +4.620 (t=1.651, insignificant); at one month, +7.572\*\*\* (t=3.132, col 5); net VRP effect turns 1.9-6.0% positive in post-Sandy period; pre-Sandy VRP negative for ~30 trading days |
| R5 | **Hurricane uncertainty affects expected returns only post-Sandy**: pre-Sandy CAR is insignificant; post-Sandy CAR is significantly positive for exposed firms | Table VII, p. 819 | Pre-Sandy: lambda = -0.022 to 0.127, all insignificant; Post-Sandy (x PostSandy): lambda = +4.599\*\*\* (t=3.157, 20-day col 1) to +6.965\*\*\* (t=4.258, 40-day col 7) |
| R6 | **Five economic channels drive hurricane uncertainty**: business interruption and physical damages dominate; insurance uncertainty is also significant | Table VI, p. 815 | 200-mile: total hurricane paragraphs +4.037\*\*\* (t=9.544); business interruption +1.157\*\*\* (t=6.131), physical damages +1.520\*\*\* (t=6.859), insurance +0.369\*\* (t=2.445) per unit of landfall exposure over 6 months |
| R7 | **Investors react to pre-landfall forecasts**: IV rises by up to 22% when storm wind speed probability reaches 50%; investors do not react to seasonal hurricane forecasts | Table VIII, p. 821; Table IX, p. 823 | Forecast exposure coefficient increases from ~4.6% (1%, one day out) to 21.6%\*\*\* (50%, one day out); seasonal forecasts: all interaction coefficients insignificant |

**Overall (paper's conclusion).** Hurricanes generate substantial, slow-to-resolve firm-level uncertainty that investors price with systematic underreaction until a particularly salient event (Sandy) improved informational efficiency. Idiosyncratic extreme weather shocks affect firms' cost of capital, which amplifies their real effects by tightening financing constraints when capital is most needed for rebuilding. Markets are unlikely to efficiently price unfamiliar climatic risks without direct experience.

## Theory / model

The paper adapts Merton (1987) to show how extreme weather uncertainty can affect expected returns even when the shock is purely idiosyncratic. This builds on the theoretical insight that idiosyncratic volatility can be priced when investors are underdiversified, formalized by Levy (1978) and Merton (1987). The theoretical framework (Internet Appendix Section I, summarized on p. 798-802) distinguishes two components of uncertainty, with uncertainty defined as expected volatility in the spirit of Bloom (2009), Pastor and Veronesi (2012), and Jurado, Ludvigson, and Ng (2015):

**Impact uncertainty.** When a firm is hit by an extreme weather event, its one-period return is (equation 3, p. 798):

$$
\tilde{R}_{i,t+1} = \bar{R}_i + b_i \tilde{Y}_{t+1} + \sigma_i \tilde{\epsilon}_{i,t+1} + \tilde{g}_{i,t+1}
$$

where $$\tilde{g}_{i,t+1}$$ is a random variable capturing the impact of the event on firm $$i$$, distributed with mean $$\mu_{g,i}$$ and variance $$\sigma_{g,i}^2$$. The term $$\sigma_{g,i}^2$$ captures impact uncertainty: the variance of the unpredictable disturbance once the event is known to have occurred.

**Incidence uncertainty.** Before the event occurs, there is also uncertainty about whether the event will hit. Expanding the return specification to include a Bernoulli hit indicator $$\tilde{\theta}_{i,t+1} \sim B(1,\phi)$$ (equation 4, p. 801):

$$
\tilde{R}_{i,t+1} = \bar{R}_i + b_i \tilde{Y}_{t+1} + \sigma_i \tilde{\epsilon}_{i,t+1} + \tilde{g}_{i,t+1} \tilde{\theta}_{i,t+1}
$$

The total return variance is then (equation 5, p. 801):

$$
\text{Var}_t(\tilde{R}_{i,t+1}) = b_i^2 + \sigma_i^2 + \sigma_{g,i}^2 \phi + \mu_{g,i}^2 \phi(1-\phi)
$$

where $$\sigma_{g,i}^2 \phi$$ is the expected impact uncertainty and $$\mu_{g,i}^2 \phi(1-\phi)$$ is the incidence uncertainty. Incidence uncertainty is highest when $$\phi = 0.5$$.

**Cost-of-capital channel.** In the extended Merton (1987) framework (Internet Appendix Section I), when investors hold underdiversified portfolios, shocks to expected idiosyncratic volatility affect the equity risk premium. The expected return on firm $$i$$ increases when the idiosyncratic variance of hit firms rises relative to control firms, providing the theoretical basis for the return premium tests in Section III.D.

**Identification logic.** The key identification assumption is that the timing and geographic incidence of hurricanes are exogenous to firm-specific conditions. Because a hurricane affects a subset of US firms (those in the landfall region), it creates a within-event treatment/control split: exposed firms experience higher uncertainty, unexposed firms serve as controls. Hurricanes are identified in real time via NOAA data, so the landfall region is known to investors as it happens.

## Method

The paper uses a **continuous-treatment difference-in-differences** (DiD) design at the firm-hurricane level, pooling 37 hurricane events from 1996-2019. The key methodological choices are:

**Firm exposure measurement.** Firm $$i$$'s exposure to hurricane $$h$$ is its share of establishments in the hurricane landfall region (equation 6, p. 802):

$$
\textit{LandfallRegionExposure}_{i,R,h} = \sum_c (\textit{FirmCountyExposure}_{i,c} \times I_{c \in L_{R,h}})
$$

where $$\textit{FirmCountyExposure}_{i,c}$$ is the share of firm $$i$$'s establishments in county $$c$$, and $$I_{c \in L_{R,h}}$$ is an indicator for county $$c$$ being within radius $$R$$ of the hurricane eye at landfall. The main analysis uses $$R = 200$$ miles, validated against smaller radii (100, 50 miles).

**Implied volatility measure.** Average implied volatility across options at the nearest-to-maturity expiry $$M$$, which adapts the measure used by Kelly, Pastor, and Veronesi (2016) for options on international stock indices to the single-stock setting (equation 1, p. 795):

$$
IV_{i,t} = IV_{i,t,M} = \frac{1}{Z} \sum_{z=1}^Z IV_{i,z,t,M}
$$

for $$Z$$ valid options satisfying: standard settlement; positive open interest; positive bid price and bid-ask spread; non-missing IV; 7-200 calendar days to expiry; $$|\delta| \in [0.2, 0.5]$$. Options are slightly out-of-the-money for liquidity. Model-based (binomial tree) implied volatilities from OptionMetrics are used; results are robust to model-free IV (Internet Appendix Section IV).

**Volatility risk premium.** VRP is the difference between ex ante implied and ex post realized volatility over the remaining life of the option (equation 2, p. 796):

$$
VRP_{i,t} = VRP_{i,t,M} = IV_{i,t,M} - RV_{i,t,M}
$$

A negative VRP for hit firms relative to controls means investors underestimate future realized volatility, consistent with underreaction.

The `difference-in-differences` and `panel-regression` estimation approaches, the `text-classification` dictionary methodology for analyst calls, and the `event-study` design for CARs are the core technical primitives.

## Empirical specifications

**Baseline implied volatility regression (Section III.A, eq. 7, p. 804).** For each hurricane entering as a separate time period:

$$
\log\!\left(\frac{IV_{i,T_L^h+\tau}}{IV_{i,T_0^h-1}}\right) = \lambda_{L,R,\tau} \cdot \textit{LandfallRegionExposure}_{i,R,h} + \pi_h + \psi_{\textit{Ind}} + \epsilon_{i,h,\tau}
$$

Dependent variable: log change in implied volatility from the day before hurricane inception ($$T_0^h - 1$$) to $$\tau$$ trading days after landfall. $$\pi_h$$ are hurricane fixed effects (equivalent to time FEs with one period per hurricane). $$\psi_{\textit{Ind}}$$ are industry FEs or industry x time FEs. Standard errors clustered by county (each firm assigned to its largest-establishment county). Ties R1 and R2.

**VRP regression (Section III.B, eq. 8, p. 807):**

$$
\overline{VRP}_{i,T_L^h+\tau} = \lambda^{VRP}_{L,R,\tau} \cdot \textit{LandfallRegionExposure}_{i,R,h} + \pi_h + \Psi_i + \epsilon_{i,h,\tau}
$$

Dependent variable: average VRP from landfall to $$\tau$$ days after. $$\Psi_i$$ is a firm fixed effect (absorbs level differences in VRP across firms; cannot use pre-inception subtraction because realized volatility over remaining option life already includes the hurricane). Ties R3-R4.

**Post-Sandy interaction (Table V):** Equation (8) augmented with $$\textit{LandfallRegionExposure}_{i,R,h} \times \textit{PostSandy}_h$$, where $$\textit{PostSandy}_h = 1$$ for hurricanes from 2013 onward.

**Abnormal return regression (Section III.D, eq. 10, p. 818):**

$$
CAR_{i,h,T_L^h+\tau:T_L^h+\tau+\textit{ReturnHorizon}} = \lambda^{Ret}_{L,R,\tau} \cdot \textit{LandfallRegionExposure}_{i,R,h} + \pi_h + \psi_{\textit{Ind}} + \epsilon_{i,h,\tau}
$$

CARs are relative to the Fama and French (2015) five-factor model, estimated in first stage using 120 trading days before hurricane inception. Starting point $$\tau = 30$$ (when IV peaks). Return horizons of 20, 30, 40 trading days. Ties R5.

**Economic channels regression (Section III.C, eq. 9, p. 814):**

$$
\textit{HurricaneDiscussions}_{i,T_L^h+120} = \lambda^{EC}_{L,R} \cdot \textit{LandfallRegionExposure}_{i,R,h} + \pi_h + \psi_{\textit{Ind}} + \varepsilon_{i,h}
$$

Dependent variable: number of analyst call paragraphs discussing hurricane and one of five economic channels (business interruption, physical damages, insurance, supply, demand) over 120 trading days after landfall. Dictionary-based classification with LDA validation. Ties R6.

**Forecast path regression (Section III.E, eq. 11, p. 820):**

$$
\log\!\left(\frac{IV_{i,T_L^h-\Gamma}}{IV_{i,T_0^h-1}}\right) = \lambda_{F,P,\Gamma} \cdot \textit{ForecastExposure}_{i,P,T_L^h-\Gamma} + \pi_h + \psi_{\textit{Ind}} + \epsilon_{i,h,\Gamma}
$$

Dependent variable: log IV change from inception to $$\Gamma$$ days before landfall. $$\textit{ForecastExposure}$$ is share of firm establishments in counties with forecast wind speed probability $$\geq P$$. Estimated for $$\Gamma \in \{1,\ldots,5\}$$ days and $$P \in \{1\%, 10\%, 20\%, 40\%, 50\%\}$$. Ties R7.

All regressions cluster standard errors by county (Petersen (2009)) based on each firm's largest establishment share.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| NOAA hurricane track data (National Hurricane Center) | Identifies 37 hurricane landfalls (1996-2019), eye location at 6-hour intervals, wind speed probability forecast advisories, seasonal outlook probabilities | [NOAA hurricanes](/wiki/datasets/noaa-hurricane/) |
| NETS (National Establishment Time Series) | Firm establishment locations by county, annual frequency; used to construct LandfallRegionExposure | [NETS](/wiki/commercial/nets/) (licensed) |
| OptionMetrics (single-stock options) | Daily implied volatilities and VRP for US-listed firms; data 1996-2019 | no page yet |
| CRSP/Compustat Merged (via WRDS) | Stock returns, market capitalization, SIC codes, share prices for sample construction and CAR estimation | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| Refinitiv (now LSEG) analyst call transcripts | Transcripts of analyst-investor-management calls; textual analysis of hurricane channels over 120 days post-landfall; data 2002-2019, 28 hurricanes | no page yet |
| S&P Global Market Intelligence | State-level property and casualty insurance premiums (Section IV.E extension only) | no page yet |

Sample: 1996-2019 (linked sample start = 1996, first year OptionMetrics data available); 3,254 unique firms; 1,799 hit firms (at least one hurricane with establishment share >= 25%); ~38,886 firm-hurricane observations at 200-mile radius baseline.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13416) if you are: examining the theoretical proof that idiosyncratic volatility can be priced (Internet Appendix Section I); studying VRP dynamics and the robustness of the model-free IV results (Internet Appendix Sections III-IV); exploring industry heterogeneity, tail effects (Section IV.D), or the insurance firm extension (Section IV.E); or extending the methodology to floods, snowstorms, or tornadoes (Table X). The locators above point to the exact tables and figures.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(2), April 2025. This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The CC BY-NC 4.0 licence permits non-commercial reproduction with attribution.

> **Attribution (CC BY-NC 4.0).** Kruttli, Mathias S., Brigitte Roth Tran, and Sumudu W. Watugala.
> "Pricing Poseidon: Extreme Weather Uncertainty and Firm Return Dynamics."
> *The Journal of Finance* 80, no. 2 (April 2025): 783-832.
> DOI: 10.1111/jofi.13416. © 2025 The Author(s).
> Licensed under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Personal Communication in an Automated World: Laudenbach & Siegel (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/laudenbach-personal-communication-automated-world-2025/
# Distilled: Personal two-way phone communication between a bank agent and a delinquent borrower increases timely repayment by 34.4 percentage points, reduces default by 23.8 percentage points, and reduces loan termination by 12.4 percentage points, identified via an IV exploiting random day-of-first-call variation. Evidence from a large German bank's early collection call center, Jan-Jun 2012, N=3,448 POS loan borrowers. J. Finance 2025, paywalled. Seven core results with source locators, datasets used, the model (IV framework), and the method (2SLS + MTE estimation).
# Tags: paper-summary, consumer-finance, household-finance, banking, lending-relationships
==============================================================================

**What this is.** The paper's core results, the institutional setup, the IV design, and the empirical specifications with the key equations: enough to know what it found and how, without reading all 45 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13388).

## TL;DR

At a large German bank in 2012, delinquent POS loan borrowers who spoke with a randomly assigned bank agent were significantly more likely to repay their overdue loan within 45 days than those who received only a standard written letter. Using the day of the first call attempt (Saturday vs. Monday) as an instrumental variable for whether a borrower actually spoke with an agent (*Talk*), the paper finds that personal two-way communication increases timely repayment by 34.4 percentage points, decreases the probability of default by 23.8 percentage points, and decreases immediate loan termination by 12.4 percentage points. The treatment effect is concentrated among borrowers who are harder to reach, marginal treatment effects increase in unobservable resistance, and average treatment effects (ATEs) are roughly half the size of the LATE. The likeability of the agent's voice is a significant predictor of payment, consistent with the prosocial or promise-keeping channel rather than a simple reminder effect. The effect persists: speaking with an agent also reduces future delinquency probability by 7.9 percentage points.

## Core results

Magnitudes and significance are as reported; \*\*\*/\*\*/\* = 1%/5%/10%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Personal communication increases timely repayment by 34.4 pp** (2SLS, N=3,448) | Table IV, Panel A, col. (1), p. 534 | Talk coef. = 0.344\*\*\* (SE=0.089); OLS = 0.201\*\*\* (SE=0.021) |
| R2 | **Personal communication reduces default by 23.8 pp** (2SLS) | Table IV, Panel A, col. (2), p. 534 | Talk coef. = -0.238\*\*\* (SE=0.077); OLS = -0.121\*\*\* (SE=0.016) |
| R3 | **Personal communication reduces immediate loan termination by 12.4 pp** (2SLS) | Table IV, Panel A, col. (3), p. 534 | Talk coef. = -0.124\*\*\* (SE=0.046); OLS = -0.072\*\*\* (SE=0.015) |
| R4 | **Average treatment effects are roughly half the LATE**: ATE for Payment = +16.6 to +20.1 pp | Table V, Panel B, p. 538 | ATE (polynomial) = 0.201\*\*\* (SE=0.053); ATE (semipar.) = 0.166\*\*\* (SE=0.046) |
| R5 | **Treatment effect on overall loan termination almost doubles** when looking beyond the initial delinquency | Table XI, Panel A, p. 553 | Talk coef. = -0.235\*\*\* (Termination: Overall, SE=0.069) vs. -0.124\*\*\* (Termination: Now); F-test p=0.021 |
| R6 | **Personal communication reduces future delinquency probability by 7.9 pp** and delays the next delinquency by ~29 days | Table XI, Panel B, p. 553 | OLS: Talk coef. = -0.079\*\*\* (SE=0.019); Oster (2019) delta=1.97 (>1, endogeneity unlikely) |
| R7 | **Agent voice likeability raises payment probability by 4 pp per 1-SD increase**, supporting the prosocial channel | Table X, p. 551 | OLS (survey sample, N=135): Likeable Voice (adjusted) coef. = 0.261\*\* (SE=0.040, wild bootstrap p) |

**Overall (paper's conclusion).** Personal, two-way communication between a bank agent and a delinquent borrower significantly increases the probability that the delinquency is resolved and that default and loan termination are avoided. The effect is unlikely to operate mainly through payment reminders or the bank's monitoring signal; evidence on agent voice likeability, German-vs.-non-German borrowers, and persistence beyond the initial delinquency all point to the personal or prosocial dimension of the phone conversation as the key channel.

The paper situates itself in a literature on banking relationships (Petersen and Rajan (1994), Drexler and Schoar (2014)) and on behavioral interventions for delinquent borrowers (Karlan, Morten, and Zinman (2015), Bursztyn et al. (2019)). Berg (2015) documents that human risk assessment improves loan quality, providing context for the value of the human element studied here.

## Theory / model

The paper has no formal structural model. The identification framework is an IV design with two instruments.

**Conceptual setup.** All borrowers in the sample receive a written letter from the bank informing them of the delinquency and requesting payment within two weeks. Treated borrowers (*Talk* = 1) additionally speak with a randomly assigned bank agent during the phone conversation. The agent may not convey additional information beyond the letter, may not change loan terms, and may not accept payment over the phone. Differences in payment behavior between treated and untreated borrowers are therefore due to the FORM of communication (personal vs. impersonal), not to information content or loan modification.

**Potential mechanisms discussed (Section III, pp. 543-552):**
1. Payment reminder / attention to the bank's enforcement process (ruled out by finding no differential effect by outstanding loan amount and no effect of borrower surprise about the call, Table VI and Table VII).
2. Personalized monitoring through a follow-up call (ruled out as the primary driver: 65-90% of the treatment effect obtains without a follow-up call, Table VIII).
3. Prosocial / promise-keeping behavior: a likeable agent voice significantly raises payment (Table X), and German borrowers (lower social distance) benefit more than non-German borrowers (Table IX, significant for *Default*). This channel is consistent with the literature on prosocial behavior and its sensitivity to social distance.

**Heterogeneous treatment effects (Section II.C, pp. 535-542).** The paper uses the Heckman-Vytlacil (1999, 2001, 2005) marginal treatment effect (MTE) framework. Borrower *i* speaks with an agent if their propensity score exceeds their unobservable resistance $$U_i$$ (prose on p. 536, not a numbered equation):

$$
Talk_i = \mathbf{1}\{f(X_i, Z_i) > V_i\}
$$

where $$Z_i$$ are the instruments (day-of-first-call indicators) and $$V_i$$ is unobservable resistance. The MTE characterizes the treatment effect as a function of unobservable resistance $$U_i$$ (the quantile of $$V_i$$), stated as equation (3) in the paper (p. 536):

$$
MTE(X_i = x, U_i = u) = \gamma + x(\beta_{T=1} - \beta_{T=0}) + E[\mathcal{E}_{T=1,i} - \mathcal{E}_{T=0,i} \mid U_i = u] \tag{3}
$$

MTEs are increasing in unobservable resistance: borrowers who are harder to reach respond more to the phone conversation than easy-to-reach borrowers, ruling out positive selection into treatment.

## Method

The paper uses linear IV (two-stage least squares) as the primary estimator, with the Heckman-Vytlacil MTE framework for heterogeneous treatment effects, and OLS for the mechanism and long-run tests where instruments are not available.

**IV approach (Section II.B, pp. 531-532).** The estimating equation (equation 1, p. 531) is:

$$
Y_i = \alpha + \gamma \, Talk_i + \beta X_i + L + M + \mathcal{E}_i \tag{1}
$$

where $$Y_i \in \{Payment_i, Default_i, Termination_i\}$$, $$Talk_i$$ is the endogenous indicator for whether the borrower spoke with a bank agent, $$X_i$$ is a vector of borrower and loan characteristics, $$L$$ are location fixed effects, and $$M$$ are month fixed effects. Standard errors are clustered at the location level.

The instruments are *First Call Saturday* (= 1 if the borrower's first call attempt falls on a Saturday) and *First Call Monday* (= 1 if Monday). The bank treats Saturday and Monday as a single day with random assignment by an auto-dialer, giving overall reachability rates of 96% (Saturday) vs. 43% (Monday) vs. 71% (Tuesday-Friday) as documented in Figure 3 (p. 529). The first-stage F-statistic is 155.92 (full sample, Table III, p. 533), well above Stock-Yogo (2005) thresholds.

**MTE estimation.** The Andresen (2018) *mtefe* estimator is used with both a parametric second-order polynomial approximation and a semiparametric approximation. Standard errors are bootstrapped with 100 repetitions, accounting for location-level clustering. ATEs are recovered by averaging MTEs over the [0, 1] interval of $$U_i$$, with the identifying assumption that MTEs are zero for $$U_i < 0.40$$ (i.e., no reversal of effect for always-takers).

## Empirical specifications

All main results use the linear IV (2SLS) specification above. Below are the key specifications tied to the core results.

**Main 2SLS (R1-R3, Table IV Panel A, p. 534):** Outcome $$Y_i \in \{Payment, Default, Termination:Now\}$$ regressed on $$Talk_i$$ instrumented by *(First Call Saturday, First Call Monday)*, controlling for installment amount, initial loan amount, interest rate, repayment term, time since origination, gender, national origin, age, age squared, employment status, month FE, and location FE. N = 3,448. Standard errors clustered at location.

**Robustness subsample (Table IV Panel B, p. 535):** Restricted to borrowers whose files enter the call center on a Thursday and who are first called on Saturday or Monday (N = 432), so the instruments are entirely generated by the auto-dialer's random Saturday/Monday assignment. Results are qualitatively similar but slightly smaller in magnitude.

**Heterogeneous treatment effects (Table V, pp. 537-538):** Equation 2 allows $$\beta$$ to interact with $$Talk$$; $$\beta_{T=1} - \beta_{T=0}$$ captures differential treatment effects by observable characteristics. No significant heterogeneity found for most observables; retired borrowers and older borrowers benefit less.

**Mechanism: loan amount split (Table VI, p. 544):** 2SLS estimated separately for borrowers with below- and above-median outstanding loan amount (N = 1,724 each). No significant difference in Talk coefficients (p-value for equality reported), ruling out a simple reminder/attention effect.

**Mechanism: follow-up call counterfactual (Table VIII, p. 548):** Two alternative outcomes (*Payment: Alt. 1* and *Payment: Alt. 2*) absorb the contribution of the follow-up call. 2SLS on these alternatives gives Talk coefficients of 0.223\*\* and 0.310\*\*\*, or 65% and 90% of the baseline 0.344\*\*\*, confirming the follow-up call accounts for at most 35% of the total effect.

**Mechanism: nationality (Table IX, p. 549):** 2SLS split by German (N = 2,745) vs. non-German (N = 703) borrowers. The Talk effect on *Default* is -0.320\*\*\* for Germans vs. 0.025 for non-Germans (p-value for equality = 0.016), consistent with social distance moderating the prosocial channel.

**Mechanism: agent voice likeability (Table X, p. 551):** OLS on survey sample (N = 135) regressing *Payment* on *Likeable Voice (adjusted)* (rater-adjusted fraction of raters rating agent's voice 4 or 5 out of 5), controlling for agent age, agent gender, borrower characteristics, loan characteristics, call duration, and whether it is a first call. Wild bootstrap t-approach with 1,000 repetitions accounts for the small number of agent clusters.

**Long-run effects (Table XI, pp. 553-554):** 2SLS for *Termination: Overall* (any termination by August 2014); OLS for *Future Delinquency*, *Time to Next Delinquency*, and *Termination: Later* on the subsample of borrowers who resolve the initial delinquency (N = 3,092), with Oster (2019) delta computed to assess endogeneity concern.

**External validity: overdraft borrowers (Table XII, p. 556):** Same IV specification applied to N = 2,499 overdraft facility borrowers (distinct from POS sample). Talk increases payment by 20.4 pp and reduces default and termination by 15.3 pp and 12.2 pp, respectively, all significant at 10%.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Proprietary administrative loan data, large German bank (POS loans, Jan-Jun 2012) | Main sample: 3,448 first-time delinquent POS borrowers; borrower characteristics, loan terms, call center records, payment outcomes through Aug 2014 | No page yet |
| Proprietary administrative loan data, large German bank (overdraft borrowers, Jan-Jun 2012) | External validity sample: 2,499 overdraft borrowers with existing bank relationship | No page yet |
| Survey of call center agents (February 2016) | Agent assessments of 245 borrower conversations; voice recordings of 8 agents rated by 6-7 raters for likeability | No page yet |

Sample scope: POS loan data from a large German bank, one of Germany's top-5 retail banks by assets, with 1,000+ branches and 10 million+ retail customers. Borrowers are first-time delinquent POS loan holders with a valid phone number, no prior bank relationship, not in bankruptcy, entering the call center pool January-June 2012. Outcome follow-up through August 2014 (~2.5 years after delinquency).

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13388) if you are:
studying the value of human vs. automated customer communication; evaluating early collection call center policies; applying the Heckman-Vytlacil MTE framework to estimate ATEs from IV complier effects; studying how social distance and agent characteristics mediate behavioral treatment effects; or comparing German POS and overdraft loan markets. Table IV (p. 534) is the headline result; Table XI (p. 553) has the long-run effects; Table X (p. 551) has the voice likeability evidence.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(1), February 2025. Published online November 28, 2024. DOI: [10.1111/jofi.13388](https://doi.org/10.1111/jofi.13388). Copyright 2024 the American Finance Association. Paywalled. This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. Extract-only; no PDF hosted here.

> Laudenbach, Christine, and Stephan Siegel. "Personal Communication in an Automated World: Evidence from Loan Repayments." *The Journal of Finance* 80, no. 1 (February 2025): 515-559. DOI: 10.1111/jofi.13388.


==============================================================================
# Regulating Over-the-Counter Markets: Lee & Wang (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/lee-regulating-counter-markets-2025/
# Distilled: Lee and Wang embed dealer cream skimming via price discrimination into a Glosten-Milgrom framework and show that restricting OTC dealer discrimination worsens aggregate volume and average spreads yet can raise utilitarian welfare whenever adverse selection risk is low, via a novel cheap-substitution mechanism. J. Finance 2025, CC BY 4.0. Six core results with source locators, datasets used (theoretical; empirical patterns in Internet Appendix), the model, and the method.
# Tags: paper-summary, market-microstructure, otc-markets, regulation, price-discrimination
==============================================================================

**What this is.** This is the distilled skeleton of Lee and Wang (2025), "Regulating Over-the-Counter Markets," J. Finance 80(4): 1929-1962. It extracts the model, propositions, and key theoretical results with PDF locators. Read the [original paper](https://doi.org/10.1111/jofi.13461) to replicate or extend.

## TL;DR

Lee and Wang (2025) embed dealer cream skimming via price discrimination into the standard Glosten and Milgrom (1985) exchange model, allowing the OTC dealer to observe a public label (Likely Informed vs. Likely Uninformed) before quoting. The venue-choice between OTC and exchange follows Seppi (1990); the label-dependent pricing and cream skimming also relate to Desgranges and Foucault (2005), who show how dealer-client relationships concentrate adverse selection on exchanges. In equilibrium LI traders go to the exchange and LU traders go to the OTC market. The paper's central finding is that restricting the dealer's ability to discriminate across labels always reduces aggregate volume and widens the average spread, yet can raise utilitarian welfare whenever adverse selection risk (the mass of informed traders) is low. The mechanism is "cheap substitution": pooling causes uninformed traders with larger hedging benefits (previously on the exchange) to enter while uninformed traders with smaller benefits (previously in the OTC market) exit, and the entrants more than offset the exiters in welfare terms when adverse selection risk is low. This contradicts Bolton, Santos, and Scheinkman (2016), who find cream skimming raises welfare via origination effort incentives. The model also contrasts with Akerlof (1970): the paper shows cheap substitution reverses in that standard framework because private and common values are perfectly correlated there. On the price discrimination side, Bergemann, Brooks, and Morris (2015) show that without adverse selection any welfare outcome is achievable; this paper shows adverse selection provides robust guidance. The paper derives an optimal Pigouvian tax on OTC trades and a simple implementable "WSR Rule" based on the ratio of exchange to OTC spreads and volume changes, without requiring knowledge of structural parameters.

## Core results

| # | Result | Locator | Magnitude as reported |
|---|---|---|---|
| R1 | Welfare effect of restricting OTC dealer: raises welfare when informed mass μ is small, lowers it when μ is large | Proposition 2, pp. 1938-1939 | Single cutoff under quasiconcave ΔW; any commonly used F satisfies conditions |
| R2 | Volume and spread always worsen under restriction | Proposition 2(e), p. 1939 | Aggregate volume V strictly falls and average spread S̄ strictly widens for all μ > 0, under any distribution with decreasing ΔV |
| R3 | Optimal Pigouvian tax dominates OTC closure | Proposition 5, p. 1944 | T\* strictly raises welfare above closing the OTC market for any (α, θ, γ); unique under U[0,1] |
| R4 | WSR Rule implements T\* without structural estimation | Proposition 6, pp. 1944-1945 | WSR(T) > 1: raise T; WSR(T) < 1: cut T; WSR = 1 at T\*; WSR ≡ \|S\_E dV\_E / (S\_O dV\_O)\| |
| R5 | High OTC market share is not evidence against restriction | Proposition 4, pp. 1942-1943 | OTC share V\_O/V is decreasing in μ; restriction raises welfare exactly where OTC share is high |
| R6 | Empirical: exchange spread and exchange share positively correlated across U.S. equities | Internet Appendix §VI.C | Positive correlation of quoted spread with exchange market share corroborates model prediction |

**Overall (paper's conclusion).** Trading costs, volumes, and market shares can all mislead regulators because cheap substitution decouples these aggregates from welfare. The adverse selection parameter β, not aggregate volume or OTC share, is the correct policy signal. When β is low, restricting OTC price discrimination strictly improves welfare and the optimal tax can be approximated by the WSR statistic computed from observable trade data.

## Theory / model

The paper has no empirical design; the identification is structural (model restrictions). The model and propositions are the core content.

**Setup.** A continuum of risk-neutral traders may trade an indivisible asset in a three-stage game (Figure 2, p. 1935). The asset pays $$v \in \{-1, +1\}$$ with equal probability. A mass μ of traders are informed (private binary signal $$s_i = v$$ with probability α and $$-v$$ otherwise; α is signal accuracy). A mass 1 of uninformed traders each draw a hedging benefit $$b_i \overset{\text{iid}}{\sim} F$$ with support $$[0,1]$$. Each trader is publicly labeled $$\ell_i \in \{\text{LI}, \text{LU}\}$$; the odds of being informed conditional on label LI exceed the unconditional odds:

$$\mathcal{O}_{LI} = \frac{\theta\mu}{1-\gamma} > \mu > \frac{(1-\theta)\mu}{\gamma} = \mathcal{O}_{LU}$$

where $$\theta < 1$$ is the probability an informed trader is labeled LI and $$\gamma < 1$$ is the probability an uninformed trader is labeled LU.

**Equilibrium.** Each trader chooses to buy, sell, or exit on the exchange, over the counter, or not trade (Definition 1, p. 1936). The OTC dealer observes the label $$\ell_i$$ before quoting; the exchange dealer does not and sets a single unconditional spread. The zero-profit condition for the dealer posting spread $$s$$ when the informed ratio in her pool is $$\beta$$ is (p. 1937, Eq. 2):

$$\underbrace{s \cdot (1 - F(s))}_{\text{Profit from uninformed}} = \underbrace{(2\alpha - 1 - s)^+ \cdot \beta}_{\text{Loss to informed traders}} \tag{2}$$

The unique solution $$S(\beta)$$ is increasing in $$\beta$$.

**Proposition 1 (equilibrium spreads, p. 1937).** The exchange spread is $$S_E = S\!\left(\frac{\theta}{1-\gamma}\mu\right)$$; the OTC spread for LU traders is $$S_O = S\!\left(\frac{1-\theta}{\gamma}\mu\right)$$ and for LI traders is $$S_E$$. LI traders choose the exchange, LU traders choose the OTC market. The exchange spread strictly exceeds the OTC spread: $$S_E > S_N > S_O$$ where $$S_N = S(\mu)$$ is the no-OTC spread.

**Welfare and volume.** Welfare $$W$$ equals the sum of hedging benefits of uninformed traders who trade. The average bid-ask spread is $$\bar{S} \propto 1/V$$. Two quantities characterize the restriction's effect (pp. 1937-1938):

$$\Delta_V(\beta) := -\!\left(\int_{S(\beta)}^1 f(s)\,ds\right)' = S'(\beta)\cdot f(S(\beta)) \tag{3}$$

$$\Delta_W(\beta) := -\!\left(\int_{S(\beta)}^1 sf(s)\,ds\right)' = \Delta_V(\beta)\cdot S(\beta) \tag{4}$$

Marginal volume $$\Delta_V$$ is the fall in uninformed trade per unit increase in $$\beta$$; marginal welfare $$\Delta_W$$ equals $$\Delta_V$$ times the marginal exiter's hedging benefit $$\bar{b}(\text{exiters}) = S(\beta)$$.

**Proposition 2 (main result, pp. 1938-1939).** For any commonly used distribution F, restricting the OTC dealer:
- (a)-(c): raises welfare W if $$\mu < \mu_l$$ (small informed mass) and lowers W if $$\mu > \mu_h$$;
- (d): if $$\Delta_W$$ is strictly quasiconcave, the two cutoffs collapse to one: restriction raises W iff $$\mu < \underline{\mu}$$;
- (e): always strictly reduces aggregate volume V and strictly widens average spread $$\bar{S}$$.

**Proposition 3 (sufficient conditions, p. 1941).** $$\Delta_W$$ is strictly quasiconcave iff

$$\frac{(2\alpha-1)(1-F(x))}{xf(x)(2\alpha-1-x)^2} - \frac{1}{2\alpha-1-x} \quad\text{is strictly quasiconvex on }(0,2\alpha-1). \tag{5}$$

Any Beta(a, b) distribution satisfies (5) for all $$a,b > 0$$.

**Proposition 4 (market shares, pp. 1942-1943).** Under condition (7) (satisfied by uniform and beta distributions), OTC share $$V_O/V$$ is strictly decreasing in $$\mu$$. Therefore restricting the OTC dealer raises welfare precisely where the OTC market share is high, directly overturning the industry argument that high OTC share signals OTC efficiency.

## Method

This is a pure theory paper with no structural estimation. The solution method is closed-form zero-profit conditions plus comparative statics via the Implicit Function Theorem and integral inequalities. Proofs are in the Appendix (pp. 1949-1962).

**Pigouvian tax characterization.** A lump-sum tax $$T$$ on OTC trades shifts the dealer zero-profit conditions to (pp. 1943-1944):

$$S_O(T)\cdot\left[1-F(S_O(T))+\beta_O\right]\gamma = (2\alpha-1)\beta_O\gamma + T \tag{8}$$

$$S_E(T)\cdot\left[1-F(S_E(T))+\beta_E\right](1-\gamma) = (2\alpha-1)\beta_E\cdot(1-\gamma) - T \tag{9}$$

All implementations that raise the same gross revenue T are equivalent. The optimal Pigouvian tax T\* maximizes welfare W. Proposition 5 (p. 1944) shows T\* strictly dominates OTC market closure.

**WSR Rule (Proposition 6, p. 1945).** Define the Weighted Spread Ratio:

$$\text{WSR}(T) := \left|\frac{S_E(T)\,dV_E(T)}{S_O(T)\,dV_O(T)}\right| \tag{10}$$

The welfare-increasing direction of T is: raise T when WSR > 1 (cheap substitution dominates), cut T when WSR < 1 (volume effect dominates). The optimal tax satisfies WSR(T\*) = 1. The ratio $$S_E/S_O$$ is observable as the ratio of exchange to OTC spreads; $$|dV_O/dV_E|$$ is the ratio of volume changes before and after a policy perturbation.

## Empirical specifications

The paper is theoretical; there is no estimation. The Internet Appendix (Sections V and VI) documents supporting empirical patterns without a causal design.

**Empirical pattern 1 (Internet Appendix §V).** The exchange's market share and spread are positively correlated because both are driven by the informed ratio β. As β rises, the OTC market absorbs more informed traders so the exchange spread rises and OTC share rises mechanically.

**Empirical pattern 2 (Internet Appendix §VI.C).** The total market share of exchanges and their quoted spreads are positively correlated across U.S.-listed equities, corroborating the model prediction (p. 1933). This is documented as a novel empirical pattern, not a causal test.

The paper explicitly leaves causal estimation of the welfare effects of OTC restrictions for future work.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| U.S. equity market microstructure data (quoted spreads, exchange market share) | Empirical corroboration in Internet Appendix §VI.C | no page yet |

The core model is theoretical; no large dataset is used for estimation.

## When to read the full paper

Read Lee and Wang (2025) when:
- Designing or evaluating OTC market regulations (post-trade transparency, name give-up rules, blockchain record-keeping proposals discussed in the Internet Appendix §I);
- Studying how to implement a Pigouvian tax on OTC trading without needing structural estimates of adverse selection risk, using the WSR Rule;
- Understanding why aggregate volume, average spread, and OTC market share are unreliable welfare indicators in two-venue markets;
- Modeling OTC-versus-exchange venue choice with imperfect trader labels in a Glosten-Milgrom framework.

The key propositions (Propositions 2, 4, 5, 6) are in §II and §III (pp. 1938-1945). Appendix proofs are pp. 1949-1962.

## Attribution and rights

This article is open access under CC BY 4.0. The canonical citation is:

> Lee, Tomy and Chaojun Wang. "Regulating Over-the-Counter Markets." *The Journal of Finance* 80, no. 4 (August 2025): 1929-1962. https://doi.org/10.1111/jofi.13461

Copyright 2025 The Authors. Published by Wiley Periodicals LLC on behalf of American Finance Association. Open access funding provided by Central European University Private University - CEU GmbH/KEMO.

This page is LLM-distilled by the IAR paper-distiller (claude-sonnet-4-6). It is not human-verified and the results have not been independently reproduced. Extract-only; the original PDF is not mirrored here (CC BY 4.0 permits redistribution by any party with proper attribution).


==============================================================================
# Decentralized Exchange: Lehar & Parlour (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/lehar-decentralized-exchange-uniswap-automated-2025/
# Distilled: Lehar and Parlour build a theoretical model of Uniswap's automated market maker (AMM), characterize equilibrium liquidity-pool size as a trade-off between fee revenue and adverse-selection (picking-off) risk, and show empirically that AMM pools are larger when volatility is lower and uninformed trading is higher, that AMM liquidity is more stable than limit-order book liquidity during extreme market events, and that Uniswap price impact is lower than Binance for low-volatility tokens. J. Finance 2025, paywalled. Four core results with source locators, datasets used, the model (constant-product AMM + limit-order-book comparison), and the estimating specifications.
# Tags: paper-summary, market-microstructure, defi, decentralized-exchange, automated-market-maker
==============================================================================

**What this is.** The paper's core results, the model (constant-product AMM and limit-order-book comparison), and the empirical specifications: enough to know what it found and how, without reading the full 54-page paper. To replicate or extend, read the original at [doi.org/10.1111/jofi.13405](https://doi.org/10.1111/jofi.13405).

This paper builds on Glosten (1994), the canonical model of limit-order-book efficiency under adverse selection, and extends the comparison to AMMs. It tests against Capponi and Jia (2021), who model AMM competition among arbitrageurs. It also draws on Angeris and Chitra (2020), who show constant-function market makers can reflect true prices. The empirical price-impact results complement Barbon and Ranaldo (2021), who compare DEX and Binance transaction costs for five token pairs. The analysis of HFT and strategic liquidity provision relates to Biais, Foucault, and Moinas (2015).

## TL;DR

Lehar and Parlour analyze Uniswap, the largest decentralized exchange, as a new model of liquidity provision. In an automated market maker (AMM), liquidity suppliers passively post capital into pools rather than actively setting prices; price impact is determined mechanically by a bonding curve. The paper develops a model showing that equilibrium pool size trades off fee revenue against picking-off risk: pools are larger when token volatility is lower and noise trading is higher. Using 95.8 million Uniswap transactions from November 2018 to December 2022 across 105,098 liquidity pools, the paper finds that (i) pool size decreases in volatility and increases in uninformed trading, consistent with theory; (ii) AMM liquidity is substantially more stable than limit-order-book liquidity during extreme market events; (iii) price impact on Uniswap is lower than on Binance (a centralized exchange) for low-volatility, noise-dominated tokens; and (iv) Binance price impact becomes less volatile and converges toward Uniswap after integrating PancakeSwap, an AMM clone, in March 2022. The paper also shows conditions under which the AMM dominates a limit-order market and documents absence of long-lived arbitrage opportunities.

## Core results

Magnitudes as reported; `\*\*\*` = 1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Pool size decreases in volatility and increases in uninformed trading** | Table II, p. 341 | Std.Dev FX Rate coefficient: -0.152\*\*\* to -0.249\*\*\* across 6 specs (col 1: -0.201; col 2: -0.249; col 3: -0.204; col 4: -0.163; col 5: -0.158; col 6: -0.152); Reversals: 49.18\*\*\* (13.68); Number Trades: 6.633\*\*\* (2.088); R2 = 0.21-0.55 |
| R2 | **AMM liquidity is stable during extreme market events**; gas fees and fee revenue both increase, discouraging withdrawal | Table III, p. 344; Figure 9, p. 343 | Only 2% of ETH-stablecoin liquidity withdrawn during 41% ETH price crash (May 19, 2021); gas fees rise 1.71 USD (10%) and fee revenue rises 17.7 USD on high-return days |
| R3 | **Uniswap price impact is lower than Binance** on average; gap narrows after Binance integrates AMM clone PancakeSwap | Table IV, p. 352 | rPI intercept significant and positive across cols (col 1: 4.080\*\*\*; col 2: 4.956\*\*\*; col 3: 3.126\*\*\*; col 8: 3.963\*\*\*); Pancake Swap dummy -2.497\*\*\* to -3.085\*\*\* across pool-FE specs; effect stronger at low volatility and medium trade sizes |
| R4 | **Price impact is more volatile on Binance than Uniswap**; volatility gap drops after PancakeSwap integration | Table V, p. 353 | Relative volatility rV intercept: col 1: 33.95\*\*\*; col 2: 39.00\*\*\*; col 8: 30.47\*\*\*; col 9: 25.91\*\*\*; Pancake Swap dummy: col 1: -13.63\*\*\*; col 8: -18.16\*\*\*; col 9: -7.756\*\*\*; pattern holds for low exchange-rate volatility and medium trade sizes |

**Overall (paper's conclusion).** Uniswap's AMM mechanism successfully provides stable, predictable liquidity at lower and less volatile price impact than a centralized limit-order book for asset pairs with lower volatility and more noise trading. The equilibrium pool size adjusts so that fee revenue compensates liquidity suppliers for adverse selection, and gas fees on the blockchain act as a commitment device that discourages strategic withdrawal during market turmoil.

## Theory / model

The model has a single asset with current value $$p_0$$. With probability $$\alpha$$, an innovation occurs and the value jumps to $$p_0 + \sigma$$ or $$p_0 - \sigma$$ with equal probability; otherwise the value remains at $$p_0$$. Three agents interact: risk-neutral liquidity suppliers, a liquidity demander (noise trader) who trades a fixed quantity $$q$$, and an informed arbitrageur who trades whenever profitable.

**Constant-product bonding curve.** For a pool with $$E_0$$ units of ETH and $$T_0$$ tokens, the bonding curve constant is (eq. 1, p. 326):

$$
k := T_1 \cdot E_1 = T_0 \cdot E_0. \tag{1}
$$

Any trade must stay on this curve. When a trader buys $$t$$ tokens by depositing $$e$$ ETH, the fee $$\tau$$ is collected, the post-trade ETH pool becomes $$E' = E + (1-\tau)e$$, and the post-trade token balance is (eq. 2, p. 327):

$$
T' = \frac{T \cdot E}{E'} = \frac{T \cdot E}{E + (1-\tau)e}. \tag{2}
$$

The token received by the trader is $$t = T - T'$$ (eq. 3), and the terms of trade in ETH per token are (eq. 4):

$$
p^{tot} = \frac{e}{t} = \frac{e}{T} + \frac{E}{(1-\tau)T}. \tag{4}
$$

In the limit as $$e \to 0$$, the spread relative to the fundamental $$p_0$$ is:

$$
\lim_{e \to 0} \frac{p^{tot}}{p_0} = \frac{ET}{ET(1-\tau)} = \frac{1}{1-\tau}. \tag{5}
$$

**Limit-order-book equilibrium.** Competing liquidity suppliers choose private investment $$\gamma_i$$ (e.g., co-location speed) at cost $$I(\gamma) = a\gamma^2$$ to become the monopolist with probability $$\gamma_i(1 - \gamma_j)$$. In symmetric equilibrium (Proposition 1, p. 330):

$$
\gamma^* = \frac{(1-\alpha)\sigma q}{2a + \sigma q(1-\alpha)}.
$$

A monopolist posts at $$p_0 + \sigma$$ (sell) and $$p_0 - \sigma$$ (buy), earning $$(1-\alpha)\sigma$$. With two competing suppliers, each earns zero and prices are $$p_0 + \alpha\sigma$$ and $$p_0 - \alpha\sigma$$ (Lemma 2, p. 330).

**AMM equilibrium pool size.** In the AMM, liquidity provision is non-rivalrous and payoffs are shared pro rata. There is no incentive for private investment. The equilibrium pool size (Proposition 2, p. 333, eq. 16) balances fee revenue from the noise trader against picking-off losses from the arbitrageur:

$$
T_0 = q \left[ \sqrt{1 + \frac{(1-\alpha)^2 \tau^2 p_0^2}{\alpha^2 \omega^2}} - \frac{(1-\alpha)\tau p_0}{\alpha \omega} \right], \tag{16}
$$

where $$\omega = \sqrt{p_0(p_0+\sigma)(1+\tau)} + \sqrt{\frac{p_0(p_0-\sigma)}{1+\tau}} - 2p_0$$. The equilibrium pool size is linear in noise-trade volume $$q$$, decreasing in innovation size $$\sigma$$, and decreasing in innovation probability $$\alpha$$ (Proposition 3, p. 339).

**AMM vs. limit-order-book trading costs.** Expected cost per unit on the limit-order book is (Proposition 4, p. 345):

$$
E(c^{\text{limit}}) = \sigma((1-\eta)\alpha + \eta),
$$

where $$\eta = 2\gamma^*(1-\gamma^*)$$ is the probability of facing a monopolist. Expected cost on the AMM is:

$$
E(c^{\text{AMM}}) = p_0 \left(\frac{(1+\tau)\lambda^b - (1-\tau)\lambda^s}{2}\right),
$$

where $$\lambda^b(\tau, \alpha, \sigma) > 1$$ and $$\lambda^s(\tau, \alpha, \sigma) < 1$$ are functions of equilibrium pool size. Proposition 5 shows the limit-order book does not universally dominate the AMM: there exists a critical innovation probability $$\alpha^*$$ below which the AMM is strictly preferred, and conditional on trade quantity, price impact is more volatile in the limit-order book.

## Method

The paper combines a stylized two-market equilibrium model with reduced-form panel regressions.

**Structural model.** Equilibrium is derived analytically for both markets. The AMM equilibrium pool size (equation 16) is the closed-form solution to the indifference condition in equation (A10) (Appendix A, p. 358). The model builds on the `amm-equilibrium-pool-size` framework, which is the paper's primary methodological contribution.

**Panel regressions for pool size (R1).** The estimating equation (eq. 17, p. 340) is at the pool-day level:

$$
\text{pool size} = a + b_1 \sigma_{fx} + b_2 \mathbf{1}_{\text{airdrop}} + b_3 \text{noise trading}, \tag{17}
$$

where pool size is daily average USD pool size, $$\sigma_{fx}$$ is the annualized block-by-block exchange-rate volatility (proxy for adverse selection), and noise trading is measured by three proxies: number of trades per day, daily volume, and immediate trade reversals (trades followed within 75% of the same size in the opposite direction). Standard errors are clustered by pool and by day. Robustness: specifications with and without pool-and-day fixed effects.

**Price impact comparison (R3, R4).** Relative price impact is defined as (eq. 18, p. 350):

$$
rPI = \frac{PI_{\text{Binance}}}{PI_{\text{Uniswap}}} - 1, \tag{18}
$$

estimated at the daily level for 43 token pairs cross-listed on both venues. The estimating equation (eq. 19, p. 350) is:

$$
rPI = a + b_1 \mathbf{1}_{\text{pancake}} + b_2 \sigma_{FX} + b_3 \text{trade size} + b_4 \text{trade size}^2 + b_5 \text{noise trading} + \epsilon. \tag{19}
$$

The PancakeSwap dummy $$\mathbf{1}_{\text{pancake}}$$ equals one after March 25, 2022 (when Binance integrated a Uniswap clone). Relative volatility of price impact is defined analogously (eq. 20, p. 351):

$$
rV = \frac{V_{\text{Binance}}}{V_{\text{Uniswap}}} - 1, \tag{20}
$$

using the same specification. Standard errors are clustered by pool and by day.

## Empirical specifications

All regressions are panel (pool-day or pool observations) with standard errors clustered by pool and by day unless noted.

**R1: Pool size and volatility/noise trading.** Pool-day level on 1,525 pools (997,507 observations). Dependent variable: daily average pool size in million USD. Key regressors: exchange-rate volatility (Std.Dev FX Rate, annualized daily sd of block-by-block price changes), airdrop dummy, and three alternative noise-trading proxies (daily USD volume, number of trades, and reversals). Six columns spanning no-FE and pool-and-day FE; R2 = 0.21-0.55 (Table II, p. 341).

**R2: Stability during extreme events.** Analyzed via Figures 9 and Table III. Gas fees and fee revenue are regressed on High Return dummy (absolute daily price change > 10%) and absolute return. Pool fixed effects. 1,303,869 observations. Gas fees rise 1.71 USD on high-return days; fee revenue rises 17.73 USD (Table III, p. 344).

**R3/R4: Relative price impact AMM vs. Binance.** 43 cross-listed token pairs; 24,963 pool-day observations in col (1) of Table IV (pool FE only, full sample without pancake dummy); 24,224 in cols with pancake dummy included; 21,409 in Table V col (1). Specifications with pool FE only and pool-and-day FE. Pre/post-PancakeSwap subsamples in columns (8) and (9) (Table IV, p. 352; Table V, p. 353).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Uniswap V1 and V2 blockchain data (Ethereum) | Primary data: 95.8 million interactions across 105,098 pools; liquidity injections, withdrawals, token swaps Nov 2018-Dec 2022 | [Uniswap on-chain](/wiki/datasets/uniswap-blockchain/) |
| Binance minute-by-minute price data | Price benchmark for USD volume conversion and price impact comparison; 43 cross-listed token pairs | No page yet |
| Ethereum blockchain gas-price data | Gas cost estimation for AMM withdrawal transactions | No page yet |

Sample: November 2, 2018 to December 21, 2022 (Uniswap V1 launch through sample end); reduced econometric sample: 1,525 pools with at least 30 trading days and 100 ETH average balance (59,606,977 observations).

## When to read the full paper

Read the original if you are: building models of AMM or decentralized exchange mechanisms; studying how blockchain-specific costs (gas fees) affect market design and liquidity stability; comparing trading costs across centralized and decentralized venues; analyzing the market-microstructure implications of DeFi protocols; or extending the model to concentrated-liquidity AMMs (Uniswap V3) or multi-pool settings. Exact tables are at pages 341 (pool size), 344 (stability), and 352-353 (price impact comparison).

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(1), February 2025, pp. 321-374. Copyright 2024 the American Finance Association. This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The paper is paywalled; only text extracts are reproduced here under fair-use principles.

> Lehar, Alfred, and Christine A. Parlour. "Decentralized Exchange: The Uniswap Automated Market Maker." *The Journal of Finance* 80, no. 1 (February 2025): 321-374. DOI: 10.1111/jofi.13405.


==============================================================================
# Women in Charge: Lewellen (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/lewellen-women-charge-evidence-hospitals-2025/
# Distilled: Female hospital CEOs make similar financial and investment decisions as male peers, respond identically to the 2008 financial crisis, but earn 32% lower pay (shrinking to 7.8% within-hospital) and face significantly higher turnover after poor performance. J. Finance 2025, paywalled. Seven core results with source locators, datasets used, and the empirical specifications.
# Tags: paper-summary, labor-careers-health, corporate-governance, gender, executive-compensation
==============================================================================

**What this is.** The paper's core results, the datasets it assembles, and the
empirical specifications behind the compensation and turnover findings: enough
to know what it found and how, without reading all 55 pages. To replicate or
extend it, read the full source at [doi:10.1111/jofi.13455](https://doi.org/10.1111/jofi.13455).

## TL;DR

The paper uses 19 years of IRS Form 990 filings for U.S. nonprofit hospitals
(2000-2018) to study female hospital CEOs: what kinds of hospitals they lead,
how they make decisions under stress, and how boards compensate and replace them.
Contrary to the gender-differences literature, female CEOs do not run hospitals
with safer financing, lower investment, or more charitable orientation, and they
respond to the 2008 financial crisis identically to male peers. However, female
CEOs earn 32% lower unconditional pay (7.8% within hospital after controls),
receive flatter pay-for-performance incentives (5.3 pp vs. 15.0 pp pay rise from
bottom to top performance quintile), and are fired more often after poor
performance (3.6 pp higher departure rate in the bottom quintile). The patterns
are consistent with hospital boards perceiving female CEOs as less productive,
whether due to true differences or biased beliefs.

## Core results

Magnitudes and significance are as reported in source tables; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **No evidence female CEOs match or shift hospitals toward safer financing or lower investment**: coefficients on leverage, cash, securities, and investment are statistically insignificant across all specifications | Table II, pp. 2212, 2214 | One-SD increase in leverage raises female CEO probability by an insignificant 0.48 pp (col. 2); can reject a decline of more than 0.70 pp; similar bounds for cash and investment |
| R2 | **No evidence female CEOs lead hospitals with greater charitable orientation**: Medicaid Ratio, Charity Care, and Medicare Ratio are all insignificant with hospital FE; Contributions positive at 10% | Table III, pp. 2213, 2215-2217 | Contributions coefficient: 0.83 pp per 1-SD increase (col. 8); all other benevolence measures insignificant within hospital |
| R3 | **Hospitals with male and female CEOs respond identically to the 2008 financial crisis**: interaction Post\_Crisis x Female is insignificant for revenue, profit margin, investment, employment, salaries, and Medicaid Ratio | Table V, pp. 2221-2222; Figure 3, p. 2222 | Female hospitals cut revenue by 3.0 pp, males by 2.9 pp; profit margin declines 1.6 pp vs. 1.3 pp; investment 7.2 pp vs. 6.9 pp; none of the differences is significant |
| R4 | **Female CEOs earn substantially lower pay**: unconditional gap is 32.2%, falling to 12.5% after controlling for hospital size, and further to 7.8% with hospital FE and full controls | Table VII, p. 2225; Table VI, p. 2224 | Female CEO coefficient: -0.322\*\*\* (col. 1), -0.125\*\*\* (col. 2), -0.078\*\* (col. 5) in log pay regressions |
| R5 | **Pay-for-performance sensitivity is significantly flatter for female than male CEOs**: men's pay rises 15.0 pp from worst to best performance quintile; women's only 5.3 pp | Table VIII, p. 2229; Figure 4, p. 2230 | Female CEO x Perf\_Quintile5: -0.096\*\* (0.046) in col. 4; -0.105\*\* (0.048) with age dummies in col. 5 |
| R6 | **Turnover-performance sensitivity is significantly higher for female CEOs**: incremental departure rate in the bottom quintile is 3.6 pp higher for women than men (forced turnover); 3.2 pp for all turnovers | Table IX, pp. 2232-2233; Figure 5, p. 2233 | Marginal effect of Female CEO x Perf\_Quintile1: 3.6 pp (col. 4, forced, significant at 5%); 3.2 pp (col. 8, all turnovers, significant at 5%) |
| R7 | **Board gender composition does not explain the pay gap**: correlation between female board share and female CEO indicator disappears with hospital FE; the gap is not systematically smaller where female trustees are more common | Table X, pp. 2238-2239 | Female Directors Non-CEO: 0.146\*\*\* (col. 2) without FE; 0.011 (col. 4) with hospital FE; Panel B interaction with gender pay gap insignificant |

**Overall (paper's conclusion).** Hospitals led by male and female CEOs look
similar on observable financial and operational measures, pursue similar
strategies, and respond similarly to external shocks. At the same time,
compensation and turnover decisions are made as though female CEOs are less
productive: they receive lower pay, flatter incentives, and steeper
performance-conditioned dismissal probabilities. The evidence is consistent with
boards holding biased beliefs about CEO ability, though true productivity
differences cannot be ruled out.

## Theory / model

The paper has no formal structural model. The tests are motivated by two classes
of theory:

**Agency and efficiency benchmark for pay.** The paper follows the assignment
models of CEO compensation (Gabaix and Landier (2008), Tervio (2008)) as an
efficiency benchmark: in competitive markets for CEO talent, pay reflects the
CEO's marginal product of skill. A gender gap in pay, under this lens, reflects
a perceived productivity difference. Boards may correctly assess CEO ability
(true skill gap) or may hold biased priors (stereotype-based beliefs), consistent
with Bayesian learning models of CEO turnover (Hermalin and Weisbach (1998),
Taylor (2010)). The unconditional gender gap documented here is broadly consistent
with the findings of Bertrand and Hallock (2001) for S&P 1500 executives, extended
here to the CEO level within a single industry over a longer panel.

**Principal-agent incentive theory.** The moral hazard framework of Holmstrom
and Milgrom (1987) predicts that optimal incentive steepness (pay-for-performance
sensitivity) is higher for less risk-averse, higher-effort, or higher-productivity
CEOs. Flatter incentives for women are consistent with boards perceiving them as
more risk-averse or less productive. The matching between CEO type and board
beliefs also predicts lower firing thresholds for lower-prior-ability CEOs, which
is consistent with the finding that female CEOs are fired more often at the bottom
of the performance distribution.

**Identification.** The core hospital-level regressions exploit within-hospital
variation over time (hospital fixed effects) to absorb time-invariant confounders.
For the financial crisis analysis, k-nearest-neighbor matching on 2007 hospital
attributes creates treatment and control samples that are balanced on observables
immediately prior to the shock, attenuating concerns about selection on trends.
The crisis itself is a plausibly exogenous, unexpected shock to hospital finances
(stock market crash, credit crunch, demand decline via unemployment) that
provides a high-stakes setting where CEO preferences, if present, should be most
visible. The data infrastructure for the crisis analysis follows Adelino,
Lewellen, and McCartney (2021) and Adelino, Lewellen, and Sundaram (2015),
who also use IRS Form 990 and AHA data to study hospital investment and governance
decisions.

The results contradict prior evidence from European private-firm samples. Faccio,
Marchica, and Mura (2016) find that firms led by female CEOs take less risk and
choose safer financing; Huang and Kisgen (2013) find that female executives make
fewer acquisitions and issue less debt. The hospital evidence shows no such
financing or investment differences, consistent with the view that self-selection
into top executive roles reduces or eliminates gender differences in behavior
documented in broader populations (Adams and Funk (2012)). Similarly, Matsa and
Miller (2013) find that Norwegian board quotas lead to labor hoarding, interpreted
as a benevolence effect of female board members; the hospital tests of Medicaid
mix, charity care, and employment response to the crisis find no analogous channel.

## Method

**Hospital fixed-effects panel regressions (Tables II, III, VII, VIII).** The
main estimating equation across all parts of the paper is:

$$
Y_{ht} = \alpha + \beta \, \text{Female}_{ht} + \gamma X_{ht-1} + \mu_h + \tau_t + \varepsilon_{ht} \tag{1}
$$

where $$Y_{ht}$$ is an outcome for hospital $$h$$ in year $$t$$ (leverage, log CEO pay,
Charity Care, etc.), $$\text{Female}_{ht}$$ is an indicator for a female CEO,
$$X_{ht-1}$$ is a vector of lagged hospital controls (log service revenues, profit
margin, growth in services, contributions, density decile of the HSA, system
membership, CEO tenure, and multiple positions), and $$\mu_h$$ and $$\tau_t$$ are
hospital and year fixed effects. Standard errors are clustered at the hospital
level throughout. The coefficient $$\beta$$ identifies the effect from
within-hospital CEO gender transitions (pp. 2213-2214, 2225).

**Matching + DiD for the financial crisis (Table V).** For the crisis test, the
paper constructs a matched sample using k-nearest-neighbor matching ($$k = 3$$,
matching with replacement) on 2007 hospital attributes: service revenues, HSA
population density rank, investment, revenue growth, and system membership.
The matched sample has 134 treated (female CEO in 2007) and 271 control (male CEO
in 2007) hospitals. The estimating equation is:

$$
Y_{ht} = \alpha + \delta \, (\text{Female}_h \times \text{Post\_Crisis}_t) + \mu_h + \tau_t + \varepsilon_{ht} \tag{2}
$$

estimated on the window 2006-2011, where $$\text{Post\_Crisis}_t = 1$$ for 2009-2011.
The coefficient $$\delta$$ is identified from the differential within-hospital change
from pre- to post-crisis between the matched treatment and control groups,
a difference-in-differences design (p. 2220).

**Probit for CEO turnover (Table IX).** The turnover regressions are probit models:

$$
\Pr(\text{Turnover}_{ht} = 1) = \Phi\!\left( \alpha + \beta_1 \, \text{Female}_{ht} + \sum_{q=1}^{4} \beta_q \, \text{Perf\_Quintile}_{qht} + \gamma \, (\text{Female}_{ht} \times \text{Perf\_Quintile1}_{ht}) + \delta X_{ht} + \mu_h + \tau_t \right) \tag{3}
$$

where $$\Phi$$ is the standard normal CDF, performance quintiles are formed by year
and hospital size bin based on the lagged profit margin (Perf\_Quintile5 = best
is excluded; Perf\_Quintile1 = worst is the main interaction), and the equation
is estimated separately for forced turnover (CEO aged 60 or younger at departure)
and all turnover. Marginal effects are reported (pp. 2231-2234).

**Pay-for-performance regressions (Table VIII).** The pay sensitivity analysis
runs OLS of log CEO pay on female dummy, performance quintile indicators, and
their interactions:

$$
\ln(\text{CEO Pay}_{ht}) = \alpha + \beta_1 \, \text{Female}_{ht} + \sum_{q=2}^{5} \beta_q \, \text{Perf\_Quintile}_{qht} + \sum_{q=2}^{5} \gamma_q \, (\text{Female}_{ht} \times \text{Perf\_Quintile}_{qht}) + \delta X_{ht} + \mu_h + \tau_t + \varepsilon_{ht} \tag{4}
$$

where $$\gamma_5$$ on the Female x Perf\_Quintile5 interaction captures the
differential pay-performance slope at the top of the distribution (pp. 2228-2230,
Table VIII).

## Empirical specifications

- **Matching and selection tests (Tables II, III, Sections IV and V).** Panel OLS
  (equation 1) with and without hospital FE; sample 23,611 to 25,513 hospital-years
  depending on completeness. Controls include log service revenue, profit margin,
  growth in services, density HSA, competition rank. Tables II and III report
  coefficients on the female CEO indicator for financial risk (leverage,
  cash/assets, securities/assets, investment) and patient orientation (Medicaid,
  Medicare, Charity Care, Contributions) respectively. The within-hospital
  specifications (cols. with hospital FE) are the headline tests.

- **Financial crisis DiD (Table V).** Matched sample, 2006-2011, hospital and year
  FE, N = 2,251-2,287 hospital-years depending on variable. Dependent variables:
  growth in services, profit margin, investment, growth in personnel, growth in
  salaries, Medicaid Ratio. The interaction Post\_Crisis x Female is the parameter
  of interest in Panel A; Panels B and C show results separately for the treatment
  and control samples.

- **Gender pay gap (Table VII).** Log CEO pay on Female CEO indicator, year FE,
  progressively adding: log service revenues (col. 2), full hospital controls (col.
  3), density and system dummies (col. 4), hospital FE (col. 5), CEO age dummies.
  N = 16,193-20,520 hospital-years depending on pay data availability. Panel B
  replicates for the post-2008 consistent-reporting subsample (2009-2018).

- **Pay-for-performance (Table VIII).** OLS (equation 4), hospital and year FE,
  full controls including tenure and multiple positions; N = 16,193-18,066. Profit
  margin quintiles defined within year x hospital size bin. Robustness: Table
  IA.V adds interactions with size and HSA density to test whether urban/size
  confounds drive the result; coefficients remain significant.

- **Turnover-performance (Table IX).** Probit (equation 3), hospital and year FE,
  full controls; forced turnover sample (age 60 or younger) N = 3,176-18,348; all
  turnover N = 4,030-24,314. Marginal effects reported for Female CEO and the key
  interaction Female CEO x Perf\_Quintile1. Robustness: Table IA.VI includes
  hospital size and density interactions; Table IA.VII documents subsequent career
  outcomes of departing CEOs (11% of male departures become CEOs elsewhere vs.
  6% for female departures, differences often significant).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| IRS Form 990 filings (via Candid/Guidestar 1999-2014; IRS website 2015-2018) | Primary source for hospital financials, CEO/officer names, titles, salaries, board composition; 2000-2018 | [IRS Form 990](/wiki/datasets/irs-form-990/) |
| AHA Annual Survey Database (Dartmouth Institute, 2000-2018) | Hospital services, operations, system affiliation | [AHA Annual Survey](/wiki/commercial/aha-annual-survey/) (licensed) |
| Healthcare Cost Report Information System (HCRIS / CMS, 2011-2018) | Charity Care and Uninsured Discounts spending | [HCRIS](/wiki/datasets/hcris/) |
| Dartmouth Atlas of Health Care | Hospital Service Area (HSA) geographic market definitions; demographic characteristics from 2010 U.S. Census | [Dartmouth Atlas](/wiki/datasets/dartmouth-atlas/) |
| CMS patient outcome metrics (mortality, readmissions, patient satisfaction) | Nonfinancial hospital performance proxies used in supplementary tests; available from 2008-2009 | [CMS quality](/wiki/datasets/cms-quality/) |
| Radaris, LinkedIn, hospital websites (manual collection) | CEO biographical data (age, education, career history) for a subsample of 2,202 CEOs | no page yet |

Sample: 1,981 hospitals and 25,762 hospital-years from 2000 to 2018; 4,353 CEOs
(819 female). Financial ratios are winsorized at the 2% level. CEO compensation
data cover approximately 80% of hospital-year observations. The matched crisis
sample covers 134 female and 271 male CEO hospitals, observed 2006-2011.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13455) if you are: researching
gender gaps in executive compensation or turnover (Tables VII, VIII, IX have
precise specifications and robustness); building on the nonprofit hospital
setting (Internet Appendix Section II has a detailed comparison with for-profit
CEO markets); extending the analysis to other nonprofit sectors; studying
how boards update beliefs about CEO quality (the discussion in Section VI.D
synthesizes the compensation and turnover patterns against learning and agency
theories); or investigating whether board gender composition affects executive
gender outcomes (Table X and the negative result on female directors, Section VII).

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(4), August 2025, pp. 2199-2253.
Publisher: Wiley on behalf of the American Finance Association. DOI:
10.1111/jofi.13455. Licensed under Wiley Terms and Conditions (paywalled; not
CC licensed). This distillation was extracted by an LLM on 2026-06-05 and is
**not human-verified or independently reproduced**. Extract-only: no PDF hosted.

> Lewellen, Katharina. "Women in Charge: Evidence from Hospitals."
> *The Journal of Finance* 80, no. 4 (August 2025): 2199-2253.
> DOI: 10.1111/jofi.13455. © 2025 the American Finance Association.


==============================================================================
# Raising Capital from Investor Syndicates: Luo (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/luo-raising-capital-investor-syndicates-2025/
# Distilled: An entrepreneur raising capital from a syndicate can use contract design to shape whether investors communicate truthfully or strategically persuade each other, explaining why flat contracts suit low-quality projects while hierarchical (differential-return) contracts suit high-quality ones. J. Finance 2025, CC BY-NC-ND 4.0. Six core results with source locators, the game-theoretic model, and the formal equilibrium characterizations.
# Tags: paper-summary, corporate-finance, syndication, venture-capital, information-asymmetry
==============================================================================

**What this is.** The paper's core propositions, the game-theoretic model, and
the formal equilibrium characterizations with exact equation locators: enough
to understand what it found and how, without reading all 55 pages. To replicate
or extend it, read the original at [https://doi.org/10.1111/jofi.13453](https://doi.org/10.1111/jofi.13453).

## TL;DR

An entrepreneur raising capital from multiple investors uses contract design to
govern how investors communicate. Each investor privately observes a signal
about the project and then sends cheap-talk messages to the others before
deciding to invest. The key insight is that the shape of the contract, flat
(identical returns) or hierarchical (differential returns), determines whether
investors have aligned interests and communicate truthfully, or divergent
interests and strategically persuade each other.

For projects with low ex ante quality, the entrepreneur prefers a flat contract:
investors truthfully share information, screening out bad projects, but never
invest when the project type is uncertain. For high-quality projects, the
entrepreneur prefers a hierarchical contract: the lead investor (promised more)
persuades others to invest even under uncertainty, raising the acceptance rate
at the cost of information rents. This provides a new motivation for investor
syndicates distinct from risk-sharing (the "second opinion" motivation of Brander, Amit, and Antweiler (2002)), or capital constraints:
allowing persuasion between investors. The paper also derives persuasion
cascades in the N-investor case and testable implications linking return
differentials to project quality and information softness.

## Core results

Locators point into the source PDF. All results are theoretical propositions.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Optimal contract is flat when ex ante quality is low, hierarchical when high**: threshold on $$P_0(\alpha-1)/(1-P_0)$$ determines the switch | Proposition 3, p. 1832 | Flat (no-enthusiast) when $$\frac{P_0(\alpha-1)}{1-P_0} \leq \frac{1}{2}\frac{2-m}{(1-m)^2}$$; hierarchical (all-enthusiast) otherwise |
| R2 | **Entrepreneur's expected profits**: hierarchical contract is preferred when surplus from uncertain projects is high enough to cover persuasion cost | Proposition 3, p. 1832 | $$U^{all} = P_0(\alpha-1) - (1-P_0)(1-m/2)$$; $$U^{no} = P_0[1-(1-m)^2](\alpha-1)$$ |
| R3 | **Return differential in hierarchical contract equals $$m(1-P_0)/P_0$$**: decreasing in prior probability of a good project and increasing in lead investor's signal precision | Proposition 3 and §VI.C, p. 1847 | Return difference $$r_1 - r_2 = m(1-P_0)/P_0$$; investor 1 gets $$r_1 = 1/P_0$$, investor 2 gets $$r_2 = 1 + (1-P_0)(1-m)/P_0$$ |
| R4 | **Communication costs raise the required return differential** and eventually make flat contracts always optimal | Proposition 11, p. 1847 | If communication cost $$\eta \geq \frac{1}{2}(1-P_0)m(1-m)$$, flat always optimal; else hierarchical return difference $$= (1-P_0)m/P_0 + 2\eta/(P_0(1-m))$$ |
| R5 | **Empirical prediction**: flat syndicates suit high-screening, low-acceptance settings (VC); hierarchical suits high-acceptance settings (syndicated loans) | Proposition 9 + §VI.A, p. 1845 | Acceptance probability ~80% for large bank loans (Berg (2018)); ~5% for VC deals (Gompers et al. (2020)); model predicts hierarchical for loans, flat for VC |
| R6 | **With N investors, persuasion cascades emerge**: returns are decreasing down the hierarchy; more investors makes no-enthusiast flat contracts optimal for more projects | Proposition 8, p. 1844 | Hierarchical: $$r_i = 1 + \frac{1-P_0}{P_0}(1-m)^{i-1}$$; $$U^{all} = P_0\left[\alpha - 1 - \frac{1-P_0}{P_0}\cdot\frac{1-(1-m)^N}{Nm}\right]$$ |

**Overall (paper's conclusion).** The model provides a unified theory of
investor syndicate formation and internal structure grounded in strategic
communication. Hierarchical contracts are optimal not because they bring in
more capital or information, but because they create conflicts of interest that
induce persuasion and raise acceptance rates on high-quality uncertain projects.
The communication channel explains why VC syndicates are flat (screening
orientation, low acceptance) and loan syndicates are hierarchical (persuasion
orientation, high acceptance), and predicts that return differentials decrease
with project quality and increase with information softness.

## Theory / model

The baseline model (Section II, p. 1821) has a penniless entrepreneur ("she")
who owns a project requiring total investment normalized to one. There are two
investors ("he"), each contributing one half. The project is either good
(generating return $$\alpha > 1$$) or bad (generating zero). The prior probability
of a good project is $$P_0$$.

Each investor $$i$$ privately observes the project type with probability $$m$$ and
nothing with probability $$1-m$$. Observation $$O_i \in \{G, B, \varnothing\}$$.
The entrepreneur publicly proposes a contract $$\{r_1, r_2\}$$ where $$r_i$$ is the
promised return to investor $$i$$ per unit of capital if the project is good.
After observing their signals, investors simultaneously send cheap-talk messages
and simultaneously make acceptance decisions.

**Investor $$i$$'s expected return from accepting** in state $$(O_i, O_j)$$ (p. 1826):

$$
\Pi_i(O_i, O_j) \triangleq \Pr[\text{project is good} \mid O_i, O_j] \cdot r_i - 1
$$

**Investor $$i$$'s expected utility from accepting** given strategies $$(\gamma_j, a_j)$$ (p. 1826):

$$
U_i(O_i, \theta_j; a_j, \gamma_i, \gamma_j) \triangleq \sum_{O_j \in \{G,\varnothing,B\}} \Pr[O_j \mid O_i, \theta_j; \gamma_j] \times \Pi_i(O_i, O_j) \times \mathbf{1}(f_i + f_j \cdot a_j(O_j, \gamma_i(O_i)) \geq 1)
$$

where $$\Pr[O_j \mid O_i, \theta_j; \gamma_j]$$ is investor $$i$$'s Bayes posterior about investor $$j$$'s observation after seeing $$j$$'s message $$\theta_j$$.

**Definition 1** (p. 1826): A contract is *flat* if all investors receive identical
returns ($$r_1 = r_2$$) and *hierarchical* if they receive different returns
($$r_1 > r_2$$).

**Equilibrium refinements** (p. 1827): The paper uses pure-strategy perfect
Bayesian equilibria (PBE) with three refinements: (SIB) an investor believes
the project is surely good (bad) upon observing $$G$$ ($$B$$), irrespective of
the other's message; (Weak Dominance) investors do not use weakly dominated
acceptance strategies; and (Pareto Dominance) investors do not play
Pareto-dominated equilibria.

**Key mechanism.** Under a hierarchical contract with $$r_1 > r_2$$, investor 1
prefers to invest even when the project type is uncertain $$(\varnothing, \varnothing)$$, but
investor 2 does not. Investor 1 therefore has an incentive to persuade investor
2 by pooling his messages for $$G$$ and $$\varnothing$$. Investor 2, receiving this pooled
message, perceives the project as more likely good and invests. The proof that
observation $$B$$ is always credibly revealed (Lemma 1, p. 1828) ensures that bad
projects are screened out in all equilibria.

## Method

This is a pure theory paper. The method is construction of a game-theoretic
model and characterization of its equilibria by backward induction and the
equilibrium refinements above. The solution builds on the `bayesian-persuasion`
framework (Kamenica and Gentzkow (2011)) and the `principal-agent` literature
on multiagent contracting (Segal (1999), Halac, Kremer, and Winter (2020)).
The persuasion-cascade mechanism differs from informational cascades driven by
planners' information design (Caillaud and Tirole (2007)), in that agents here
communicate strategically via cheap talk rather than Bayes-rational observational
learning.

**Step 1.** For any proposed contract, characterize the unique investment outcome
$$\hat{I}$$ (the set of observation states in which the project is implemented) via
Proposition 1 (p. 1829). The set $$\hat{I}$$ is:

$$
\hat{I} \triangleq \{(O_1, O_2) \mid [a_1^*(O_1, \gamma_2^*(O_2)), a_2^*(O_2, \gamma_1^*(O_1))]\cdot[f_1,f_2]^T \geq 1\}
$$

Under the refinements, Proposition 1 shows: if $$r_1 < 1/P_0$$, $$\hat{I} = \{(O_1,O_2) \mid O_1=G \text{ or } O_2=G\}$$ (screening);
if $$r_1 \geq 1/P_0$$ and $$r_2 < 1 + \frac{1-P_0}{P_0}(1-m_1)$$, $$\hat{I} = \{(O_1,O_2) \mid O_2=G\}$$;
if $$r_1 \geq 1/P_0$$ and $$r_2 \geq 1 + \frac{1-P_0}{P_0}(1-m_1)$$, $$\hat{I} = \{(O_1,O_2) \mid O_1 \neq B \text{ and } O_2 \neq B\}$$ (persuasion).

**Step 2.** Characterize which investors are *enthusiastic* (Definition 3, p. 1830):
investor $$i$$ is enthusiastic if in all equilibria he accepts whenever no investor
observes $$B$$. Proposition 2 (p. 1830) pins down $$\hat{N}$$, the number of
enthusiastic investors, as a function of promised returns.

**Step 3.** Show (Lemma 3, p. 1831) that the optimal contract must be either
all-enthusiast or no-enthusiast; partial-enthusiast contracts are dominated.

**Step 4.** Characterize the optimal contract (Proposition 3) by comparing the
entrepreneur's expected profits from each type. Define $$U^{all}$$ and $$U^{no}$$:

$$
U^{all} \triangleq P_0(\alpha-1) - (1-P_0)\!\left(1-\frac{m}{2}\right), \qquad U^{no} \triangleq P_0\!\left[1-(1-m)^2\right](\alpha-1)
$$

The entrepreneur prefers the hierarchical all-enthusiast contract
$$\left(\frac{1}{P_0},\ 1+\frac{1-P_0}{P_0}(1-m)\right)$$ when $$U^{all} \geq U^{no}$$, i.e., when

$$
\frac{P_0(\alpha-1)}{1-P_0} \geq \frac{1}{2}\cdot\frac{2-m}{(1-m)^2},
$$

and the flat no-enthusiast contract $$(1,1)$$ otherwise (Proposition 3, p. 1832).

## Empirical specifications

This is a theory paper with no estimation. The paper derives testable
implications in Section VI (pp. 1844-1848), connecting the model to empirical
regularities in syndicate structures.

**Application to syndicate structures (Proposition 9, p. 1845; R5 above).** The
model predicts that as project acceptance probability increases (higher $$P_0$$
or $$\alpha$$), optimal syndicate structure switches from flat to hierarchical.
According to Berg (2018), acceptance probability for large bank loans is near
80%; Gompers et al. (2020) find that only 5 of 101 VC deals considered advance
to due diligence, suggesting acceptance near 5%. The model predicts hierarchical for
bank loan syndicates and flat for VC syndicates, consistent with empirical
observation: most VC syndicates invest at the same valuation (flat), while loan
syndicates have lead arrangers earning higher fees (hierarchical).

**Underwriting and strategic communication (§VI.B, p. 1846).** The model
interprets underwriting fees as the mechanism implementing hierarchical contracts.
The prediction that hierarchical structures generate communication problems
between lead investors and others is consistent with legal cases: the IFE Fund
v. Goldman Sachs International (Petkovic (2008)) case documents Goldman's
suppression of unfavorable reports to maintain syndicate participation, consistent
with the equilibrium pooling behavior of the enthusiastic investor in the model.

**Testable implications for return differences (Propositions 10-12, pp. 1847-1848).**

Proposition 10: Return difference $$r_1 - r_2 = m(1-P_0)/P_0$$ is decreasing in
the prior probability $$P_0$$ of a good project and increasing in the lead
investor's signal probability $$m$$. This implies that high-risk projects with low
acceptance probability should have larger fee differentials between lead and
participant lenders.

Proposition 11: With communication cost $$\eta$$, return difference in hierarchical
contracts equals $$(1-P_0)m/P_0 + 2\eta/(P_0(1-m))$$, which is increasing in $$\eta$$.

Proposition 12: With information softness parameter $$\lambda$$ (fraction of $$G$$
observations that cannot be credibly revealed), return difference equals

$$
\frac{1-P_0}{P_0}\cdot\frac{m\lambda}{m\lambda + 1 - m},
$$

which is increasing in $$\lambda$$: softer information makes persuasion more
effective, requiring a smaller return premium to make the lead enthusiastic,
but investor 2 demands less in equilibrium because investor 1's persuasion is
more informative.

## Datasets used

This is a theory paper with no primary dataset. Empirical regularities cited
for motivation and discussion:

| Source | Role in paper | Wiki page |
|---|---|---|
| DealScan (Pitchbook) | Syndicated loan and VC deal statistics: "78% of loans in the DealScan universe were syndicated, 65% of VC deals" (§VI.A, p. 1845) | [DealScan](/wiki/commercial/dealscan/) / [PitchBook](/wiki/commercial/pitchbook/) (licensed) |
| Berg (2018) (large German bank) | Loan acceptance probability ~80% for loans above EUR 1 million | no page yet |
| Gompers et al. (2020) | VC deal acceptance statistics: 101 deals considered, 5 advanced to due diligence (implying ~5% acceptance) | no page yet |

No estimation, no regression, no data sample. All results are derived from the
theoretical model.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13453) if you are:
building a model of multiagent contracting with cheap talk;
studying the economic rationale for hierarchy in syndicates or inside firms;
extending the persuasion-cascade framework to more-than-two agents; or
testing the model's predictions on return differentials across syndicate types.
The proofs for all propositions are in Appendix A (pp. 1849-1869).

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(3), June 2025, pp. 1815-1869.
This distillation was extracted by an LLM on 2026-06-06 and is **not
human-verified or independently reproduced**. The CC BY-NC-ND 4.0 licence
permits sharing with attribution but prohibits derivatives and commercial reuse;
the verbatim PDF is not hosted here.

> Luo, Dan. "Raising Capital from Investor Syndicates with Strategic Communication."
> *The Journal of Finance* 80, no. 3 (June 2025): 1815-1869.
> DOI: 10.1111/jofi.13453. © 2025 The Author(s).
> Published by Wiley on behalf of the American Finance Association.
> Licensed under [CC BY-NC-ND 4.0](https://creativecommons.org/licenses/by-nc-nd/4.0/).
> This page is an **extract and summary** by the Institute for Automated Research;
> no modifications to the results or equations are represented as the author's words.


==============================================================================
# In the Red: Di Maggio, Ma & Williams (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/maggio-red-overdrafts-payday-lending-2025/
# Distilled: Banning high-to-low transaction reordering (HTLR) by banks reduces low-income consumers' payday borrowing by $85 per quarter (11%), improves credit scores, and raises consumption of essential goods, while also triggering bank branch closures in low-income areas. J. Finance 2025, paywalled. Seven core results with source locators, datasets used, the identification strategy (class-action lawsuits as natural experiment), and the estimating specifications.
# Tags: paper-summary, household-finance, consumer-finance, banking, overdraft
==============================================================================

**What this is.** The paper's core results, the identification strategy (class-action lawsuits banning high-to-low transaction reordering as a natural experiment), and the estimating specifications with exact table locators: enough to know what was found and how, without reading all 48 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1111/jofi.13447).

## TL;DR

Banks' practice of reordering transactions from high to low (HTLR) maximizes overdraft fees earned from low-income depositors. The paper exploits 37 class-action lawsuits that forced some banks to cease HTLR, comparing treated zip codes (with banned banks) to nearby control zip codes (also with sued banks but not banned) in a difference-in-differences design. After HTLR bans, low-income consumers reduce payday borrowing by $85 per borrower per quarter (11%), reduce installment borrowing by $358 per borrower per quarter (8%), improve credit scores, gain access to cheaper mainstream credit, and increase essential consumption. But HTLR bans also cause banks to close branches, especially in low-income areas, highlighting a tension between fee limits and financial access.

## Core results

Magnitudes as reported; `\*\*`, `\*\*\*` = 5%, 1% significance. Standard errors clustered at the neighborhood level (consumer specs) or bank and zip-code level (branch specs). All consumer specifications use zip code x quarter observations in below-median income zip codes.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | HTLR bans reduce **payday loan borrowing** per borrower | Table V col. (3), p. 1711 | DiD coeff. = -$84.84\*\*\* (se 31.47); 11% reduction relative to mean |
| R2 | HTLR bans reduce **installment loan borrowing** per low-income borrower | Table VI col. (3), p. 1714 | DiD coeff. = -$358.3\*\*\* (se 135.8); 8% reduction relative to mean |
| R3 | Alternative-credit declines **persist for at least 3 years** | Table VIII, p. 1717 | Payday: -$84.84 (yr 1), -$72.02 (yr 2), -$50.47\*\* (yr 3); Installment: -$358.3, -$293.2\*\*, -$276.5\*\* |
| R4 | Consumers are **more likely to experience a 50+ point credit score increase** | Table IX Panel B col. (1), p. 1719 | 0.0288 pp\*\*\* (se 0.0111) at 1 year; 0.0315 pp\*\*\* at 3 years |
| R5 | **Credit card balances and limits rise**, indicating improved access to mainstream credit | Table IX Panel A col. (1-2), p. 1719 | Credit card balance: +$33.96\*\* (yr 1); credit card limit: +$40.10\*\* (yr 1) per low-income borrower |
| R6 | Consumers **increase essential consumption** (durable and nondurable essential goods) | Table X col. (1)-(4), p. 1721 | Durable spending: +$45.18\*\* per household; nondurable essential: +$15.57\*\*; nonessential nondurable: insignificant |
| R7 | HTLR-banned banks are **significantly more likely to close branches**, concentrated in low-income areas | Table XI col. (1) and (5), p. 1724 | Branch exit probability: +0.90 pp\*\*\* overall; interaction with Low-Income dummy: +0.33 pp\*\* additional |

**Overall (paper's conclusion).** Aggressive overdraft pricing via HTLR creates demand for alternative credit, trapping low-income consumers in a cycle of high-cost borrowing. Ceasing HTLR reduces payday and installment borrowing, improves consumer credit health, and unlocks access to cheaper mainstream credit. However, HTLR bans also prompt banks to exit low-income neighborhoods, an unintended spillover that may offset some consumer gains and raise new concerns about financial access.

## Theory / model

The paper has no formal model. It develops two competing hypotheses for the relationship between overdraft pricing and demand for alternative credit, and tests them empirically.

**Hypothesis 1 (complements / debt-trap):** Consumers improperly estimate overdraft costs (as in Bertrand and Morse (2011)) and turn to payday lenders to repay unaffordable overdraft balances and fees. Morgan, Strain, and Seblani (2012) and Melzer and Morgan (2015) document that overdraft providers and payday lenders compete and that consumers use both as substitutes; the paper's findings challenge that view, showing they are better described as complements. A reduction in overdraft costs reduces payday demand.

**Hypothesis 2 (substitutes / price-reduction supply shift):** Overdraft credit and payday loans are substitutes for short-term liquidity. If HTLR bans reduce the supply of overdraft services (banks become less willing to provide the service at lower fee revenue), excess demand for short-term credit shifts to payday lenders. In this case, payday borrowing rises after HTLR bans.

Related literature: Melzer (2011) and Skiba and Tobacman (2019) provide evidence on the real costs of payday loan access. Dlugosz, Melzer, and Morgan (2021) study overdraft fee ceilings and the unbanked. Gabaix and Laibson (2006) formalize how shrouded attributes enable add-on pricing to persist in competitive markets, motivating why HTLR pricing is not competed away.

**Identification logic.** The key source of variation is the lawsuit outcome: within the same geographic neighborhood, some zip codes contain branches of banks required to cease HTLR, and neighboring zip codes contain branches of banks that were sued but not required to stop. The two groups operate in similar local economic conditions and share similar consumer demand dynamics. Within-neighborhood-quarter fixed effects additionally absorb any time-varying shocks (such as local unemployment changes) correlated with both HTLR bans and credit demand. The paper argues that lawsuit outcomes are quasi-exogenous because they depend on contract-level arbitration clauses and litigation strategy, not on bank or consumer characteristics, and tests this with a battery of balance checks (Table III Panel B, Internet Appendix Tables IV-V, pp. 1706, 1728-1729).

## Method

The primary estimator is a zip-code-by-quarter-level difference-in-differences regression. The treatment group is zip codes containing branches of banks required to cease HTLR; the control group is zip codes within seven miles of a treated zip code that contain branches of sued banks not required to cease HTLR. This design restricts variation to local areas subject to similar lawsuit contexts.

The paper builds on the `difference-in-differences` and `panel-regression` primitives. The key structural feature is that the control group is not the general population of zip codes but rather nearby zip codes sharing the same lawsuit environment, selected to match on local consumer demand dynamics.

For colocation motivating evidence (Table IV, p. 1707), a conditional logit regression is estimated at the bank branch by year level:

$$
\Pr(\text{AltFin within } d)_{bz} = F\left(\gamma \cdot \text{HTLR}_{b} + \eta_z\right) \tag{4-IV}
$$

where $$\text{HTLR}_b$$ is an indicator for whether branch $$b$$ belongs to an HTLR bank, $$\eta_z$$ are zip code fixed effects, and $$F(\cdot)$$ is the logistic CDF. This within-zip-code test avoids comparing geographic areas with different demographic compositions.

For bank-level first-stage evidence (Table A.I, p. 1735), the estimating equation is a bank by quarter DiD:

$$
\Delta \text{BankOutcome}_{bt} = \beta \cdot \text{HTLRBan}_b \cdot \text{Post}_t + \eta_b + \eta_t + \varepsilon_{bt} \tag{FS}
$$

where bank and quarter fixed effects are included and standard errors are clustered at bank and quarter levels. This confirms HTLR bans reduced overdraft-related revenue by 6.44% (Table A.I col. 1, p. 1735).

## Empirical specifications

All consumer-outcome specifications are estimated at the zip-code by quarter level, restricted to zip codes with below-median income.

**Main DiD for payday and installment borrowing (R1, R2; equations 1 and 2 from the paper, p. 1712 and p. 1718):**

$$
\text{PaydayBorrowing}_{zt} = \beta \cdot \text{HTLRBan}_z \cdot \text{Post}_t + \eta_{nt} + \varepsilon_{zt} \tag{1}
$$

$$
\text{FinancialHealth}_{zt} = \beta \cdot \text{HTLRBan}_z \cdot \text{Post}_t + \eta_{nt} + \varepsilon_{zt} \tag{2}
$$

where $$\text{PaydayBorrowing}_{zt}$$ is average dollars of payday loans disbursed per payday borrower in zip code $$z$$ in quarter $$t$$; $$\text{HTLRBan}_z$$ is a dummy for whether the zip code contains branches of a bank required to cease HTLR; $$\text{Post}_t$$ is a dummy for quarters after the HTLR ban; and $$\eta_{nt}$$ are neighborhood-by-quarter fixed effects (the most conservative specification, where neighborhood = all zip codes within seven miles). Standard errors are clustered at the neighborhood level. The window is two quarters before and four quarters after the HTLR ban.

The coefficient of interest $$\beta$$ measures the differential effect of the ban on consumer alternative borrowing in treated relative to control zip codes within the same neighborhood. Tables V (Clarity payday) and VI (Equifax installment) present six columns each, varying between neighborhood FE, quarter FE, and neighborhood-by-quarter FE.

**Long-term horizon test (R3; Table VIII, p. 1717):** The same DiD specification but extending the post window to one, two, or three years. The results show that the decline in both payday and installment borrowing persists without reverting to pre-ban levels, ruling out a temporary substitution explanation.

**Credit health and consumption (R4, R5, R6; Tables IX and X, pp. 1719, 1721):** The same zip-code-by-quarter DiD structure (equation 2) applied to credit card balances, credit card limits, probability of a 50+ point credit-score increase, total balance in good standing, and household expenditure (durable, essential nondurable, nonessential nondurable) from Earnest transaction data. Credit card variables come from Equifax; consumption variables are from Earnest Research's transaction-level data for 6 million U.S. households. Consumption tests use an 8-quarter post window and 4-quarter pre window.

**Branch exit spillovers (R7; Table XI, equation 4 from the paper, p. 1723):**

$$
\text{Exit}_{izt} = \beta \cdot \text{HTLRBan}_i \cdot \text{Post}_t + \eta_{zt} + \varepsilon_{izt} \tag{4}
$$

where $$\text{Exit}_{izt}$$ is a dummy for bank $$i$$ exiting zip code $$z$$ in year $$t$$; the specification includes zip-code-by-year and bank-by-zip-code fixed effects; $$\text{Post}_t$$ covers up to three years after the HTLR ban. The interaction $$\text{HTLRBan}_i \cdot \text{Post}_t \cdot \text{LowIncome}_{zt}$$ shows that exits are concentrated in below-median income zip codes. Standard errors are clustered at the bank and zip-code level.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Clarity Services (alternative credit bureau) | Primary outcome: payday and alternative installment loans disbursed 2013-2019; random sample of 171,445 alternative borrowers | No page yet |
| Equifax (traditional credit bureau) | Installment loan borrowing, credit card balances and limits, credit scores, total balances in good standing; representative 10% sample of 680,856 borrowers, 2005-2018 | No page yet |
| Pew Charitable Trusts bank study (2012-2015) | HTLR practice indicator for the largest 50 US banks; four annual waves | No page yet |
| FDIC Summary of Deposits | Bank branch locations and deposit-level data for 50 largest banks | [FDIC Summary of Deposits](/wiki/datasets/fdic-summary-of-deposits/) |
| FR Y-9C call report data (FFIEC 031/041) | Bank-level overdraft revenue, all-other-loans balances, and income statement items | No page yet |
| Earnest Research expenditure data | Household consumption: credit/debit card transaction-level data for 6 million US households; durable, essential nondurable, and nonessential nondurable spending | No page yet |
| Infogroup Historical Business Database | Payday lender and check-casher locations (SIC codes 609903 and 614113), 1997-2018 | No page yet |
| Hand-collected lawsuit data set | 37 class-action lawsuits against HTLR banks: event dates, settlement terms, behavior relief (HTLR ban or not), cash settlement amounts | No page yet |
| American Community Survey (Census) | Zip-code-level demographics: age, race, income, poverty, housing, employment, 2011-2018 | No page yet |
| HMDA and SBA lending data | Bank branch exit spillovers to mortgage and small business lending (Table XII) | [HMDA](/wiki/datasets/hmda/) (no page yet) |

Sample: Clarity data 2012-2019; Equifax data 2005-2018; bank lawsuit events concentrated 2008-2015. Analysis at the zip-code by quarter level; 6,975 zip-code-quarter observations for Clarity (main spec), 30,487 for Equifax.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13447) if you are:
studying the link between bank overdraft practices and the alternative lending market; estimating causal effects of fee policies on consumer financial health; interested in spillover effects of banking regulation (branch closures, financial deserts); or evaluating the "banked vs. underbanked" distinction as a policy target. The Appendix (Tables A.I-A.III, pp. 1735-1736) contains the first-stage bank-level DiD results. The Internet Appendix contains the full lawsuit data set (Table I), balance tests (Tables IV-V), and additional robustness tables (Tables VI-XIV).

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(3), June 2025. This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The paper is paywalled; extract-only redistribution applies.

> Di Maggio, Marco, Angela Ma, and Emily Williams. "In the Red: Overdrafts, Payday Lending, and the Underbanked." *The Journal of Finance* 80, no. 3 (June 2025): 1691-1738. DOI: 10.1111/jofi.13447. © 2025 the American Finance Association.


==============================================================================
# Creating Controversy in Proxy Voting Advice: Malenko, Malenko & Spatt (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/malenko-creating-controversy-proxy-voting-2025/
# Distilled: A profit-maximizing proxy advisor optimally produces fully informative research reports but partially informative, asymmetrically biased vote recommendations that favor the a priori unlikely alternative, increasing the incidence of close, contentious votes to enhance the value of its advice. J. Finance 2025, CC BY-NC-ND 4.0. Seven core results with source locators, the information-design model, and the Bayesian persuasion method with its defining equations.
# Tags: paper-summary, corporate-governance, proxy-voting, information-design
==============================================================================

**What this is.** The paper's core results, the information-design model, and the Bayesian persuasion method it builds on, with the defining equations: enough to know what it found and how, without reading all 52 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13438).

## TL;DR

A profit-maximizing proxy advisory firm sells research reports to institutional investors and issues public vote recommendations. The paper shows that the advisor's optimal strategy combines two distinct elements: a fully informative (unbiased) private research report for subscribers, and a partially informative, asymmetrically biased public vote recommendation that favors the a priori unlikely alternative too often. By recommending against the likely outcome more often than its probability warrants, the advisor induces close, contentious shareholder votes, raising the value of its advice and thereby the willingness to pay for the research report. The paper calls this "creating controversy." The model is cast as an information design problem following Rayo and Segal (2010) and Kamenica and Gentzkow (2011). The result rationalizes the proxy advisory industry's one-size-fits-all approach, explains the rubber-stamping pattern documented in Malenko and Shen (2016) (positive ISS recommendations receive 93% average support on say-on-pay), explains why negative recommendations generate dispersed votes, and suggests that the active voting behavior of large institutional investors studied by Iliev and Lowry (2015) is consistent with rational adjustment for recommendation bias.

## Core results

Magnitudes and significance are as reported. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Optimal public recommendation is binary and partially informative**: the advisor recommends against the a priori likely alternative too often, inducing a controversial posterior near 0.5 | Proposition 4, pp. 2325-2326 | When mu >= mu_0(q), recommendation s = 0 is given with probability (1-mu)/(1-mu_0(q)) > Pr(theta=0) = 1-mu; more often than warranted by the true probability the proposal is value-decreasing |
| R2 | **Maximizing the average probability of a split vote is the objective**: controversial recommendations raise the split-vote probability, increasing subscribers' willingness to pay for the report | Eq. (12)-(14), pp. 2321-2322; Figure 3, p. 2323 | Concavification of Pr(Piv\|q, mu_s) over mu_s, subject to the Bayes plausibility constraint, yields the advisor's optimal policy; for q = 0.01 and mu = 0.8, partially informative recommendation raises split-vote probability above that of any other policy |
| R3 | **Near mu = 0.5, optimal recommendation is uninformative**: close votes are already likely without intervention, so information design has no benefit | Proposition 5, p. 2327; Section III.E, p. 2334 | ISS board-declassification positive recommendations: 100% of cases 2010-2019 (zero negative), yet shareholder support averages 73.7%, consistent with uninformative design prediction |
| R4 | **Subscriber demand rises with controversy**: the fee the advisor can charge equals V(q, S) H^{-1}(1-q); a higher average split-vote probability V raises willingness to pay | Proposition 3, eq. (11), p. 2321; eq. (9)-(10), p. 2320 | Shareholder i subscribes iff v_i >= f/V(q, S); the controversy mechanism converts a higher split-vote probability directly into higher revenue per subscriber |
| R5 | **Fully informative private research report is optimal**: noise in the report weakly lowers subscribers' willingness to pay | Proposition 8, p. 2330 | Any signal that induces a posterior of exactly 1/2 for the subscriber is worthless; the advisor can always combine signals to eliminate such posteriors, and the remaining signals achieve full informativeness |
| R6 | **Rubber-stamping of pro-prior recommendations, dispersion under anti-prior recommendations**: ISS say-on-pay positive recommendations receive 93% average support with no failures; negative recommendations receive 69% average support with 11% failure probability | Section III.G, p. 2335; Table IA.I (Internet Appendix) | About 500 close votes per year (40%-60% support) occur when ISS recommends against management; close votes are extremely rare when ISS agrees with management |
| R7 | **Ban on recommendations has ambiguous welfare effects**: it removes the bias but reduces information for nonsubscribers; whether the net effect is positive depends on the distribution of shareholder valuations v_i | Proposition A1, p. 2343; Example A1, p. 2344 | For v_H = 5: ban raises correct-decision probability from 92.5% to 95.65% (positive effect dominates); for v_H = 7: ban lowers it to 90.3% < 92.5% (negative effect dominates) |

**Overall (paper's conclusion).** Proxy advisors have a structural incentive to design biased vote recommendations that create controversy, because close votes increase the relevance and value of their research reports. This incentive is inherent to the core business model of selling information and is independent of any consulting conflict of interest. At the same time, research reports remain fully informative and valuable to subscribers. Proxy advisors' recommendations should therefore not be treated as the appropriate benchmark for evaluating institutional investors' voting behavior; the votes of large, engaged asset managers are a more suitable reference.

## Theory / model

### Setup (Section I, pp. 2311-2316)

A firm has $$N \geq 3$$ (odd) shareholders, each owning one share. A proposal is approved if at least $$\frac{N+1}{2}$$ shareholders vote in favor. Let $$d \in \{0,1\}$$ denote the decision (1 = approve). The proposal's payoff to shareholder $$i$$ is (p. 2312, eq. 1-2):

$$
u_i(d, \theta) = v_i \cdot u(d, \theta), \qquad u(1, \theta) = \begin{cases} 1 & \text{if } \theta = 1 \\ -1 & \text{if } \theta = 0 \end{cases}, \quad u(0, \theta) = 0 \tag{1-2}
$$

where $$\theta \in \{0,1\}$$ is the unknown state and $$v_i \geq 0$$ is shareholder $$i$$'s concern (drawn i.i.d. from distribution $$H(\cdot)$$ on $$[\underline{v}, \bar{v}]$$). The prior is $$\Pr(\theta=1) = \mu \in (0,1)$$.

### Information structure (Section I.B, pp. 2312-2313)

The proxy advisor designs two signals. The **private signal** (research report) $$\mathcal{R} = (R, \{\phi(\cdot|\theta)\}_{\theta \in \{0,1\}})$$ is available only to subscribers. The **public signal** (vote recommendation) $$\mathcal{S} = (S, \{\gamma(\cdot|r)\}_{r \in R})$$ maps the report realization $$r$$ to a public recommendation $$s \in S$$, observable by all shareholders. The paper shows (Proposition 8) that the optimal private signal is fully informative: $$R = \{0,1\}$$ and $$r = \theta$$, so subscribers learn the state with certainty.

### Voting equilibrium (Section II.A, pp. 2316-2319)

Given recommendation $$s$$ (inducing posterior $$\mu_s = \Pr(\theta=1|s)$$) and fraction $$q$$ of subscribers, a nonsubscribing shareholder's equilibrium probability of voting "for" is (Proposition 2, eq. 7, p. 2318):

$$
\pi(q, \mu_s) = \frac{z_s(1-2q) - 1 + \sqrt{(z_s-1)^2 + 4q^2 z_s}}{2(z_s-1)(1-q)}, \qquad z_s \equiv \left(\frac{\mu_s}{1-\mu_s}\right)^{\frac{2}{N-1}} \tag{7}
$$

for $$\mu_s \neq \frac{1}{2}$$. Subscribers vote according to the state: $$a_i = \theta$$. Nonsubscribers condition their vote not only on $$\mu_s$$ but also on the information revealed by the event of being pivotal ("strategic voting").

The value of the report to shareholder $$i$$, conditional on recommendation $$s$$, is $$\frac{v_i}{2} \Pr(\text{Piv}|q, \mu_s)$$, where (eqs. 8-10, p. 2320):

$$
V(q, \mathcal{S}) = \frac{1}{2} \sum_{s \in S} \Pr(\text{Piv}|q, \mu_s) \tau_s, \qquad \tau_s \equiv \mu \gamma(s|1) + (1-\mu)\gamma(s|0) \tag{10}
$$

is the average (pre-recommendation) probability of a shareholder being pivotal, weighted by the frequency of each recommendation.

## Method

The paper solves the advisor's problem by **Bayesian persuasion** (Kamenica and Gentzkow (2011)) applied to a multi-agent, multi-audience setting. The advisor maximizes expected profit $$N q f$$, where $$q$$ is the equilibrium fraction of subscribers and the optimal fee (eq. 11, p. 2321) is:

$$
f = V(q, \mathcal{S}) \cdot H^{-1}(1-q) \tag{11}
$$

Substituting into the profit expression, the advisor solves (eq. 12, 15, pp. 2321-2328):

$$
\max_{q, \mathcal{S}} \; q H^{-1}(1-q) \left( \sum_{s \in S} \Pr(\text{Piv}|q, \mu_s) \tau_s \right) \tag{12}
$$

subject to the **Bayes plausibility constraint**:

$$
\sum_{s \in S} \mu_s \tau_s = \mu \tag{13}
$$

The key step is the **concavification** of the function $$\Pr(\text{Piv}|q, \cdot)$$ over posterior beliefs. For a given $$q$$, the optimal public recommendation design is found by taking the concave closure $$P(q, \mu_s)$$ of $$\Pr(\text{Piv}|q, \mu_s)$$ (Figure 3, p. 2323). Because $$\Pr(\text{Piv}|q, \mu_s)$$ is (by Lemma 1, p. 2322) strictly convex near $$\mu_s = 0$$ and $$\mu_s = 1$$ (for small $$q$$), and strictly concave near $$\mu_s = 1/2$$, the concave closure is achieved by a **binary recommendation** that places mass at $$\mu_0 \in (0, 1/2)$$ (the controversial posterior) and $$\mu_1 = 1$$ (full certainty in the likely direction), or symmetrically.

The method builds on the `principal-agent` framework for the advisor's optimization, and the `bayesian-persuasion` technique for solving the optimal information design. The timeline-consistent property (p. 2313) distinguishes this paper from most Bayesian persuasion models: because the advisor maximizes ex ante profits and has no stake in the vote outcome, the optimal policy is dynamically consistent.

## Empirical specifications

This is a pure-theory paper. The paper does not estimate any econometric specification. Section III (pp. 2330-2336) presents anecdotal and survey evidence to corroborate the mechanism, drawing on:

- **Ertimur, Ferri, and Oesch (2013, 2018)**: variability in ISS research-report severity when negative recommendations are issued; shareholders less likely to vote against management when the report conveys less severe concerns.
- **Case studies**: the 2024 Tesla say-on-pay vote (ISS issued a negative recommendation but the report was more positive; Vanguard and BlackRock voted for, consistent with large-v_i shareholders voting on the report rather than the recommendation; p. 2331).
- **ISS board declassification data**: zero negative ISS recommendations on shareholder proposals to declassify boards from 2010 to 2019, yet average support of 73.7%, consistent with Proposition 5 (uninformative recommendation near mu = 0.5; p. 2334).
- **ISS say-on-pay voting outcomes**: 93% average support with zero failures on positive recommendations; 69% average support and 11% failure rate on negative recommendations (Table IA.I; p. 2335), consistent with rubber-stamping (Proposition 4) and close-vote prediction.
- **Hayne and Vance (2019) interview evidence**: proxy advisor employees confirm that maintaining a consistent proportion of negative recommendations is viewed as a way to stay relevant (Section III.F, p. 2334-2335).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| ISS Voting Analytics (say-on-pay and director elections, 2011-2019) | Empirical corroboration of rubber-stamping and close-vote patterns (Section III.G) | No page yet |
| ISS board-declassification recommendations (2010-2019) | Corroboration of uninformative-recommendation prediction (Section III.E) | No page yet |
| Ertimur, Ferri, and Oesch (2013, 2018) hand-collected ISS reports | Content of research reports vs. recommendations (Section III.A) | No page yet |

The paper has no original data collection. Empirical illustrations use published sources and aggregate statistics from the literature.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13438) if you are: building a structural model of information intermediaries in financial markets; studying proxy advisory regulation (the ban-on-recommendations analysis in Appendix A); analyzing the information content of voting recommendations vs. research reports; or extending the model to costly information acquisition, ideological shareholders, or multi-firm settings (Internet Appendix Sections IV.C, IV.F). The locators above point to the exact propositions and figures.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(4). This distillation was extracted by an LLM on 2026-06-05 and is **not human-verified or independently reproduced**. The CC BY-NC-ND 4.0 licence permits sharing with attribution but does not permit adaptations or commercial use; the verbatim PDF is not hosted here.

> Malenko, Andrey, Nadya Malenko, and Chester Spatt. "Creating Controversy in Proxy Voting Advice." *The Journal of Finance* 80, no. 4 (August 2025): 2303-2354. DOI: 10.1111/jofi.13438. © 2025 The Author(s). Licensed under [CC BY-NC-ND 4.0](https://creativecommons.org/licenses/by-nc-nd/4.0/). This page is a distillation by the Institute for Automated Research: core results extracted and re-expressed; extract-only.


==============================================================================
# Does Saving Cause Borrowing: Medina & Pagel (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/medina-saving-cause-borrowing-implications-2025/
# Distilled: A large-scale field experiment with 3.1 million Mexican bank customers shows that saving nudges increase savings and reduce spending but leave credit card borrowing unchanged, evidence more consistent with self- or partner-control explanations for the coholding puzzle than with transactions-convenience models. J. Finance 2025, paywalled. Seven core results with source locators, datasets used, the conceptual models, and the causal-forest method with its estimating equations.
# Tags: paper-summary, household-finance, coholding, nudges, savings, credit-cards
==============================================================================

**What this is.** The paper's core results, the theoretical frameworks it tests (transactions-convenience and self/partner-control models), and the causal-forest method with its estimating equations: enough to know what it found and how, without reading the full 50 pages. To replicate or extend it, read the original at [doi.org/10.1111/jofi.13466](https://doi.org/10.1111/jofi.13466).

## TL;DR

Using a randomized field experiment in which 3.1 million customers of Mexican bank Banorte were sent SMS saving nudges for seven weeks, this paper asks whether encouraging savings causes individuals to also borrow more. The authors first develop two classes of theoretical models (transactions-convenience and self/partner-control) and show they make distinguishable predictions about the joint responses of spending, saving, and borrowing to a patience shock. Using causal forests to handle treatment effect heterogeneity, the most responsive individuals (top quartile) reduce monthly spending by 7.2% (roughly 2,524 MXN) and increase checking account balances by 4.9% (roughly 1,932 MXN). However, credit card interest charges change by a precisely estimated zero: less than 14% of the spending reduction is reflected in lower credit card interest payments, and the confidence interval rules out a borrowing increase of more than 11 MXN for every 1,932 MXN saved. This null borrowing response is more consistent with self- or partner-control models than with transactions-convenience models, and it implies that saving nudges exacerbate coholding of low-interest savings and high-interest debt.

## Core results

Magnitudes and significance are as reported. All treatment effects are proportional (exp(beta)-1), estimated via Poisson regression with strata fixed effects and robust standard errors. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Saving nudges **reduce spending** for all subjects and for credit card holders | Table IV, Panels A and B, p. 2713 | Full sample: -0.8% from base 17,870 MXN (\*); credit card holders: -1.9% from base 31,998 MXN (\*\*\*) |
| R2 | Saving nudges **increase checking account balances** (savings) | Table IV, Panels A and B, p. 2713 | Full sample: +0.6% from base 19,913 MXN (\*\*); credit card holders: +1.1% from base 35,657 MXN (\*\*) |
| R3 | **Borrowing is unchanged**: monthly credit card interest is a precisely estimated zero for the whole sample | Table IV, Panel B, col. (3), p. 2713 | -0.002, SE 0.008, base 215.91 MXN; upper CI rules out increase >1.4% or decrease >1.8% |
| R4 | In the top quartile of predicted spending treatment effects, **spending falls 7.22%** | Table V, Panel A, col. (1), p. 2718 | -7.22%\*\*\* (SE 1.33%) from base 34,969 MXN; decrease of ~2,524 MXN; N=150,177 |
| R5 | In the same subgroup, **savings rise 4.93%** | Table V, Panel A, col. (2), p. 2718 | +4.93%\*\*\* (SE 1.28%) from base 39,174 MXN; increase of ~1,932 MXN |
| R6 | **Credit card interest is a null in the high-response subgroup**: upper CI rules out more than 6.98 MXN increase per cycle; null holds also for the top savings quartile | Table V, Panel A, col. (3); Table VI, Panel A, col. (3), pp. 2718, 2721 | -1.01% (SE 2.21%), base 210 MXN; <14% of spending reduction reflected in lower interest; null for rolled-over debt and payments |
| R7 | **Message 4 (mental accounting/lock-away) produces the largest effects** on spending and saving; borrowing is null for every individual message | Table VIII, pp. 2726, and p. 2727 | Msg 4: -9.0%\*\*\* spending (SE 2.3%), +6.9%\*\*\* savings (SE 2.4%), -0.6% interest (SE 2.8%, insignificant); significantly larger than short-term and long-term pooled at 10% |

**Overall (paper's conclusion).** Saving nudges increase savings and reduce spending among the most responsive individuals, but the additional savings are not used to repay credit card debt; the null effect on borrowing is consistent with self- or partner-control explanations for coholding and inconsistent with predictions from transactions-convenience or credit-limit-chasing models. This contrasts with automatic enrollment interventions studied by Beshears et al. (2022) and Beshears et al. (2024), where pension auto-enrollment raises credit card and mortgage debt; the absence of such an offset here aligns with Gathergood and Olafsson (2024), who document that most coholding is modest and relatively short-lived. The causal-forest approach also illustrates that traditional strata-based heterogeneity analysis is subject to severe overfitting bias, overstating borrowing reductions by more than an order of magnitude.

## Theory / model

The paper develops two classes of two-period conceptual frameworks to map theoretical predictions onto empirical moments and formalize the hypothesis tests. No formal structural estimation is performed; the models are used to derive the sign of comparative statics that the experiment can test. The coholding puzzle was first documented by Gross and Souleles (2002). The liquidity-premia class encompasses the transactions-convenience model of Telyukova (2013) and the credit-limit-chasing models of Druedahl and Jorgensen (2018), Gorbachev and Luengo-Prado (2019), and Fulford (2015). The self-or-partner-control class follows Bertaut, Haliassos, and Reiter (2009) and Vihriala (2019).

**Liquidity-premia model (transactions convenience).** An agent allocates initial endowment $$x_1$$ to period-1 consumption $$c_1$$ and is required to carry a minimum cash amount $$x$$ for transaction purposes. Borrowing $$b_1$$ is available at interest rate $$r$$. The period-1 maximization problem (p. 2695) is:

$$
\max_{c_1} \{ \log(c_1) + \delta \log(x_1 - c_1 - r b_1) \} \tag{Transactions model}
$$

where $$b_1 := f_{b_1}(c_1) = \begin{cases} c_1 - x_1 + x & \text{if } c_1 - x_1 + x > 0 \\ 0 & \text{otherwise} \end{cases}$$ and $$\delta \in (0,1)$$ is the discount factor. The agent coholds when

$$
x_1 - \frac{1}{\delta+1} x_1 + \frac{r}{(\delta+1)(1+r)} x < x \tag{1}
$$

and the optimal consumption and borrowing satisfy $$c_1^* = \frac{1}{\delta+1} x_1 - \frac{r}{(\delta+1)(1+r)} x$$ and $$b_1^* = c_1^* - x_1 + x$$.

Proposition 1 (p. 2695): if agents cohold and become more patient (higher $$\delta$$), consumption decreases and debt decreases by the same amount: $$\frac{\partial b_1^*}{\partial \delta} = \frac{\partial c_1^*}{\partial \delta} < 0$$. Proposition 2 (p. 2696): if cash needs $$x$$ increase, borrowing increases by almost the same amount: $$\frac{\partial b_1^*}{\partial x} = 1 - \frac{r}{(\delta+1)(1+r)} > 0$$. Thus in any liquidity-premia or credit-limit-chasing model, saving nudges interpreted as a patience shock predict lower spending AND lower debt, while cash-need shocks predict higher debt.

**Self- or partner-control model.** A patient party (discount factor $$\delta$$) hides an amount $$x \geq 0$$ from an impatient party (discount factor $$\beta \in (\beta, \delta)$$). The impatient party perceives their period-1 endowment as $$x_1 - ax$$, where $$a \in (0,1]$$ is the fraction of hidden cash that is invisible to them. The impatient party maximizes (p. 2698):

$$
\max_{c_1} \{ \log(c_1) + \beta \log(x_1 - ax - c_1 - r b_1) \} \tag{Self-control model}
$$

subject to $$b_1 := f_{b_1}(c_1) = \begin{cases} c_1 - x_1 + x & \text{if } c_1 - x_1 + x > 0 \\ 0 & \text{otherwise} \end{cases}$$. If the impatient agent coholds, optimal consumption is $$c_1^* = \frac{1}{\beta+1} x_1 - \frac{r+a}{(\beta+1)(1+r)} x$$ and borrowing is $$b_1^* = c_1^* - x_1 + x$$.

The patient party chooses $$x$$ to maximize $$\max_x \{ \log(f_{c_1^*}(x)) + \delta \log(x_1 - f_{c_1^*}(x) - r f_{b_1^*}(x)) \}$$. The coholding condition is:

$$
x_1 - x_1 \frac{1}{1+\delta} \left( \delta \frac{1+r}{r+a} - \frac{\beta + r\beta}{a - r\beta} \right) - \frac{1}{\beta+1} x_1 + \frac{r+a}{(\beta+1)(1+r)} x_1 \frac{1}{1+\delta} \left( \delta \frac{1+r}{r+a} - \frac{\beta+r\beta}{a-r\beta} \right) < 0 \tag{2}
$$

and the optimal hidden cash is $$x^* = x_1 \frac{1}{1+\delta} \left( \delta \frac{1+r}{r+a} - \frac{\beta+r\beta}{a-r\beta} \right)$$.

Proposition 3 (p. 2699): if the patient self becomes more patient (higher $$\delta$$), hidden assets increase ($$\frac{\partial x^*}{\partial \delta} > 0$$) and if the impatient self becomes more impatient (lower $$\beta$$), hidden assets also increase ($$\frac{\partial x^*}{\partial \beta} < 0$$). Proposition 4 (p. 2700): if the patient self increases hidden assets $$x$$, the impatient party's consumption decreases, especially when more assets can be hidden ($$\frac{\partial c_1^*}{\partial x} < 0$$ and $$\frac{\partial^2 c_1^*}{\partial x \partial a} < 0$$), while the sensitivity of borrowing to hidden cash is $$\frac{\partial b_1^*}{\partial x} = \frac{\partial c_1^*}{\partial x} + 1$$, which is less than one and can be near zero if $$\frac{\partial c_1^*}{\partial x} \approx -1$$.

The key distinguishing prediction: in the self- or partner-control model, a nudge that increases patience or hidden cash raises savings but produces far less comovement between saving and borrowing (slope of borrowing on savings ranges from 0.55 to 0.77 across calibrations) than the transactions-convenience model (slope always close to one).

## Method

The identification design is a large-scale randomized controlled trial at Banorte, a top-five Mexican bank. From a pool of 3,054,503 customers meeting three eligibility requirements (payroll account, average daily balance at least 50 MXN over two prior months, valid cell phone), 357,567 were randomly assigned to a control group. The remaining 2,696,936 were assigned to one of seven saving-nudge SMS messages sent bi-weekly over seven weeks (September 13 to November 1, 2019). Randomization was stratified on income quartile, age quartile, bank tenure, baseline savings, digital-banking dummy, ATM transaction median, credit card dummy, and debit card transaction terciles, balancing on 161 pretreatment variables (Table IA.I).

**Aggregate treatment effects.** The primary estimating equation (equation 3, p. 2711) is a Poisson regression for proportional effects:

$$
Y_i = \exp\{\alpha_s + \beta \cdot \text{treatment}_i + \epsilon_i\} \tag{3}
$$

where $$\alpha_s$$ are randomization-block fixed effects, $$\text{treatment}_i$$ is a dummy for receiving any of the seven messages, and $$\beta$$ yields proportional treatment effects $$\exp(\beta) - 1$$. Following Chen and Roth (2024) and Cohn, Liu, and Wardlaw (2022), this handles zero outcomes consistently. For binary variables, a linear probability model is used. Robust standard errors throughout.

**Heterogeneous treatment effects via causal forests.** To avoid overfitting when searching for the most responsive subgroup, the paper applies generalized random forests following Athey, Tibshirani, and Wager (2019). For each of the three outcomes (spending, savings, borrowing), a pilot forest with 2,000 trees is trained on all 161 pretreatment variables; a second forest is trained on variables with importance above 1%. Treatment heterogeneity is tested using the Chernozhukov et al. (2018) calibration test. Cross-fitted rankings over five folds assign each observation to a quartile of predicted treatment effects, and actual treatment effects are estimated using Poisson regression within each quartile. The test detects significant heterogeneity for spending and saving but no heterogeneity for borrowing across all subgroups, supporting the null effect on credit card debt.

## Empirical specifications

**Main specification (aggregate effects, Table IV).** Equation (3) is estimated for monthly spending (ATM withdrawals + card transactions + transfers), checking account balance (average daily balance), and monthly credit card interest charges. For credit card outcomes, the dependent variable is the average of interest charges over the two billing cycles intersecting the treatment window (September and October 2019); a robustness check uses November and December cycles for carryover effects. The specification includes strata fixed effects; robust standard errors are clustered at the individual level. Sample: all 3,054,503 individuals (Panel A), 362,223 credit card holders (Panel B), 152,016 individuals with a credit card who paid interest at baseline (Panel C).

**High-response subgroup specification (Tables V and VI).** Equation (3) is re-estimated within the top quartile of the distribution of predicted treatment effects on spending (N=150,177) or saving (N=151,834), identified by the causal forest. This is the horse race: do those who most reduce spending or increase savings show any change in borrowing? The confidence intervals are compared directly to the spending/saving changes to bound the fraction of the saving/spending effect that could be explained by debt.

**Spending composition (Table VII).** The spending reduction is decomposed into deposits, ATM withdrawals, card spending, and outgoing transfers, with the same Poisson specification. About half of the reduction comes from ATM withdrawals (consistent with hiding cash from family sharing pressures) and half from card spending.

**Message-level specification (Table VIII).** Equation (3) is run separately for each of the seven messages, then messages are grouped into short-term, long-term, and mental-accounting (Message 4) categories. Differences in effects across categories are tested via interaction terms.

**Overfitting illustration (Table X).** The sample is divided into 6,104 strata-blocks. Treatment effects on spending are computed per block, then the top quartile of blocks (by treatment effect) is identified. Causal forest estimates for the same top-quartile subgroup are compared to the block-sorted estimates to show that the block-sort inflates the spending reduction from 7% to 38% and spuriously shows decreased borrowing.

All specifications use robust standard errors. No clustering by branch or region (individual-level randomization). Winsorization at 1st and 99th percentiles for continuous variables; credit card balances, interest charges, and credit limits winsorized only for credit card holders.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Banorte proprietary bank-account panel | Checking account balances, credit card interest charges, transaction-level spending (ATM, card, transfers), credit card balance and limit, deposits, payments; 161 pretreatment variables | No page yet |
| Banorte credit bureau pulls (bimonthly) | Non-Banorte credit card balances for substitution check; aggregate credit exposure | No page yet |
| OECD PPP conversion rates (2019) | Convert MXN to USD PPP for reporting; 1 MXN = 0.107 USD PPP | No page yet |

Sample: 3,054,503 individual bank customers at Banorte, Mexico; experiment conducted September 13 to November 1, 2019 (seven weeks). Credit card subsample: 362,223 holders; interest-paying subsample: 152,016. All financial variables measured in Mexican Pesos (MXN); 1 MXN = 0.107 USD PPP.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13466) if you are:
- building or testing models of the coholding puzzle in other settings (the theoretical framework in Section I with Propositions 1-4 is the formal foundation);
- applying causal forests to heterogeneous treatment effects in large RCTs (Sections II.D and III.F provide the most detailed methodological discussion and the overfitting comparison);
- studying the effectiveness of saving nudges or SMS-based financial interventions, especially unintended balance-sheet effects;
- replicating or extending the empirical results (the Internet Appendix contains additional robustness tables IA.I through IA.XII referenced throughout).

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(5), October 2025, pp. 2689-2738. DOI: 10.1111/jofi.13466. This distillation was extracted by an LLM (claude-sonnet-4-6) on 2026-06-05 and is **not human-verified or independently reproduced**. The paper is paywalled; no CC licence was found in Crossref metadata. Extract-only: no PDF is hosted here.

> Medina, Paolina C., and Michaela Pagel. "Does Saving Cause Borrowing? Implications for the Coholding Puzzle." *The Journal of Finance* 80, no. 5 (October 2025): 2689-2738. DOI: 10.1111/jofi.13466. Published by Wiley on behalf of the American Finance Association. All rights reserved.


==============================================================================
# Anomalies and Their Short-Sale Costs: Muravyev, Pearson & Pollet (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/muravyev-anomalies-short-sale-costs-2025/
# Distilled: across 162 asset pricing anomalies, average long-short abnormal returns of 0.14%/month vanish once stock borrow fees are accounted for, either by fee adjustment or by dropping high-fee stocks; the result holds for subsets including microcaps, high-fee anomalies, and factor-mimicking portfolios. J. Finance 2025, CC BY 4.0. Nine core results with source locators, datasets used, the identification strategy, and the estimating specifications with exact panel-regression equations.
# Tags: paper-summary, anomalies, short-selling, asset-pricing, limits-to-arbitrage, portfolio-sort, panel-regression, open-access, cc-by, peer-reviewed, unreplicated, data:wrds, data:open-source-asset-pricing, data:markit-securities-finance
==============================================================================

**What this is.** The paper's core results, datasets, identification strategy, and the estimating equations: enough to know what it found and how without reading all 56 pages. To replicate or extend it, read the original at
[doi.org/10.1111/jofi.13501](https://doi.org/10.1111/jofi.13501).

## TL;DR

Using 162 anomalies from Chen and Zimmermann (2021) and stock borrow fee data
from Markit (July 2006 to December 2020), the paper shows that the average
long-short abnormal return of 0.14%/month is entirely due to high-fee stocks
(borrow fee greater than 1%/year, roughly 12% of stock-month observations).
Once high-fee stocks are excluded, or once returns are adjusted for the borrow
fee, the average long-short abnormal return collapses to 0.04% or -0.01%/month,
respectively, neither significantly different from zero. The result holds across
microcap stocks, the 20 anomalies with the highest fees, four factor-mimicking
portfolios (momentum, profitability, investment, book-to-market), and five
individually named anomalies. Portfolios sorted on theoretically grounded risk
measures (CAPM beta, tail-risk beta) are unaffected, serving as a placebo.
A short-interest-to-institutional-holdings ratio (from Compustat/13F) works as
a publicly available proxy for borrow fees, eliminating the need for Markit data
to test exploitability.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%.
Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Average long-short **abnormal return across 162 anomalies is 0.14%/month before fees**, significantly positive | Table III Panel A, p. 3659 | Mean = 0.14%/mo, t-stat = 2.87\*\*\* |
| R2 | **High-fee stocks drive the long-short return**: excluding stocks with borrow fee > 1%/yr, long-short abnormal return drops to 0.04% and is insignificant | Table III Panel B, p. 3659 | Mean = 0.04%/mo, t-stat = 0.84 |
| R3 | **Fee-adjusted long-short return is near zero** (borrow fee added back to short-side returns for full sample) | Table III Panel C, p. 3659 | Mean = -0.01%/mo, t-stat = -0.24 |
| R4 | **Unadjusted decile 1 abnormal returns are approximately linear in the borrow fee**, slope -0.088 (approx. -1/12 per 1 ppt fee) | Table V col. 1, p. 3671 | Slope = -0.0883, t-stat = -1.96\* |
| R5 | **Microcap anomaly returns are also entirely due to high-fee stocks**: before fees decile 1 = -0.48%/mo (t = -2.92); excluding high-fee stocks decile 1 = +0.09% (t = 0.76); fee-adjusted = +0.08% (t = 0.50) | Table IV, p. 3669 | Before-fee L/S = +0.36% (t = 5.06); after drop = +0.11% (t = 1.39); net-of-fee = -0.05% (t = -0.62) |
| R6 | **High-fee anomaly subset (20 anomalies, avg fee > 4%/yr)**: before-fee long-short = 0.41%/mo (t = 2.08); excluding high-fee stocks = -0.04%/mo; fee-adjusted long-short = 0.02%/mo (near zero, insignificant) | Table VI, p. 3673 | Exclud. high-fee L/S = -0.04% (t = -0.21); net-of-fee = 0.02% (t = 0.10) |
| R7 | **Momentum and profitability factor long-short returns are eliminated by borrow fees**: momentum L/S drops from 0.21% to 0.03%/mo after excluding high-fee stocks, profitability from 0.61% to 0.09%/mo; net-of-fee momentum = -0.01%, profitability = 0.21%/mo (annualizes to 2.52%/yr, insignificant) | Table XI, p. 3688 | Momentum net-of-fee L/S: -0.01% (t = -0.11); profitability net-of-fee: 0.21% (t = 0.59) |
| R8 | **CAPM-beta and tail-risk-beta sorted portfolios are unaffected** by borrow fee exclusion or adjustment (placebo): CAPM L/S net-of-fee = 0.45%/mo (t = 0.81), tail-risk L/S net-of-fee = 0.43%/mo (t = 0.97) | Table XII, p. 3690 | Net-of-fee CAPM L/S = 0.45%, tail-risk L/S = 0.43%; both insignificant |
| R9 | **Short-interest-to-institutional-holdings ratio (Compustat/13F, no Markit needed) replicates the main result**: excluding stocks with SI/IO > 18% reduces decile 1 abnormal return by 0.27%/mo relative to Panel A | Table IX Panel D, p. 3682 | SI/IO exclusion decile 1: 0.03% (t = 0.47) vs. -0.24% (t = -2.92) in Panel A |

**Overall (paper's conclusion).** Stock borrow fees function as a common limit
to arbitrage that is sufficient to explain the persistence of anomaly returns
for marginal investors. The average anomaly cannot be profitably exploited via
long-short strategies once shorting costs are incorporated. The residual puzzle
is why long-side investors continue to hold high-fee stocks despite bearing
negative expected returns relative to low-fee benchmarks.

## Theory / model

The paper has no original structural model. It is an applied empirical study
operating in the limits-to-arbitrage tradition (Lee, Shleifer, and Thaler (1991);
Nagel (2005); Stambaugh, Yu, and Yuan (2012)). The tested hypothesis is:

> Stock borrow fees are a common, binding limit to arbitrage that prevents
> exploitation of cross-sectional return anomalies and explains their
> apparent out-of-sample persistence.

**Conceptual identity: borrow fee as shadow dividend.** A short seller pays the
daily borrow fee for every day a short position is open. A long-side investor
whose shares are lent receives the fee less prime-broker intermediation spreads,
but only on the fraction of shares actually borrowed. The fee is therefore a
shadow dividend not recorded in CRSP stock returns. The paper's identification
logic follows directly (pp. 3640-3641, 3650):

$$
\begin{aligned}
R^{\text{short,adj}} &= R_{\text{stock}} + f \\
R^{\text{long,adj}}  &= R_{\text{stock}} + \text{utilization} \times (1 - \text{spread\_fraction}) \times f \\
                     &\approx R_{\text{stock}} + \text{utilization} \times 0.7 \times f
\end{aligned}
$$

- $$f$$ = indicative borrow fee (annualized, converted to monthly)
- $$\text{utilization}$$ = shares on loan / lendable shares
- D'Avolio (2002) estimates the spread fraction at approximately 0.3

In contrast to Drechsler and Drechsler (2021), who find positive net-of-fee
returns on eight anomalies using lender-side fees, this paper uses buy-side
fees over a longer and more recent sample and finds near-zero net-of-fee
returns.

**Identification design.** Two complementary approaches test whether anomaly
returns survive after accounting for fees (p. 3640):

1. Exclude high-fee stocks: drop stock-month observations with indicative borrow
   fee > 1%/year from the sorted portfolios (without resorting). Approximately
   12% of stock-months qualify as high-fee; 21% of decile 1 observations do.
2. Fee-adjust returns: add the full fee to returns for short-side portfolios
   (deciles 1 and 2); add $$\text{utilization} \times 0.7 \times f$$ to returns for long-side
   portfolios (deciles 3-10).

Both approaches use a DGTW characteristics-matched benchmark that excludes
high-fee stocks from the benchmark portfolios to avoid contaminating the
abnormal-return calculation (p. 3654 fn. 16).

**Placebo test.** Portfolios sorted by CAPM beta and tail-risk beta (Kelly and
Jiang (2014)) are theoretically grounded risk measures, not behavioral anomalies.
If borrow fees reflect arbitrage frictions specific to mispriced stocks, these
sorted portfolios should be insensitive to borrow-fee adjustments. The
insensitivity result (Table XII, p. 3690) provides placebo support.

## Method

The paper applies `portfolio-sort` and `panel-regression` methods. There is no
newly proposed estimator; the methodological contribution is the systematic
application of buy-side borrow fees to a comprehensive, out-of-sample anomaly
universe.

**DGTW abnormal return construction (pp. 3654-3655).** For each stock $$i$$ and
month $$t$$, the abnormal return is:

$$
\text{DGTW\_ETB}_{i,t} = R_{i,t} - R_{\text{benchmark},i,t}
$$

- $$R_{\text{benchmark},i,t}$$ is the equal-weighted return of the DGTW characteristics-matched portfolio, constructed excluding stocks with borrow fee > 1%/year (to prevent high-fee benchmark contamination). The benchmark portfolios match on market capitalization, book-to-market, and prior six-month momentum, following Daniel et al. (1997).

**Portfolio-level aggregation.** The abnormal return on a sorted decile
portfolio in month $$t$$ is the cross-sectional average of stock-level abnormal
returns within that decile. For each anomaly, the time-series average is
computed over the performance evaluation period (July 2006 to December 2020,
14.5 years). The cross-sectional mean across the 162 anomalies is the headline
statistic (p. 3654).

**Fee adjustment procedure (pp. 3650, 3666-3667).** The monthly fee is the
simple average of daily indicative fees over the 21-trading-day return
evaluation window. For decile 1 stocks (short side), the full fee is added to
the stock return. For decile 3-10 stocks (long side, held or lent), the
expected fee received is:

$$
\text{fee\_received}_{i,t} = \text{utilization}_{i,t} \times 0.7 \times f_{i,t}
$$

This is added to the stock return. Decile 2 uses the same adjustment as decile
1 in both the fee-exclusion analysis and the fee-adjustment analysis (i.e., the
full fee is added to decile 2 stock returns, treating it as short-side).

**Standard-error treatment.** All t-statistics on cross-sectional averages are
computed using a panel regression in which the monthly portfolio return for each
anomaly is regressed onto decile fixed effects; standard errors are
double-clustered by anomaly and month. The estimate for each decile fixed effect
is the average return on the corresponding decile with the appropriate
t-statistic (pp. 3655-3656).

## Empirical specifications

The paper's results come from three types of portfolio-performance constructions
rather than a single OLS regression. Each is described below with the estimating
equation.

**Specification 1: Cross-sectional mean of decile abnormal returns (R1-R3,
R5-R6, Table III p. 3659).**

For each anomaly $$a$$, the time-series average abnormal return on decile $$d$$ is:

$$
\mu_{a,d} = \frac{1}{T} \sum_{t=1}^{T} \text{DGTW\_ETB}_{a,d,t}
$$

The cross-sectional mean across anomalies (the headline statistic) and the
corresponding t-statistic are extracted from the panel regression:

$$
\text{DGTW\_ETB}_{a,d,t} = \alpha_d + \epsilon_{a,d,t}
$$

- LHS: monthly decile portfolio abnormal return for anomaly $$a$$, decile $$d$$, month $$t$$
- RHS: decile fixed effects $$\alpha_d$$ (one per decile 1-10)
- SE: double-clustered by anomaly $$a$$ and month $$t$$
- Sample: 162 anomalies, July 2006 to December 2020, varying N per anomaly

The long-short result is $$\alpha_1 - \alpha_{10}$$ (decile 1 minus decile 10, since
anomaly signals are signed so decile 1 is the short side). The procedure is
applied in three versions: (A) all stocks, no fee adjustment; (B) excluding
high-fee stocks (fee > 1%/yr); (C) fee-adjusted returns.

**Specification 2: Relation between anomaly returns and borrow fees (R4,
Table V p. 3671).**

For each anomaly, the average decile 1 borrow fee $$f_a$$ is regressed on the
average abnormal return to show the approximately linear relationship:

$$
\mu_{a,1} = \beta_0 + \beta_1 f_a + u_a
$$

- LHS: time-series average abnormal return on decile 1 portfolio, anomaly $$a$$
- RHS: average indicative borrow fee for decile 1 stocks of anomaly $$a$$ at portfolio formation date
- SE: double-clustered by anomaly and month (panel version with month controls)
- Key result: $$\beta_1 = -0.0883$$ (t = -1.96*), approximately $$-1/12$$

The slope of approximately $$-1/12$$ means a 1 percentage-point increase in the
annual fee corresponds to roughly $$-1/12$$ percentage points per month in abnormal
return, consistent with complete fee absorption of the anomaly signal (p. 3671).

**Specification 3: SI/IO exclusion as Markit substitute (R9, Table IX Panel D
p. 3682).**

The same Specification 1 panel regression is applied after excluding
stock-months where short interest divided by institutional holdings (SI/IO,
from Compustat monthly short interest and 13F filings) exceeds 18%:

$$
\text{DGTW\_ETB}_{a,d,t} = \alpha_d + \epsilon_{a,d,t} \quad \text{[restricted sample]}
$$

- Restriction: drop stock-months with $$(\text{short\_interest}_{i,t-1} / \text{inst\_holdings}_{i,t-1}) > 0.18$$
- Cutoff 18% chosen to match approximately 12% of stock-months excluded by fee > 1% cutoff
- SE: double-clustered by anomaly and month
- Key result: decile 1 $$\alpha$$ changes by +0.27%/mo relative to unadjusted baseline, decile 1 $$\alpha$$ = +0.03% (t = 0.47) vs. -0.24% (t = -2.92) with all stocks

Robustness: microcap extension (all stocks, market cap below 20th percentile
NYSE, Table IV p. 3669), which tests the claim of Hou, Xue, and Zhang (2020)
that anomaly returns are concentrated in microcaps and finds that even microcap
anomaly returns disappear once borrow fees are accounted for; 20 highest-fee anomalies (Table VI p. 3673); subsets
by t-statistic, pre-sample Sharpe ratio, and publication venue (Table VII
p. 3674); long-side investor perspective with varying intermediation fractions
(Table VIII p. 3678); five specific anomalies (Table X p. 3685); four
factor-based long-short portfolios (Table XI p. 3688).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Chen and Zimmermann (2021) anomaly signals (openassetpricing.com), Jan 2000 to Dec 2020; 202 anomalies, 162 retained | Anomaly signal construction; decile portfolio assignment | [Open Source Asset Pricing](/wiki/datasets/open-source-asset-pricing/) |
| Markit Securities Finance Buy Side Analytics Data Feed, daily from Jun 28 2006 | Stock borrow fees (indicative fee = buy-side expected borrow cost); utilization | no page yet |
| CRSP (via WRDS) common stocks, returns, delisting returns | Stock returns, market cap filters, sample construction | [WRDS / CRSP / Compustat](/wiki/commercial/wrds/) (licensed) |
| Compustat (via WRDS) short interest + 13F institutional holdings | Proxy for borrow fee (SI/IO ratio) for researchers without Markit access | [WRDS / CRSP / Compustat](/wiki/commercial/wrds/) (licensed) |

Sample: 554,253 stock-months (162 anomaly signals, Jul 2006 to Dec 2020),
after dropping stocks below $1 price or $50 mn market cap and requiring at
least 4 days of borrow fee observations per month.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13501) if you are: replicating
(code in the journal's Supporting Information); extending the borrow fee
adjustment methodology; examining the specific anomaly-by-anomaly results in
the Internet Appendix; or auditing factor return attrition (Section VI). The
locators above point to the exact tables. For "what did this paper find," the
table above is the intended default.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(6). This distillation was
extracted by an LLM on 2026-05-31 and augmented on 2026-06-01; it is **not
human-verified or independently reproduced**. CC BY 4.0 permits this adaptation;
the verbatim PDF is not hosted in this batch but CC permits mirroring.

> **Attribution (CC BY 4.0).** Muravyev, Dmitriy, Neil D. Pearson, and Joshua
> M. Pollet. "Anomalies and Their Short-Sale Costs." *The Journal of Finance*
> 80, no. 6 (December 2025): 3639-3694. DOI: 10.1111/jofi.13501.
> Copyright 2025 The Author(s). Licensed under
> [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# OTC Markets for Nonstandardized Assets: Nozawa & Tsoy (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/nozawa-counter-markets-nonstandardized-assets-2025/
# Distilled: Nozawa and Tsoy build a search-and-bargaining model of OTC markets for nonstandardized assets, deriving that bargaining delays are hump-shaped in unobserved asset quality and asset turnover is U-shaped. Empirical tests on corporate bonds (TRACE, 2002-2020) and commercial real estate (CoStar, 1998-2022) confirm the U-shaped liquidity pattern; a placebo test on agency MBS finds no such pattern. J. Finance 2025, paywalled. Seven core results with source locators, datasets used, the model, and the method.
# Tags: paper-summary, market-microstructure, otc-markets, search-frictions
==============================================================================

**What this is.** The paper's core results, the structural model it proposes (OTC search and bargaining over nonstandardized assets), and the empirical tests with the key equations: enough to know what it found and how, without reading all 43 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13483).

## TL;DR

Nozawa and Tsoy study over-the-counter markets where assets differ in unobserved quality. They extend the Duffie, Garleanu, and Pedersen (2005) model (and its risk-aversion variant in Duffie, Garleanu, and Pedersen (2007)) with two-sided private information during bargaining: once matched, buyers and sellers privately learn their own values (correlated with asset quality) but not their counterparty's. This two-sided information structure generates bargaining delays that are hump-shaped in the unobserved quality of the asset, which translates into a novel, testable prediction: asset turnover is U-shaped in quality. Extreme-quality assets trade quickly (both sides concede early), while intermediate-quality assets experience long negotiations (both sides wait for a better offer). The theory also implies that the opaque OTC structure supports trading of a wide variety of assets (extensive margin), while centralization of trading concentrates liquidity in the most-liquid assets and can harm the extensive margin. Empirical tests using corporate bond TRACE data (2002-2020) and commercial real estate CoStar data (1998-2022) confirm the U-shaped pattern; a placebo on agency MBS (highly standardized) finds no such pattern.

## Core results

Magnitudes and significance are as reported. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Corporate bond turnover is U-shaped** in excess bond premiums: extreme deciles have significantly higher turnover than middle deciles | Figure 2, p. 2858 | t(10-min)=7.99; t(1-min)=3.37; consistent U-shaped pattern |
| R2 | **Corporate bond trading volume is U-shaped** in excess bond premiums: extreme deciles trade more per month in dollar terms | Figure 2, p. 2858 | t(10-min)=5.43; t(1-min)=1.50 (marginal); U-shape visible |
| R3 | **Zero-trade days are hump-shaped**: middle decile bonds spend more days without transactions than extreme deciles | Figure 2, p. 2858 | t(1-max)=-2.41; t(10-max)=-4.84; hump pattern mirrors U-shaped turnover |
| R4 | **Bid-ask spreads are U-shaped** in excess bond premiums: extreme deciles have higher spreads than middle | Figure 2, p. 2858 | t(10-min)=9.15; t(1-min)=2.88 |
| R5 | **Post-Volcker period shows more pronounced U-shaped volume** than pre-Volcker; as dealer intermediation capacity shrinks, bilateral negotiation matters more | Figure 3, p. 2859 | Post-Volcker (Apr 2014-Dec 2020) volume shows sharper U; turnover and bid-ask similar across subperiods |
| R6 | **Commercial real estate days to sell is hump-shaped** in unobserved property quality: mid-quality properties take longest to sell | Figure 4, p. 2861 | 7th decile max minus 1st decile = 24 days (t=5.65); 7th minus 10th decile = 32 days (t=9.46); both significant |
| R7 | **Placebo test on agency MBS finds no U-shaped pattern**: volume, turnover, and zero-trade days show no significant variation across unexplained yield spread deciles | Figure 6, p. 2864 | All t-stats below 2.0 in absolute value for volume and turnover; confirms nonstandardization drives the nonlinearity |

**Overall (paper's conclusion).** The results support the model's central prediction that two-sided private information about asset quality in OTC markets generates hump-shaped bargaining delays and U-shaped liquidity. The evidence holds for corporate bonds and CRE (both highly nonstandardized) but not for agency MBS (highly standardized), providing a clean placebo test. The model implies that centralization of trading should accompany standardization to avoid liquidity concentrating on a narrow set of assets.

## Theory / model

The model is a continuous-time search-and-bargaining general equilibrium. Asset quality $$\theta \in [0,1]$$ is unobserved and captures heterogeneous payoffs. A unit mass of infinitely lived, risk-neutral investors hold assets with flow payoff $$1 + k\theta$$ ($$k > 0$$ captures payoff heterogeneity; larger $$k$$ means less standardized). Investors experience liquidity shocks that convert high-type (asset holder, no holding cost) to low-type (asset holder, holding cost $$\delta > 0$$) at Poisson rate $$\lambda_d$$, and recover at rate $$\lambda_u$$ (p. 2836-2837).

Low types (sellers) and high types without assets (buyers) search for counterparties. Buyers and sellers meet with Poisson contact intensity $$2\lambda m_b m_s$$. The steady-state balance conditions (p. 2838, equations 1-2) are:

$$
\mu_{lu}(\theta) + \mu_{hu}(\theta) + \mu_{lm}(\theta) = sf(\theta), \quad \forall \theta \in [0,1] \tag{1}
$$

$$
M_{lu}(\varnothing) + M_{hu}(\varnothing) + M_{hm}([0,1]) = 1 - s \tag{2}
$$

Bellman equations for unmatched and matched investors are given by equations (3)-(8) on pp. 2838-2840, where $$rV_{hu}(\varnothing)$$, $$rV_{lu}(\theta)$$, $$rV_{lu}(\varnothing)$$, $$rV_{hu}(\theta)$$ capture the continuation values of buyers without assets, sellers, unmatched low types, and high types with asset $$\theta$$, respectively.

The key innovation is the bargaining structure. Once matched, buyers and sellers privately observe their own types but not their counterparty's. The buyer's type $$\theta^b = \theta + \varepsilon^b$$ and seller's type $$\theta^s = \theta + \varepsilon^s$$ are affiliated, converging to the true quality $$\theta$$ in the limit as noise vanishes. Drawing on Tsoy (2018, 2019), in the limit of frequent offers ($$\Delta \to 0$$) the bargaining game admits a limit of Perfect Bayesian Equilibria with closed-form price and delay outcomes (p. 2841, equations 12-13):

$$
p(\theta) = qv(\theta) + (1-q)c(\theta), \quad \forall \theta \in [0,1] \tag{12}
$$

$$
t(\theta) = \begin{cases} \frac{1}{\rho} \int_0^{\theta} \frac{dp(\tilde\theta)}{p(\tilde\theta) - c(\tilde\theta)}, & \theta \le \theta^* \\ \frac{1}{\rho} \int_\theta^1 \frac{dp(\tilde\theta)}{v(\tilde\theta) - p(\tilde\theta)}, & \theta \ge \theta^* \end{cases} \tag{13}
$$

where $$q \in (0,1)$$ is the seller's bargaining power, $$v(\theta)$$ is the buyer's value, $$c(\theta)$$ is the seller's cost, and $$\theta^*$$ is the quality where seller and buyer roles reverse in who drives the bargaining dynamics (p. 2841). The bargaining delays $$t(\theta)$$ in equation (13) are hump-shaped in $$\theta$$: extreme-quality assets are traded quickly because one side has a dominant incentive to concede; intermediate-quality assets near $$\theta^*$$ involve longer negotiations.

**Proposition 3** (p. 2848) establishes that the effective discount factor $$x(\theta) = \exp\{-\rho t(\theta)\}$$ and the asset turnover $$w(\theta)$$ are U-shaped in $$\theta$$ (equation 29):

$$
w(\theta) = \frac{\lambda_d \Lambda_s \sigma(\theta)}{\lambda_u + \lambda_d + \Lambda_s \sigma(\theta)} x(\theta)^{(\lambda_u + \lambda_d)/\rho} \tag{29}
$$

where $$\Lambda_s$$ is market thickness for sellers and $$\sigma(\theta) \in [0,1]$$ is the probability that a buyer proceeds to bargaining with asset $$\theta$$.

The equilibrium is characterized by (p. 2844-2847): (i) steady-state market thickness $$\Lambda_s$$ from equation (14), (ii) buyers' cutoff on acceptable effective discount factors $$\underline{x} = \bar{x}\Lambda_b / (\rho + \Lambda_b)$$ (equation 24), and (iii) the U-shaped structure of $$x(\theta)$$ from Lemma 3.

Asset prices (Proposition 2, p. 2847, equation 28) embed an illiquidity discount relative to full-information frictionless prices:

$$
p(\theta) = U_s(\theta) + \frac{q\delta}{\rho}\!\left(1+\frac{\lambda_d}{r}\right) - \frac{\delta}{r}\!\left[(1-q)\frac{\lambda_u}{\rho}\frac{\Lambda_b \bar{x}}{\rho+\Lambda_b} + q\frac{\lambda_d}{\rho}\!\left(1 - \frac{\sigma(\theta)\Lambda_s x(\theta)}{\rho + \sigma(\theta)\Lambda_s}\right)\right] \tag{28}
$$

where $$U_s(\theta) \equiv \frac{1}{r}(1+k\theta) - \frac{\delta}{\rho}(1+\lambda_d/r)$$ is the autarky value for the seller.

## Method

The paper is primarily a structural theory paper with an empirical validation component. On the theory side, the equilibrium solution method proceeds in three steps (pp. 2843-2847):

**Step 1.** Derive the steady-state distribution $$\mathbf{M}$$ in closed form via balance equations (A3)-(A4) in the appendix. Lemma 1 establishes that $$\Lambda_s$$ is the unique positive solution to equation (27) (the uniform-distribution specialization of Lemma 1's general equation (14)):

$$
\frac{\Lambda_s}{2\lambda} = \frac{\lambda_u(1-s)}{\lambda_u + \lambda_d} - \frac{\lambda_d s}{\lambda_u + \lambda_d} I(L, \Lambda_s) \tag{27}
$$

where $$I(L, \Lambda_s)$$ is given in the proof.

**Step 2.** Buyers' optimal strategy takes a cutoff form on the effective discount factor $$x(\theta)$$: buyers accept assets with $$x(\theta) > \underline{x}$$, reject those with $$x(\theta) < \underline{x}$$, and mix at equality (Lemma 2, p. 2844, equation 24).

**Step 3.** Lemma 3 (p. 2846, equation 25) pins down the effective discount factor $$x(\theta)$$ as:

$$
x(\theta) = \begin{cases} \exp\{-\rho(c(\theta) - c(0))/(q\delta)\} & \theta \in [0, \check\theta] \\ \exp\{-\rho(v(1) - v(\theta))/((1-q)\delta)\} & \theta \in [\hat\theta, 1] \end{cases} \tag{25}
$$

The equilibrium is then pinned down by solving equations (14) and (24) jointly for $$\underline{x}$$ and $$\Lambda_s$$.

On the empirical side, the paper does not estimate the structural model; it tests the model's reduced-form qualitative prediction (U-shaped liquidity in quality) in two nonstandardized asset classes. The empirical strategy is:

1. Classify assets by a proxy for unobserved quality (excess bond premium for corporate bonds, regression residuals of the price-assessment ratio for CRE, unexplained yield spread for MBS).
2. Form 10 decile portfolios by the quality proxy.
3. Compute equal-weighted average liquidity measures within each decile.
4. Plot and test whether the pattern is U-shaped (bonds) or hump-shaped (days to sell for CRE).

The proxy for unobserved corporate bond quality $$\theta$$ follows Gilchrist and Zakrajsek (2012): the excess bond premium is the difference between observed credit spreads and spreads predicted by public information. For CRE, unobserved quality translates into residual variation in price premiums relative to the tax assessment benchmark: Garmaise and Moskowitz (2004) show that a property's tax assessment serves as a public benchmark for its value (p. 2860). The price-assessment ratio residual comes from estimating equation (30):

$$
P_i / B_i = a + b X_i + \varepsilon_i \tag{30}
$$

where $$P_i$$ is the sale price, $$B_i$$ is the tax assessment, and $$X_i$$ includes property age, floors, square footage, property-type dummies, submarket dummies, building condition, and buyer/seller location dummies (p. 2860). Standard errors in the CRE analysis are clustered at the quarter-submarket-property-type level, computed by block bootstrap (Figure 4 caption, p. 2861). Corporate bond standard errors use Newey and West (1987) with 12 lags (Figure 2 caption, p. 2858).

## Empirical specifications

**Corporate bonds (R1-R5).** TRACE monthly corporate bond transactions (July 2002-December 2020) are merged with Mergent FISD for amount outstanding and credit rating. Nonfinancial issuer firms only. Bonds are sorted monthly into 10 portfolios by their excess bond premium (Gilchrist and Zakrajsek (2012)). Within each portfolio, four liquidity measures are computed: (i) equal-weighted average dollar volume per month, (ii) turnover rate (monthly volume divided by amount outstanding), (iii) number of zero-trade days in the following 12 months, and (iv) bid-ask spread (average sell minus average buy price within a day, scaled by the average). The time-series average of each measure within each decile is then plotted against the decile rank (Figure 2). Subperiod analysis splits at April 2014 following Bao, O'Hara, and Zhou (2018) (Figure 3).

**Commercial real estate (R6).** CoStar data on completed CRE sales across 15 US cities from 1998Q1 to 2022Q3. The liquidity proxy is days to sell (listing date to delisting date). Unobserved quality is the OLS residual from equation (30) estimated city by city. Deals are then classified into 10 decile bins, and the median (not mean, due to skewness) days to sell within each decile is computed. Standard errors are estimated by block bootstrap, clustered at the quarter-submarket-property-type level (Figure 4).

**MBS placebo (R7).** TRACE MBS data (July 2013-July 2021) for Fannie Mae specified-pool 30-year fixed-rate MBS, merged with Refinitive Eikon for MBS characteristics. Unobserved quality proxied by residuals from a regression of yield spreads on MBS characteristics. 10 equal-weighted portfolios of MBS formed analogously to bonds. Liquidity measures as in the bond analysis. No significant U-shaped pattern emerges (Figure 6), consistent with agency MBS being highly standardized.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| TRACE (corporate bonds) | Corporate bond transaction prices, volumes, and dates; Jul 2002-Dec 2020 | [TRACE](/wiki/commercial/trace/) (licensed) |
| Mergent FISD | Amount outstanding and credit rating for corporate bonds | [no page yet] |
| Gilchrist-Zakrajsek excess bond premium | Proxy for unobserved corporate bond quality (theta); computed from TRACE and public information | [no page yet] |
| CoStar CRE | Commercial real estate sale transactions in 15 US cities; 1998Q1-2022Q3; includes price, assessment value, property characteristics | [CoStar](/wiki/commercial/costar/) (licensed) |
| TRACE (agency MBS) | MBS transaction prices and volumes for Fannie Mae specified-pool MBS; Jul 2013-Jul 2021 | [TRACE](/wiki/commercial/trace/) (licensed) |
| Refinitive Eikon | MBS characteristics (issue amount, date) for the placebo MBS analysis | [no page yet] |

Sample coverage: corporate bonds monthly, Jul 2002-Dec 2020; CRE quarterly, 1998Q1-2022Q3; MBS monthly, Jul 2013-Jul 2021.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13483) if you are: building or testing search-and-bargaining models of OTC markets for heterogeneous assets; studying how asset standardization affects market liquidity (including the trade-off between intensive and extensive liquidity margins); analyzing the impact of market reforms (TRACE, Volcker rule, CDS Big Bang) on OTC liquidity; or extending the model to portfolio choice, dynamic standardization, or other nonstandardized asset classes (derivatives, structured products). The Internet Appendix contains additional proofs, robustness checks, and subsample analyses referenced from the main text.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(5). Copyright 2025 the American Finance Association. Published by Wiley on behalf of the AFA. This distillation was extracted by an LLM on 2026-06-05 and is **not human-verified or independently reproduced**. Reproduction rights: extract-only (paywalled; the Wiley VOR licence does not permit redistribution).

> Nozawa, Yoshio, and Anton Tsoy. "Over-the-Counter Markets for Nonstandardized Assets." *The Journal of Finance* 80, no. 5 (October 2025): 2831-2873. DOI: 10.1111/jofi.13483.


==============================================================================
# What Is the Cost of Privatization for Workers?: Olsson & Tag (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/olsson-what-cost-privatization-workers-2025/
# Distilled: Using Swedish administrative data covering two decades, this paper shows that privatization of state-owned enterprises imposes wage losses of 5-9% and raises unemployment by 12%, while firm-level productivity rises 35.7%; government transfers offset roughly half the worker income losses. J. Finance 2025, CC BY-NC 4.0. Eight core results with source locators, datasets used, the identification strategy, and the estimating equations.
# Tags: paper-summary, privatization, labor-economics, firm-dynamics, panel-regression
==============================================================================

**What this is.** The paper's core results, identification strategy, and estimating equations extracted from the source: enough to know what privatization costs workers and why, without reading all 45 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13462).

## TL;DR

Using Swedish administrative data from 1990 to 2017, the paper traces workers and firms through 553 privatization events involving 70,079 incumbent workers. Privatization raises unemployment by 12% and cuts wages by 5-9%, with losses persisting up to eight years. The Swedish social safety net cushions roughly half the income loss through increased transfers (unemployment benefits and activity support). At the firm level, employment falls 16% and the job destruction rate rises 11 percentage points, while productivity increases 35.7% and profitability improves by 2.1 percentage points. Productivity gains occur only when the CEO is replaced, consistent with governance changes breaching implicit labor contracts and enabling workforce reallocation. Rough cost-benefit calculations show productivity gains from privatization outweigh worker income losses by a factor of two to six.

## Core results

Magnitudes and significance as reported; `*`/`**`/`***` = 10%/5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Workers experience persistent wage declines** after privatization; wages fall 5.8% short-run, 9.3% medium-run, and 8.4% long-run | Table II, p. 2124 | Avg. effect -7.9% (t = -2.96); short run -5.8% (t = -3.47); medium run -9.3% (t = -4.13); long run -8.4% (t = -2.23) |
| R2 | **Unemployment incidence rises persistently** by ~12-14% relative to pre-privatization levels | Table II, p. 2124 | Full period +1.3 pp (t = 5.16), +12.4% vs baseline; short run +1.1 pp (10.5%); medium run +1.2 pp (11.4%); long run +1.5 pp (14.3%) |
| R3 | **Government transfers offset roughly half the wage income loss**; total income drops only 3.5% despite a 7.9% wage cut | Table II, p. 2124 | Transfers up 11.9% (t = 5.82); total income effect -3.5% (t = -1.52) full period; -2.9% to -4.3% in sub-periods |
| R4 | **Firm productivity rises 35.7%** (or 11.5% using log specification); gains driven by the top 75th-90th percentiles | Table V, p. 2134; Table IA.VI | Full period avg. effect +109.5 KSEK (t = 2.77); log-productivity +11.5% (t = 2.03) |
| R5 | **Firm employment falls 16.3%** through higher job destruction, with total payroll down 12.2% | Table V, p. 2134 | Employee count -16.3% (t = -2.82); job destruction rate +10.9 pp (t = 4.92); job creation rate unchanged (t = -1.40); payroll -12.2% (t = -2.15) |
| R6 | **Firm profitability (ROA) improves by 2.1 percentage points** from a near-zero baseline | Table V, p. 2134 | ROA +2.1 pp (t = 1.82), +321.3% from mean of 0.007 |
| R7 | **Business ownership among former SOE workers doubles**, rising ~97-132% over eight years; effect driven by limited liability company formation, not self-employment | Table III, p. 2127 | Business owner rate +6.3 bps (96.8%, t = 3.61) full period; +20 bps long run (131.7%, t = 3.52) |
| R8 | **Productivity gains occur only when the CEO is replaced**; firms where the CEO stays show no productivity improvement | Table VI, p. 2138 | Productivity: CEO leaves +31.4% (t = 2.00); CEO stays +12.8% (t = 0.81); CEO replacement also associated with greater employment and payroll reductions (Table IA.VII) |

**Overall (paper's conclusion).** Privatization leads to a reallocation of human capital that contributes to both an increase in firm-level productivity and losses in income for workers. The productivity gains from privatization exceed the associated worker costs (pre-government transfers) by a factor of two to six. Government transfers total 10-30% of the per-worker productivity gains, implying workers could receive full compensation with a residual surplus remaining for distribution between the new firm owners and the government.

## Theory / model

The paper has no formal structural model. The identification logic rests on an event study difference-in-differences framework under a parallel trends assumption.

**Key hypotheses tested.** Privatization has ex ante ambiguous effects on workers. On one hand, better governance raises labor demand, increasing wages and reducing unemployment, consistent with the firm-level survey evidence in Megginson and Netter (2001). On the other hand, new private owners face a profit motive to replace workers who enjoyed state protection, and the ownership change may breach implicit contracts between managers and workers (Shleifer and Summers (1988)), making labor cost reductions more likely. Prior work documents wage declines in Ukraine (Brown, Earle, and Vakhitov (2006)) and Brazil (Arnold (2022)), while Bastos, Monteiro, and Straume (2014) find wage increases in Portugal due to different wage-setting institutions. The productivity gain mechanism is tested by examining whether productivity increases only when the CEO is replaced, consistent with the governance-and-implicit-contract channel. The matching-plus-stacked-DiD approach extends the design used in Olsson and Tag (2017) on private equity buyouts.

**Identification strategy.** The primary concern is selection bias: SOEs that are privatized may differ systematically from those that remain public. The paper addresses this via:

1. A matched comparison group: each treated worker (or firm) is matched one year before privatization to a similar worker (or firm) that remains state-owned, on age, gender, industry, region, and wage (for workers) or industry, region, size, and ROA sign (for firms).
2. A stacked difference-in-differences design (de Chaisemartin and D'Haultfoeuille (2020), Goodman-Bacon (2021)) to handle heterogeneous treatment effects arising from staggered privatization timing across 1997-2017.
3. Pre-trend testing: event-time plots confirm parallel trends in all outcomes for three years before privatization (Figures 3, 4, 5).

A residual concern is anticipation bias (workers sorting out of SOEs before privatization). The paper addresses this by matching one year before the event, and confirms robustness when matching three years earlier (Table IA.VIII).

## Method

The core estimator is a stacked difference-in-differences regression run in event time, combined with cell matching without replacement.

**Matching.** For workers, each of the 70,079 treated workers is matched to one control worker from the 3.3 million non-privatized SOE worker pool on the Cartesian product of age quartile, gender, industry (four broad NACE groups), NUTS1 region, and wage quartile. The match is one-to-one without replacement, yielding 63,231 pairs.

**Worker-level estimating equation.** Let $$Y_{i,f,k,t}$$ be an outcome for worker $$i$$ at firm $$f$$ in event year $$k$$ and calendar year $$t$$, where $$k = 0$$ is the privatization year. The stacked DiD model (equation 1, p. 2121) is:

$$
Y_{i,f,k,t} = \alpha + \pi \, After_k + \gamma D_i + \beta \, After_k \times D_i + \omega_t + X_i + X_f + \varepsilon_{i,f,k,t} \tag{1}
$$

where $$After_k = 1$$ for $$k \geq 0$$; $$D_i = 1$$ for workers in a privatized SOE (treated) and 0 for matched controls; $$\omega_t$$ is calendar year fixed effects; $$X_i$$ includes individual controls (age, gender, immigrant status, labor market experience, tenure, education, municipality, industry, calendar year, and privatization year fixed effects); $$X_f$$ includes firm age, industry, and region fixed effects. $$\beta$$ is the average intention-to-treat effect.

**Dynamic specification.** To trace the time profile, $$After_k$$ is replaced by event-time dummies $$\tau_k$$ for $$k = -3$$ to $$k = +8$$ (equation 2, p. 2121):

$$
Y_{i,f,k,t} = \alpha + \tau_k + \gamma D_i + \sum_{k=-3}^{k=8} \beta_k \, \tau_k \times D_i + \omega_t + X_i + X_f + \varepsilon_{i,f,k,t} \tag{2}
$$

where $$k - 0$$ is the reference period, so $$\beta_k$$ is the average intention-to-treat effect at event time $$k$$.

**Firm-level equation.** The firm-level analog (equation 3, p. 2122) replaces worker subscripts with firm subscripts; $$X_f$$ includes firm age, industry fixed effects, region fixed effects; standard errors are clustered at the firm level:

$$
Y_{f,t} = \alpha + \pi \, After_k + \gamma D_f + \beta \, After_k \times D_f + \omega_t + X_f + \varepsilon_{f,t} \tag{3}
$$

**Employment growth rates.** Following Davis, Haltiwanger, and Schuh (1998), job flows are computed as (equations 4-6, p. 2133):

$$
g_{f,t} = \frac{E_{f,t} - E_{f,t-1}}{0.5 \times (E_{f,t} + E_{f,t-1})}, \quad JDR_{f,t} = |min\{g_{f,t}, 0\}|, \quad JCR_{f,t} = max\{g_{f,t}, 0\} \tag{4-6}
$$

bounded between -2 (exits) and 2 (entries).

Standard errors are clustered at the municipality level for worker-level regressions and at the firm level for firm-level regressions.

## Empirical specifications

All worker-level results (R1-R3, R7) use equation (1) / (2) on 1,414,270 worker-year observations from 63,231 matched pairs (p. 2124, Table II). Firm-level results (R4-R6, R8) use equation (3) on 4,804 firm-year observations from 368 privatized firms matched to 368 control firms (p. 2134, Table V).

**Wage income (R1).** Dependent variable is the log of annual gross salary income via the inverse hyperbolic sine transformation. The pre-privatization mean is 271,080 SEK (~27,108 USD). The full-period DiD coefficient is -0.079 (t = -2.96), -7.9%. Short-run coefficients: -0.058 (t = -3.47, -5.8%), medium-run: -0.093 (t = -4.13, -9.3%), long-run: -0.084 (t = -2.23, -8.4%). The mechanical unemployment channel accounts for only 16.4% of the observed wage cut.

**Unemployment incidence (R2).** Dependent variable is a binary indicator equal to one if the worker was unemployed at any time during the year. The full-period effect is +0.013 (t = 5.16, +1.3 pp, +12.4% vs baseline mean of 0.105). Short-run: +0.011 pp (t = 4.02); medium-run: +0.012 pp (t = 4.27); long-run: +0.015 pp (t = 5.27).

**Government transfers (R3).** Dependent variable is the log of annual gross government transfers (unemployment benefits, activity support, social benefits) via inverse hyperbolic sine. Full-period effect: +0.119 (t = 5.82, +11.9%). Total income (wages plus transfers): -0.035 (t = -1.52, -3.5%), roughly half the wage loss. Transfers break down as: unemployment benefits +11.1% (t = 7.45), activity support payments +4.3% (t = 6.69), social benefit payouts essentially unchanged (-0.02%, t = 0.23) (Table IA.III).

**Firm productivity and employment (R4-R6).** Productivity (value-added per employee) full-period effect: +109.5 KSEK (t = 2.77, +35.7%), driven by the top 75th-90th percentile; log-productivity: +11.5% (t = 2.03). Employment: -16.3% (t = -2.82). Payroll: -12.2% (t = -2.15). ROA: +2.1 pp (t = 1.82). All effects are present in both short and medium run.

**CEO channel (R8).** The sample is split into privatizations where the CEO remains (column 5, Table VI, N = 2,433 observations) and where the CEO departs (column 6, N = 2,278). Productivity effect: CEO stays +12.8% (t = 0.81, insignificant); CEO leaves +31.4% (t = 2.00, significant). The same split is applied to employment and payroll (Table IA.VII).

**Heterogeneity (Section V, Tables VII-IX).** Partial privatizations (28,694 treated workers) show larger effects: wages -16.5% (t = -4.11), unemployment +2.5 pp (42%, t = 5.77). Foreign buyers lead to unemployment increases of +3.2 pp (32%, t = 8.59) versus +0.5 pp for domestic buyers; total income falls -16.6% with foreign buyers (t = -6.07) versus no change for domestic. MBO privatizations show no statistically significant adverse effects. Triple-difference regressions find no differential effects across industries, high-unemployment regions, or recession years (Table IX).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| LISA database (Statistics Sweden, longitudinal integration database for health insurance and labor market studies) | Individual-level annual data on wages, unemployment, transfers, family structure, education, wealth for all Swedish residents aged 15+, 1990-2017 | no page yet |
| Structural Business Statistics (FEK) database (Statistics Sweden) | Firm-level data on employees, payroll, productivity (value-added per employee), ROA, investment ratio, leverage, 1997-2017 | no page yet |
| Wealth Register (Statistics Sweden) | Individual total wealth, risky assets, cash, and debt, 1999-2007; used for household finance outcomes | no page yet |
| Swedish military draft cognitive/noncognitive skill scores | Quality-of-hire measure for male employees; stanine scales for cognitive and noncognitive ability at enlistment | no page yet |

Sample: Sweden, 1990-2017 (individual data) and 1997-2017 (firm data). 553 privatization events identified between 1997 and 2017. 63,231 matched worker pairs (70,079 treated workers, matched from 3.3 million non-privatized SOE workers). 368 matched privatized firms (368 matched control firms).

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13462) if you are: (i) designing privatization policy and want the cost-benefit arithmetic (Table IA.XVI) and the policy intervention discussion (Section V.D); (ii) studying the role of CEO governance changes in driving firm performance post-ownership-change (Table VI and Table IA.VII); (iii) replicating with the stacked DiD design, which requires the event-specific data sets and the matched control construction; (iv) extending the analysis to household finance and entrepreneurship outcomes (Tables III-IV and Section III.B-B.4); or (v) assessing heterogeneity by privatization type, buyer nationality, or macroeconomic conditions (Tables VII-IX).

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(4), August 2025. This distillation was extracted by an LLM on 2026-06-05 and is **not human-verified or independently reproduced**. The CC BY-NC 4.0 licence permits non-commercial reuse with attribution; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY-NC 4.0).** Olsson, Martin, and Joacim Tag.
> "What Is the Cost of Privatization for Workers?"
> *The Journal of Finance* 80, no. 4 (August 2025): 2107-2151.
> DOI: 10.1111/jofi.13462. (c) 2025 The Author(s).
> Licensed under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Green Window Dressing: Parise & Rubin (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/parise-green-window-dressing-2025/
# Distilled: ESG mutual funds strategically inflate their ESG factor loadings in the 10 days before mandatory portfolio disclosure, then revert to higher-return, lower-ESG holdings after filing. Three complementary tests (factor loadings, return gap, and stock-level CARs) all confirm the behavior, which boosts Morningstar sustainability ratings and attracts investor flows. J. Finance 2025, CC BY-NC 4.0. Eight core results with source locators, datasets used, the identification design, and the estimating specifications.
# Tags: paper-summary, esg, sustainable-investing, mutual-funds, window-dressing
==============================================================================

**What this is.** The paper's core results, the identification design, and the
estimating specifications: enough to know what it found and how, without reading
all 34 pages. To replicate or extend it, read the full source at the
[original](https://doi.org/10.1111/jofi.13499).

## TL;DR

ESG mutual fund managers inflate their portfolio ESG exposure in the 10 days
before mandatory quarterly disclosure, only to reverse those positions
afterward. The paper calls this "green window dressing." Three independent
empirical tests all point to the same behavior: (1) fund ESG betas (loadings on
the Morningstar U.S. Sustainability Total Return Index) rise by 0.12 in the
10 days before filing and revert to prior levels immediately after; (2) funds
earn a negative return gap of -1.4 bps per day before disclosure (underperforming
the portfolio they are about to report) but a positive gap of +1.0 bp per day
after (outperforming the just-disclosed portfolio); and (3) ESG stocks earn
cumulative abnormal returns of +0.20% in the three days before quarter-end
filings, which reverse completely afterward. Green window dressing positively
affects Morningstar sustainability ratings and attracts net fund flows,
particularly from institutional investors. The behavior is absent before March
2016 (when Morningstar introduced sustainability ratings) and is not present for
passive ESG vehicles, confirming active management is required.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%.
Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | ESG funds increase their ESG beta by 0.12 in the 10 days before mandatory disclosure, reverting to baseline immediately after | Table II, p. 3564 | Pre-disclosure: $$\overline{\Delta\beta}_{-}^{ESG}$$ = 0.120\*\*\* (SE 0.044) for 10-day window; post-disclosure: 0.026 (SE 0.050), not significant |
| R2 | The pre-disclosure ESG beta rise corresponds to a 46% increase over the baseline ESG exposure from the prior month (estimated at 0.26) | Table II footnote, p. 3565 | $$0.12 / 0.26 \approx 46\%$$; market beta falls by 0.107\*\*\* (SE 0.039) concurrently |
| R3 | Funds earn a negative pre-disclosure return gap of -1.4 bps/day, then a positive post-disclosure gap of +1.0 bp/day, with post-disclosure gap driven by lower ESG beta on realized vs. counterfactual returns | Table IV, p. 3570 | Pre: GAP = -0.014\*\*\* (SE 0.003); Post: +0.010\*\*\* (SE 0.004); post-disclosure ESG beta of actual returns (0.127) is 0.047\*\*\* lower than counterfactual ESG beta (0.174) |
| R4 | ESG stocks generate cumulative abnormal returns of +0.20% in the three days before fund portfolio disclosure, which reverse completely after | Figure 4, p. 3575 | CARs peak at roughly +0.20% at event time 0 in the [-3, +3] window; decline to near zero by +3 days; 95% CIs reported |
| R5 | A one-standard-deviation increase in pre-disclosure ESG exposure raises the probability of receiving a five-globe Morningstar rating by 2.1 pp (between funds) and 1.5 pp (within fund) | Table VI, p. 3579 | 2.1 pp (SE 0.009)\*\* without fund FE; 1.5 pp (SE 0.006)\*\*\* with fund FE; reduces probability of one-globe rating by 1.1 pp\*\* (without fund FE) |
| R6 | Green window dressers attract significantly more fund flows: a one-standard-deviation increase in pre-disclosure ESG exposure raises three-month net flows by 0.48 pp, driven entirely by institutional-client funds | Table VII, p. 3580 | All funds: 0.482\*\* (SE 0.201); Institutional: 0.653\*\* (SE 0.275); Retail: 0.152 (SE 0.358), not significant |
| R7 | High-fee funds are 6.7 pp more likely and low-fee funds are 5.2 pp less likely to window dress; star and laggard funds are each more likely to window dress than mid-ranked funds | Table V, p. 3577 | High fees: 0.067\*\*\* (SE 0.021); Low fees: -0.052\*\*\* (SE 0.013); Star fund: 0.027\*\* (SE 0.011); Laggard: 0.077\*\*\* (SE 0.015); PRI signatory: 0.035\*\* (SE 0.014) |
| R8 | Green window dressing is absent before March 2016 (no Morningstar sustainability ratings) and absent for ESG ETFs and index funds; ESG funds also reduce CO2 (pollution) exposure before disclosure | Internet Appendix Tables IA.I (col. 2-4), p. 3565 | Pre-2016 ESG beta change: not significant; ESG index funds: no significant increase; pollution factor beta also declines pre-disclosure (Table IA.X) |

**Overall (paper's conclusion).** Fund managers use quarterly disclosure as a
coordination device: by holding greener portfolios only when portfolios are
publicly observable, they can earn higher sustainability ratings and attract
more capital while holding less constrained, higher-yielding portfolios
between disclosures. Green window dressing does not generate CAPM alpha and
effectively delegates the ESG constraint to disclosure dates only.

## Theory / model

The paper has no formal theoretical model. It offers an economic rationale
based on two incentives for fund managers operating under imperfect monitoring.

The identification relies on the fact that Morningstar sustainability ratings
are determined primarily by disclosed quarterly portfolio holdings: a fund's ESG
profile is assessed based on the ESG scores of the stocks it holds at the
disclosure date. Because funds only need to disclose once per quarter, the
mandate to hold ESG stocks is effectively binding four times a year. This
asymmetric observability creates an incentive to hold ESG-aligned stocks
precisely when positions are observable and to substitute toward higher-yielding
non-ESG assets when they are not. Hartzmark and Sussman (2019) document that
investor flows strongly respond to Morningstar sustainability ratings, providing
the economic motive.

Prior work yields conflicting results. Muñoz, Ortiz, and Vicente (2022) compare
end-of-month to quarter-end holdings and find no window dressing, but the paper
shows their approach relies on the invalid assumption that voluntarily disclosed
month-end holdings are free of strategic manipulation. Kempf and Osthoff (2008)
find no green window dressing in the period before Morningstar sustainability
ratings; this paper reconciles that finding by showing that the behavior emerged
only after March 2016, when Morningstar introduced its globe ratings.

**Null hypothesis (eq. 4, p. 3562).** The identifying assumption is that in the
absence of green window dressing, ESG betas are stable from the pre-event
control window to the pre-event window:

$$
H_0: \; \Delta\beta_{-,i,e}^{ESG} \equiv \beta_{-,i,e}^{ESG} - \beta_{0,i,e}^{ESG} = 0, \quad \forall \; i = 1, \ldots, N \text{ and } e = 1, \ldots, E \tag{4}
$$

**Placebo and falsification evidence.** Three falsification tests validate the
design: (i) random disclosure dates produce no ESG beta increase (Internet
Appendix Table IA.1, col. 1); (ii) the period January 2010 to February 2016,
before Morningstar sustainability ratings, shows no significant effect (col. 2);
and (iii) passive ESG ETFs and index funds, which do not actively manage
portfolios, show no pre-disclosure ESG beta increase (cols. 3-4, p. 3565-3566).
Together these confirm the behavior is driven by active management responding
to the incentive created by Morningstar ratings.

## Method

The method applies three independent identification strategies to a sample of
223 U.S.-domiciled ESG active equity mutual funds over March 2016 to December
2022.

**Test 1: Factor-loading comparison around disclosure (Section I.B, pp. 3561-3563).**
This is the central design. The paper builds on `panel-regression` by estimating
fund-level two-factor regressions on symmetric windows around each mandatory
quarterly filing date $$t_e$$. The pre-disclosure regression (eq. 1, p. 3561):

$$
R_{i,t} = \alpha_{-,i,e} + \beta_{-,i,e}^{MKT} MKT_t + \beta_{-,i,e}^{ESG} ESG_t + \varepsilon_{i,t}, \quad t \in [t_e - n, \; t_e - 1] \tag{1}
$$

and the post-disclosure regression (eq. 2, p. 3561):

$$
R_{i,t} = \alpha_{+,i,e} + \beta_{+,i,e}^{MKT} MKT_t + \beta_{+,i,e}^{ESG} ESG_t + \varepsilon_{i,t}, \quad t \in [t_e + 2, \; t_e + n + 1] \tag{2}
$$

where $$R_{i,t}$$ is the daily return of fund $$i$$ on day $$t$$, $$MKT_t$$ is
the Kenneth French daily market return, and $$ESG_t$$ is the return on the
Morningstar U.S. Sustainability Total Return Index. Both regressions exclude the
disclosure date itself and the following trading day. The control window is the
entire second month of the fiscal quarter, excluding its first and last trading
days (eq. 3, p. 3562):

$$
R_{i,t} = \alpha_{0,i,e} + \beta_{0,i,e}^{MKT} MKT_t + \beta_{0,i,e}^{ESG} ESG_t + \varepsilon_{i,t}, \quad t \in \mathcal{T}_{0,e} \tag{3}
$$

The test statistic is the cross-fund, cross-event average change in ESG beta
(eq. 5, p. 3563):

$$
\overline{\Delta\hat{\beta}_{-}^{ESG}} \equiv \frac{1}{N \cdot E} \sum_{i=1}^{N} \sum_{e=1}^{E} \Delta\hat{\beta}_{-,i,e}^{ESG} \tag{5}
$$

where $$\Delta\hat{\beta}_{-,i,e}^{ESG} \equiv \hat{\beta}_{-,i,e}^{ESG} - \hat{\beta}_{0,i,e}^{ESG}$$. Inference uses a wild bootstrap (Internet Appendix Section I) to account for residual cross-sectional correlation.

The baseline uses $$n = 10$$ trading days; robustness uses $$n = 5$$ and $$n = 15$$.

**Test 2: Return gap (Section II.B, pp. 3569-3572).** Building on
Kacperczyk, Sialm, and Zheng (2008) and the portfolio-disclosure timing design
of Agarwal, Gay, and Ling (2014), the daily return gap for fund $$i$$ is
(eq. 7, p. 3570):

$$
GAP_{i,t} = R_{i,t} - \underbrace{\bigl(R_{i,t}^H - EXP_{i,t}\bigr)}_{R_{i,t}^C} \tag{7}
$$

where $$R_{i,t}$$ is the net-of-fees realized fund return, $$R_{i,t}^H$$ is the
return on a hypothetical buy-and-hold portfolio invested in the disclosed
quarter-end positions, and $$EXP_{i,t}$$ are fund fees. A fund that holds its
disclosed positions exactly earns $$GAP_{i,t} = 0$$; negative values indicate
pre-disclosure positioning away from what will be reported.

**Test 3: ESG stock event study (Section II.C, pp. 3574-3575).**
The paper estimates cumulative abnormal returns on ESG stocks around the four
annual filing dates (March 31, June 30, September 30, December 31). Abnormal
returns use a market model estimated on a 100-day window ending 50 days before
each event; CARs are computed in the [-3, +3] day window. The equally-weighted
portfolio covers all ESG-eligible stocks disclosed by sample funds.

**Trading cost estimation (Section II.A, pp. 3566-3569).** Trading costs are
measured using the Abdi and Ranaldo (2017) "CHL" two-day corrected effective
spread (eq. 6, p. 3566):

$$
\hat{\kappa}_{two\text{-}day,j,t} = \frac{1}{D_t} \sum_{d=1}^{D_t} \hat{\kappa}_{j,d}, \quad \hat{\kappa}_{j,d} = \sqrt{\max\!\bigl\{4(\text{cls}_{j,d} - \text{mid}_{j,d})(\text{cls}_{j,d} - \text{mid}_{j,d+1}),\, 0\bigr\}} \tag{6}
$$

High-ESG stocks have average effective spreads of 0.86% versus 0.98% for
non-high-ESG stocks, making them about 12% cheaper to trade.

## Empirical specifications

All main results use OLS on the panel of 223 ESG active equity mutual funds
over March 2016 to December 2022 (223 funds, 5,793 non-ESG funds available as
comparison group; Table I, p. 3560). The identifying variation is temporal
(within-fund, across-event), not cross-sectional.

**Factor-loading specification (R1, R2, R8).** Separate time-series regressions
(eqs. 1-3) are estimated per fund per event. The cross-event average change
(eq. 5) is the test statistic. Bootstrap inference accounts for cross-sectional
correlation. Sample: 4,063 fund-event observations (Panel A, Table II, p. 3564).
Standard errors in parentheses from the wild bootstrap.

**Return gap specification (R3).** OLS of $$GAP_{i,t}$$ on event-time indicators,
separately for the 10-day pre-disclosure window $$[-10, -1]$$ and post-disclosure
window $$[2, 11]$$. Bootstrapped standard errors. 23,320 fund-day observations.
ESG beta comparison: two-factor model on $$[2, 11]$$ post-disclosure window,
comparing betas on realized and counterfactual returns (2,332 observations,
Table IV, p. 3570).

**Rating specification (R5).** Linear probability model regressing $$I(\text{Five globes})_{t+2}$$ and $$I(\text{One globe})_{t+2}$$ on standardized $$\Delta\hat{\beta}_{-,i,t}^{ESG}$$:

- Column (1): time fixed effects only; column (2): time + fund fixed effects.
- Standard errors clustered at fund level. 2,656 observations (Table VI, p. 3579).

**Fund characteristics and window dressing propensity (R7).** Linear probability
model regressing $$\text{Window dresser}_{t+1}$$ on fund characteristics (fees,
past performance decile, size, disclosure frequency, PRI status, retail investor
share). Time fixed effects throughout. Standard errors clustered at fund level.
3,519 observations (Table V, p. 3577).

**Fund flow specification (R6).** OLS of three-month net flows (in %) on
standardized $$\Delta\hat{\beta}_{-,i}^{ESG}$$, fund size, family size, expense
ratio, and lagged performance. Time fixed effects. Split by retail vs. institutional
clientele (50% retail asset cutoff). 3,795 observations (Table VII, p. 3580).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CRSP mutual fund data (returns, portfolio holdings, TNA, turnover) | Fund returns, portfolio holdings, quarterly disclosure events, fund characteristics | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| Morningstar sustainability ratings and fund identifiers | ESG fund classification, five-globe sustainability rating outcome variable, fund category | No page yet |
| Morningstar U.S. Sustainability Total Return Index (MSEGUSTU) | ESG factor ($$ESG_t$$) in two-factor fund-return regressions | No page yet |
| Trucost CO2 emissions | Pollution factor (robustness: pollution exposure before disclosure, Table IA.X) | [Trucost](/wiki/commercial/trucost/) (licensed) |
| Kenneth French data library | Daily market factor ($$MKT_t$$), Fama-French factors in robustness tests | [Ken French library](/wiki/datasets/ken-french/) |
| SEC EDGAR N-PORT filings | Mandated quarterly portfolio disclosure dates | [SEC EDGAR](/wiki/datasets/edgar/) |

Sample: 223 ESG active equity mutual funds, March 2016 to December 2022. Comparison group: 5,793 non-ESG U.S. domestic equity mutual funds (Internet Appendix Section III). Portfolio holdings matched with CRSP daily stock returns for 89% of fund assets (median fund, p. 3570 fn.18).

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13499) if you are:
replicating the bootstrap inference procedure (Internet Appendix Section I);
applying trade imputation to recover intra-quarter holdings (Internet Appendix
Section II, using the Bongaerts, van Brakel, and van Dijk (2024) methodology);
testing for green window dressing in non-ESG funds (Internet Appendix Table
IA.XVII); studying the asymmetry in ESG beta changes for small vs. large funds
(Section IV.C, Table IA.XII); or studying broader ESG fund performance using
the Fama and French (1993) five-factor model (Internet Appendix Table IA.XVI).
Locators above point to the exact tables and figures.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(6). This distillation was
extracted by an LLM on 2026-06-03 and is **not human-verified or independently
reproduced**. The CC BY-NC 4.0 licence permits sharing and adaptation for
non-commercial purposes; the verbatim PDF is not hosted here.

> **Citation.** Parise, Gianpaolo, and Mirco Rubin. "Green Window Dressing."
> *The Journal of Finance* 80, no. 6 (December 2025): 3555-3588.
> DOI: 10.1111/jofi.13499. © 2025 The Author(s).
> Licensed under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/).
> This page is an extract by the Institute for Automated Research:
> core results and specifications summarized; **changes were made**.


==============================================================================
# Household Portfolios and Retirement Saving: Parker, Schoar, Cole & Simester (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/parker-household-portfolios-retirement-saving-2025/
# Distilled: Using account-level data on millions of U.S. middle-class investors from 2006 to 2018, this paper documents that equity shares rose 10 percentage points relative to the 1990s, became hump-shaped over the life cycle, and were driven largely by the Pension Protection Act of 2006 and the adoption of target date funds as default options. Retirement contribution rates, by contrast, changed little. J. Finance 2025, CC BY 4.0. Seven core results with source locators, datasets used, the identification design (difference-in-differences around PPA 2006), and the estimating equations.
# Tags: paper-summary, household-finance, life-cycle, retirement, pension, target-date-funds
==============================================================================

**What this is.** The paper's core results, the identification design exploiting the Pension
Protection Act of 2006 as a natural experiment, and the estimating equations: enough to know
what it found and how, without reading all 49 pages. To replicate or extend the analysis, read
the full source at the [original](https://doi.org/10.1111/jofi.13473).

## TL;DR

Using account-level data from a large U.S. financial institution on millions of middle-class
investors (2006 to 2018), the paper documents three facts. First, middle-class investors hold
roughly 71% of their investable wealth in equities, about 10 percentage points more than in the
1990s. Second, the life-cycle profile of equity shares is now hump-shaped: investors increase
their stock allocation from age 25 to around 50 and then reduce it as they approach retirement,
a pattern absent before 2000. Third, retirement contribution rates have increased steadily with
age by about 4% to 5% over working lives and are stable across cohorts. The PPA of 2006, which
permitted target date funds (TDFs) as qualified default investment alternatives, is shown to
have caused the portfolio changes, particularly for younger and lower-income workers, while
having little persistent effect on contribution rates. The savings behavior documented here is
broadly consistent with prescriptive life-cycle portfolio models, unlike the 1990s patterns.

## Core results

Magnitudes and significance are as reported; locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Average equity share of investable wealth is 71%, 10pp above the 1990s, and inversely related to age in the cross-section** | Table III Panel A, p. 2753; Figure 5, p. 2757 | Mean equity share 71.0% (median 77.3%) in 2016; cross-sectional age gradient from ~74% at ages 25-27 to ~55% at ages 64-65; SCF shows only 54.5% (underreporting argued) |
| R2 | **Within-person, equity shares follow a hump-shaped life-cycle: up ~7% from age 25 to 50, down ~7% from 50 to 65**; in the 1990s the profile was flat or upward-sloping | Figure 6, p. 2759; Figure 9 bottom panel, p. 2763 | Person fixed effects regression shows monotone rise then fall; 1990s comparison (Ameriks and Zeldes (2004)) shows opposite pattern; McKenzie (2006) double-differencing confirms the hump |
| R3 | **Younger cohorts hold higher equity shares at every overlapping age**; each cohort born later has 15-20pp more in equities by 2018 than those born around 1945 | Figure 7, p. 2760; Figure 8, p. 2761 | Monotone cohort shift in equity shares for births before age 40; cohorts with high initial TDF share show steeper hump-shaped profiles |
| R4 | **TDF adoption as QDIA (PPA 2006) raises equity share of young new enrollees by ~5.5pp (age 25-35) and ~6pp for low-income workers** | Table IV col. (1) and (3), pp. 2766-2767 | Treated coefficient 0.0552 (SE 0.0907) for full sample; 0.0599 bottom income tercile; <2% for top income tercile; firm fixed effects DiD, two-year window |
| R5 | **Same TDF default reduces equity share of older new enrollees by ~13pp (age 55-65)**, consistent with TDF glidepath moving near-retirees out of stocks | Table IV col. (1), pp. 2766-2767 | Interaction Age 55 to 65 x Treatment: -0.1325 (SE 0.0012); effect is robust to controlling for income and restricting to those with no prior retirement assets |
| R6 | **Medium-run (5-year) effect: treated young workers' 5-year coefficient 0.0134 (~1.3pp above control); treated older workers' 5-year coefficient -0.0564 (~5.6pp below control)**, declining from year-1 peak/trough as control group converges voluntarily | Table V col. (4) and (7), pp. 2770-2771 | Year-0 through Year-5 treatment interaction coefficients trace declining but persistent gap; paper text approximates these as "~3%" (young) and "~2%" (old), likely based on Figure 10 predicted-equity-share comparisons |
| R7 | **Retirement contribution rates rise ~4-5pp over working lives and are stable across cohorts; PPA had a transitory negative effect of -0.4% to -1.2% on contribution rates, fading to near zero after five years** | Figure 11, pp. 2774-2775; Table VI col. (1), p. 2778; Table VII col. (1), p. 2780 | Realized contribution rate increases from ~4.6% at age 25 to ~8.5% at age 65 (cross-section); within-person increase ~5%; PPA DiD coefficient -0.0043 at year of treatment, converges to ~0 by year 5 |

**Overall (paper's conclusion).** The rise of TDFs following the PPA of 2006 caused middle-class
American investors to hold more equity earlier in their careers and to progressively de-risk as
they approach retirement, bringing the life-cycle portfolio profile closer to prescriptive
life-cycle models. This rebalancing occurred largely via portfolio composition changes, not
changes in saving rates. The PPA provisions intended to raise contribution rates had little
lasting effect, leaving open whether regulatory design of retirement plans is an effective tool
for increasing saving quantities.

## Theory / model

The paper has no formal utility or equilibrium model; it instead tests hypotheses derived from
the life-cycle portfolio-choice literature against administrative panel data. The paper's
identification logic and the hypotheses it tests are as follows.

**Life-cycle portfolio hypotheses.** The paper is motivated by Campbell (2016), who surveys how
financial product design and regulation can improve household financial well-being.
Classical models (Merton (1969), Samuelson (1969)) imply
constant portfolio allocations in scale-invariant settings. Richer models with non-tradable
human capital that declines with age (Viceira (2001), Heaton and Lucas (2000), Campbell and
Viceira (2002), Gomes, Michaelides, and Zhang (2020)) recommend reducing risky asset holdings
over the working life. TDFs embed exactly this prescription: a glidepath that holds ~90% in
equity roughly 20 years before retirement and reduces to 40-50% at target retirement date.
The paper's first hypothesis is that current life-cycle profiles of equity shares are broadly
consistent with these prescriptions, in contrast to the 1990s patterns documented by Ameriks
and Zeldes (2004).

**PPA identification assumption.** The Pension Protection Act of 2006 permitted TDFs as
Qualified Default Investment Alternatives (QDIAs) in employer-sponsored defined-contribution
plans. The key assumption is that employees' employment decisions were not affected by whether
their employer adopted a TDF as QDIA: workers were typically unaware of plan feature changes
at the time of hiring, chose jobs on many other dimensions, and employer selection of a TDF
default was driven by liability concerns (the safe harbor provision), not worker characteristics
(pp. 2764-2765). Workers at the same employer enrolling in the two years before the switch serve
as the control group; those enrolling in the two years after serve as the treated group.
The design follows the tradition of Madrian and Shea (2001), who showed that default investment
options have large effects on 401(k) choices. Mitchell and Utkus (2022) use Vanguard data to
document similar TDF effects; this paper extends that work using a different large institution
and focusing on life-cycle patterns and causal PPA effects.

**Contribution rate hypotheses.** Optimal saving models (Gomes et al. (2018), Poterba (2014))
predict significant shortfalls in U.S. retirement saving. The paper tests whether the PPA's
provisions intended to raise saving (autoenrollment, autoescalation) had measurable effects on
the realized rate at which employees saved. Broader aggregate implications of TDF growth are
analyzed in Parker, Schoar and Sun (2023).

## Method

The main empirical strategy is a difference-in-differences (DiD) design exploiting the Pension
Protection Act of 2006. The two estimating equations are given in Sections III.A and III.B of
the paper.

**Short-run DiD (equation 2, p. 2765).** For workers starting a new job between 2005 and 2008
and observed for the first two years, the specification is:

$$
y_{ift} = \beta_1 \times D_{\text{treated}} + \beta_2 \times D_{\text{treated}} \times \text{AgeEnrolled}_i
+ \beta_3 \times \text{AgeEnrolled}_i + \lambda_f + \epsilon_{ift}, \tag{2}
$$

where $$y_{ift}$$ is the portfolio equity share (or reported contribution rate) of individual
$$i$$ starting at firm $$f$$ in year $$t = 0$$; $$D_{\text{treated}}$$ equals one if the worker
enrolled after the firm adopted a TDF as its default (in 2007 or 2008); $$\text{AgeEnrolled}_i$$
is a vector of 10-year age-group indicators at enrollment, included both alone and interacted
with the treatment to capture the age-heterogeneous TDF glidepath effect; and $$\lambda_f$$ is
a firm fixed effect so the comparison is within-employer across the two cohorts. Standard errors
are clustered at the household level.

**Medium-run DiD (equation 3, p. 2769).** To track dynamics over five years after enrollment:

$$
y_{ift} = \beta_1 \times D_{\text{treated}} + \beta_2 \times D_{\text{treated}} \times \text{AgeEnrolled}_i
+ \beta_3 \times \text{AgeEnrolled}_i
+ \beta_4 \times D_{\text{treated}} \times \lambda_t + \beta_5 \times \lambda_t + \lambda_f + \epsilon_{ift}, \tag{3}
$$

where $$\lambda_t$$ are year fixed effects and their interaction with treatment captures how the
treatment gap evolves year by year. Separate regressions by age group replace the interacted
$$\text{AgeEnrolled}$$ terms.

**Life-cycle descriptive regressions (equation 1, p. 2756).** Cross-sectional and within-person
regressions of equity share on age-group indicators document the cross-sectional age profile and
the within-cohort life-cycle hump:

$$
y_{it} = \boldsymbol{\beta}_1' \mathbf{Age}_{it} + \beta_2 \, Inc_{it} + \epsilon_{it}, \tag{1}
$$

where $$\mathbf{Age}_{it}$$ is a vector of three-year age-group indicators and $$Inc_{it}$$ is
the log deviation of the individual's income from the sample mean income. Person fixed effects
are added for the within-person version (Figure 6, Figure 11 Panel B). The McKenzie (2006)
double-differencing method is also applied as a robustness check to recover pure age effects
free of cohort and time effects (Internet Appendix Section II).

## Empirical specifications

**Descriptive life-cycle profiles (R1, R2, R3).** The full RI sample of investors aged 25-65
observed annually from 2006 to 2018 (millions of individual-year observations). The cross-
sectional regression (eq. 1) uses age-group dummies with and without a log-income control;
the within-person version adds person fixed effects. Cohort analysis splits by 10-year birth
cohorts (Figure 7). For R2, the McKenzie (2006) approach differences within-cohort over
adjacent ages, then differences again over time, to eliminate cohort and time effects and
identify the second partial derivative of the age profile (Internet Appendix II). Standard
errors are clustered at the individual level (and at the employer level in robustness checks,
Internet Appendix Table IA.XI).

**Short-run PPA DiD (R4, R5).** Sample restricted to workers who started a new job 2005-2008,
observed for two years. Treatment defined as enrollment in 2007 or 2008 at an employer that
switched its default to a TDF. Control: enrollment in 2005 or 2006 at the same employer.
Firm fixed effects absorb all employer-level differences. The key identifying variation is the
timing of the plan-level switch relative to enrollment date. Columns (3)-(4) of Table IV
repeat the analysis by income tercile; columns (5)-(6) restrict to those with no prior
retirement assets at the institution. The portfolio equity share is the dependent variable
for R4 and R5; the reported contribution rate for R7.

**Medium-run PPA DiD (R6, R7).** Same base sample, extended to track individuals for five
years after enrollment. Year-of-treatment and year-after interactions (eq. 3) trace convergence
between treated and control. Columns (4) and (7) of Table V split by age at enrollment (25-34
vs. 55-65); columns (4) and (7) of Table VII do the same for contribution rates. Standard
errors are clustered at the household level throughout.

**Robustness checks.** (i) Restricting the contribution-rate PPA sample to 2007 enrollees
only (Table IA.XXIV) to reduce financial-crisis confounds. (ii) Using price-constant equity
shares that ignore passive price appreciation (Internet Appendix Table IA.IX). (iii) Repeating
DiD with standard errors clustered at the employer (Table IA.XVII). (iv) The ex-ante designated
equity share as an alternative equity-share measure (Table IA.X).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Proprietary account-level administrative data from a large U.S. financial services company | Primary data: individual portfolios, contribution rates, demographics, employer identifier, 2006-2018 | No page yet |
| Survey of Consumer Finances (SCF), 2016 wave | Comparison sample of U.S. retirement investors (RIs) for calibration and validation of representativeness | No page yet |
| Administrative data from Ameriks and Zeldes (2004) | 1990s comparison for life-cycle equity share patterns | No page yet |

The paper's headline results are derived entirely from proprietary account-level data provided by
a single unnamed financial services firm. Reproducing the results requires access to that
confidential dataset, which is not publicly available.

Sample scope: investors aged 25-65 with retirement wealth in the middle 80% of the age-adjusted
distribution (retirement investors, RIs), observed annually from December 2006 to December 2018.
Sample covers millions of individuals and trillions of dollars in investable wealth (Table II,
p. 2750).

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13473) if you are:
studying the design of 401(k) plans and the life-cycle portfolio implications of TDF defaults;
calibrating or testing life-cycle portfolio-choice or saving models against administrative
microdata; investigating the effects of the Pension Protection Act of 2006 on investor behavior;
or interested in the representativeness of SCF survey data relative to administrative data on
portfolio equity shares. The Internet Appendix (available on the Journal of Finance website)
contains 25+ robustness tables, estimation moments, and the McKenzie (2006) double-differencing
reconstruction.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(5). This distillation was extracted by an
LLM on 2026-06-05 and is **not human-verified or independently reproduced**. The CC BY 4.0
licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Parker, Jonathan A., Antoinette Schoar, Allison Cole, and
> Duncan Simester. "Household Portfolios and Retirement Saving over the Life Cycle."
> *The Journal of Finance* 80, no. 5 (October 2025): 2739-2787.
> DOI: 10.1111/jofi.13473. (C) 2025 The Author(s).
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Designing Stress Scenarios: Parlatore & Philippon (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/parlatore-designing-stress-scenarios-2025/
# Distilled: Parlatore and Philippon model the optimal design of bank stress test scenarios as an information-acquisition problem, solving it via a Kalman filter. Capital requirements cover losses under an adverse scenario while targeted interventions depend on covariances among residual exposures; calibration shows information is far more valuable for targeted interventions than for broad capital requirements. J. Finance 2025, paywalled. Five core results with source locators, the model, and the method.
# Tags: paper-summary, banking, stress-testing, financial-regulation, prudential-regulation
==============================================================================

**What this is.** The paper's core results, the model (a Kalman filter learning problem with linear-quadratic regulator preferences), and the method it contributes (optimal stress scenario design as information acquisition) with defining equations: enough to know what it found and how, without reading all 42 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13422).

## TL;DR

Parlatore and Philippon develop the first formal theory of how a financial regulator should design stress test scenarios. They model stress testing as a two-stage process: a risk-discovery stage (the regulator learns banks' hidden exposures from reported losses under hypothetical scenarios) and a risk-mitigation stage (the regulator chooses capital requirements or targeted interventions). By mapping the scenario-design problem into a Kalman filter information-acquisition problem, they derive optimal scenarios and show that their design depends critically on what the regulator plans to do with the information. Information from stress tests is only modestly valuable for setting broad capital requirements (worth about 20 bps of welfare gain), but is four to five times more valuable when the regulator can make targeted interventions such as loan-to-value limits or supervisory matters requiring attention. The paper also calibrates the model to U.S. bank data (DFAST 2015, 1991-2013 quarterly NCO rates) and finds that optimal scenarios focus on factors with correlated exposures across banks and have a hump-shaped dependence on prior mean exposures.

## Core results

Magnitudes and significance as reported; `\*` = 10%, `\*\*` = 5%, `\*\*\*` = 1%. Locators point to the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Optimal capital requirements are linear in expected losses under distress and set to cover the adverse scenario | Lemma 2, eq. (20), p. 848 | $$\overline{W}^* = \sum_i \mathbb{E}[y_i \mid \text{Distress}, \mathcal{S}] + \tilde{W} - \frac{\kappa - p\theta}{p\gamma}$$; requirements increase in risk aversion and estimated exposures |
| R2 | Welfare gain from learning under pure capital requirements is modest: ~20 bps, significantly below the gain from a 10% reduction in capital costs | Figure 4, p. 861 | Stress tests distinguishing adverse from severely adverse scenario raise welfare about 20 bps; this is less than one-quarter of the gain from a 10% lower capital cost |
| R3 | Welfare gains from stress testing with targeted interventions are 4-5x larger than under pure capital requirements | Figure 4, p. 861-862 | At high prior uncertainty, targeted interventions raise welfare gains to the same order of magnitude as a 10% decrease in the cost of bank capital; robust across calibrations |
| R4 | Optimal scenario weight on a factor is hump-shaped in the regulator's prior mean exposure | Section IV.C.1, p. 855; Figure 2 Panel A | Weight increases in expected exposure at low values (intervention more likely and information more valuable) but decreases at high values (posterior anchored to prior, reducing learning value) |
| R5 | Optimal scenario stresses factors with correlated exposures across banks more; if correlation is high enough, only the correlated factor is stressed | Section IV.C.2, p. 856; Figure 2 Panel B | Cross-bank correlated factors are more systemic and receive outsized scenario weight; specialization may be complete when correlation is sufficiently high |

**Overall (paper's conclusion).** Stress tests are best understood as an information tool whose value depends on what the regulator plans to do. When limited to setting broad capital requirements, scenarios should match the plausible adverse state and scenario design matters little for learning. When targeted interventions are available, optimal scenarios deviate from the average bad state to elicit information about specific exposures, and their design can generate welfare gains comparable to a significant reduction in the cost of bank capital. This is consistent with the sequential recapitalization characterization of Orlov, Zryumov, and Skrzypacz (2023).

## Theory / model

The model builds on the information-acquisition framework of Van Nieuwerburgh and Veldkamp (2010), the disclosure literature of Goldstein and Leitner (2018), and the stress-test design context studied by Faria-e-Castro, Martinez, and Philippon (2017) and Shapiro and Skeie (2015). It has three stages: scenario design, stress testing, and intervention. There are $$N$$ banks indexed by $$i \in \{1,\dots,N\}$$, each exposed to $$J$$ macroeconomic risk factors gathered in the $$J \times 1$$ state vector $$\mathbf{s}$$ with $$\mathbb{E}[\mathbf{s}] = 0$$ (p. 838). Bank $$i$$'s cumulative losses in state $$\mathbf{s}$$ are (eq. 2, p. 838):

$$
y_i(\mathbf{s}) = \sum_{j=1}^{J} x_{i,j} s_j, \tag{2}
$$

where $$x_{i,j}$$ is bank $$i$$'s unobserved exposure to factor $$j$$. Exposures are stacked in the $$NJ \times 1$$ vector $$\mathbf{x}$$. The aggregate banking system capital is (eq. 3, p. 839):

$$
W(\mathbf{s}) \equiv \sum_{i=1}^{N} w_i(\mathbf{s}) = \overline{W} - \sum_{i=1}^{N} y_i(\mathbf{s}). \tag{3}
$$

The regulator has prior beliefs $$\mathbf{x} \sim \mathcal{N}(\overline{\mathbf{x}}, \Sigma_x)$$ over the $$NJ \times 1$$ vector of exposures (p. 840). A stress scenario $$\hat{\mathbf{s}} = (\hat{s}_1,\dots,\hat{s}_J)'$$ is a realization of the state vector (Definition 1, p. 840). A stress test is a collection of $$M$$ scenarios $$\{\hat{\mathbf{s}}^m\}_{m=1}^M$$ and reported losses $$\{\hat{y}_i^m\}$$ (Definition 2, p. 841). Bank $$i$$'s estimated loss under scenario $$\hat{\mathbf{s}}^m$$ is (eq. 5, p. 841):

$$
\hat{y}_i(\hat{\mathbf{s}}^m, M) = \hat{\mathbf{s}}^m \cdot \mathbf{x}_i + \hat{\varepsilon}_{i,m}(\|\hat{\mathbf{s}}\|, M), \tag{5}
$$

where $$\hat{\varepsilon}_i$$ captures measurement error whose variance increases in scenario severity $$\|\hat{\mathbf{s}}\|$$. In state-space form (eq. 9, p. 844):

$$
\hat{\mathbf{y}} = \hat{\mathbf{S}} \mathbf{x} + \hat{\boldsymbol{\varepsilon}}, \tag{9}
$$

where $$\hat{\mathbf{S}} \equiv (\mathbf{I}_N \otimes \hat{S})$$ stacks the scenario matrix across banks.

The regulator has pseudo mean-variance (linear-quadratic) preferences over aggregate banking system wealth $$W$$. With probability $$p$$ the economy lands in a distress region around $$\tilde{W}$$, and the regulator uses a second-order approximation of marginal utility (eq. 15, p. 847):

$$
U'(W) = \begin{cases} 1 & \text{with probability } 1-p \\ 1 + \theta - \gamma(W - \tilde{W}) & \text{with probability } p, \end{cases} \tag{15}
$$

where $$\theta \equiv U'(\tilde{W}) - 1 > 0$$ and $$\gamma \equiv -U''(\tilde{W}) > 0$$. The regulator designs $$\hat{S}$$ to maximize ex-ante expected utility given that she will choose optimal actions $$(\overline{W}, \mathbf{a})$$ after observing stress test results (eq. 8, p. 843):

$$
\mathbb{E}_{\mathcal{S}}\left[\mathbb{E}\!\left[U\!\left(W\!\left(\mathbf{s},\mathbf{x};\mathbf{a}^*(\mathcal{S}),\overline{W}^*(\mathcal{S})\right)\right)\!\Big|\mathcal{S}\right] - \mathcal{C}(\mathbf{a}^*(\mathcal{S})) - \mathcal{K}\!\left(\overline{W}^*(\mathcal{S})\right)\right]. \tag{8}
$$

## Method

**Learning via the Kalman filter.** The model's key analytical insight is that stress test results $$\hat{\mathbf{y}}$$ are signals about latent exposures $$\mathbf{x}$$ in the linear Gaussian system (eq. 9). By Lemma 1 (p. 844), the posterior beliefs are:

$$
\mathbf{x} \mid \hat{\mathbf{y}} \sim \mathcal{N}(\hat{\mathbf{x}}, \hat{\Sigma}_{\mathbf{x}}), \tag{Lemma 1}
$$

with posterior mean $$\hat{\mathbf{x}}$$ and posterior covariance $$\hat{\Sigma}_{\mathbf{x}}$$ given by (eqs. 10-11, p. 844):

$$
\hat{\mathbf{x}} = \left(\mathbf{I}_{NJ} - K\hat{\mathbf{S}}\right)\overline{\mathbf{x}} + K\hat{\mathbf{y}} \quad \text{and} \quad \hat{\Sigma}_{\mathbf{x}} = \Sigma_x - K\hat{\mathbf{S}}\Sigma_x, \tag{10-11}
$$

where the $$NJ \times MN$$ Kalman gain matrix is $$K = \Sigma_x \hat{\mathbf{S}}' \left(\hat{\mathbf{S}}\Sigma_x\hat{\mathbf{S}}' + \Sigma_\varepsilon\right)^{-1}$$. The expected learning is $$\Sigma_{\hat{x}} \equiv \Sigma_x - \hat{\Sigma}_x = K\hat{\mathbf{S}}\Sigma_x$$ (eq. 12, p. 845).

**Optimal interventions.** Under Assumption L (linear costs: $$\mathcal{K}(\overline{W}) = (1+\kappa)\overline{W}$$ and $$\mathcal{C}(\mathbf{a}) = \Phi'\mathbf{a}$$), optimal capital requirements are given by Lemma 2 (eq. 20, p. 848):

$$
\overline{W}^* = \sum_{i=1}^{N} \mathbb{E}[y_i \mid \text{Distress}, \mathcal{S}] + \tilde{W} - \frac{\kappa - p\theta}{p\gamma}. \tag{20}
$$

Optimal targeted interventions are given by (eq. 22, p. 850):

$$
\mathbf{a}^* = (p\gamma\tilde{\mathbb{V}})^{-1}\!\left(\kappa(\mathbf{1}_{N\times 1} \otimes \tilde{\mathbf{s}}) \circ \hat{\mathbf{x}} - \Phi + p\gamma\tilde{\mathbb{V}}\mathbf{1}_{NJ\times 1}\right), \tag{22}
$$

where $$\tilde{\mathbb{V}} \equiv \mathbb{C}\mathbb{O}\mathbb{V}[(\mathbf{1}_N \otimes \mathbf{s}) \circ \mathbf{x} \mid \mathcal{S}, \mathcal{D}=1]$$ is the distress uncertainty matrix (eq. 23, p. 850). The distress uncertainty matrix $$\tilde{\mathbb{V}}$$ decomposes into two terms: uncertainty about the macro state under distress interacted with expected exposures, and residual uncertainty about bank exposures from the Kalman filter.

**Optimal scenario design.** The scenario design problem for pure capital requirements reduces to (Lemma 3, eq. 25, p. 852):

$$
\min_{\hat{\Sigma}_{\mathbf{x}} \in \Sigma} \mathbf{1}_{1\times NJ} \mathbb{E}[\tilde{\mathbb{V}}] \mathbf{1}_{NJ\times 1}, \tag{25}
$$

that is, the regulator minimizes residual uncertainty, choosing the posterior covariance matrix $$\hat{\Sigma}_{\mathbf{x}}$$ in the feasible set $$\Sigma$$ implied by the Kalman filter. With targeted interventions, by Lemma 4, the design problem is (eq. 28, p. 853):

$$
\min_{\hat{\Sigma}_{\mathbf{x}} \in \Sigma} \mathbb{E}_{\hat{\mathbf{x}}}\!\left[\kappa\overline{W}^* + \Phi'\mathbf{a}^*\right], \tag{28}
$$

which trades off capital costs against the cost of targeted actions, weighted by intervention responsiveness to new information.

## Empirical specifications

The paper's quantitative results come from calibration of the model to U.S. bank-level and macroeconomic data, not from panel regressions. The calibration targets moments from the 2015 Dodd-Frank Act Stress Test (DFAST) and quarterly bank NCO rates from 1991 to 2013 (Table I, p. 856; Table II, p. 857; Appendix D, p. 870-871).

**Step 1: Identifying macro risk factors.** Following Hirtle et al. (2014), the authors regress aggregate banking system NCO rates on standard macroeconomic variables (GDP growth, short-term and long-term interest rates, unemployment, housing prices, equity prices, and credit spreads) over 1991-2013 at the quarterly frequency. GDP growth and a real estate price index (equal-weighted average of residential and commercial) explain more than 80% of the variation in NCO rates. The two risk factors are standardized to have mean zero and unit variance (Table D.1, p. 871).

**Step 2: Calibrating exposure priors.** The regressor coefficients are approximately $$-0.357$$ (GDP) and $$-0.303$$ (real estate). The asymptotic variances of the coefficient estimates ($$0.0037$$ and $$0.0035$$ respectively) are used to calibrate the regulator's prior uncertainty. Prior mean exposures are set to $$\bar{x}_1 = \bar{x}_2 = 0.015$$ and prior standard deviation $$\Sigma_x^{1/2} = \text{diag}(0.006, 0.006)$$.

**Step 3: Calibrating preferences.** The adverse scenario corresponds to $$\tilde{W} = 10\%$$ of RWA and $$\overline{W}^* = 13\%$$. Distress probability $$p = 0.1$$, marginal cost of capital $$\kappa = 0.3$$ (matching shadow-cost estimates of Kisin and Manela (2016)), marginal value of capital in distress $$\theta = 3$$, and risk-aversion curvature $$\gamma = 100$$.

**Step 4: Calibrating measurement error.** The standard deviation of bank model errors $$\sigma_\varepsilon = \alpha + \beta\|\mathbf{s}\|^2$$ is estimated by regressing the CLASS model's cross-sectional forecasting errors against the squared norm of macro factors, giving $$\alpha = 0.55\%$$, $$\beta = 0.11\%$$.

**Welfare comparisons (Figure 4, p. 861).** The authors solve four problems: (1) two scenarios freely chosen optimally, (2) one scenario fixed at the adverse state, one chosen optimally, (3) one fixed at adverse, two chosen optimally, and (4) two freely chosen optimally with targeted interventions. Welfare gains are normalized by the gain from a 10% reduction in the cost of bank capital ($$\kappa' = 0.9\kappa$$). Problems 1-3 yield welfare gains less than one-quarter of this benchmark; Problem 4 yields gains of the same order of magnitude.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| DFAST (Dodd-Frank Act Stress Test) 2015 summary statistics | Calibration targets: Tier 1 capital ratio (ex-ante 13.5%, adverse 10.4%, severely adverse 8.4%), Tier 1 leverage, loan loss rate (Table I, p. 856) | No page yet |
| Quarterly aggregate U.S. bank NCO rates (1991-2013) | Identifying macro risk factors and calibrating exposure priors; regression of NCO rate on GDP growth, real estate prices (Appendix D, p. 870-871) | No page yet |
| Capital Loss Assessment under Stress Scenarios (CLASS) model, Hirtle et al. (2014) | Bank-level panel regressions to calibrate regulator priors and measurement error (Table D.1, p. 871) | No page yet |

Sample: quarterly, 1991 Q1 to 2013 Q4 for the macro calibration; one representative bank (aggregate U.S. banking system) plus DFAST cross-section.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13422) if you are: designing stress test frameworks and want the formal characterization of optimal scenario choice (Section IV, pp. 851-856); interested in the calibration details and the four welfare comparison problems (Section V, pp. 856-862); extending the model to multiple banks, trading losses, interest rate risk, or contagion; or reading the proofs of Lemmas 1-4 in Appendices A-C (pp. 863-867).

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(2), April 2025, pp. 833-873. DOI: 10.1111/jofi.13422. Published under the Wiley VOR licence (paywalled, not CC). This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. Extract-only: the verbatim PDF is not hosted here.

> Parlatore, Cecilia, and Thomas Philippon. "Designing Stress Scenarios."
> *The Journal of Finance* 80, no. 2 (April 2025): 833-873.
> DOI: 10.1111/jofi.13422. © 2025 the American Finance Association.


==============================================================================
# Presidential Address: Housing Betas: Piazzesi (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/piazzesi-presidential-address-housing-betas-2025/
# Distilled: Housing betas (the stock-market beta of housing returns) were negative before the Global Financial Crisis and turned positive after it, despite highly correlated cashflow growth rates across the two asset classes. A heterogeneous-agent model with segmented and credit-connected markets explains the pre-GFC puzzle. J. Finance 2025, paywalled. Eight core results with source locators, the model (OLG segmented-markets Lucas tree), and the propositions on negative comovement.
# Tags: paper-summary, asset-pricing, housing, macro, real-estate, comovement
==============================================================================

**What this is.** The paper's core results, the stylized facts on housing betas,
and the segmented-markets model that explains the pre-GFC puzzle: enough to know
what it found and how, without reading all 34 pages. To replicate or extend it,
read the full source at the [original](https://doi.org/10.1111/jofi.70000).

## TL;DR

The paper documents that housing returns had a *negative* stock-market beta from
the postwar period until the Global Financial Crisis, and a *positive* beta
thereafter, while the cashflow growth rates of stocks and housing comove
positively throughout. This "housing beta puzzle" is a challenge for
representative-agent models, which predict positive return comovement for assets
with similar cashflows. A two-type OLG model with segmented stock and housing
markets connected through a collateralized bond market resolves the puzzle:
homebuyers are poorer per unit of their future housing dividends than
stockholders, so they borrow using houses as collateral. Aggregate bad news
triggers either a credit demand channel (unconstrained homebuyers cut borrowing
sharply while stockholders cut supply only modestly, raising house prices and
cutting stock prices) or a credit supply channel (constrained homebuyers are
pinned at the collateral limit while stockholders flee to safety, expanding
credit supply, cutting the interest rate, and raising house prices as stock
prices fall). Both channels produce negative return comovement. Post-GFC shifts
(wealthier homebuyers, institutional housing investors, foreign Treasury demand)
weaken credit and market segmentation, moving the economy toward positive
comovement.

## Core results

Magnitudes are as reported; locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Housing betas (slope from regression of 10-year housing returns on 10-year stock returns, two-sided exponential decay kernel) were **mostly negative or zero before the GFC** and turned strongly positive after it | Figure 2, p. 3109 | Housing beta: roughly -0.2 to 0 in 1960s-1990s; peaks near +1 around 2020; levered housing beta even more negative pre-GFC (below -1 during 1980s-1990s) |
| R2 | **Cashflow growth rates of stocks and housing comove positively** throughout the sample, with correlation 80% pre-GFC and 59% for the full sample | Figure 3, p. 3110 | Correlation of 10-year real cashflow growth rates: 80% for 1930-GFC subsample, 59% full sample 1939-2024 |
| R3 | **Levered housing return has an even more negative beta** before the GFC than unlevered housing, because borrowing costs are lower during housing booms | Figure 2, p. 3109; Figure 5, p. 3113 | Levered housing beta reaches below -1 during the 1980s-1990s (Figure 2, green line); Panel A of Figure 5 shows real 10-year Treasury returns are negative during the 1950s, 1970s, and 2020s when house prices rise |
| R4 | Homebuyers' **idiosyncratic housing risk is large**: cross-sectional standard deviation of annualized capital gains spans -40% to +60% per year in San Francisco and similar volatility in Huntsville, AL | Figure 6, p. 3114 | Distribution is right-skewed and more dispersed post-GFC; average capital gain 6.7%/yr (San Francisco) vs 40 bp (Huntsville) over the last decade per CoreLogic data |
| R5 | **Proposition 1** (p. 3121): if homebuyers have lower savings per unit of cashflows than stockholders (Assumption A), homebuyers borrow from stockholders in equilibrium; credit market connects the two segmented markets | p. 3121 | Qualitative; follows from $$f(0) > 0$$ under Assumption A: excess credit demand is positive at the no-credit benchmark |
| R6 | **Proposition 2** (p. 3122): bad news about the aggregate economy generates negative return comovement. (i) Unconstrained homebuyers: higher uncertainty $$\sigma^2$$ raises the stock price-dividend ratio and lowers the housing price-dividend ratio, with credit declining; (ii) constrained homebuyers: bad news shifts credit supply down (flight to safety by stockholders), increasing credit, raising house prices, and lowering stock prices | p. 3122 | Qualitative proposition; mechanism is the asymmetric sensitivity of levered homebuyers vs. unleveraged stockholders to aggregate uncertainty |
| R7 | **Proposition 3** (p. 3125): forces that weaken the credit channel generate positive comovement. Equal savings-to-dividends ratios across types eliminate credit and negative comovement (part i); scaling household savings by $$\lambda > 1$$ raises both price-dividend ratios (part ii); population growth $$n$$ reduces the effective discount rate and raises both price-dividend ratios (part iii) | p. 3125-3126 | Qualitative; wealthier homebuyers post-GFC, demographic aging, institutional and foreign investors are the candidate empirical counterparts |
| R8 | **In the 2000s housing boom**, negative comovement is consistent with the data: house prices rose while stock returns declined (Figure 1, p. 3108); both the credit demand channel (laxer collateral, buyer optimism expanding credit demand) and a credit supply shifter (securitization, subprime expansion) were active | Figure 1, p. 3108; §III.B, p. 3124 | Annual data 1955-2024 show the shaded house-price boom episodes (1970s, 2000s, 2020s) coincide with stock market slumps |

**Overall (paper's conclusion).** Representative-agent models predict positive
return comovement for assets with similar cashflows; the data show the opposite
before the GFC. A segmented-markets model with collateralized credit explains
the puzzle. Post-GFC positive comovement reflects weakened credit and
segmentation from demographics, institutional investors, and foreign capital.

## Theory / model

The paper builds on a two-asset Lucas-tree benchmark and then extends it to a
heterogeneous-agent, segmented-markets economy. The motivation for studying
stocks and housing jointly draws on Piazzesi and Schneider (2016), who survey
the housing and macroeconomics literature. Landvoigt, Piazzesi and Schneider
(2015) develop a quantitative predecessor model for the San Diego housing
market with segmented markets and credit. Piazzesi and Schneider (2007) model
momentum traders in housing with a search framework. Piazzesi and Schneider
(2008) document the inflation illusion and credit channel for the 1970s
housing boom and stock slump across 12 OECD countries. Jorda, Schularick and
Taylor (2019) document a Sharpe ratio near one for aggregate housing returns
in U.S. postwar data, used here to contextualize idiosyncratic risk.
Mankiw and Weil (1989) linked Baby Boom demographics to the 1970s housing
surge, a candidate for post-GFC comovement via higher savings. Iacoviello
(2005) studies house prices and borrowing constraints in a monetary business
cycle model, cited as a quantitative single-market predecessor.

**Benchmark: representative-agent Lucas trees (Section I.C, p. 3109).**
Aggregate output $$Y_t$$ grows at log rate (equation 1, p. 3109):

$$
g_t := \log Y_t - \log Y_{t-1}, \quad g_t \stackrel{\text{iid}}{\sim} N(\mu, \sigma^2). \tag{1}
$$

The representative agent has log utility $$\sum_{t=0}^\infty \beta^t \log C_t$$.
Two long-lived assets pay cashflows $$D_t^s$$ (stocks) and $$D_t^h$$ (housing); the
Euler equation for asset $$i = s, h$$ is (equation 2, p. 3110):

$$
P_t^i = E_t\!\left[\beta \frac{C_t}{C_{t+1}}\bigl(D_{t+1}^i + P_{t+1}^i\bigr)\right]. \tag{2}
$$

Since both assets share the same cashflow growth, the model implies perfectly
positively comovving returns, in contradiction to Figure 1. Time-varying
discount rates (incorporating Cochrane (2011), Bansal and Yaron (2004), and
habit formation) generate positive comovement in valuation ratios and therefore
do not resolve the puzzle (p. 3111).

**Segmented-markets OLG model (Section III, p. 3115).** There are two types of
households: fraction $$\delta_s$$ trade only stocks plus bonds; fraction $$\delta_h$$
trade only housing plus bonds, with $$\delta_s + \delta_h = 1$$. In an OLG
framework, young households of type $$i$$ receive labor income $$w_{i,t} Y_t$$, save
their entire income, and consume only when old. Stocks and housing trade only
within type. Each young household solves (equation 3, p. 3116):

$$
\max E[\log c_{i,t+1}] \quad \text{s.t.} \quad p_{i,t}\theta_{i,t} + b_{i,t} = w_{i,t}, \tag{3}
$$
$$
c_{i,t+1} = (p_{i,t+1} + d_{i,t+1})\theta_{i,t} + b_{i,t} R,
$$
$$
\phi\, p_{i,t+1}\theta_{i,t} \geq -R\, b_{i,t},
$$

where $$\theta_{i,t}$$ is the share of the Lucas tree purchased, $$b_{i,t}$$ is bond
holdings, $$R$$ is the gross interest rate, and the last inequality is the
collateral constraint (the loan cannot exceed fraction $$\phi$$ of the future asset
value). Rewriting in portfolio weights $$\alpha_{i,t} = p_{i,t}\theta_{i,t}/w_{i,t}$$
(equation 4, p. 3117):

$$
\max_{\alpha_{i,t}\leq R/(R-\phi)} E_t\!\left[\log\!\left(\tilde{R}_{t+1}^i\, \alpha_{i,t} + R(1-\alpha_{i,t})\right)\right]. \tag{4}
$$

The optimal portfolio weight (equation 5, p. 3117) is:

$$
\alpha_{i,t} \approx \min\!\left\{\frac{\frac{1}{v^i} + \mu - r}{\sigma^2},\; \frac{R}{R-\phi}\right\}. \tag{5}
$$

When the collateral constraint does not bind, the weight equals the Merton
weight (Campbell and Viceira (1999)). The **Gordon Growth formula** for the
price-dividend ratio (equation 7, p. 3118) is:

$$
v_i = \frac{1}{r - \mu + \sigma^2 \alpha_i}, \tag{7}
$$

where $$\sigma^2 \alpha_i$$ is the risk premium. The **credit supply function**
(equation 8, p. 3118) is obtained from the stock Gordon Growth formula:

$$
r_s(Q) = \frac{d_s}{\delta_s w_s - Q} + \mu - \sigma^2 \frac{\delta_s w_s - Q}{\delta_s w_s}, \tag{8}
$$

and the **credit demand function** from the housing Gordon Growth formula
(equation 9, p. 3119) is:

$$
r_h(Q) = \frac{d_h}{\delta_h w_h + Q} + \mu - \sigma^2 \frac{\delta_h w_h + Q}{\delta_h w_h}. \tag{9}
$$

Equilibrium credit $$Q^*$$ solves $$f(Q) = r_h(Q) - r_s(Q) = 0$$ (equation 10,
p. 3119). When homebuyers hit the collateral constraint, credit demand becomes
(equation 11, p. 3120):

$$
r_h(Q) = \phi\, \frac{\delta_h w_h + Q}{Q} - 1. \tag{11}
$$

**Assumption A** (p. 3120): homebuyers have lower savings per unit of cashflows
than stockholders: $$\delta_h w_h / d_h < \delta_s w_s / d_s$$.

**Proposition 1** (p. 3121): if Assumption A holds, homebuyers borrow from
stockholders in equilibrium ($$Q^* > 0$$).

**Proposition 2** (p. 3122): under Assumption A, bad news about the whole
economy generates negative return comovement in housing and stock price-dividend
ratios, through either a credit demand channel (part i, unconstrained
homebuyers) or a credit supply channel (part ii, constrained homebuyers).

**Proposition 3** (p. 3125): forces that reduce credit and segmentation move
comovement toward positive: equal savings-to-dividends ratios, higher overall
savings (part ii), or population growth (part iii) all raise both price-dividend
ratios together.

## Method

This is primarily a theoretical paper. The model is solved analytically in
steady state. Key steps:

1. Impose constant price-dividend ratios in steady state (all shares of
   aggregate output are constant: cashflows, labor income).
2. Derive the Gordon Growth formulas (equations 7-9) for each asset class
   from the market-clearing conditions and the optimal portfolio weights.
3. Derive the credit supply function from stockholders' Gordon Growth formula
   and the credit demand function from homebuyers' Gordon Growth formula;
   show these are respectively upward- and downward-sloping in credit $$Q$$
   (Figure 7, p. 3119).
4. Establish existence and uniqueness of equilibrium using the intermediate
   value theorem and monotonicity of the excess demand function $$f(Q)$$
   (Appendix A, pp. 3129).
5. Characterize comparative statics via the implicit function theorem, with
   formal proofs in Appendices B-D (pp. 3129-3133).

The paper builds on `overlapping-generations` (two-period OLG with constant
savings rates) and `life-cycle-model` (portfolio choice over a finite
horizon). The housing beta is measured empirically using a two-sided
exponential-decay kernel regression of 10-year real housing returns on a
constant and 10-year real stock returns (10% weight on observations five years
in the past or future), as described on p. 3108 and displayed in Figure 2.

## Empirical specifications

There is no formal econometric specification in the traditional sense; the
empirical contribution is measurement and stylized facts. The key estimating
procedure is:

**Housing beta (R1, R3, Figure 2):** The stock-market beta of housing is
estimated as the slope from a local regression:

$$
\tilde{r}_{t}^h = a + \beta_t^h\, \tilde{r}_{t}^s + \varepsilon_t,
$$

where $$\tilde{r}_t^h$$ and $$\tilde{r}_t^s$$ are real 10-year geometric mean returns
on housing and stocks, the regression is estimated with a two-sided exponential
decay kernel (10% weight at horizon 5 years), and $$\beta_t^h$$ is the
time-varying slope. Sample: annual, 1955-2024 (Figure 2, p. 3109).

**Cashflow growth correlation (R2, Figure 3):** The correlation of real
10-year cashflow growth rates on stocks (S&P 500 dividends) and housing (NIPA
housing services expenditure) is computed over the same annual sample 1930-2024.
The 10-year rates are geometric means reported per year (p. 3107).

**Levered housing return (R3, Figure 5):** Levered housing return equals the
housing return minus borrowing costs, with borrowing costs proxied by real
10-year Treasury returns scaled by 80% (reflecting the typical down payment)
to match the leverage of mortgage-financed housing (p. 3107, p. 3113).

**Idiosyncratic capital gains (R4, Figure 6):** Cross-sectional distribution
of idiosyncratic capital gains on individual houses in San Francisco, CA and
Huntsville, AL from CoreLogic individual transaction data. Idiosyncratic gain
= individual house gain minus the location-specific average gain over the
same holding period (p. 3114).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| S&P 500 index (Shiller website: shillerdata.com, file ie_data.xls) | Real 10-year stock returns and dividend growth (cashflows); end-of-year values 1930-2024 | [Shiller data](/wiki/datasets/shiller-data/) |
| Financial Accounts of the United States (Fed, table B101/B104) | Value of residential real estate held by households; capital gains computed net of residential fixed investment (table F6); annual 1946-2024 | no page yet |
| NIPA (BEA, tables 2.4.4 line 50, 2.4.5 lines 25 and 47) | Housing cashflows (dollar expenditure on housing services, including imputed rents); price index for nondurables and services; annual 1929-2024 | [NIPA / FRED](/wiki/datasets/fred/) |
| 10-year Treasury returns (Shiller website) | Proxy for mortgage borrowing costs (multiplied by 0.8 for typical down payment); annual 1940-2024 | [Shiller data](/wiki/datasets/shiller-data/) |
| American Housing Survey (HUD) | Average homeowner tenure (15.1 yr in 1980, 11.5 yr in 2021), share of buyers with mortgage, share of first-time buyers, mortgage payment as share of income (Figure 4, p. 3112) | no page yet |
| CoreLogic individual transaction data | Cross-sectional distribution of idiosyncratic capital gains on individual houses in San Francisco, CA and Huntsville, AL (Figure 6, p. 3114) | [CoreLogic](/wiki/commercial/corelogic/) (licensed) |

Sample: primarily annual, U.S., 1930-2024 (returns and cashflows); housing leverage data 1975-2022 (Figure 4).

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.70000) if you are: studying
why housing and stock returns move in opposite directions before 2008; building
models of segmented housing and stock markets connected through credit; analyzing
how demographics, institutional investors, or foreign capital affect the
stock-housing return relationship; or extending the theoretical channels
(defaultable debt, idiosyncratic risk, multi-cohort OLG) identified on
p. 3128. The appendices (pp. 3129-3133) contain formal proofs of all
propositions.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(6). This distillation was
extracted by an LLM on 2026-06-03 and is **not human-verified or independently
reproduced**. The paper is paywalled (Wiley VOR terms); extract-only.

> Piazzesi, Monika. "Presidential Address: Housing Betas."
> *The Journal of Finance* 80, no. 6 (December 2025): 3103-3136.
> DOI: 10.1111/jofi.70000. © 2025 the American Finance Association.
> Paywalled; this page contains only extracted summary and analysis.


==============================================================================
# Simplicity and Risk: Puri (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/puri-simplicity-risk-2025/
# Distilled: This paper introduces and axiomatizes a preference for simplicity in choice under risk, showing that participants' measured risk aversion and dominance violations increase with lottery complexity (number of outcomes), holding moments fixed, and that no canonical behavioral theory fully captures this. J. Finance 2025, paywalled. Six core results with source locators, the simplicity representation model with axioms, and the experimental design.
# Tags: paper-summary, behavioral-finance, decision-theory, risk-preferences
==============================================================================

**What this is.** The paper's core results, the simplicity representation model with its axioms, and the experimental design: enough to know what it found and how, without reading all 52 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13417).

## TL;DR

The paper introduces a preference for simplicity in choice under risk: people value having fewer outcomes in a lottery above and beyond what moments capture. In a preregistered online experiment (n = 582, Amazon Mechanical Turk, September to October 2022), participants provided certainty equivalents for lotteries with 2, 4, 8, and 16 outcomes, held similar in mean, variance, skewness, and range. Estimated CRRA risk aversion nearly triples from two to eight outcomes and nearly quadruples from two to sixteen outcomes. Dominance violations increase with complexity. An axiom characterizing simplicity (differences-in-differences complexity aversion) is satisfied by 84 to 90 percent of participants. None of the canonical theories tested (CPT, PT, salience, sparsity, rational inattention, cognitive noise, varying probability weighting) fully captures these patterns. The paper also axiomatizes and characterizes the simplicity representation and generalizes it to capture obfuscation, computation, and language effects.

## Core results

Magnitudes and significance are as reported; `\*` = 5%/10% level. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Measured CRRA risk aversion increases with lottery complexity**, holding moments fixed | Figure 1, p. 1030; §I.C, p. 1037 | CRRA alpha: 0.67 [0.38, 1.00] at 2 outcomes; 1.38 [1.05, 1.71] at 4; 3.21 [2.68, 3.84] at 8; 5.93 [5.06, 7.07] at 16 outcomes |
| R2 | **Dominance violations increase with lottery complexity** | Figure 2, p. 1040; §I.E | 3.9% at 2 outcomes; 4.4% at 4; 5.2% at 8; 6.4% at 16 outcomes; increase from 2 to 16 marginally significant (p = 0.08) |
| R3 | **Characterizing axiom satisfied by 84-90% of participants** (differences-in-differences complexity aversion) | Table I, p. 1039; §I.D | 50.4% satisfy axiom in 3/3 pairs (2 vs 4), 52.5% (2 vs 8), 59.4% (2 vs 16); 57-59% strictly satisfy for at least 1/3 pair; significantly above 50% random rate (p < 0.01) |
| R4 | **Complexity aversion is monotonic and economically large**: average c - c' spread exceeds twice the study's average relative risk premium for 2 vs 16 | §I.D, p. 1039 | \$0.005 for 2 vs 4 outcomes; \$0.012 for 2 vs 8; \$0.146 for 2 vs 16; difference between 2 vs 4 and 2 vs 8 significant at p < 0.01 |
| R5 | **Cognitive ability and complexity aversion are negatively correlated**: low-cognitive-ability individuals show much larger increases in measured risk aversion as complexity increases | Figure 3, p. 1042; §I.F | CRRA gap (low vs high cognitive ability): 0.86 at 2 outcomes (p < 0.1); 2.02 at 8 outcomes (p < 0.05); 3.83 at 16 outcomes (p < 0.05) |
| R6 | **Risk aversion and complexity aversion are largely separable**: once complexity costs are explicitly controlled, estimated CRRA alpha is stable across complexity levels | §I.G, p. 1043 | With C(n) controlled: alpha = 0.11 (SE 0.07) at 8 outcomes, 0.21 (SE 0.17) at 16; numerically similar and not statistically distinguishable; far below the 2.7-point gap when complexity is ignored |

**Overall (paper's conclusion).** People have preferences for simplicity distinct from other drivers of choice under risk. Observed risk aversion increases with lottery complexity, and dominance violations rise with complexity. Measured cognitive ability and complexity aversion are negatively correlated at higher complexity levels. None of the canonical behavioral-finance models tested (Tversky and Kahneman (1992) CPT; Kahneman and Tversky (1979) PT; Bordalo, Gennaioli, and Shleifer (2012) salience; Sims (2003) / Woodford (2012) rational inattention/cognitive noise; Gabaix (2014) sparsity) fully captures simplicity-seeking behavior. The simplicity representation is axiomatized and generalizes to nonrisk settings including computation and language complexity. Goodman and Puri (2023) and Fudenberg and Puri (2022a) are companion papers applying this framework to binary options and heterogeneous-agent settings respectively.

## Theory / model

The model has no formal model of asset markets; instead, it axiomatizes a preference over money lotteries. The underlying space is simple money lotteries $$\Delta\mathbb{R}$$. A preference $$\succeq$$ on $$\Delta\mathbb{R}$$ is assumed to be transitive, complete, and defined on simple lotteries (no compound lotteries in the space; Definition 1, p. 1034).

**Definition 1 (Simplicity Representation, p. 1034).** A preference $$\succeq$$ on money lotteries $$\Delta\mathbb{R}$$ is said to have a **simplicity representation** if, for any two lotteries $$p, q$$,

$$
p \succ q \iff \sum u(x)p(x) - C(\#p) > \sum u(x)q(x) - C(\#q), \tag{1}
$$

where $$u$$ is a continuous, strictly increasing, unbounded below Bernoulli utility, $$\#p$$ is the number of outcomes in lottery $$p$$, and $$C : \mathbb{N} \to \mathbb{R}$$ is a weakly increasing complexity cost function. This is an as-if representation and makes no claims about mechanism or welfare.

**Characterizing Axiom: Differences-in-Differences Complexity Aversion (Axiom 1, p. 1038).** The main axiom states that, when comparing a larger and smaller support lottery, the larger support lottery's appeal should increase by more when complexity differences are removed. Formally, consider lotteries $$p, q \in \Delta X$$, $$\#p \geq \#q$$, with certainty equivalents (CEs) $$\delta_p, \delta_q$$. For any $$\alpha \in (0, 1)$$ and any lottery $$r$$ whose support includes all outcomes in $$p, q, \delta_p$$, and $$\delta_q$$:

$$
\gamma\!\left(\tfrac{1}{2}p + \tfrac{1}{2}\delta_q\right) + (1-\gamma)r \succeq \gamma\!\left(\tfrac{1}{2}q + \tfrac{1}{2}\delta_p\right) + (1-\gamma)r. \tag{2}
$$

Mixing equalizes the complexity of both sides; the axiom says the more complex lottery $$p$$ benefits more from this equalization. It is a joint test of complexity aversion and separability.

**Theorem 1 (Characterization, p. 1048).** A preference $$\succeq$$ on $$\Delta\mathbb{R}$$ admits a Simplicity Representation if and only if it satisfies:
- Axiom 1 (Differences-in-differences complexity aversion)
- Axiom 2 (Monotonicity): $$x > y \Rightarrow \delta_x \succ \delta_y$$
- Axiom 3 (Same-Support Independence and Continuity): for lotteries with the same support, EU-type mixing axioms hold
- Axiom 4 (Singleton Continuity)
- Axiom 5 (Singleton Unboundedness)

The proof proceeds in three steps (p. 1048): (i) fix support $$Z$$ and apply Herstein and Milnor (1953) results to obtain EU restricted to same-support lotteries; (ii) use the same-support axiom to show the Bernoulli utility is support-independent ($$u_Z = u_{Z'}$$ for all $$Z, Z'$$); (iii) construct $$C$$ via an iterative algorithm using Axiom 1 to show complexity cost is well-defined independent of the specific lottery used.

**Uniqueness (Proposition 1, p. 1049).** The representation is unique up to an affine transformation: if $$(u, C)$$ and $$(u', C')$$ are two simplicity representations, then there exist $$\xi > 0, \beta \in \mathbb{R}$$ such that $$u' = \xi u + \beta$$ and $$C' = \xi C$$.

**Generalization (Definition 3, p. 1050).** The paper generalizes to cognitive tiers $$\mathbb{T} \subseteq \mathbb{R}$$: a **general simplicity representation** replaces $$\#p$$ with $$\text{Tier}(p)$$, capturing obfuscation, computation, and language effects. The representation is:

$$
p \succ q \iff \sum u(x)p(x) - C(\text{Tier}(p)) > \sum u(x)q(x) - C(\text{Tier}(q)). \tag{3}
$$

## Method

The experiment uses a preregistered design (AEA RCT Registry Trial ID 10136) run on Amazon Mechanical Turk in September to October 2022. There are two modules:

**Risk-aversion module (§I.C, p. 1037).** Participants provide certainty equivalents (CEs) for 32 lotteries using the standard multiple price list (MPL) procedure with enforced single switching (Holt and Laury 2002). Each MPL has six evenly spaced choices from 50 cents below the lottery's lowest payoff to the lottery's highest payoff. Participants are randomly assigned to this module or the axiom module. Lotteries have $$N \in \{2, 4, 8, 16\}$$ outcomes (8 lotteries per complexity level), generated to have similar means, variances, skewness, and ranges across complexity levels (Appendix A.1).

The CRRA utility is $$U(p) = \sum p(x) \frac{x^{1-\alpha}}{1-\alpha}$$, $$\alpha \geq 0$$. The econometric model follows Bruhin, Epper, and Fehr-Duda (2010): observed CE for individual $$i$$ on lottery $$l$$ is

$$
ce_{i,l} = \hat{ce}_l(\theta) + \epsilon_{i,l}, \quad \epsilon_{i,l} \sim N(0, \sigma_{i,l}), \quad \sigma_{i,l} = \xi_i |\text{range}(l)|,
$$

where each individual has their own error variance term $$\xi_i$$. Estimated using maximum likelihood via expectation-maximization (Appendix A.3, p. 1056). Confidence intervals via bootstrap with 1,000 iterations.

**Axiom module (§I.D, p. 1038).** Participants test Axiom 1 (differences-in-differences complexity aversion) via a five-step procedure: elicit CEs for two lotteries, construct a mixed lottery equating their complexity, elicit CEs for both mixed versions, and test whether $$c \geq c'$$. Three pairs each of 2 vs 4, 2 vs 8, and 2 vs 16 outcome lotteries.

**Cognitive ability.** Participants complete Raven's Advanced Progressive Matrices (APM), Set I, at end of survey. High vs low cognitive ability split at median Raven's score (10/12 in both modules). Paid \$0.25 per correct Raven's question.

**Data quality.** Workers in the US with at least 95% approval rating and at least 100 prior completed tasks; three comprehension checks; average payment \$9.33 for 23 minutes. Final sample: 582 participants (48% female, 54% below age 40, 60% college educated).

## Empirical specifications

The headline empirical results all come from the preregistered experimental design. There are no panel regressions with fixed effects; the primary inferential procedure is structural CRRA estimation by complexity level plus OLS checks on residuals of alternative models.

**R1: CRRA estimation by complexity level.** For each number of outcomes $$n \in \{2, 4, 8, 16\}$$, fit the CRRA model to the CEs of the $$n$$-outcome lotteries only, pooling individuals. Standard errors clustered by individual; bootstrap confidence intervals. Report $$\alpha$$ separately for each $$n$$ and test whether $$\alpha$$ is constant across $$n$$; the increase from 2 to 16 is the main finding.

**R2: Dominance violations.** Code dominance violation as 1 if the participant's stated CE for lottery $$l$$ is strictly less than the lowest possible outcome of $$l$$. Regress violation indicator on number of outcomes, clustering standard errors by individual. Separately, regress on log(number of outcomes). Marginal significance at $$p = 0.08$$ for the increase from 2 to 16.

**R3: Axiom satisfaction rate.** For each pair of 2 vs $$n$$-outcome lotteries, code whether the participant satisfies $$c \geq c'$$ (Axiom 1 test). Compute fraction of participants satisfying axiom in 0/3, 1/3, 2/3, 3/3 pairs. Test whether fraction satisfying axiom for $$\geq 2/3$$ pairs differs from 50% (random-clicking) and from 100% (expected utility), using a two-sided binomial test at $$p < 0.01$$.

**R5: Cognitive ability heterogeneity.** Split sample at median Raven's score. Re-estimate CRRA $$\alpha$$ separately by cognitive ability group for each complexity level. Test whether the difference between low and high groups is statistically significant at $$p < 0.1$$ and $$p < 0.05$$.

**R6: Separability test.** From the axiom module, recover dollar utility spreads $$c - c'$$ for pairs at $$n = 8$$ and $$n = 16$$ vs $$n = 2$$. Convert to complexity cost differences $$C(8) - C(2)$$ and $$C(16) - C(2)$$ (Appendix A.8). Use maximum likelihood to jointly estimate $$C(2)$$ and $$\alpha$$ controlling for complexity cost. If separable, estimated $$\alpha$$ should be stable across $$n$$; report $$\alpha$$ and $$C(2)$$ for $$n = 8$$ and $$n = 16$$.

**Alternative theory tests (§II, pp. 1043-1047).** For CPT, PT, salience, rational inattention, sparsity, and cognitive noise: fit the model to the risk-aversion module data, extract residuals, and regress residuals on number of outcomes. A positive and significant slope rejects the alternative. Also test whether the probability weighting parameter varies significantly by complexity (it does not). These tests are described fully in §II and appendices A.9-A.13.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Author-collected experimental data (AMT, preregistered, September to October 2022) | Certainty equivalents and axiom-module choices for 582 participants across 2/4/8/16-outcome lotteries | No page yet |

Sample: 582 US participants on Amazon Mechanical Turk, September to October 2022. Two modules: risk-aversion (32 lotteries, 8 per complexity level) and axiom module (12 lotteries, 3 per complexity level).

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13417) if you are: (i) axiomatizing or extending simplicity preferences to new settings (language, computation, obfuscation: §III.C, Definitions 2-3); (ii) applying simplicity theory to financial product design, portfolio choice, or incentive schemes (§IV); (iii) seeking the full suite of alternative-theory tests against CPT, PT, salience, sparsity, rational inattention, and cognitive noise (§II and Appendices A.9-A.13); or (iv) looking for the formal proofs and generalized representation theorems (Internet Appendix IA.A). The locators above point to the exact figures and tables.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(2), April 2025. This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The paper is paywalled (Wiley VOR licence; not CC); only extracts reproduced here under fair use.

> Puri, Indira. "Simplicity and Risk." *The Journal of Finance* 80, no. 2 (April 2025): 1029-1080. DOI: 10.1111/jofi.13417. © 2024 the American Finance Association. Paywalled; extract-only.


==============================================================================
# The Actual Retail Price of Equity Trades: Schwarz, Barber, Huang, Jorion & Odean (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/schwarz-actual-retail-price-equity-2025/
# Distilled: A controlled trading experiment across six brokerage accounts at five brokers finds that mean account-level round-trip costs range from 7 to 46 basis points for identical simultaneous market orders, and that the entire cross-broker execution difference is attributable to market centers giving systematically different execution to different brokers for the same trades, not to broker venue-routing choices or payment for order flow. J. Finance 2025, CC BY-NC 4.0. Six core results with source locators, datasets used, the empirical design, and the regression specifications.
# Tags: paper-summary, market-microstructure, payment-for-order-flow, retail-trading
==============================================================================

**What this is.** The paper's core results, the trading experiment design, and the regression
specifications that identify the source of price execution variation across brokers: enough to
know what was found and how, without reading all 35 pages. To replicate or extend, read the
full source at the [original](https://doi.org/10.1111/jofi.13467).

## TL;DR

The paper runs a controlled trading experiment opening six brokerage accounts at five major
U.S. retail brokers (TD Ameritrade, Fidelity, E\*Trade, Robinhood, Interactive Brokers Pro
and Lite) and placing approximately 85,000 simultaneous identical market orders in 128 stocks
over six months (December 2021 to June 2022). Mean price improvement (PI) relative to the
National Best Bid and Offer (NBBO) ranges from 47% of the NBBO spread (TD Ameritrade) to 19%
(IBKR Lite), and average round-trip costs range from 7 basis points to 46 basis points. The
key finding is that all of this variation is attributable to the same market centers (off-exchange
wholesalers) giving systematically different execution to different brokers for identical trades,
not to broker routing decisions or payment for order flow (PFOF). PFOF, whose variation across
brokers is an order of magnitude smaller than PI variation, explains almost none of the
cross-broker execution differences.

## Core results

Magnitudes and significance are as reported; `**` = 1%, `*` = 5%. Locators point into the
source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Cross-broker PI ranges from 19% to 47% of NBBO**, with mean round-trip costs from 7 to 46 bps; off-exchange execution averages 31% of NBBO vs. 17% for exchange trades | Table IV, p. 2527; Figure 1, p. 2510 | TD: PI\% = 47.2%, cost = -7.2 bps; FD: 35.8%, -19.7 bps; ET: 36.1%, -23.4 bps; RH: 26.8%, -31.4 bps; IB Lite: 19.5%, -44.3 bps; IB Pro: 18.8%, -46.2 bps |
| R2 | **Round-trip cost differences are economically large**: TD-to-worst-broker gap is 39 bps; 1 bp of aggregate retail trading cost equals ~$2.8 billion annually | Table IV, p. 2527; p. 2529 | Round-trip costs: TD -7.2 bps, IBKR Pro -46.2 bps; midpoint-execution benchmark = 0 bps; NBBO execution = -61.9 bps |
| R3 | **All pairwise broker differences are statistically significant** at the 1% level (except FD vs. ET, insignificant); execution quality varies highly significantly across all six accounts | Table V, p. 2529 | TD-RH diff = 20.5 pp (t=63.10\*\*); TD-IBKR Pro = 28.5 pp (t=47.36\*\*); FD-ET = -1.4 pp (t=-1.29, n.s.) |
| R4 | **The same market center (venue) gives systematically different execution to different brokers** for the exact same trade; routing venue choice does not explain the execution gap | Table VII, p. 2532 | TD vs. RH at same venue: PI diff = 22.5 pp; at different venues: PI diff = 22.5 pp; identical across venue conditions for all broker pairs |
| R5 | **OIB (order flow toxicity) can generate variation consistent with observed execution differences**, but broker fixed effects remain after controlling for OIB and venue/stock fixed effects | Table VIII, p. 2536 | Broker-FE model: R2 = 15.5%; OIB adds 10 pp; FE on RH vs. TD baseline = -0.204\*\* (robust to venue and stock FEs) |
| R6 | **PFOF does not explain cross-broker PI variation**: the rank order of execution quality is unrelated to PFOF levels, and PFOF per share is ~1/30th the size of PI per share | Figure 4, p. 2535; Table IV, p. 2527 | PFOF range: $0.000-$0.215 cents/share; PI range: $2.78-$7.84 cents/share; Fidelity (PFOF=0) has better execution than Robinhood (PFOF=$0.215/share) |

**Overall (paper's conclusion).** Retail investors trading "free" through commission-zero brokers
face real execution costs that differ by as much as 39 basis points per round trip across
identical simultaneous orders. The source of this dispersion is not broker routing but
wholesaler-level pricing: the same off-exchange market centers give different execution to
different brokers for the same trades. Payment for order flow explains almost none of the
variation. Better disclosure of execution quality at the broker level, particularly through
an expansion of SEC Rule 605 and Rule 606 reporting, is the proposed remedy.

## Theory / model

The paper has no formal model. The theoretical framework draws on three mechanisms from the
market microstructure literature:

**Adverse selection and order toxicity (Kyle (1985), Battalio and Holden (2001)).** Off-exchange
wholesalers profit from uninformed retail order flow. If order flow from a given broker is more
"toxic" (more directional, more informed), wholesalers face higher inventory risk and protect
themselves by offering lower PI. The paper tests whether observed OIB variation can account for
the PI variation across brokers.

**Inventory management (Stoll (1978)).** Dealers face costs from absorbing order imbalances. High
one-sided order flow from a broker creates inventory risk, leading to higher effective spreads.

**Competitive incentives and scale.** Wholesalers compete for order flow and cater to broker
objectives. Brokers with larger order-flow volumes (e.g., TD Ameritrade, which has more than
double the daily average trades of other brokers in the experiment) may receive better execution
because wholesalers compete more aggressively for their flow.

**Prior experimental evidence.** Bakos et al. (2005) ran an earlier experiment at three brokers
in 1999 and found no difference in overall PI, though total trading costs (including commissions)
differed. Kothari, So & Johnson (2021) study Robinhood trades in TAQ and report average PI of
33% NBBO for small trades using the actual known trade direction. Levy (2022) compares 1,000
trades at Robinhood and TD Ameritrade, finding a similar ranking to this paper. Eaton et al. (2022)
provide independent evidence that broker-level order flow toxicity affects market quality using
Robinhood platform outages as a natural experiment.

**Identification strategy.** The paper explicitly trades simultaneously across all broker accounts
in the same stock at the same time, so any latency difference is controlled by construction (and
verified empirically). This design isolates broker-level treatment effects from stock and time
variation. The within-venue test (Table VII) separates venue-routing from within-venue execution
differences.

## Method

The core identification relies on a **controlled trading experiment** rather than observational
data.

**Stock selection.** The CRSP universe (4,037 names as of June 2021) is stratified into 128 bins
by market capitalization, liquidity (share turnover), volatility, and price, with one stock
randomly selected per bin. Stocks with share price below $1 are excluded. The paper also
includes four high-retail-interest stocks (AMC, Tesla, Nio, Aurora Cannabis), several mega-caps
(Apple, Bank of America, NVIDIA, ExxonMobil, Google, Visa), and the top four Robinhood "mover"
stocks each week (pp. 2521-2522).

**Trade execution.** For API-accessible brokers (TD Ameritrade, Robinhood, E\*Trade, IBKR Pro),
trades are placed programmatically, randomizing submission order across buy and sell sides to
remove latency bias. For Fidelity and IBKR Lite, trades are placed manually in parallel.
Target order size is $100 per trade (full shares rounded to nearest whole share); robustness
checks use $1,000 and $5,000 targets. Positions are closed within 30 minutes. The experiment
runs from December 21, 2021 to June 9, 2022 (113 trading days). After filters (TAQ
matchability, 2-second simultaneity window, price $>$1), the final sample is 74,801 trades
(pp. 2522-2524).

**Price improvement (PI) measures** (equations 1a-b, p. 2518):

$$
\text{PI\$}_{\text{buy}} = \text{NBO} - P \quad \text{and} \quad \text{PI\$}_{\text{sell}} = P - \text{NBB} \tag{1a}
$$

$$
\text{PI\%} = \frac{\text{PI\$}}{\text{NBBO Spread}} \tag{1b}
$$

**Effective spread** (equations 2a-b, p. 2518):

$$
\text{ES\$}_{\text{buy}} = 2 \times (P - P_{\text{mid}}) \quad \text{and} \quad \text{ES\$}_{\text{sell}} = 2 \times (P_{\text{mid}} - P) \tag{2a}
$$

$$
\text{ES\%} = \frac{\text{ES\$}}{\text{NBBO Spread}} \tag{2b}
$$

where $$P$$ is the execution price, NBO (NBB) is the national best offer (bid), and $$P_{\text{mid}}$$
is the NBBO midpoint. Round-trip transaction costs are computed using midpoint prices at entry
and exit as the benchmark.

**Routing-venue data.** Under SEC Rule 606(b)(1), brokers must provide customer-specific routing
data upon request. The paper requests and obtains these data, identifying each trade's execution
venue, enabling the within-venue test in Table VII (p. 2532).

## Empirical specifications

**Pairwise broker comparison (R1-R3).** The headline result compares unconditional means of
PI\% across broker accounts (Table IV, p. 2527). Statistical significance is assessed via
Table V (pairwise differences), which reports mean PI\% differences for matched parallel
trades with standard errors clustered by stock:

$$
\text{PI\%}_{i,b} - \text{PI\%}_{i,b'} \quad \text{for each matched trade pair } (b, b') \tag{}
$$

The two-sample test uses only matched (simultaneous) trades between each broker pair.

**Same-venue vs. different-venue decomposition (R4).** Table VII splits all pairwise matched
trades into "Same" (both brokers routed to the same wholesale venue) and "Different" (routed to
distinct venues). The comparison tests whether the PI gap is driven by venue routing or by
within-venue differences:

$$
\overline{\text{PI\%}}_{\text{Broker A}} - \overline{\text{PI\%}}_{\text{Broker B}} \bigg|_{\text{Same venue}} \quad \text{vs.} \quad \overline{\text{PI\%}}_{\text{Broker A}} - \overline{\text{PI\%}}_{\text{Broker B}} \bigg|_{\text{Different venue}}
$$

Standard errors are clustered by stock (Table VII, p. 2532).

**Multivariate regression (R5).** Table VIII (p. 2536) regresses PI\% on broker account
dummies (TD Ameritrade omitted) and deciles of off-exchange order imbalance (OIB), separately
for buys and sells, with progressive addition of venue and stock fixed effects:

$$
\text{PI\%}_{i,j,t} = \sum_{b \neq \text{TD}} \beta_b \cdot \mathbf{1}[\text{Broker} = b]
+ \sum_{k=1}^{10} \gamma_k^{\text{Buy}} \cdot \text{OIB}_{k,t}^{\text{Buy}}
+ \sum_{k=1}^{10} \delta_k^{\text{Sell}} \cdot \text{OIB}_{k,t}^{\text{Sell}}
+ \text{FE} + \varepsilon_{i,j,t} \tag{}
$$

where OIB is computed as buy minus sell off-exchange orders in the same stock during the minute
of the trade, scaled by their sum (from TAQ, code "D"), assigned to deciles. Four models are
estimated: (1) broker FEs only, (2) broker FEs + OIB deciles, (3) + venue FEs, (4) + stock FEs.
Standard errors are clustered by stock.

**PFOF analysis (R6).** PFOF is obtained from SEC Rule 606 reports. The per-share PFOF is
plotted against per-share PI (Figure 4, p. 2535) to assess rank-order and magnitude
correspondence. No regression is estimated; the test is visual and descriptive, motivated by
the finding that PFOF ($0.001-$0.003/share) is an order of magnitude smaller than PI
($0.028-$0.078/share).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Authors' own trading experiment (85,417 trades in 6 broker accounts) | Primary outcome: execution price, PI, round-trip cost per trade | No page yet |
| TAQ (Trade and Quote Database) | Match trades to exchange/off-exchange execution; compute OIB; obtain execution venue codes | [TAQ](/wiki/commercial/taq/) (licensed) |
| CRSP (Center for Research in Security Prices) | Stock selection: universe stratification by market cap, liquidity, volatility, price | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| SEC Rule 606 reports (broker routing reports) | PFOF per share by broker and venue; routing shares to each wholesale venue | No page yet |
| SEC Rule 605 reports (market center execution reports) | Venue-level execution quality benchmarks (fraction with PI, average PI amounts) | No page yet |

Sample: December 21, 2021 to June 9, 2022 (113 trading days). 128 stocks; 74,801 trades in
the final cleaned sample ($15.4 million notional). Six broker accounts: TD Ameritrade, Fidelity,
E\*Trade, Robinhood, IBKR Lite, IBKR Pro. Order target: $100 per trade, market orders only.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13467) if you are:
studying retail market microstructure or the effects of PFOF; evaluating broker execution
quality methodologies or designing similar trading experiments; assessing the policy implications
of SEC Rule 605/606 disclosure reform; or interested in the within-venue vs. between-venue
decomposition of execution quality (Table VII is the key table). The Internet Appendix contains
additional robustness results (trade size sensitivity, latency tests, stock-by-stock
distributions).

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(5), October 2025. This distillation was
extracted by an LLM on 2026-06-05 and is **not human-verified or independently reproduced**.
The CC BY-NC 4.0 licence permits sharing with attribution for non-commercial use; the verbatim
PDF is not hosted in this batch.

> Schwarz, Christopher, Brad Barber, Xing Huang, Philippe Jorion, and Terrance Odean.
> "The 'Actual Retail Price' of Equity Trades."
> *The Journal of Finance* 80, no. 5 (October 2025): 2507-2541.
> DOI: 10.1111/jofi.13467. © 2025 The Author(s).
> Licensed under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Segmented Arbitrage: Siriwardane, Sunderam & Wallen (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/siriwardane-segmented-arbitrage-2025/
# Distilled: across 32 arbitrage spreads in equity, fixed income, and foreign exchange markets (2010-2020), the average pairwise correlation is only 22%, far below what canonical intermediary asset pricing models predict. The paper attributes this to two distinct forms of segmentation: funding segmentation (certain trades rely on specific unsecured vs. secured funding sources) and balance-sheet segmentation (intermediaries specialize in certain trades, so idiosyncratic balance-sheet shocks move specific spreads). J. Finance 2025, paywalled. Nine core results with source locators, datasets used, and the theory tested.
# Tags: paper-summary, asset-pricing, arbitrage, limits-to-arbitrage, intermediary-asset-pricing, fixed-income, foreign-exchange, equities, panel-regression, event-study, svar, peer-reviewed, unreplicated, data:bloomberg, data:cftc-cot, data:markit-cds, data:crane-mmf, data:preqin, data:gsw-yields, data:cboe-options, data:wrds
==============================================================================

**What this is.** The paper's core results, datasets, and theory: enough to
know what it found without reading all 48 pages. To replicate or extend it,
obtain the full source via the
[DOI](https://doi.org/10.1111/jofi.13469) (paywalled).

## TL;DR

Using daily data on 32 no-arbitrage spreads across equity, fixed income, and
foreign exchange markets over January 2010 to February 2020, the paper
documents that the average pairwise correlation of arbitrage spreads is only
0.22, far below what integrated-intermediary models predict. The paper argues
this reflects two types of segmentation: (i) funding segmentation, where
certain trades (equity spot-futures, equity options box, CIP) rely on
unsecured funding while others rely on secured (repo) funding, so that shocks
to unsecured funding markets raise unsecured spreads but not secured ones; and
(ii) balance-sheet segmentation, where intermediaries specialize in certain
trades, so idiosyncratic balance-sheet shocks (JPMorgan London Whale 2012,
Deutsche Bank CDS exit 2014) move specific arbitrage spreads but not others.
A sign-restricted SVAR shows the high-dimensional factor structure of spreads
is driven largely by weakly correlated supply shocks on the arbitrageur side.

## Core results

Magnitudes and significance are as reported; `*`/`**` = 10%/5%. Locators
point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Average pairwise correlation of arbitrage spreads is 22%**, far below the single-factor benchmark | Table II Panel A, p. 2560 | Mean ρ = 0.22, SD = 0.30; 75th pctile = 0.42; 90% of pairs reject H₀: ρ > 0.67 (p = 0.00); N = 496 pairs |
| R2 | **Low correlations persist even within same-tenor trades**, ruling out measurement error and noise-trader risk | Table II Panels B–D, pp. 2560–2561; Figure 3, p. 2562 | Short-tenor mean ρ = 0.19; after 1-month moving average, 10 PCs needed to explain 90% of variation; overnight IOER–GCF pair ρ = 23% |
| R3 | **Supply shocks from sign-restricted SVAR are weakly correlated** (avg 16% pairwise), not demand shocks | Figure 4, p. 2567 | Average ρ of supply shocks across all futures-based trades = 16%; average ρ of demand shocks also 16%; supply shocks within equity SF cluster higher at 62%; 1% upper bound on quarterly supply-shock correlations = 37% |
| R4 | **Unsecured arbitrages load strongly on TED spread; secured arbitrages do not** | Table III, p. 2569 | Unsecured β(TED) = 0.49\*\* (t = 4.58); secured β(TED) = 0.07 (t = 1.33, insignificant); unsecured spreads approximately 7× more sensitive to TED than secured |
| R5 | **2016 MMF reform raised unsecured arbitrage spreads by ~12 bps; secured spreads unaffected** | Table IV col. (1), p. 2573; Figure 5, p. 2571 | β = 11.77\*\* (t = 2.47); dynamic estimates show initial spike of 18.03\*\* at reform month, elevated for 3+ months; pass-through ≈ 0.59, matching OLS TED coefficient |
| R6 | **Fidelity IPrime MMF outflows specifically move equity spot-futures spreads but not other unsecured or secured spreads** | Table V, p. 2575 | OLS: Fidelity flows coef = −0.55\*\* (t = −3.86) for equity SF; IV estimate = −1.09\*\* (t = −2.25); CIP/Box coef = −0.14\* (t = −1.84, significant at 10% only); secured coef = 0.01 (insignificant) |
| R7 | **JPMorgan London Whale balance-sheet shock raised equity spot-futures spreads relative to other unsecured spreads** | Figure 7C, p. 2582 | Equity SF spreads significantly higher vs. other unsecured arbitrages following March 1, 2012 and June 13, 2012; widening persisted for several months; JPM CP rates unchanged, ruling out funding channel |
| R8 | **Deutsche Bank's 2014 CDS market exit raised CDS-bond arbitrage spreads relative to other secured spreads** | Figure 8B, p. 2584 | Effect significant at 5%; relative widening persisted over 5 months; other secured and unsecured arbitrage spreads unaffected |
| R9 | **Fixed-income hedge fund losses predict future increases in secured (not unsecured) arbitrage spreads** | Table VII, p. 2586 | Secured β(FI Arb HF Return\_{t-1}) = −0.66\*\* (t = −3.04); unsecured β = 0.00 (t = 0.01); driven by Treasury-swap and CDS-bond sub-strategies |

**Overall (paper's conclusion).** Riskless arbitrage is segmented. Both
funding segmentation (unsecured vs. secured funding markets) and balance-sheet
segmentation (intermediary specialization) drive low correlations across
arbitrage spreads. The evidence implies that intermediary asset pricing models
most naturally describe individual market segments rather than capital markets
as a whole.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Bloomberg (spot rates, FX forwards, OIS rates, futures prices, Treasury yields, inflation swaps, CDS via Markit) | Construction of all 32 arbitrage spreads; FX CIP, equity spot-futures, Treasury spot-futures, Treasury-swap, TIPS-Treasury series | [Bloomberg](/wiki/commercial/bloomberg/) (licensed) |
| van Binsbergen, Diamond & Grotteria (2019) box rates extended by authors using CBOE SPX options data | Equity options (box arbitrage) spreads at 6-, 12-, 18-month tenors | No page yet |
| Markit (cash bond and CDS pricing) | CDS-bond basis construction for IG and HY bonds | [Markit bond pricing](/wiki/commercial/markit/) / [Markit CDS](/wiki/commercial/markit-cds/) (licensed) |
| CFTC Commitments of Traders (Traders in Financial Futures, weekly) | Quantities data on positions by dealer, hedge fund, and asset-manager type for futures-based trades | [CFTC COT](/wiki/datasets/cftc-cot/) |
| Crane data / SEC Form N-MFP | MMF total net assets and holdings, for MMF reform analysis | [Crane Data](/wiki/commercial/crane-mmf/) (licensed) |
| Preqin Pro Hedge Fund Database | Hedge fund returns data; fixed-income arbitrage strategy flag | [Preqin](/wiki/commercial/preqin/) (licensed) |
| Federal Reserve yield curve models (Gurkaynak, Sack, Wright 2007/2010) | Zero-coupon constant-maturity Treasury and TIPS yields for TIPS-Treasury arbitrage | No page yet |
| CRSP / Compustat (via WRDS) | Supporting equity holdings data (Y-9C bank trading book filings cited; CRSP implied for stock characteristics) | [WRDS / CRSP / Compustat](/wiki/commercial/wrds/) (licensed) |
| Coalition Greenwich / S&P (qualitative) | JPMorgan equity derivatives market share since 2015 (cited contextual fact) | No page yet |

The CIP spread construction and FX arbitrage measurement follow the
methodology of Du, Tepper, and Verdelhan (2018).

Sample: 32 arbitrage spreads, daily, January 1, 2010 to February 29, 2020
(post-GFC, pre-Covid). CDS-bond and Treasury-swap series start Sep 2011 for
some maturities. CFTC quantity data weekly from July 2010.

## Theory / model

The paper develops a stylized model (Section I, pp. 2548-2553) in which a
unit measure of competitive, atomistic arbitrageurs (intermediaries) trade
N riskless arbitrage trades. The arbitrageur's objective is (eq. 1, p. 2549):

$$
\max_{q_{n,t},\, V_{k,t}} \sum_{n=1}^{N} q_{n,t} \left( s_{n,t} - \sum_l w_{n,l} f_{l,t} \right)
- \frac{1}{2} \sum_{k=1}^{K} c_{k,t} V_{k,t}^2
$$

where $$s_{n,t}$$ is the arbitrage spread on trade $$n$$ at time $$t$$, $$q_{n,t}$$
is the quantity supplied, $$w_{n,l}$$ is the fraction of trade $$n$$ funded from
source $$l$$ (with cost $$f_{l,t}$$ in excess of zero), $$V_{k,t}$$ is the
aggregate scale of activities under balance-sheet constraint $$k$$, and
$$c_{k,t}$$ is the marginal cost of meeting constraint $$k$$. Market clearing
requires $$q_{n,t} = a_{n,t}$$ (inelastic outside demand).

**Canonical benchmarks.** Under balance-sheet and funding integration with a
single balance-sheet constraint ($$c_{k,t} = 0$$ for $$k > 1$$, $$f_{l,t} = 0$$
for all $$l$$), the equilibrium spread is (eq. 2, p. 2550):

$$
s_{n,t} = v_{n,1} c_{1,t} V_{1,t} = v_{n,1} c_{1,t} \left( \sum_n a_{n,t} v_{n,1} \right)
$$

All spreads move with the single factor $$c_{1,t} V_{1,t}$$ and are perfectly
correlated. This is the prediction the data contradict: the canonical
single-constraint intermediary model of He and Krishnamurthy (2013) implies
perfect spread correlation, whereas the mean pairwise correlation is 0.22.
Under a single frictional funding factor with balance-sheet
integration ($$c_{k,t} = 0$$, $$v_{n,k} = 0$$, $$f_{n,1} > 0$$, $$f_{n,l} = 0$$
for $$l > 1$$), spreads are $$s_{n,t} = w_{n,1} f_{1,t}$$ -- again a one-factor
structure. Under integration with many constraints ($$L = 1$$, $$K > 0$$) spreads
have a K+1 factor structure (eq. 3, p. 2550). The margin-based asset pricing
of Garleanu and Pedersen (2011), in which integrated funding implies a one- or
two-factor spread structure, is contradicted by the high-dimensional factor
structure documented in the data:

$$
s_{n,t} = w_{n,1} f_{1,t} + \sum_{k=1}^{K} v_{n,k} c_{k,t} V_{k,t}
$$

**Funding segmentation.** When trades $$n = 1,\ldots,N_1$$ can use only source
$$l = 1$$ and trades $$n = N_1+1,\ldots,N$$ can use only source $$l = 2$$, the
equilibrium is (eq. 4, p. 2551):

$$
s_{n,t} = \begin{cases} w_{n,1} f_{1,t} & \text{if } n \leq N_1 \\ w_{n,2} f_{2,t} & \text{if } N_1 < n \end{cases}
$$

Cross-group correlation equals only $$\rho(f_{1,t}, f_{2,t})$$ (eq. 5, p. 2551):

$$
\rho(s_{n_1,t},\, s_{n_2,t}) = \begin{cases} 1 & \text{if } n_1, n_2 \leq N_1 \text{ or } N_1 < n_1, n_2 \\ \rho(f_{1,t}, f_{2,t}) & \text{if } n_1 \leq N_1,\; n_2 > N_1 \end{cases}
$$

**Balance-sheet segmentation.** When arbitrageurs in group I specialize in
trades $$n = 1,\ldots,N_1$$ and group $$\tilde{I}$$ in trades $$n = N_1+1,\ldots,N$$, with
different marginal balance-sheet costs, equilibrium spreads are (eq. 6,
p. 2552). The premise that intermediaries specialize draws on evidence of
intermediary specialization in credit derivatives in Siriwardane (2019):

$$
s_{n,t} = \begin{cases} \epsilon_{n,i} + v_{n,1} c_{1,t}^{I} V_{1,t}^{I} & \text{if } n \leq N_1 \\ \epsilon_{n,j} + v_{n,1} c_{1,t}^{\tilde{I}} V_{1,t}^{\tilde{I}} & \text{if } N_1 < n \end{cases}
$$

The correlation between spreads of the two groups depends on (i) correlation
of balance-sheet shocks across groups, (ii) correlation of demand shocks, and
(iii) cross terms (eq. 7, p. 2552):

$$
\rho(s_{1,t}, s_{2,t}) = \rho(c_{1,t}^{I},\, c_{1,t}^{\tilde{I}}) \times \rho(a_{1,t}, a_{2,t})
+ \rho(c_{1,t}^{I},\, a_{2,t}) \times \rho(c_{1,t}^{\tilde{I}},\, a_{1,t})
$$

The model is not estimated structurally; it organizes the empirical tests by
providing testable signatures: funding segmentation implies covariance between
certain spreads and certain funding rates; balance-sheet segmentation implies
covariance between certain spreads and specific intermediary balance-sheet
costs.

## Method

The paper applies four methods, each addressing a different identification
challenge, building on `sign-restricted-svar`, `panel-regression`,
`differences-in-differences`, and `instrumental-variables`.

**Sign-restricted SVAR (supply vs. demand decomposition, Section II.C).** For
each futures-based trade $$i$$, let $$Y_t = [s_t \;\; q_t]'$$ be the vector of the
spread and quantity (gross open interest). The structural VAR is (eq. 8,
p. 2565):

$$
B Y_t = A_0 + A_1 Y_{t-1} + \epsilon_t, \qquad \epsilon_t = [\epsilon_{s,t} \;\; \epsilon_{d,t}]'
$$

The reduced form is $$Y_t = \Phi_0 + \Phi_1 Y_{t-1} + u_t$$ where
$$\Phi_0 = B^{-1} A_0$$, $$\Phi_1 = B^{-1} A_1$$, and the residual covariance
$$\Sigma_u$$ depends on $$B$$. Sign restrictions on the impact matrix identify
the structural shocks (eq. 9, p. 2565):

$$
\begin{bmatrix} u_{s,t} \\ u_{q,t} \end{bmatrix}
= \underbrace{\begin{bmatrix} - & + \\ + & + \end{bmatrix}}_{B^{-1}}
\begin{bmatrix} \epsilon_{s,t} \\ \epsilon_{d,t} \end{bmatrix}
$$

A supply shock ($$\epsilon_s$$) lowers spreads and raises quantities; a demand
shock ($$\epsilon_d$$) raises both. The sign-restricted SVAR used to separate
supply from demand shocks builds on Uhlig (2005). Estimation follows Arias,
Rubio-Ramirez, and
Waggoner (2018): Bayesian Normal-Wishart prior, 1,000 draws from the posterior
using Cholesky decomposition of $$\Sigma_u$$. The model is estimated separately
for each trade; the correlations of supply and demand shocks across trades are
computed from the median-draw shock series.

**Panel OLS: funding sensitivity (Section III.B, Table III).** The baseline
funding regression relates monthly changes in arbitrage-implied riskless rates
to Treasury yield changes and TED spread changes (eq. 10, p. 2569):

$$
\Delta r_{i,j,t} = \alpha_{i,j} + \beta_1 \Delta y_{j,t} + \beta_2 \Delta \text{TED}_t + \epsilon_{i,j,t}
$$

- $$r_{i,j,t}$$ is the implied riskless rate for trade $$i$$ in strategy $$j$$
- $$y_{j,t}$$ is the maturity-matched Treasury yield
- $$\text{TED}_t$$ is the maturity-matched LIBOR minus Treasury spread (proxy for unsecured funding costs)
- Standard errors are clustered by strategy-month.

## Empirical specifications

**Spec 1: MMF reform DiD (R5, eq. 11, p. 2572).** Baseline differences-in-
differences estimating the 2016 MMF reform impact on unsecured vs. secured
spreads, using daily data, trade and time fixed effects, clustered by trade
and date. The 2016 MMF reform event study design builds on Anderson, Du, and
Schlusche (2019):

$$
s_{i,t} = \alpha_i + \alpha_t + \beta \cdot \mathbf{1}[i \in \text{Unsecured}] \cdot \mathbf{1}[t \geq \text{October2016}] + \epsilon_{i,t}
$$

- $$s_{i,t}$$ is the absolute value of the arbitrage spread for trade $$i$$ on date $$t$$
- $$\mathbf{1}[i \in \text{Unsecured}]$$ equals 1 for CIP, box, and equity spot-futures trades
- $$\mathbf{1}[t \geq \text{October2016}]$$ equals 1 on or after the reform month
- Fixed effects: trade ($$\alpha_i$$) and time ($$\alpha_t$$); SE clustered by trade and date

Column (1) of Table IV (p. 2573) reports $$\beta = 11.77^{**}$$ (t = 2.47).
A dynamic version replaces the single post-reform indicator with monthly
leads and lags to trace the time profile of adjustment.

**Spec 2: IV for equity repo funding (R6, Table V, p. 2575).** The baseline
augments eq. (10) with flows into Fidelity IPrime MMFs. The IV instrument is
passive flows (eq. on p. 2575):

$$
Z_t = F_t \times L_{t-3}^{I}
$$

- $$F_t$$ is total flow into all Fidelity MMFs
- $$L_{t-3}^{I}$$ is the lagged share of Fidelity MMF assets that are IPrime
- Standard errors are clustered by strategy-month

The IV estimate (column 4, Table V) of the
equity spot-futures spread on Fidelity flows is $$\beta = -1.09^{**}$$ (t = -2.25);
the CIP/box and secured spread coefficients are indistinguishable from zero.

**Spec 3: London Whale dynamic DiD (R7, eq. 12, p. 2583).** In a weekly panel
of unsecured arbitrage spreads, the event study estimates relative widening of
equity spot-futures vs. other unsecured spreads around the March 1 and June
13, 2012 event dates:

$$
s_{i,t} = \alpha_i + \alpha_t + \sum_{j=-4}^{24} \beta_j \cdot \mathbf{1}[i \in \text{Equity SF}] \cdot \mathbf{1}[t = j] + \epsilon_{i,t}
$$

- $$j$$ indexes weeks since the first event date
- $$\alpha_i$$, $$\alpha_t$$ are trade and time fixed effects

Panel C of Figure 7 (p. 2582) shows equity
spot-futures spreads were significantly elevated relative to other unsecured
spreads for several months following each event date.

**Spec 4: Hedge fund balance-sheet forecasting regression (R9, eq. 13,
p. 2585).** Monthly changes in spread levels on lagged hedge fund returns:

$$
\Delta s_{i,t} = \alpha + \beta\, r_{t-1}^{H} + \epsilon_{i,t}
$$

- $$r_{t-1}^{H}$$ is the lagged monthly return of Barclay's fixed-income arbitrage hedge fund index (standardized to mean zero, unit variance)

Table VII (p. 2586) shows $$\beta = -0.66^{**}$$ (t = -3.04) for secured spreads and
$$\beta = 0.00$$ (t = 0.01) for unsecured spreads; the effect is concentrated
in Treasury-swap and CDS-bond sub-strategies.

## When to read the full paper

Obtain the article via [DOI 10.1111/jofi.13469](https://doi.org/10.1111/jofi.13469)
if you are: evaluating the cross-market structure of limits to arbitrage;
testing intermediary asset pricing models across market segments; studying
the 2016 MMF reform's effects on derivatives markets; replicating the SVAR
decomposition or the event studies; or checking specific coefficient estimates
in the Internet Appendix. The locators above point to the exact tables and
figures. For "what did this paper find," the table above is the intended
default.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(5), October 2025,
pp. 2543–2590. DOI: 10.1111/jofi.13469. © 2025 the American Finance
Association. Published by Wiley under the Wiley VOR terms; paywalled.

This distillation was extracted by an LLM on 2026-05-31 and is **not
human-verified or independently reproduced**. Extraction is extract-only:
core results and locators reproduced for research commentary; no verbatim
reproduction of substantial portions. Contact the publisher for reuse rights.

> Siriwardane, Emil N., Adi Sunderam, and Jonathan Wallen. "Segmented
> Arbitrage." *The Journal of Finance* 80, no. 5 (October 2025): 2543–2590.
> DOI: 10.1111/jofi.13469.


==============================================================================
# Imperfect Intermediation of Money-Like Assets: Stein & Wallen (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/stein-imperfect-intermediation-money-like-2025/
# Distilled: T-bill rates fall below the Fed's RRP rate because money funds substitute imperfectly between T-bills and RRP, with heterogeneous and state-dependent elasticity, and because corporate treasurers demand T-bills as pledgeable collateral. When T-bill supply shrinks enough to drive elastic funds to a corner, remaining less-elastic funds become marginal, and supply shocks have an order-of-magnitude larger impact on T-bill rates. J. Finance 2025, paywalled. Eight core results with source locators, the theoretical model, and the empirical specifications.
# Tags: paper-summary, money-markets, fixed-income, intermediary-asset-pricing
==============================================================================

**What this is.** The paper's core results, the theoretical model of heterogeneous-elasticity intermediation, and the empirical specifications that test it: enough to know what was found and how, without reading the full 38 pages. To replicate or extend, read the original at [doi.org/10.1111/jofi.13500](https://doi.org/10.1111/jofi.13500).

## TL;DR

Stein and Wallen (2025) study why the one-month T-bill rate regularly falls below the rate on the Fed's reverse repurchase (RRP) facility, even though both are overnight-equivalent, credit-risk-free, government-backed instruments. The RRP-bill spread is a stark violation of the near-money premium logic of Nagel (2016), who shows T-bills command a liquidity premium precisely because they are more money-like. Stein and Wallen build and test a model with two types of investors: money funds (which have direct RRP access) and corporate treasurers (which do not). The model has three frictions: (i) money fund AUM is exogenously fixed (segmentation from outside investors, consistent with Bech and Klee (2011) showing limited access creates persistent rate wedges), (ii) corporate treasurers value T-bills for pledgeability as derivatives collateral, related to the Treasury richness documented by Fleckenstein and Longstaff (2024), and (iii) money funds substitute imperfectly and heterogeneously between T-bills and RRP, echoing the slow-moving capital mechanism of Duffie (2010). As the most-elastic money funds exhaust their T-bill holdings and reach a corner, the remaining less-elastic funds become marginal. Supply shocks then hit T-bill rates seven times harder than when funds are not constrained. Related evidence on law-of-one-price violations across segmented markets is provided by Siriwardane, Sunderam and Wallen (2025). The broader intermediary asset pricing literature, including He and Krishnamurthy (2013) who model aggregate intermediary wealth as the key state variable, typically focuses on a single representative intermediary; this paper shows that heterogeneity across intermediaries is equally important. Supply-side T-bill scarcity and its pricing consequences are documented in d'Avernas and Vandeweyer (2024). The mechanism described here was the key driver of the large yield dislocations during the 2023 debt-ceiling episode, when more than 90% of money fund AUM was already at a corner.

## Core results

Magnitudes and significance are as reported. `\*`, `\*\*`, `\*\*\*` = 10%, 5%, 1%.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | One-month T-bill rates fell well below expected RRP rates for most of June 2021 to May 2023, with the gap frequently exceeding 50 bps and spiking to over 160 bps during the March-April 2023 debt-ceiling uncertainty | Figure 1, p. 3186; Figure 3, p. 3199 | Average RRP-bill spread 6 bps in Region 2 (non-scarce periods); 37 bps in scarce period (April 2022 to April 2023); spike to ~160 bps in March-April 2023 |
| R2 | Flows from outside investors into money funds explain very little of the variation in aggregate T-bill portfolio weight; virtually all substitution from T-bills to RRP is driven by active portfolio rebalancing by money fund managers | Figure 4, p. 3201 | Counterfactual (investor-flows-only) portfolio share is an order of magnitude less variable than actual share; the two series are nearly uncorrelated |
| R3 | Money fund investor flows are almost insensitive to fund returns: a 100 bp increase in returns raises fund AUM by only 5.9 percentage points per quarter | Table II, column (4), p. 3202 | Coefficient 5.885\*\*\* (0.98); adjusted R-squared 14%; N = 4,449 fund-quarters |
| R4 | Corporate derivatives exposure is associated with a significantly larger Treasury share in liquid assets: a one-standard-deviation increase in derivatives exposure raises the Treasury share by 1.8 percentage points, relative to a 7.6% mean | Table III, column (3), p. 3203 | Coefficient 3.481\*\* (1.57); adjusted R-squared 0.15; N = 2,730 firm-years, 2004-2021 |
| R5 | IV estimate of aggregate money fund elasticity of substitution: a 1 bp increase in the RRP-bill spread due to a supply-driven decrease in T-bills causes money funds to decrease their T-bill portfolio weight by 6.2%, or 79 bps of portfolio weight | Table IV, column (3), p. 3206 | IV coefficient -6.19\*\*\* (2.24); first stage: 1% increase in Treasury supply decreases spread by 2.20\*\*\* (0.59) bps; N = 90 monthly observations |
| R6 | Fund-level elasticities are highly heterogeneous: median elasticity is 7.4 (interquartile range 4.4 to 10.1); Treasury funds (most restrictive mandate) have average elasticity 5.4 vs. government funds 7.2 and prime funds 8.6 | Figure 6, p. 3209 | Kernel density of fund-level IV elasticities estimated on September 2013 to March 2021; Treasury funds significantly less elastic than government and prime funds |
| R7 | When T-bills are scarce (April 2022 to April 2023), a 1% decrease in instrumented T-bill supply causes the RRP-bill spread to increase by 4.5 bps, about seven times the unconditional estimate of 0.66 bps | Table V, column (4), p. 3212 | Interaction coefficient Scarce T-Bills x Delta T-Bill: -4.484\*\* (1.83); unconditional IV: -0.656\*\*\* (0.21); full sample September 2013 to June 2024 |
| R8 | T-bills maturing before June 1, 2023 (the projected debt-ceiling date) yielded on average 153 bps less over May 2023 than T-bills maturing on June 1, 2023; average RRP-bill spread for pre-June maturities was 81 bps | Figure 10, p. 3215 | Discontinuity in T-bill yield curve at June 1, 2023 maturity cutoff; T-bills maturing in May 2023 averaged ~4.4%, vs. RRP rate ~5.05%; June maturities showed a positive RRP-bill spread reflecting default risk |

**Overall (paper's conclusion).** Even in a simple and transparent setting, frictions in financial intermediation create economically significant and time-varying spreads between money-like assets. Heterogeneity in intermediary elasticity is the central mechanism: as more-elastic intermediaries are driven to corners, their departure leaves only less-elastic intermediaries as the marginal market participants, amplifying the rate impact of supply and demand shocks. This state-dependence has broad implications for evaluating QE programs and other policy interventions whose market impact is implicitly assumed to be constant over time.

## Theory / model

The model (Section I, pp. 3191-3195) has a total supply of T-bills $$S$$ held by money funds or corporate treasurers. Money funds also have access to the RRP facility; treasurers cannot access RRP and can only hold T-bills or private repo. Aggregate money fund AUM is exogenously fixed at $$A$$.

**Treasurers' collateral demand.** Corporate treasurers derive collateral benefits from T-bills. Their demand for T-bills is an increasing function of the RRP-bill spread (p. 3192):

$$
X(r_b - r_p), \quad \text{with } X(\cdot) \text{ increasing}
$$

where $$r_b$$ is the T-bill rate and $$r_p$$ is the (exogenous) RRP rate. Because private repo and RRP rates are nearly identical in the data, the model writes treasurer demand as $$X(r_b)$$ for simplicity.

**Money fund preferences.** Fund $$i$$ splits its portfolio between T-bills (weight $$T_i \in [0,1]$$) and RRP (weight $$1 - T_i$$) with utility (equation (1), p. 3192):

$$
U_i = r_b T_i + r_p (1 - T_i) + V_i(T_i) \tag{1}
$$

The term $$V_i(T_i)$$ is an increasing concave nonpecuniary benefit from holding T-bills, capturing idiosyncratic preferences (fund mandates, window-dressing, etc.). The paper uses the functional form (equation (2), p. 3192):

$$
V_i(T_i) = \frac{T_i - \frac{1}{2} T_i^2}{b_i} \tag{2}
$$

where $$b_i$$ is the elasticity parameter. Less elastic funds have smaller $$b_i$$; more elastic funds have larger $$b_i$$. The elasticity parameter $$b_i$$ is distributed uniformly on $$[b_L, b_H]$$ with $$b_L > 0$$.

**Optimal portfolio.** The first-order condition for an interior optimum is (equation (3), p. 3193):

$$
(r_p - r_b) = V_i'(T_i) = \frac{1 - T_i}{b_i} \tag{3}
$$

yielding fund $$i$$'s optimal T-bill share in an interior solution (equation (4), p. 3193):

$$
T_i^* = 1 - b_i (r_p - r_b) \tag{4}
$$

**Three regions.** The model's solution is characterized by three regions as T-bill supply $$S$$ shrinks:

- *Region 1 (ample supply):* $$r_b \geq r_p$$; all money funds hold only T-bills; market-clearing pins the spread at zero or negative: $$X(r_b) = S - A$$ (equation (5), p. 3193).
- *Region 2 (moderate scarcity):* Some funds are at an interior; the market-clearing condition is $$X(r_b) = S - A\theta(r_b) T^{**}$$ (equation (11), p. 3194), where $$\theta(r_b)$$ is the fraction of funds still holding T-bills (equation (12), p. 3194):

$$
\theta(r_b) = \frac{b_U - b_L}{b_H - b_L} \tag{12}
$$

The sensitivity of the T-bill rate to supply changes in Region 2b (Case 2b, where the most elastic fund is already at the corner) is (equation (13), p. 3194):

$$
\frac{dr_b}{dS} = \frac{1}{X'(r_b) + A\theta\!\left(\frac{b_U + b_L}{2}\right)} \tag{13}
$$

This sensitivity increases continuously ($$\frac{d^2 r_b}{dS\, d\theta} < 0$$) as fewer funds remain in the T-bill market.

- *Region 3 (extreme scarcity):* All money funds are at corners holding only RRP. Market clearing is entirely determined by treasurer collateral demand: $$X(r_b) = S$$ (equation (14), p. 3195), and the sensitivity is:

$$
\frac{dr_b}{dS} = \frac{1}{X'(r_b)} \tag{15}
$$

The transition from Region 2 to Region 3 is the key amplification mechanism documented empirically.

## Method

The paper combines a partial-equilibrium theoretical model with instrumental-variable panel regressions. It builds on `panel-regression` and `instrumental-variables` as its core estimation primitives.

**Instrument for the RRP-bill spread.** The key endogeneity concern is that T-bill supply may respond to money fund demand. The paper instruments monthly changes in the RRP-bill spread using monthly log differences in the privately-held supply of all Treasuries (not just T-bills), on the grounds that Treasury auction timing follows a "regular and predictable" schedule (citing Garbade (2007)) and total Treasury supply is not driven by money fund demand dynamics (p. 3206-3207, Table IV).

First-stage specification (equation (23), p. 3207):

$$
\Delta \text{Spread}_t = \alpha + \beta \Delta \text{Treasury}_t + \epsilon_t \tag{23}
$$

**Aggregate elasticity regression.** The estimating equation for the aggregate T-bill portfolio weight of the money fund sector (equation (22), p. 3205):

$$
\Delta w_{\text{Bills},t} = \alpha - \beta \Delta \text{Spread}_t + \epsilon_t \tag{22}
$$

where $$\Delta w_{\text{Bills},t}$$ is the log difference in the aggregate T-bill portfolio weight. The IV estimate (Table IV, column (3)) delivers $$\hat\beta = 6.19$$.

**Supply-shock amplification regression.** The headline test of state-dependence (equation (24), p. 3212):

$$
\Delta \text{Spread}_t = \alpha + \beta_0 \Delta \text{TBill Supply}_t + \beta_1 \text{Scarce TBills}_t + \gamma \Delta \text{TBill Supply}_t \times \text{Scarce TBills}_t + \varepsilon_t \tag{24}
$$

The interaction term $$\gamma$$ captures the incremental sensitivity when T-bills are scarce. A continuous version replaces the dummy with *Constrained Share* (equation (25), p. 3213):

$$
\Delta \text{Spread}_t = \alpha + \beta_0 \Delta \text{TBill Supply}_t + \beta_1 \text{Constrained Share}_t + \gamma \Delta \text{TBill Supply}_t \times \text{Constrained Share}_t + \varepsilon_t \tag{25}
$$

where *Constrained Share* is the AUM-weighted fraction of funds with less than 5% T-bill weight and at least 10% weight in either RRP or T-bills.

**Flow decomposition.** To test the segmentation assumption, the paper decomposes dollar changes in T-bill holdings into an investor-flow component and a managerial rebalancing component (equations (16)-(20), pp. 3200-3201):

$$
\Delta D_{\text{Bill},i,t} = w_{\text{Bill},i,t} A_{i,t} - w_{\text{Bill},i,t-1} A_{i,t-1} \tag{16}
$$

$$
\text{IFlow}_{\text{Bill},i,t} = w_{\text{Bill},i,t-1} (A_{i,t} - A_{i,t-1}) \tag{17}
$$

$$
\text{MFlow}_{\text{Bill},i,t} = \Delta D_{\text{Bill},i,t} - \text{IFlow}_{\text{Bill},i,t} \tag{18}
$$

The counterfactual T-bill portfolio share driven only by investor flows is constructed as (equation (20), p. 3201):

$$
\widetilde{\text{Ratio}}_t = \frac{\tilde{D}_{\text{Bill},t}}{\tilde{D}_{\text{Bill},t} + D_{\text{RRP},t}} \tag{20}
$$

**Corporate collateral regression.** For the collateral-demand channel (equation (21), p. 3202):

$$
w_{UST,i,t} = \alpha_i + \beta \, \text{Deriv}_{i,t-1} + \varepsilon_{i,t} \tag{21}
$$

where $$w_{UST,i,t}$$ is firm $$i$$'s Treasury holdings share and $$\text{Deriv}_{i,t-1}$$ is lagged derivatives exposure (absolute value of P&L on derivatives / total assets). Standard errors are clustered by firm; OLS with time, industry, and time-by-industry fixed effects.

## Empirical specifications

**Aggregate elasticity (R5, Table IV, p. 3206).** Monthly time-series, September 2013 to March 2021 (N = 90). LHS: log change in aggregate money fund T-bill portfolio weight. RHS: change in the RRP-bill spread (instrumented by log change in privately-held Treasury supply). First stage F is implicit in the significant first-stage coefficient (-2.20 bps per 1% Treasury supply increase). Standard errors are robust to heteroskedasticity. The IV coefficient of -6.19 rises by almost an order of magnitude relative to OLS (-0.94), consistent with demand-side endogeneity biasing the OLS toward zero.

**Supply-shock amplification (R7, Table V, p. 3212).** Monthly time-series, September 2013 to June 2024 (N = 127; excludes May-June 2023 for debt-ceiling effects). LHS: change in RRP-bill spread. RHS: instrumented change in T-bill supply, scarce-T-bills indicator (April 2022 to April 2023), and their interaction. Unconditional IV: -0.656\*\*\* (0.21) bps per 1% supply change. Interaction with scarce indicator: -4.484\*\* (1.83). Total effect when scarce: -(0.656 + 4.484) = -5.1 bps per 1% supply decrease (i.e., a 7x amplification). Column (5) with continuous constrained-share interaction shows that when 91% of funds are constrained (April 2023), a 1% supply decrease increases the spread by 6.0 bps, vs. 1.5 bps when 24% are constrained (May 2022).

**Corporate collateral demand (R4, Table III, p. 3203).** Annual firm-year panel, 193 large U.S. non-financial corporates, 2004-2021 (N = 2,730). LHS: Treasury holdings share (Treasury + agency securities / Treasury + agency + cash equivalents + money fund shares). RHS: lagged derivatives exposure (absolute P&L / total assets), firm size, industry and time fixed effects. One-standard-deviation increase in derivatives exposure (0.516%) is associated with a 1.8 percentage-point higher Treasury share (column (3)). Standard errors are clustered by firm.

**Investor flow sensitivity (R3, Table II, p. 3202).** Monthly and quarterly, September 2013 to June 2024 (N = 13,434 monthly; 4,449 quarterly). LHS: investor flows as a percentage of lagged fund AUM. RHS: contemporaneous fund return. Coefficient in AUM-weighted quarterly specification (column 4): 5.885\*\*\* (0.98). The economic implication: during the scarce period (April 2022 to April 2023) when the spread averaged 37 bps, a fund invested entirely in RRP would receive only 2.2 percentage points more quarterly inflows than a T-bill-invested fund.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Crane Data LLC (monthly money fund holdings and AUM) | T-bill and RRP portfolio weights, fund-level elasticity estimation, AUM decomposition | [Crane Data](/wiki/commercial/crane-mmf/) (licensed) |
| Federal Reserve (NY) - RRP program data | RRP counterparty identities, RRP take-up amounts, administered RRP rate | No page yet |
| Bloomberg (secondary market yields) | One-month T-bill yields, OIS rates for maturity adjustment | [Bloomberg](/wiki/commercial/bloomberg/) (licensed) |
| Federal Reserve (effective Fed Funds rate) | Maturity adjustment for expected one-month RRP return; monetary policy benchmarks | No page yet |
| U.S. Treasury / Federal Reserve (Treasury supply) | Privately-held outstanding Treasuries (instrument for T-bill supply shocks) | No page yet |
| Compustat (annual) via WRDS | Derivatives P&L, firm size, corporate bond and Treasury holdings for large non-financial firms, 2001-2021 | [WRDS](/wiki/commercial/wrds/) (licensed) |
| Darmouni and Mota (2024) - corporate securities holdings | Cash and securities holdings for 200 largest U.S. public non-financial firms, 2001-2021 | No page yet |

Sample for the main money-market analysis: monthly, September 30, 2013 to June 30, 2024 (130 months). Sample for corporate collateral analysis: annual firm-year panel, 2004-2021.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13500) if you are: studying the T-bill market microstructure or the Fed's RRP facility design; analyzing how intermediary heterogeneity shapes asset price sensitivity to supply shocks; calibrating the impact of Treasury debt management decisions on short-term rates; or extending the framework to other asset classes where identifying which intermediaries are at corners is feasible. The exact model equations and Appendix A derivation of equation (13) are the load-bearing technical content.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(6), December 2025. Wiley, paywalled. This distillation was extracted by an LLM on 2026-06-03 and is **not human-verified or independently reproduced**. Extract-only; the verbatim PDF is not hosted here.

> Stein, Jeremy C., and Jonathan Wallen. "The Imperfect Intermediation of Money-Like Assets."
> *The Journal of Finance* 80, no. 6 (December 2025): 3185–3221.
> DOI: 10.1111/jofi.13500. © 2025 the American Finance Association.


==============================================================================
# Banks, Low Interest Rates, and Monetary Policy Transmission: Wang (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/wang-banks-low-interest-rates-2025/
# Distilled: A structural model of banks as dual credit and liquidity providers shows that secular declines in nominal interest rates compress deposit spreads, tighten banks' financial constraints, and reduce long-run bank credit supply, with loan spreads rising to offset lost deposit income. Cross-sectional bank-level evidence from U.S. Call Reports (2000-2014) confirms the mechanism. J. Finance 2025, paywalled. Four core results with source locators, datasets used, the model, and the empirical specifications.
# Tags: paper-summary, banking, monetary-policy, interest-rates, credit-supply
==============================================================================

**What this is.** The paper's core results, the structural model it builds (banks as dual credit and liquidity providers), and the empirical specifications: enough to know what it found and how, without reading all 38 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1111/jofi.13436).

## TL;DR

Olivier Wang develops a tractable general-equilibrium model in which commercial banks earn income from two distinct spreads: a loan spread (credit provision) and a deposit spread (liquidity provision). Because deposits compete with publicly issued money (cash/currency), lower nominal interest rates reduce the opportunity cost of holding cash, compress the deposit spread that banks earn, and reduce their retained earnings and equity. Once the nominal rate falls below a threshold $$\tilde{i}$$ (calibrated at roughly 8% in the baseline), the economy enters a "constrained lending regime" where lower deposit income tightens banks' leverage constraints, shrinks loan supply, and forces loan spreads to rise. The paper provides aggregate time-series evidence that the U.S. loan-deposit spread has been stable while its composition shifted (deposit spread down, loan spread up) and cross-sectional bank-level evidence that banks with stickier deposit rates (lower deposit beta, meaning more spread compression) experienced lower retained earnings, equity, loan growth, and higher loan spreads between 2000 and 2014.

## Core results

Magnitudes are as reported. `***`/`**`/`*` = 1%/5%/10%. Locators point to the source PDF (pages are printed page numbers 1379-1416).

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | The maturity-adjusted loan-deposit spread is stable but its composition shifted: deposit spread fell ~1 pp, loan spread rose ~1 pp over 1997-2018 | Figure 8, p. 1409 | Total spread ~4-6%, stable from 1997Q2-2018Q2; deposit spread fell from ~2% to ~1%; loan spread rose by ~1 pp |
| R2 | Banks with lower deposit-rate pass-through (lower "expense beta") experienced significantly lower retained earnings and equity growth 2000-2014 | Table II, p. 1413 | Coefficient on predicted liability spread decline: retained earnings 0.177\*\*\* (SE 0.027); equity 0.340\*\*\* (SE 0.072); N = 4,387 |
| R3 | Low-deposit-beta banks also had significantly lower loan growth 2000-2014, consistent with the leverage-constraint channel | Table II, p. 1413 | Coefficient on predicted liability spread decline: loans 0.291\*\*\* (SE 0.067); N = 4,387 |
| R4 | Low-deposit-beta banks also had significantly larger increases in loan spreads 2000-2014 | Table II, p. 1413 | Coefficient on predicted liability spread decline: loan spread -0.317\*\*\* (SE 0.045); N = 4,387 |

**Overall (paper's conclusion).** The secular decline in nominal interest rates hurts long-run bank credit supply even well above the zero lower bound. Because deposits compete with publicly issued money, lower nominal rates compress banks' deposit-spread income, weakening their financial constraints and reducing lending. The short-run and long-run effects of rate cuts differ: a rate cut can stimulate lending on impact via capital-gain revaluation of long-term assets, but permanently reduces lending in the long run once deposit incomes fall. The model has normative implications for the optimal inflation target, calling for a departure from the Friedman rule.

## Theory / model

The model features discrete time $$t = 0, 1, \ldots$$. Banks intermediate between two household types: savers (unconstrained) and borrowers (bank-dependent). Banks hold two types of assets (bonds and loans of maturity $$k \in \{1,\ldots,K\}$$) and two types of liabilities (deposits). Savers maximize lifetime utility and solve (p. 1385):

$$
\max_{c_t, a_{t+1}, m_{t+1}, d_{t+1}} \sum_{t=0}^{\infty} \beta^t U(c_t, x(m_t, d_t))
$$

$$
\text{s.t.} \quad c_t + \frac{1}{1+r_t}\left[\Omega_{t+1} + i_t m_{t+1} + s_t^d d_{t+1}\right] \leq A_t \bar{n}^s + \Omega_t + \text{Div}_t + T_t^s, \tag{1}
$$

where $$\Omega_t = a_t + m_t + d_t$$ is total financial wealth, $$i_t$$ is the nominal rate (the liquidity premium on money), $$s_t^d = \frac{1+r_t}{1+r_t^d} - 1$$ is the deposit spread. The liquidity aggregator $$x(m,d)$$ is strictly increasing, homothetic, and concave (Assumption 1, p. 1386). The standard CES specification is:

$$
x(m, d) = \left[\alpha^{1/\epsilon} m^{\frac{\epsilon-1}{\epsilon}} + (1-\alpha)^{1/\epsilon} d^{\frac{\epsilon-1}{\epsilon}}\right]^{\frac{\epsilon}{\epsilon-1}}, \tag{9}
$$

where $$\epsilon$$ is the elasticity of substitution between money and deposits.

Bank equity at the beginning of period $$t$$ is (p. 1387, eq. 2):

$$
e_t(i) \equiv \sum_{k=0}^{K-1} \frac{l_{t,t+k}(i)}{(1+r_{t,t+k}^l)^k} + a_t(i) - d_t(i), \tag{2}
$$

where the first term values outstanding loans at market yields. Banks are subject to a leverage constraint $$\text{liab}_t \leq \bar{\phi}_t e_t$$ (eq. 4), arising from limited pledgeability (Assumption 4: banks can pledge only a fraction $$\theta < 1$$ of date-$$t+1$$ assets). By Lemma 1 (p. 1388), in equilibrium the leverage ratio satisfies:

$$
\bar{\phi}_t = \frac{\theta(1+r_t^l)/(1+r_t^d)}{1 - \theta(1+r_t^l)/(1+r_t^d)}.
$$

The excess return on bank equity is the key object linking spreads to lending capacity (p. 1389, eq. 7):

$$
\nu_t = \underbrace{\bar{\phi}_t \frac{s_t^d}{1+s_t^d}}_{\text{excess return from liquidity provision}} + \underbrace{(1+\bar{\phi}_t) s_t^l}_{\text{excess return from credit provision}}. \tag{7}
$$

The steady-state real rate is pinned by fundamentals: $$r^* = G/\beta - 1$$ (p. 1393). The Fisher equation (p. 1394, eq. 13) implies:

$$
1 + i = (1 + \pi)(1 + r^*) = (1 + \pi)\frac{G}{\beta}. \tag{13}
$$

**Proposition 1** (p. 1394): When money and deposits are strict gross substitutes, there exists a positive threshold $$\tilde{i}$$ such that the economy is in the constrained lending regime if and only if $$i < \tilde{i}$$. In the constrained regime, a decline in $$i$$ lowers bank equity, deposits, leverage, and lending, while the deposit spread $$s^d$$ falls and the loan spread $$s^l$$ rises.

**Proposition 2** (p. 1395): The threshold $$\tilde{i}$$ satisfies:

$$
\tilde{i} = \frac{\nu}{1+\nu}\frac{1-\theta}{\theta} \Phi\!\left(\frac{\chi_{\min}}{\chi}\right), \tag{14}
$$

where $$\Phi$$ is an increasing function of the liquidity aggregator. The threshold increases with banks' required excess return $$\nu$$, decreases with pledgeability $$\theta$$, and decreases with liquidity benefits $$\chi$$. In the baseline calibration (Table I, p. 1398: $$G=1.02$$, $$\beta=0.98$$, $$\pi=2\%$$, $$\epsilon=8$$, $$\rho=0.2$$, $$\theta=0.85$$), $$\tilde{i} \approx 8\%$$.

**Proposition 3** (p. 1396): For $$i < \tilde{i}$$, both the deposit-rate and loan-rate pass-throughs are increasing functions of $$i$$ (convexity): at lower rates, a further rate decline compresses deposit spreads more per unit because the deposit beta $$\beta^d(i) = 1 - s^d(i)/i$$ is an increasing function of $$i$$.

**Proposition 4** (p. 1404): With only short-term loans ($$K=1$$), a permanent decline in $$i$$ always initially reduces lending relative to the previous steady state, because there is no maturity-mismatch-driven capital-gain revaluation channel.

**Proposition 5** (p. 1406-1407): With heterogeneous deposit markets differing in the relative preference for cash vs deposits (parameters $$\alpha, \chi$$), the long-run pass-through of the common nominal rate to the deposit rate is lower in markets with a stronger relative preference for cash, and a common rate decline leads to a larger increase in loan spreads and larger contraction in lending in those markets.

## Method

The model is a structural general equilibrium model with flexible prices and full employment (Definition 1, p. 1389). It is studied analytically via steady-state conditions and propositions proved in the Internet Appendix, and numerically via a baseline calibration (Table I, p. 1398). Transitional dynamics are obtained numerically.

The aggregate equity return identity (eq. 7) is the unifying device: it decomposes the excess return on bank equity into the deposit-spread leverage component and the loan-spread leverage component. Because the net payout rate $$\rho$$ pins the steady-state return $$\nu = \beta/(1-\rho) - 1$$ (eq. 12, p. 1393), an exogenous decline in $$s^d$$ must be offset by an endogenous rise in $$s^l$$ or a fall in leverage and equity.

The model builds on `panel-regression` for the empirical part and on the bank capital framework of Gertler and Kiyotaki (2010). The deposit competition mechanism follows the deposit-spread framework of Drechsler, Savov, and Schnabl (2017), extended to a two-sided bank balance sheet and a financial-constraint channel. The paper relates to Drechsler, Savov, and Schnabl (2021) on NIM stability, which it decomposes into offsetting loan and deposit spreads. It contrasts with Abadi, Brunnermeier, and Koby (2023), where money is irrelevant at positive rates and harmful lending effects arise only at negative rates. Di Tella and Kurlat (2021) study related bank exposure to monetary policy with maturity mismatch and liquidity premia on deposits.

## Empirical specifications

**Aggregate evidence (Section III.A, pp. 1409-1411).** To decompose the total loan-deposit spread into a loan spread and a deposit spread, the paper constructs a Treasury replicating portfolio that matches the maturity structure of the bank loan portfolio, using Call Report data on repricing maturity bins. The yield on the replicating portfolio at date $$t$$ is (p. 1410):

$$
R_t^{\text{Treas}} = y_{t-1}^{ST} \omega_{t-1}^{ST} + y_{t-10}^{LT}\left(1 - \omega_{t-1}^{ST}\right),
$$

where $$\omega_{t-1}^{ST}$$ is the share of loans repricing within one year and $$y^{ST}$$ ($$y^{LT}$$) is the one-year (ten-year) Treasury yield. The "loan spread" (red area in Figure 8, p. 1409) is the difference between the effective loan yield and the replicating Treasury yield, capturing credit and liquidity premia net of duration. The "deposit spread" (blue area) is the replicating Treasury yield minus the effective deposit rate. Sample: 1997Q2-2018Q2, all U.S. commercial banks (Call Reports and FDIC Quarterly Banking Profile).

**Cross-sectional evidence (Section III.B, pp. 1411-1413).** The paper constructs a bank-level "expense beta" $$\beta_i$$ by estimating a separate time-series regression for each bank $$i$$:

$$
\Delta \text{IntExp}_{it} = \alpha_i + \sum_{\tau=0}^{3} \beta_{i,\tau} \Delta ffr_{t-\tau} + \epsilon_{it}, \tag{implied by eq. 18}
$$

in the 1984-2000 pre-period, where $$\text{IntExp}_{it}$$ is interest expense over total assets and $$ffr$$ is the Fed funds rate. The "expense beta" $$\beta_i = \sum_{\tau=0}^{3} \beta_{i,\tau}$$ measures historical deposit-rate sensitivity: a low $$\beta_i$$ bank has stickier deposit rates and thus experiences more spread compression when rates fall. Using this, the paper constructs the predicted change in liability spread from the 2000-2014 rate decline:

$$
\widehat{\Delta \text{LiabilitySpread}}_{i,00-14} = (1 - \beta_i)(ffr_{2014} - ffr_{2000}).
$$

The main cross-sectional regression is then (p. 1412, eq. 18):

$$
\frac{y_{i,2014} - y_{i,2000}}{y_{i,2000}} = \alpha + \delta\, \widehat{\Delta\text{LiabilitySpread}}_{i,00-14} + \Gamma' \text{controls}_i + \epsilon_i, \tag{18}
$$

for outcomes $$y \in \{\text{retained earnings}, \text{equity}, \text{loans}, \text{loan spread}\}$$ (for spreads, the left-hand side is $$y_{i,2014} - y_{i,2000}$$). Controls include 2000Q4 leverage and the deposit-asset ratio. Standard errors are block-bootstrapped by quarter (1,000 iterations). Sample: 4,387 U.S. commercial banks with at least 20 quarterly observations for $$\text{IntExp}_{it}$$ in 1984-2000.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Call Reports (Federal Reserve / FFIEC) | Quarterly income and balance sheet data for all U.S. commercial banks; repricing maturity structure for replicating portfolio construction; bank-level interest expense and loan/deposit rates | no page yet |
| FDIC Quarterly Banking Profile | Aggregate U.S. commercial bank data cross-checking the Call Report series | [FDIC QBP / financials](/wiki/datasets/fdic/) |
| Federal Reserve H.15 / Fed funds rate | Policy rate series for expense-beta estimation and spread decomposition | no page yet |

Sample: 4,387 U.S. commercial banks; aggregate quarterly series 1997Q2-2018Q2; expense-beta pre-period 1984-2000.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13436) if you are: building a model of bank credit supply with liquidity frictions; studying the transmission of secular interest-rate declines to credit markets; interpreting NIMs and loan-deposit spreads in the data; or analyzing the optimal long-run inflation target from a banking-system perspective. Internet Appendix Section II details the calibration (Table I). Internet Appendix Section III contains extensions (role of money, endogenous equity issuance, operating costs, firm investment).

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(3), June 2025. Paywalled; no Creative Commons licence detected. This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. Extract-only: reproducing the verbatim text requires a subscription or library access.

> Wang, Olivier. "Banks, Low Interest Rates, and Monetary Policy Transmission."
> *The Journal of Finance* 80, no. 3 (June 2025): 1379–1416.
> DOI: [10.1111/jofi.13436](https://doi.org/10.1111/jofi.13436).
> © 2025 the American Finance Association.


==============================================================================
# Working More to Pay the Mortgage: Zator (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/zator-working-more-pay-mortgage-2025/
# Distilled: Using Polish administrative tax records linked to floating-rate mortgage payments (2005-2015), Zator shows households increase labor income by roughly PLN 0.35 for each PLN 1 rise in mortgage interest, with an asymmetric response that is two to three times stronger following payment increases than decreases. J. Finance 2025, paywalled. Seven core results with source locators, datasets used, the identification strategy, and the estimating equations.
# Tags: paper-summary, household-finance, labor-supply, mortgage, interest-rates
==============================================================================

**What this is.** The paper's core results, the identification strategies, and the key
estimating equations: enough to know what it found and how, without reading all 37 pages.
To replicate or extend it, read the full source at
[10.1111/jofi.13413](https://doi.org/10.1111/jofi.13413).

## TL;DR

Using a panel of Polish income tax records linked to floating-rate mortgage interest
deductions (2005-2015), Zator shows that households increase their gross labor income
by roughly PLN 0.35 for each PLN 1 increase in mortgage interest payments. The effect
is causal: identification exploits within-household variation in mortgage size interacted
with the reference rate (WIBOR or LIBOR), instrumented to remove endogenous prepayment.
The response is sizable and asymmetric: households react two to three times more
strongly to payment increases than to decreases. Secondary earners (often women) and
dual-earner childless households drive the increase response, consistent with the fixed
costs of entering the labor market and consumption commitment models.

## Core results

Magnitudes and significance are as reported; `\*\*\*` = 1%, `\*\*` = 5%. Locators
point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Labor income rises PLN 0.30-0.35 for each PLN 1 increase in mortgage interest (OLS)** | Table III, p. 1185 | Coefficients 0.337 to 0.348\*\*\* (mortgage holders, cols 1-3); 0.297\*\*\* (full population, col 4); se 0.020-0.021 |
| R2 | **IV estimates confirm OLS: PLN 0.35-0.46 response; reference-rate x mortgage-size instrument** | Table III, p. 1185 | IV col 5 (mortgage holders): 0.459\*\*\* (se 0.043); IV col 6 (full pop): 0.348\*\*\* (se 0.037); OLS and IV confidence intervals overlap |
| R3 | **Currency-comparison design (PLN vs CHF loans) gives a smaller estimate of PLN 0.15, consistent with asymmetry** | Table IV, p. 1188 | IV cols 4-5: 0.148\*\*\* (se 0.034) mortgage holders, 0.145\*\*\* (se 0.016) all; smaller because the CHF/WIBOR divergence covers only the declining-rate 2013-2015 period |
| R4 | **Response to payment increases is 2-3x larger than to decreases (asymmetric labor supply)** | Table V, p. 1192; Figure 3, p. 1193 | Average coefficient 0.106-0.283\*\*\*; interaction on increases 0.233\*\*\* to 0.313\*\*\* (six specs); slope for increases roughly 2-3x slope for decreases across all six specifications |
| R5 | **Dual-earner households and secondary (often female) earners drive the increase response; women show more pronounced asymmetry** | Figure 6, p. 1198; Figure 7, p. 1199 | Dual-earner coefficient on payment increase ~0.48 vs single-earner ~0.14; secondary earner coefficient substantially higher than primary on increases; women increase income strongly on increases but do not reduce significantly on decreases |
| R6 | **Higher mortgage payments reduce the probability of remaining a single-earner household by 0.1 pp per PLN 1,000** | Table VI, col 4, p. 1201 | Coefficient -0.113\*\*\* (se 0.021); mean dep var 27.5%; consistent with spousal labor market entry as a key mechanism |
| R7 | **Higher mortgage payments raise the probability of changing jobs by 0.104 pp per PLN 1,000** | Table VI, col 6, p. 1201 | Coefficient 0.104\*\*\* (se 0.013); mean dep var 21.9%; consistent with households accepting higher-paying, less-preferred jobs to service debt |

**Overall (paper's conclusion).** Households with floating-rate mortgages adjust their
labor supply substantially in response to interest rate changes. Income adjusts by roughly
35% of the payment change on average, with a markedly stronger response to increases.
The adjustment operates through spousal labor market entry, supplemental gig income,
and job changes toward higher-paying positions. The asymmetry helps reconcile the wide
dispersion of labor supply elasticity estimates in the literature: lottery-win studies
such as Imbens, Rubin, and Sacerdote (2001) and Cesarini et al. (2017) capture responses
to budget-loosening shocks (comparable to payment decreases here), while Di Maggio et al.
(2017) document consumption responses to ARM resets consistent with the labor supply
responses this paper finds. Brown and Matsa (2020) and Bernstein and Koudijs (2020)
show correlated patterns in the relationship between household debt and labor market
outcomes.

## Theory / model

The paper has no formal structural model. It operates from a standard household
labor-leisure model in which the household equates the marginal disutility of work
with the marginal utility of consumption (p. 1184). When mortgage payments rise, the
available budget for non-housing consumption falls, raising the marginal utility of
consumption and, under concave utility, inducing greater labor supply. This mechanism
implies the response to payment increases should be stronger than to decreases when
utility is concave, which is the core theoretical prediction tested in Section IV.

The paper invokes Chetty and Szeidl (2007) to argue that consumption commitments
(housing being the canonical commitment) amplify the asymmetry. With a commitment
that cannot be adjusted downward quickly, an increase in the committed payment has
a large liquidity effect with no offsetting consumption substitution on the downside.

## Method

The paper uses two complementary identification strategies, both estimating
panel regressions with individual fixed effects and rich controls.

**Strategy 1: Reference-rate x Mortgage-size variation (equation (1), p. 1180).** The
main estimating equation is:

$$
Y_{i,t} = \alpha \cdot (\text{Interest}_{i,t} = \text{RefRate}_t \cdot \text{LoanSize}_i) + \mu_i + \xi X_{i,t} + \epsilon_{i,t} \tag{1}
$$

where $$Y_{i,t}$$ is gross household income for person $$i$$ in year $$t$$,
$$\text{Interest}_{i,t}$$ is annual mortgage interest deducted from taxable income,
$$\mu_i$$ is an individual fixed effect, and $$X_{i,t}$$ includes year-by-age-by-previous-income
bin fixed effects and county-year fixed effects. To remove endogenous variation from
prepayment decisions, interest is instrumented with the product of the reference rate
(WIBOR for PLN loans, LIBOR CHF for CHF loans) and an estimate of initial mortgage
size (the household's second observed positive interest payment). The instrument is
strong (Kleibergen-Paap F-statistics of 5.5-7.6 x 10^4 in Table III, p. 1185).

This approach resembles a shift-share design (Borusyak, Hull, and Jaravel (2022)):
the endogenous exposure measure (mortgage size) is interacted with several exogenous
shocks (11 years of reference rate changes from two correlated rates). Standard errors
are clustered at the household level throughout.

**Strategy 2: Currency-comparison design (equation (2), p. 1181).** The second design
compares PLN-indexed (WIBOR) and CHF-indexed (LIBOR) mortgages of the same size after
2012, when LIBOR hit the zero lower bound while WIBOR continued to decline:

$$
Y_{i,t} = \beta \cdot (\text{Interests}_{i,t} = \text{LoanSize}_i \cdot \text{Post2013}_t \cdot \text{PLNLoan}_i)
+ \phi \cdot \!\!\left(\text{LoanSize}_i \cdot \sum_{t=2006}^{2015} \text{Year}_t\right)
+ \mu_i + \xi X_{i,t} + \epsilon_{i,t} \tag{2}
$$

The coefficient $$\beta$$ captures the differential income change for PLN-loan households
relative to CHF-loan households of the same size after 2013. Since LIBOR was flat and
WIBOR declined, this design identifies the labor supply response to payment decreases.

**Asymmetry test (equation (3), p. 1182).** To test asymmetry, the base specification
is augmented with an interaction between the interest payment and a binary indicator
for years in which payments increased:

$$
Y_{i,t} = \alpha \cdot \text{Interests}_{i,t} + \beta \cdot \text{Interests}_{i,t} \cdot \text{Increase}_{i,t} + \mu_i + \beta X_{i,t} + \epsilon_{i,t} \tag{3}
$$

The coefficient $$\alpha$$ captures the response to payment decreases and $$\alpha + \beta$$
captures the response to payment increases; Table V (p. 1192) shows $$\beta > 0$$ and
significant across three alternative definitions of the increase indicator.

## Empirical specifications

All regressions use a strongly balanced panel with observations weighted by the inverse
of household size (one or two). Standard errors are clustered at the household level
(or by year for robustness).

**First stage (Table II, p. 1185).** Reference rate interacted with mortgage size
strongly predicts interest payments: coefficient 0.223\*\*\* (se 0.001 household-cluster;
0.012 year-cluster) on the full mortgage-holder sample. WIBOR drives PLN-loan
interest (0.145\*\*, se 0.053); LIBOR CHF drives CHF-loan interest (0.234\*\*\*, se 0.041).

**Main income regressions (Table III, p. 1185).** Specification columns:
- Col 1-3: Mortgage holders, progressively adding previous-income-year FE and county-year FE. Coefficient stable at 0.337-0.348.
- Col 4: Full population (OLS); coefficient 0.297\*\*\*.
- Col 5-6: IV (instrument = RefRate x LoanSize); coefficient rises slightly to 0.459\*\*\* (mortgage holders) and 0.348\*\*\* (full population).

**Currency-comparison regressions (Table IV, p. 1188).** First stage: PLN loan x
post-2012 x loan size reduces interest paid by PLN 0.231\*\*\* per unit loan size
(se 0.010). Second stage income coefficients: 0.148\*\*\* and 0.145\*\*\* (IV). Pre-2012
coefficients insignificant, confirming parallel trends.

**Mechanism regressions (Table VI, p. 1201).** Specifications analogous to equation
(1) with alternative outcomes: log wages (0.0037\*\*\*, se 0.0002), log pensions
(-0.0001, insig.), log business profits (0.0107\*\*\*, se 0.0012), single-earner
probability (-0.113\*\*\*, se 0.021 per PLN 1,000), supplemental income indicator
(0.013\*, se 0.007), job-change indicator (0.104\*\*\*, se 0.013).

**Consumption and savings proxies (Table VII, p. 1203).** Interest payments reduce
charitable donations (asinh coefficient -0.0012\*\*\* to -0.0020\*\*\*), private pension
contributions (-0.0004\*\* to -0.0006\*\*\*), and internet-access spending
(-0.0019\*\*\* to -0.0018\*\*\*), confirming the income response reflects labor supply
rather than differential macroeconomic sensitivity of mortgage holders.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Polish Ministry of Entrepreneurship and Technology: income tax declarations (PIT), 2005-2015 | Universe of Polish taxpayers; individual income sources (wages, business profits, pensions), mortgage interest deductions, demographics (age, sex, county), consumption proxies (charitable donations, pension contributions, internet access) | No page yet |
| Three-month WIBOR (Warsaw Interbank Offered Rate) | Reference rate for PLN-denominated mortgages; the main shock variable interacted with mortgage size | No page yet |
| Three-month LIBOR CHF | Reference rate for CHF-denominated mortgages; used in currency-comparison design | No page yet |

Sample: Strongly balanced panel, 9.8 million individuals, 2005-2015 annual (100+
million observations). Mortgage-holder subsample: 171,445 individuals (1,714,450
observations). The data are obtained from the Polish government and are not publicly
available; replication requires an access agreement with the Polish Ministry of
Entrepreneurship and Technology.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13413) if you are:
studying the labor supply channel of monetary policy (Section III-IV give the clearest
panel IV evidence in a floating-rate mortgage setting); interested in the asymmetry of
labor supply responses to income shocks (Section IV and Table V); studying intrahousehold
labor supply and the role of secondary earners or women in debt-adjustment (Sections V-VI);
or using Poland as a case study for a nearly universal floating-rate mortgage market
(99.8% floating rate as of 2016, p. 1176).

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(2), April 2025, pp. 1171-1207.
DOI: [10.1111/jofi.13413](https://doi.org/10.1111/jofi.13413).
This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified
or independently reproduced**. The article is paywalled; extract-only redistribution.

> Zator, Michal. "Working More to Pay the Mortgage: Household Debt, Interest Rates,
> and Family Labor Supply." *The Journal of Finance* 80, no. 2 (April 2025): 1171-1207.
> DOI: 10.1111/jofi.13413. © 2024 the American Finance Association.
> This page is an **extraction** by the Institute for Automated Research:
> core results summarized with source locators; **not a substitute for the original**.


==============================================================================
# Carbon Returns across the Globe: Zhang (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2025/zhang-carbon-returns-globe-2025/
# Distilled: After correcting for the data release lag of carbon emissions, the brown-minus-green return (the carbon premium) turns significantly negative in the United States and insignificant globally, overturning prior findings; the apparent premium stems from forward-looking sales information embedded in emissions data rather than a true risk premium. J. Finance 2025, CC BY 4.0. Eight core results with source locators, datasets used, and the empirical specifications.
# Tags: paper-summary, asset-pricing, esg, climate-finance, carbon-risk, factors
==============================================================================

**What this is.** The paper's core results, the data methodology, and the empirical specifications that overturn prior findings on the carbon premium: enough to know what it found and how, without reading all 31 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13402).

## TL;DR

The paper revisits the carbon return, defined as the return spread between high-carbon-intensity (brown) and low-carbon-intensity (green) firms. Prior studies find a positive carbon premium. Zhang (2025) shows this premium arises from forward-looking sales information embedded in emissions data rather than a true ex ante risk premium. After applying the actual data release lag (median 10 months for U.S., 12 months for international), the carbon return turns significantly negative in the United States (brown firms underperform green firms) and is insignificant globally. Developed markets exhibit more negative carbon returns due to stronger growth in climate concerns; countries with tighter climate policies show higher carbon returns consistent with pricing policy risk.

## Core results

Magnitudes and significance are as reported in the paper; `\*\*`/`\*\*\*` = 5%/1%. The sample is June 2009 to December 2021.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **U.S. carbon return is significantly negative**: brown firms underperform green firms after applying the data release lag | Table IV, Panel A, p. 628 | Value-weighted H-L return: -0.39%/month (scope 1, t=-2.47); -0.27%/month (scope 2, t=-1.87) |
| R2 | **FF6-adjusted alphas are significantly negative** for U.S. carbon-sorted portfolios | Table IV, Panel A, p. 628 | Alpha: -0.40%/month (scope 1, t=-2.51); -0.34%/month (scope 2, t=-2.40) |
| R3 | **WLS regression confirms negative carbon return** in the U.S. cross-section | Table VI, cols. 1-2, p. 632 | Scope 1: -0.19% per SD (t=-2.52); scope 2: -0.21% per SD (t=-2.46) |
| R4 | **Global carbon return is insignificant** on average across all countries | Table VII, Panel A, p. 633 | Value-weighted alpha: -0.06% (scope 1, t=-0.74); -0.03% (scope 2, t=-0.43) |
| R5 | **Developed markets (DM) have more negative carbon returns than emerging markets (EM)** | Table X, Panel A, p. 640 | DM alpha: -0.40% (scope 1, t=-9.52); EM alpha: +0.20% (scope 1, t=3.55) |
| R6 | **Positive carbon premium in prior studies is explained by forward-looking sales information**: once contemporaneous sales growth is controlled, positive emissions-return relation disappears | Table VIII Panel A and Table IX, pp. 635-637 | Contemporaneous scope 1 delta-emissions return: +0.47%/month (H-L, t=3.25); after controlling for same-period sales growth: insignificant (-0.03%, -0.24%, +0.10% within sales terciles) |
| R7 | **Climate concern shocks explain cross-country carbon return variation**: a 1-SD increase in sustainable flow lowers scope 1 carbon return by 0.10%/month | Table XI, Panel A, p. 642 | Sustainable flow: -0.10% (scope 1, t=-1.37); climate concern: -0.11% (scope 1, t=-1.68); earnings-day return: +0.79% (scope 1, t=8.53) |
| R8 | **Tighter climate policies associated with higher carbon returns**: countries with higher policy tightness earn 0.13% more per SD in carbon returns | Table XI, Panel B, p. 642 | Policy tightness: +0.13% per SD (t=2.12); renewable energy share: +0.20% per SD (t=2.60); civil law dummy: +0.55% (t=3.38) |

**Overall (paper's conclusion).** The positive carbon premium documented in prior studies arises from forward-looking firm sales information contained in emissions rather than a risk premium for carbon transition exposure. After properly accounting for data release lags, brown stocks underperform green stocks in the United States, consistent with ongoing investor shifts toward carbon-aware investing. Globally, carbon returns vary substantially across countries as a function of climate concern shocks, cash flow news, and local policy tightness.

## Theory / model

The paper has no formal equilibrium model. Instead, it operates within two theoretical frameworks that motivate the empirical tests.

**Framework 1: Carbon risk premium in equilibrium.** Following Bolton and Kacperczyk (2021) and Pastor, Stambaugh and Taylor (2021), brown firms face greater policy exposure during the transition to net zero and should in equilibrium earn higher expected returns. Empirically, this predicts a positive carbon premium. The paper tests this prediction using point-in-time emissions data. Bolton and Kacperczyk (2023) interpret cross-country variation in carbon returns as expected return differences driven by the global pricing of carbon transition risk; this paper reinterprets that variation as reflecting climate concern shocks and in-sample cash flow news rather than ex ante risk premia.

**Framework 2: Transition period with investor preference shifts.** During an ongoing transition to a carbon-aware equilibrium, investor preference shifts and unanticipated demand for green assets can generate negative realized carbon returns. The paper's cross-country evidence is consistent with this channel: developed markets where climate concerns have grown most have the most negative realized carbon returns.

**Identification of forward-looking bias.** The key identification idea is that emissions are derived from firm sales via IPCC emission factors (p. 624, equation 1):

$$
\text{Emissions} = \text{Activity Data} \times \text{Emission Factor} \tag{1}
$$

Because emissions scale nearly linearly with sales (regression R-squared up to 71% for U.S. scope 1, Table III), emissions released during fiscal year $$t$$ contain information about firm sales in year $$t$$. Prior studies that link returns to contemporaneous or one-month-lagged emissions effectively exploit this forward-looking sales signal. The paper's correction is to use point-in-time emissions based on actual Trucost release dates, with a median lag of 10 months (U.S.) and 12 months (international) from the fiscal year-end (Figure 1, p. 624).

## Method

The paper applies two standard empirical approaches from the asset pricing toolkit: portfolio sorts and panel regressions. Both are applied using point-in-time emissions data based on actual Trucost data release dates to eliminate forward-looking bias.

**Emission scale regressions** (Table III). To document the link between emissions and sales, the paper runs at the firm-year level (p. 624, equation 2):

$$
\log \text{Emission}_{it} = \alpha + \beta \log \text{Sales}_{it} + \varepsilon_{it}, \qquad
\Delta \text{Emission}_{it} = \alpha + \beta \Delta \text{Sales}_{it} + \varepsilon_{it} \tag{2}
$$

Standard errors are double-clustered at firm and year levels.

**Carbon intensity and firm characteristics regression** (Table III, Panel B, p. 626, equation 3):

$$
\text{Intensity}_{it} = \alpha + \beta \cdot \text{Characteristics}_{it} + \varepsilon_{it} \tag{3}
$$

where $$\text{Intensity}_{it}$$ is the scope 1 or 2 log carbon intensity available to investors at time $$t$$ and $$\text{Characteristics}_{it}$$ includes beta, size, book-to-market, ROA, asset growth, momentum, leverage, log PPE, IVol, sales growth, EPS growth, and commodity exposures.

**Portfolio sorts** follow Bolton and Kacperczyk (2021). For each month $$t$$, stocks are sorted into tercile portfolios by carbon intensity. Value-weighted monthly returns at $$t+1$$ are calculated. The high-minus-low (H-L) portfolio takes a long position in the most carbon-intensive tercile (H) and short in the least (L). Alphas are obtained by regressing H-L returns on FF6 factors (Fama and French (2018): market, SMB, HML, RMW, CMA, MOM). The U.S. sample screens common stocks following Fama and French (1992) conventions. For the international sample, stocks are screened following Hou, Karolyi and Kho (2011) to minimize outlier effects.

**Country-level portfolio sorts** (Section IV, equation 7, p. 639): for the cross-country analysis, each country's long-short return is obtained by regressing on regional FF6 factors (Fama and French (2017)):

$$
r_{it} = \alpha_i + \beta_i \text{factors}_{it} + \varepsilon_{it} \tag{7}
$$

where $$r_{it}$$ is the value-weighted long-short carbon return in country $$i$$ and $$\text{factors}_{it}$$ are FF6 factors for each region.

## Empirical specifications

**Baseline U.S. return regression** (equation 4, p. 630):

$$
r_{it} = \alpha + \beta \text{Intensity}_{it-1} + \gamma \text{Controls}_{it-1} + \nu_t + \varepsilon_{it} \tag{4}
$$

Run at the firm-month level with time fixed effects. Standard errors are double-clustered at firm and month levels. Weighted least squares is used to reduce influence from small stocks. Carbon measures are standardized to zero mean and unit variance so coefficients represent the monthly return change per one-SD increase in carbon footprint. Controls include beta, size, book-to-market, ROA, asset growth, momentum, leverage, log PPE, IVol, sales growth, EPS growth, and exposures to oil, natural gas, and commodity returns estimated over a 60-month rolling window. Results in Table VI (p. 632): scope 1 coefficient -0.19 (t=-2.52), scope 2 coefficient -0.21 (t=-2.46).

**Contemporaneous emissions regression** replicating Bolton and Kacperczyk (2021) (equation 5, p. 636):

$$
r_{it} = \alpha + \beta \text{Carbon}_{it} + \gamma \text{Controls}_{it-1} + \delta_k + \nu_t + \varepsilon_{it} \tag{5}
$$

where $$\text{Carbon}_{it}$$ is contemporaneous (same-year) log emissions growth or log emissions, with industry fixed effects $$\delta_k$$. Emissions growth is strongly positively associated with contemporaneous stock returns (Table IX, column 1: scope 1 coefficient +0.28, t=5.98).

**Sales-controlled contemporaneous regression** (equation 6, p. 636):

$$
r_{it} = \alpha + \beta \text{Carbon}_{it} + \beta \mathbf{Sales}_{it} + \gamma \text{Controls}_{it-1} + \nu_t + \varepsilon_{it} \tag{6}
$$

where $$\mathbf{Sales}_{it}$$ includes log sales and sales growth during the same emission period. After controlling for sales information, carbon emissions and emissions growth are no longer positively associated with returns and tend to be negative, consistent with the baseline result (Table IX, columns 5-6).

**Cross-country carbon return variation regression** (equation 8, p. 641):

$$
r_{it}^s = a + b \cdot X_{it-1} + \kappa \cdot Y_{it} + \nu_t + e_{it} \tag{8}
$$

where abnormal carbon returns $$r_{it}^s = \alpha_i + \varepsilon_{it}$$ are from equation (7), $$X_{it-1}$$ is lagged country characteristics (log GDP per capita, sustainable flow, climate concern), and $$Y_{it}$$ is contemporaneous cash flow shocks (earnings-day return, analyst EPS revision, sales growth). Standard errors are clustered at the monthly level. Results in Table XI (p. 642).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Trucost (S&P) | Firm-level annual carbon emissions (scope 1 and 2, tCO2e), with actual data release dates; primary emissions source | [Trucost](/wiki/commercial/trucost/) (licensed) |
| CRSP | Monthly stock returns, market capitalization, share prices; U.S. equities sample | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| Compustat (U.S.) | Firm accounting fundamentals: book-to-market, ROA, asset growth, leverage, PPE, EPS growth, sales growth | [WRDS / Compustat](/wiki/commercial/wrds/) (licensed) |
| Compustat Global | Firm accounting fundamentals for international sample; primary security on primary exchange | [WRDS / Compustat](/wiki/commercial/wrds/) (licensed) |
| Ken French Data Library | FF6 factor returns (market, SMB, HML, RMW, CMA, MOM) for factor adjustment; regional factors | [Ken French library](/wiki/datasets/ken-french/) |
| FRED (St. Louis Fed) | Natural gas price, Brent oil price, and commodity index used to estimate commodity exposures | [FRED](/wiki/datasets/fred/) |
| World Bank | Country-level GDP per capita and socioeconomic controls | No page yet |
| World Risk Poll (Lloyd's Register Foundation 2020) | Country-level climate concern measure (fraction perceiving climate change as very or somewhat serious threat) | No page yet |
| Climate Change Performance Index | Country-level climate policy tightness score | No page yet |
| Morningstar Sustainable Funds | Country-level quarterly sustainable investor flows as fraction of market cap | No page yet |

Sample: June 2009 to December 2021 (U.S. and global). U.S. sample: 211,495 firm-month observations; global: ~92,790 (country-industry-time regressions).

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13402) if you are: studying or replicating the carbon premium literature and need the full methodology for correcting the data release lag (Section I.C, pp. 622-627); conducting international/cross-country analysis of carbon returns and need the country-level dispersion results (Section IV, pp. 639-643); assessing whether prior evidence on the carbon premium is robust to proper timing of emissions data; or working on climate policy's effect on asset prices. Table IV (p. 628) is the key U.S. result; Table VII (p. 633) is the global result; Table XI (p. 642) is the cross-country driver analysis.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 80(1), February 2025. This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Zhang, Shaojun. "Carbon Returns across the Globe."
> *The Journal of Finance* 80, no. 1 (February 2025): 615-645.
> DOI: 10.1111/jofi.13402. © 2024 The Author(s).
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# An Economic View of Corporate Social Impact: Allcott, Montanari, Ozaltun & Tan (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2026/allcott-corporate-social-impact-2026/
# Distilled: a welfare-economics framework defines corporate social impact as the social welfare loss from a firm's exit; applied to 74 firms in 12 industries, consumer surplus dominates all other components, and ESG ratings are essentially unrelated to the resulting estimates. J. Finance 2026, open access (Wiley/AFA terms). Eight core results with source locators, datasets used, and the theory tested.
# Tags: paper-summary, esg, corporate-social-responsibility, welfare-economics, consumer-surplus, impact-investing, structural-estimation, panel-regression, open-access, peer-reviewed, unreplicated, data:nielseniq, data:wrds, data:acs, data:infousa, data:rystad, data:us-epa-supply-chain
==============================================================================

**What this is.** The paper's core results, datasets, and theory: enough
to know what it found without reading all 44 pages. To replicate or extend
it, read the full source at the
[original](https://doi.org/10.1111/jofi.70004) (open access on the Wiley
platform under AFA/Wiley terms). Replication files and the survey instrument
are available at allcott.stanford.edu/research.

## TL;DR

The paper proposes an economic definition of corporate social impact as the
social welfare loss that would result from a firm's exit. Using a 3,500-person
survey (fielded July and November 2021) combined with standard industrial
organization and labor economics methods, it estimates social impact for 74
large firms across 12 U.S. industries. Consumer surplus is by far the largest
component, dwarfing profits, worker surplus, and externalities. Firm size is
the strongest driver of total impact; product differentiation (own-price
elasticity) drives impact per dollar of revenue. Existing ESG ratings from
CSRHub and Just Capital are essentially uncorrelated with these
welfare-grounded estimates. Cigarette companies have negative social impact
in the model; Walmart's grocery business has by far the largest positive impact
($151 billion/year).

## Core results

Magnitudes and significance are as reported. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Consumer surplus dominates all other components of corporate social impact** | §VI.C, Figure 7, p. 317–318 | Consumer surplus accounts for the overwhelming share of weighted individual impact per dollar of revenue across all differentiated industries; profits receive much less weight (welfare weight on profits ≈ 0.12 because high-income people own most equity); worker surplus is small because average total compensation is only about 22% of revenues |
| R2 | **Firm size is highly correlated with social impact** (R² = 0.89) | Figure 5, p. 316 | Log-log plot of unweighted individual impact vs. revenue across 74 firms; firms excluded are the two cigarette companies plus Frontier and Spirit Airlines (negative impact) |
| R3 | **Product differentiation (own-price inelasticity) drives impact per dollar of revenue** | Figure 6, p. 316 | Unweighted impact/revenue ranges from about 0.2 to over 1.0 across firms; much of this variation explained by own-price elasticity from the survey |
| R4 | **Walmart's grocery business has by far the highest individual social impact in the sample** | Table VI Panel A, p. 319 | Weighted individual impact: Walmart $150.54 billion/year (rank 1); Philip Morris: -$16.78 billion/year (rank 74, most harmful) |
| R5 | **Oil companies have the highest social impact per dollar of revenue** due to global supply inelasticity | Table VI Panel B, p. 319 | Weighted impact/revenue: Conoco, Eni, Total, Shell, Chevron all at 1.50–1.51; the large consumer surplus arises because oil exit raises prices substantially given inelastic global demand (elasticity ≈ -0.14) |
| R6 | **Cigarette companies have negative social impact** due to large internalities ($2.77/\$ sales) | §VI.A, Table IV, pp. 311, 314 | Internality for cigarettes ≈ $2.77 per dollar of sales (vs. externality ≈ $0.12/$ sales); internality-adjusted consumer surplus is negative; Philip Morris -$16.78 billion/year, Reynolds -$13.72 billion/year (Table VI) |
| R7 | **Shares of industry impact are considerably larger than individual firm impacts**, especially in industries with inelastic aggregate demand | §VI.C, p. 318 | When all auto firms exit, consumers must find entirely different forms of transportation; individual BMW exit allows substitution to other auto firms; the gap is largest for toothpaste, groceries, and smartphones (most inelastic aggregate demand per Figure 2) |
| R8 | **ESG ratings from CSRHub and Just Capital are essentially unrelated to the welfare-grounded impact estimates** | Figure 9, §VII, pp. 322–323 | Scatterplot of weighted individual impact/revenue vs. CSRHub and Just Capital ratings shows little relationship; Internet Appendix Table IA.VII also shows other rating systems (Refinitiv, S&P) are not closely correlated with each other or with the paper's estimates |

**Overall (paper's conclusion).** Consumer surplus is the primary driver of
corporate social impact. Impact investors should consider devoting more
attention to firms that deliver more consumer surplus, especially for lower
income people. Making more differentiated products that more consumers want
to buy is the key to social impact in this framework (p. 324). Where Chatterji
et al. (2015) document that ESG ratings disagree substantially with each other,
this paper shows that they are also uncorrelated with its welfare-grounded
estimates.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Original 3,500-person consumer and worker survey (Lucid and Cint panels, July and November 2021) | Firm-level price response, aggregate price response, worker price response, satisfaction, income | Proprietary/custom survey; not a public dataset; no page |
| NielsenIQ Homescan and Statista Consumer Market Outlook | Consumer packaged goods revenues and market shares | [no page yet](/wiki/datasets/) |
| U.S. Department of Transportation DB1B | Airline revenues | no page yet |
| Wards (auto revenues) | Auto revenues | no page yet |
| Winsight (grocery revenues) | Grocery revenues | no page yet |
| Technomic (restaurant revenues) and Statista / Statcounter (smartphone revenues) | Restaurant and smartphone revenues | no page yet |
| Compustat | Revenues and employment for publicly traded firms | [WRDS / Compustat](/wiki/commercial/wrds/) (licensed) |
| InfoUSA | Firm-level county employment counts | no page yet |
| American Community Survey (ACS), 2010-2019 | Occupation and county employment distributions for worker surplus estimation | [ACS](/wiki/datasets/acs/) |
| Rystad Energy | Oil production and operating expenses for the seven oil supermajors across all oil fields worldwide (2018) | [Rystad Energy](/wiki/commercial/rystad/) (licensed) |
| U.S. EPA supply-chain CO2 emission factors (Ingwersen and Li 2020) | Production externalities from CO2 emissions | [EPA Supply Chain GHG](/wiki/datasets/us-epa-supply-chain/) |
| Piketty, Saez and Zucman (2018) distributional national accounts | After-tax income distribution for social marginal welfare weights | no page yet |
| C-corporation ownership data (Piketty et al. 2018) | Profit distribution across income percentiles | no page yet |
| National Household Travel Survey (2017) | U.S. gasoline consumption by income (for oil consumer surplus welfare weights) | no page yet |

Sample: 74 firms; 12 industries (autos, airline, beer, cereal, cigarettes, grocery, oil, restaurant, smartphone, soda, toothpaste, yogurt); survey n = 3,544 valid respondents after screening.

## Theory / model

Building on the concept of enterprise impact in Brest and Born (2013), the paper
operationalizes social impact as the welfare loss from firm exit. The paper builds
a micro-founded partial equilibrium model. There are $$N$$ people
indexed by $$i$$ with income-earning ability $$\theta_i$$. Each person $$i$$ has quasilinear
utility additively separable in consumption, labor, and externality (eq. 1,
p. 292):

$$
U_i(y; p, w(\theta_i)) = U_i\!\left( \sum_m \sum_{t \in T_m} \sum_{j \in J_m} (u_{ijt} - p_j)\, y_{ijt}
  + \pi_i
  + \sum_{fl} (u_{ifl} + w_{ifl}(\theta_i))\, y_{ifl}
  - \Phi \right)
$$

where $$y_{ijt}$$ are binary purchase indicators, $$y_{ifl}$$ are binary employment
indicators, $$\pi_i$$ is person $$i$$'s share of redistributed profits, and $$\Phi$$ is the
per-capita externality. For markets with behavioral biases (cigarettes, soda),
consumers maximize perceived utility $$\tilde{U}_i$$ where $$u_{ijt}$$ is replaced by
$$u_{ijt} + \gamma_j$$; $$\gamma_j < 0$$ is a negative internality. Consumer choice is
(eq. 2, p. 292):

$$
y^* = \operatorname*{argmax}\, \tilde U_i(y; p, w(\theta_i))
$$

Firm $$f$$'s profits are (eq. 3, p. 292):

$$
\Pi_f(p) = \sum_{j \in J_f} \left[ p_j q_j(p) - C_j(q_j) \right]
$$

Per-capita externality (eq. 4, p. 292):

$$
\Phi = \frac{1}{N} \sum_m \sum_{j \in J_m} q_j(p)\, \phi_j
$$

Social welfare is the Pareto-weighted sum of indirect utilities (eq. 5, p. 293):

$$
W(p, w) = \sum_i \omega_i V_i(p, w(\theta_i))
$$

**Individual impact.** Firm $$f$$'s individual impact is the welfare loss from its
exit if all other firms remain (eq. 6-7, p. 293):

$$
\Delta W_f(X) := W(p^{X_0}, w^{X_0}) - W(p^{X_0 \setminus f}, w^{X_0 \setminus f})
$$

$$
\Delta W_f^{\text{Individual}} = \Delta W_f(F)
$$

**Share of industry impact.** Defined as firm $$f$$'s Shapley value for the social
welfare loss if the entire industry exited. With $$R_m$$ the set of orderings of
firms in market $$m$$ and $$P_f^R$$ the set of firms preceding $$f$$ in ordering $$R$$
(eq. 8, p. 293):

$$
\Delta W_f^{\text{Shapley}} = \frac{1}{F_m!} \sum_{R_n} \Delta W_f(P_f^R)
$$

**Social marginal welfare weights.** Following Saez (2002), welfare weights are
inversely proportional to after-tax income (eq. 9, p. 294):

$$
g_i = \kappa\, a(z_i)^{-\rho}, \qquad \kappa = \frac{N}{\sum_i a(z_i)^{-\rho}}
$$

where $$a(z_i)$$ is after-tax income and $$\rho = 1$$ as benchmark (log utility).
When $$\rho = 0$$, all people receive equal weight and $$W$$ is total surplus.

**Identification.** This is a structural quantification exercise, not a
quasi-experimental design. Demand parameters are identified by survey moments
(substitution, income-firm, aggregate price response), combined with aggregate
market-share matching (BLP contraction). The partial equilibrium framework
assumes each firm is a small share of the labor market and that intermediate
inputs are produced at constant marginal cost.

## Method

The estimation proceeds in three parallel modules: (i) differentiated product
markets via BLP demand, (ii) oil market via price-taking competitive fringe, and
(iii) labor markets via linear probability model. It builds on `blp-demand`,
`method-of-simulated-moments`, and `shapley-value-allocation`. The
differentiated-product demand follows the random-coefficient logit framework of
Berry, Levinsohn, and Pakes (1995). The estimation strategy for differentiated-product
markets using micro data builds on Berry, Levinsohn, and Pakes (2004).

**Differentiated product markets (Sections IV.A-B, pp. 302-305).** Representative
utility for income group $$z$$ and firm $$f$$ is (eq. 12, p. 303):

$$
\begin{aligned}
V_{zf}(p_f, v_i) &= \eta(-p_f + u_{ift}) - \epsilon_{ift} \\
  &= -\eta p_f + \xi_f + \gamma_f + A_i \zeta_f + \sigma_f v_{if} + \sigma_n v_{in}
\end{aligned}
$$

where $$\eta$$ is a market-level price scaling factor, $$\xi_f + \gamma_f = \delta_f$$ is
firm $$f$$'s mean utility, $$\zeta_f$$ is an income-firm interaction parameter, and
$$\sigma_f$$, $$\sigma_n$$ are standard deviations of firm-specific and inside-good random
coefficients. Income group $$z$$'s choice probability (eq. 13, p. 303):

$$
P_{zf}(p) = E_v\!\left[ \frac{\exp(V_{zf}(p_f, v_i))}{1 + \sum_{k \in F_m} \exp(V_{zk}(p_k, v_i))} \right]
$$

Consumer surplus loss from firm $$f$$'s exit (eq. 15, p. 304):

$$
\Delta CS_f(X_0) = N \sum_z \mu_z\, g(z)\, T_m \left[ \widetilde{CS}_{zm}(p^{X_0}) - \widetilde{CS}_{zm}(p^{X_0 \setminus f})
  - \sum_f \gamma_f \left( P_{zf}(p^{X_0}) - P_{zf}(p^{X_0 \setminus f}) \right) \right]
$$

**Estimation moments (MSM, p. 305).** Three sets of micro-moments identify the
structural parameters ($$\Theta^m = \{\eta, \zeta, \sigma_f, \sigma_n\}$$):

Income-firm moments (informative about $$\zeta_f$$):

$$
g_f^{\text{inc}} = \left( \sum_i \omega_i \chi_{im} \right)^{-1} \sum_i \omega_i \chi_{im}
  \left[ (A_i F_{if} - B_i F_{if}) - \frac{\mu_A P_{Af}(p^0) - \mu_B P_{Bf}(p^0)}{1 - P_0(p^0)} \right]
$$
(eq. 16, p. 305)

Substitution moments (informative about $$\eta$$ and $$\sigma_f$$):

$$
g_f^{\text{sub}} = \left( \sum_i \omega_i \chi_{im} F_{if} \right)^{-1} \sum_i \omega_i \chi_{im} F_{if}
  \left[ H_{if} - \frac{P_f(p_f')}{P_f(p^0)} \right]
$$
(eq. 17, p. 305)

Outside-good moments (informative about $$\sigma_n$$):

$$
g^{\text{out}} = \left( \sum_i \omega_i \chi_{im} \right)^{-1} \sum_i \omega_i \chi_{im}
  \left[ O_i - \frac{1 - P_0(p')}{1 - P_0(p^0)} \right]
$$
(eq. 18, p. 305)

Parameters are estimated by minimizing $$G^m(\Theta^m)' G^m(\Theta^m)$$. Marginal
costs are backed out from Nash-Bertrand first-order conditions (eq. 10, p. 302):

$$
p_f - C'_f = \frac{q_f}{-\,\dfrac{\partial q_f(p)}{\partial p_f}}
$$

Counterfactual equilibrium prices $$p^X$$ are found by fixed-point iteration
(Conlon and Gortmaker 2020).

**Oil market (Section IV.D, pp. 308-309).** Oil is treated as an undifferentiated
globally traded commodity with price-taking firms. Market clearing (eq. 19):

$$
D(p^X) = S(p^X; F)
$$

Consumer surplus loss from firm $$f$$'s exit under linear demand (eq. 20):

$$
\Delta CS_f(X_0) = \tfrac{1}{2}\left( D(p^{X_0 \setminus f}) + D(p^{X_0}) \right)\left( p^{X_0 \setminus f} - p^{X_0} \right)
$$

**Labor markets (Section V, pp. 311-314).** Worker surplus per worker assuming
linear labor supply. Normalized surplus relative to outside option (eq. 22-23,
pp. 311-312):

$$
\frac{u_{ifl} + w_{ifl} - u_{i0} - w_{i0}}{w_{ifl}} = \frac{\epsilon_{ifl}}{\alpha x_{ifl}}, \qquad \epsilon \sim U(0,1)
$$

$$
E_i[ WS_{ifl} ] = \int_0^1 \frac{w_{ifl}\, \epsilon}{\alpha x_{ifl}}\, d\epsilon = \frac{w_{ifl}}{2 \alpha x_{ifl}}
$$

Total worker surplus loss from firm $$f$$'s exit (eq. 24, p. 312):

$$
\Delta WS_f = \sum_{l \in L_f} \sum_{i \in fl} \frac{w_{ifl}}{2 \alpha x_{ifl}}
$$

## Empirical specifications

**Labor supply regression (Table V, p. 313).** The labor supply arc elasticity
$$\alpha$$ is estimated from a linear probability model of whether workers leave if
their employer cuts salaries by 10% (eq. 27, p. 312):

$$
\Pr(L_i = 1) = (0.1\,\alpha)\, x_{ifl}
$$

where $$x_{ifl}$$ includes: annual earnings $$w_{ifl}$$ (from survey), college degree
indicator, major occupation indicators (management/business/science reference),
natural log of firm total employment in county (from InfoUSA), natural log of
labor market size (employment in occupation-county cell, from ACS), and a
constant. Standard errors in parentheses; n = 1,302 employed non-self-employed
respondents. Wages divided by 0.69 to convert from compensation to wages
(U.S. DOL 2023).

Key estimates from column (3): constant = 0.448 (SE 0.079)\*\*\*, total
compensation ($10,000) = -0.014 (SE 0.002)\*\*\*, college degree = -0.078
(SE 0.032)\*\*, ln(firm employees in county) = 0.025 (SE 0.006)\*\*\*;
$$R^2 = 0.064$$. (Table V, p. 313.)

**Product market estimation (Section IV.B-C, pp. 304-306).** MSM estimation per
market. Sample restricted to firms with at least 25 survey respondents as
customers; all other firms in market pooled as "other" firm. Baseline prices
$$p^0 = 1$$. Parameter identification: firm $$\zeta$$ estimated from share of purchases
by high- vs. low-income consumers (income-firm moments); market $$\eta$$ and $$\sigma_f$$
from share of customers who would still buy after a 25% price increase
(substitution moments); $$\sigma_n$$ from share of inside-good consumption retained
if all prices double (outside moments). Berry (1994) contraction mapping used
to match aggregate market shares in every iteration. (p. 305.)

**Externality and internality calibration (Section IV.F, Table IV, pp. 310-311).**
No regression; values imported from prior literature. Production externalities:
U.S. EPA supply-chain CO2 emission factors valued at $190/metric ton (U.S.
government social cost of carbon, 2020). Consumption externalities: beer $33.60
per liter of pure alcohol (Herrnstadt et al. 2015); cigarettes $0.64/pack
(DeCicca et al. 2020); soda 0.85 cents/oz; autos and oil include lifetime
CO2 emissions discounted at 3%. Cigarette internality $$= (1 - \beta) \times H^c$$ = (1 - 0.67) * $44.40 per pack $$\approx$$ $14.65/pack (Chaloupka et al. 2019; Gruber
and Koszegi 2001); soda internality 0.93 cents/oz (Allcott et al. 2019a).

**Robustness checks (Section VI.E, Figure 8, p. 320).** Six panels vary: social
cost of carbon doubled to $380/ton; cigarette internality halved; soda internality
doubled; more inelastic labor supply assumed. Main results are qualitatively
unchanged across all panels.

## When to read the full paper

Use the [original article](https://doi.org/10.1111/jofi.70004) if you are:
constructing your own social impact measure and need the full derivations
(Sections I-V); replicating the demand or labor surplus estimates (replication
code at allcott.stanford.edu/research); extending the framework to new
industries or markets; auditing a specific firm-level estimate (Internet
Appendix Table IA.V has all 74 firms); or comparing against the Harvard
Business School Impact Weighted Accounts methodology (Section VII). Where
Serafeim, Trinh, and Zochowski (2020) use accounting methods to monetize impact,
this paper uses demand estimation instead, yielding different results. The
locators above point to the exact table or figure.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 81(1), February 2026.
Copyright © 2025 the American Finance Association. Open access on the Wiley
platform under Wiley/AFA terms; no Creative Commons licence was confirmed
(Crossref returns only the Wiley VOR terms URL). This distillation was
extracted by an LLM on 2026-05-31 and is **not human-verified or
independently reproduced**. The PDF sidebar carries a "Creative Commons
License" watermark, but no CC attribution block appears in the article
and no CC URL appears in Crossref metadata; the rights signal is therefore
flagged as conflicted (`rightsSignalConflict: true`).

> Allcott, Hunt, Giovanni Montanari, Bora Ozaltun, and Brandon Tan.
> "An Economic View of Corporate Social Impact." *The Journal of Finance*
> 81, no. 1 (February 2026): 285–328. DOI: 10.1111/jofi.70004.
> © 2025 the American Finance Association.
> This page is an extract-only distillation by the Institute for Automated
> Research: core results re-expressed; **changes were made**. No verbatim
> PDF is mirrored here.


==============================================================================
# The Benefits of Access: Becht, Franks & Wagner (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2026/becht-private-meetings-portfolio-firms-2026/
# Distilled: Using GPT-4 to parse 4,700 private meeting notes from a large active asset manager and its UK portfolio firms (2007-2015), the paper shows that meetings convey predominantly soft information that is associated with fund-manager trading, generates risk-adjusted outperformance of 180 bps/month for a combined FM+GS meeting portfolio, and in only 0.4% of cases involves material nonpublic information. J. Finance 2026, CC BY 4.0. Ten core results with source locators, datasets used, the identification strategy, and the estimating specifications.
# Tags: paper-summary, corporate-governance, institutional-investors, asset-pricing
==============================================================================

**What this is.** The paper's core results, identification strategy, and estimating equations for how private meetings between an active asset manager and its UK portfolio firms affect trading: enough to know what was found and how, without reading all 51 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1111/jofi.13495).

## TL;DR

Using proprietary data from Standard Life Investments (SLI, the UK's largest active asset manager in 2015), the paper analyses 4,700 private meetings between the asset manager's fund managers (FMs) and governance specialists (GSs) and their FTSE All-Share portfolio firms over 2007-2015. GPT-4 is used to classify the information content of detailed meeting notes. The paper finds that meetings convey predominantly soft information (about management quality, strategy, and culture) rather than hard financial metrics, that this soft information is associated with fund-manager trading, and that meeting-informed portfolios generate risk-adjusted outperformance of about 180 basis points per month. Only 0.4% of meetings contain material nonpublic information (MNPI), and possession of MNPI is associated with no trading.

## Core results

Magnitudes and significance as reported; `\*\*\*`/`\*\*`/`\*` = 1%/5%/10%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **FM meetings trigger significant net buying**: funds increase positions by 3.2% per day in the [0,5]-day window around FM meetings | Table IV, Panel A, col.(1), p. 761 | FM_{0,5} coeff. = 3.197\*\*\* [0.660] |
| R2 | **GS meetings trigger significant net selling**: funds decrease positions by 2.3% per day in the same window | Table IV, Panel A, col.(1), p. 761 | GS_{0,5} coeff. = -2.272\*\*\* [0.727] |
| R3 | **Soft information drives both trading directions**: above-median soft FM meetings cause 1.8% net buying; above-median soft GS meetings cause 2.6% net selling; hard-information FM meetings cause larger buys of 7.5% | Table V, col.(1), pp. 764-766 | FM Soft_{0.5}(>=x-bar) = 1.810\*\*\* [0.772]; FM Hard_{0.5}(>=x-bar) = 7.526\*\*\* [1.373]; GS Soft_{0.5}(>=x-bar) = -2.566\*\*\* [0.827] |
| R4 | **High-quality FM meetings are associated with the strongest buy trades**: high-rating meetings generate 9.5% net position increase per day; low-rating meetings generate net selling of 5.5% | Table VIII, Panel A, col.(1), p. 771 | FM High Rating_{0,5} = 9.466\*\*\* [1.585]; FM Low Rating_{0,5} = -5.485\*\*\* [1.729] |
| R5 | **Consensus about meeting information is required to trade**: consensus FM meetings cause 3.4% net buying; no-consensus meetings do not trigger significant aggregate trading | Table IX, col.(1), p. 773 | FM Cons_{0,5} = 3.412\*\*\* [0.890]; FM No Consensus_{0,5} = 0.123 (n.s.) [1.801] |
| R6 | **Analyst upgrades combined with consensus generate the largest buy response** (15.5% per day); downgrades with consensus generate the largest sell response (-19.2% per day) | Table IX, col.(4), p. 773 | Upgrade_{0,5} x Consensus_{-30,0} = 15.48\*\*\* [3.983]; Downgrade_{0,5} x Consensus_{-30,0} = -19.15\*\*\* [2.871] |
| R7 | **Only 0.4% of meetings discuss MNPI; possession of MNPI is associated with no trading** | Figure 5; p. 775 | 17 of 4,700 meetings flagged as MNPI after manual review; all event-time coefficients around MNPI meetings are statistically indistinguishable from zero |
| R8 | **FM meeting long-short portfolio alpha is 198 bps/month**; GS meeting alpha is 83 bps/month (insignificant); combined all-meetings alpha is 180 bps/month | Table XI, Panel A, cols.(1)-(3), p. 778 | FM LS: Constant = 1.983\*\* (0.695); GS LS: 0.828 (n.s.) (0.495); All LS: 1.802\*\* (0.610); No-meeting LS: 0.493\*\*\* (0.105) |
| R9 | **Money made per position from active FM trading is 19 bps (fast) or 6 bps (slow)** in the [0,5]-day window; GS meetings generate 7 bps (fast) or -1 bps (slow) | Table XI, Panel B, p. 779 | FM: 19 bps fast / 6 bps slow; GS: 7 / -1 bps; All meetings: 15 / 4 bps |
| R10 | **FM meeting cumulative market-adjusted long-short return reaches 213 bps over 20 trading days and 314 bps over 40 trading days** after the meeting; GS meetings generate 121 bps / 128 bps; No-meeting stocks generate 30 / 52 bps | Figure 6, p. 780 | FM Meeting LS: 213 bps (day 20), 314 bps (day 40); GS Meeting LS: 121 / 128 bps; No Meeting LS: 30 / 52 bps |

**Overall (paper's conclusion).** Private meetings between active asset managers and portfolio firms convey predominantly soft information that significantly influences fund manager trading decisions and generates risk-adjusted outperformance. The information advantage from private access is based on soft, not material nonpublic, information. The paper informs the regulatory debate on the balance between transparency and the value of private, candid interactions between asset managers and portfolio firms.

## Theory / model

The paper has no formal structural model. It tests four interconnected hypotheses about the information economics of private meetings:

1. **Information content**: meetings convey soft information (non-quantitative, requiring interpretation) rather than hard information (quantitative, self-evident). The LLM-based classification following Liberti and Petersen (2019) operationalises this distinction.
2. **Trading response**: soft information obtained in FM meetings should generate buying (FM meetings are interpreted as positive on average), while GS meetings focusing on ESG governance risks should generate selling.
3. **MNPI absence**: if meetings convey material nonpublic information, funds should not trade at the time of those meetings (UK "parity of information" law prohibits trading while in possession of MNPI). The LLM is used to test whether MNPI meetings are associated with zero trading.
4. **Performance**: meeting-informed trades should earn positive risk-adjusted returns if private access provides an information advantage.

**Identification.** The key empirical challenge is that meetings may not cause trading; they may instead proxy for pre-existing concerns shared between FMs and GSs. The event-time plots (Figure 3, p. 763) address this: for FM meetings, trading spikes only on the meeting day, supporting a meeting-driven interpretation. For GS meetings, elevated trading precedes the meeting, consistent with shared concerns driving both the GS engagement and earlier FM trading. Meetings are timed relative to earnings announcements: FM meetings occur systematically after earnings reports (Figure IA.3), deliberately avoiding insider-information risk. GS meetings are scheduled independently of earnings dates (p. 758).

## Method

The paper combines three methodological building blocks, each with its own specification:

**LLM text classification.** Meeting notes (up to 2,200 words) are parsed using OpenAI's GPT-4 Turbo, GPT-4, and GPT-4o via the API. Separate prompts (reported in the Internet Appendix) instruct the model to act as a finance expert and assess: (i) the fraction of hard vs soft information in each note; (ii) the subject focus of soft information (firm, industry, market); (iii) the probability that a note discusses MNPI on a five-point Likert scale; and (iv) for FM notes with a "Conclusion" section, a no-consensus likelihood score. This builds on `text-classification` and `llm-text-classification` (proposed vocab). The LLM assessment of hard/soft information is validated against human-labelled random samples and representative quote examples (Table II).

**Panel regression (baseline trading).** The main estimating equation (p. 760, equation 1) is:

$$
\text{Trade}_{ijt} = \alpha + \beta' X_{it} + \gamma' Y_{ijt} + \Lambda + e_{ijt}
$$

- $$\text{Trade}_{ijt}$$ = daily % change in shares of company i held by fund j on day t (or a 0/1 indicator for sell/buy trade)
- $$X_{it}$$ = meeting indicators $$\text{FM}_{0,5}$$, $$\text{GS}_{0,5}$$ (= 1 if meeting in [0,5]-day window)
- $$Y_{ijt}$$ = controls: Ln(Shrout), Stake Held, Day Return
- $$\Lambda$$ = stock fixed effects + date fixed effects
- SE = clustered at fund and trading-day level
- Sample = 2007-2015, FTSE All-Share; restricted to positions with at least one FM meeting, at least one GS meeting, and at least one nonzero trade during the sample period

For the soft/hard information split (Table V), $$X_{it}$$ is replaced by $$\text{FM Soft}_{0,5}(\geq \bar{x})$$, $$\text{FM Hard}_{0,5}(\geq \bar{x})$$, $$\text{GS Soft}_{0,5}(\geq \bar{x})$$, and $$\text{GS Hard}_{0,5}(\geq \bar{x})$$, where $$\bar{x}$$ is the sample median LLM-assessed soft-information share. An alternative split at 50% is also used.

**Calendar-time portfolio performance (Table XI).** The long-short portfolio construction follows the methodology of Bradley, Jame, and Williams (2022), here extended with actual meeting notes and daily trade data. Monthly time-series Fama-MacBeth-style regressions on long-short portfolios:

$$
R^{\text{LS}}_{t} = \text{Constant} + \beta (\text{Mkt-Rf})_t + s\,\text{SMB}_t + h\,\text{HML}_t + u\,\text{UMD}_t + \varepsilon_t
$$

- $$R^{\text{LS}}_{t}$$ = monthly return on a long-short portfolio; Long: top tercile of buy-trade stocks after meeting day; Short: bottom tercile of sell-trade stocks after meeting day; Portfolios rebalanced daily; monthly returns compounded
- Factors = Fama-French (1993) + Carhart (1997) momentum, UK version from Gregory, Tharayan & Christidis (2013)
- $$\text{Constant}$$ = risk-adjusted alpha in % per month
- SE = robust standard errors; N = 108 monthly observations

## Empirical specifications

**Baseline meetings-and-trading (Table IV, Panel A).** The sample is the full unbalanced panel of fund-stock-day positions, 2007-2015 (N = 10,436,084 for the full-sample columns; 240,058 for the trades-only column). All specifications include stock and date fixed effects; standard errors are clustered at fund and trading-day level. Columns (1)-(2) use $$\text{Trade}$$ (net % change in shares) as the dependent variable; columns (3)-(4) use linear probability models for $$P(\text{Sell Trade})$$ and $$P(\text{Buy Trade})$$. Results are robust to longer [6,20]-day windows (Panel B). The trading-response estimates are about seven times larger than the corporate-jet-visit evidence in Bushee, Gerakos, and Lee (2018), which the paper attributes to its use of daily trade data rather than inferred trading.

**Soft and hard information by meeting type (Table V).** Same specification as Table IV but with meeting-type indicators split by LLM-assessed soft/hard content above and below the sample median. Separately estimated for FM and GS meetings. N = 240,058 (trades) and 10,450,063 (full panel).

**Meeting quality and sentiment (Table VIII, Panel A).** FM meetings are split by the analyst's human-assigned quality rating (Low 1-2, Average 3, High 4-5) and by the ratio of negative to positive words using the Loughran-McDonald (2011) dictionary (Neg Tone: ratio > 0.5; Pos Tone: ratio <= 0.5). Each indicator interacted with the [0,5]-day meeting window.

**Consensus and recommendation changes (Table IX).** For consensus, meeting notes are redacted of any recommendation language, then the LLM assesses the no-consensus likelihood. The interaction $$\text{Upgrade}_{0,5} \times \text{Consensus}_{-30,0}$$ combines a recommendation upgrade within [0,5] days with an FM meeting assessed as leading to consensus within the prior 30 trading days. Standard errors clustered at fund and trading-day level; stock and date fixed effects.

**MNPI event-time analysis (Figure 5, p. 777).** For the 17 confirmed MNPI meetings, event-time coefficients of the Table IV Panel A specification are estimated with event indicators from t-5 to t+5 around the MNPI meeting date. The dependent variable is the daily % change in shares held; no coefficient is statistically significant.

**Long-short portfolio performance (Table XI, Panel A).** Monthly OLS of long-short portfolio returns on the four-factor model. Long portfolio includes stocks with FM (or GS or both) meeting within 20 trading days in the top tercile of aggregate net buy trades on that day; short portfolio includes stocks in the bottom tercile of sell trades. Portfolios containing fewer than three stocks are replaced with the market return. N = 108 months (2007-2015).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| SLI private meeting notes (Standard Life Investments / abrdn) | 4,700 meeting notes (3,410 FM + 1,285 GS) with detailed text, attendee info, analyst ratings; the core proprietary dataset | No page yet |
| SLI fund holdings and daily trades | Daily fund-level positions and trades across ~40-50 funds covering FTSE All-Share stocks, 2007-2015 (10.45 million fund-stock-day observations) | No page yet |
| FTSE All-Share Index constituents and daily returns | Universe of eligible stocks, market capitalisation, daily return; used as controls and to define the sample universe | No page yet |
| SLI analyst recommendations | Internal Buy/Hold/Sell recommendations and recommendation changes (upgrades/downgrades) | No page yet |
| Loughran-McDonald (2011) financial dictionary | Positive and negative word lists for meeting-note tone measurement | No page yet |
| Fama-French + Carhart UK factors | Monthly size, value, and momentum factors for the UK, from Gregory, Tharayan & Christidis (2013); used for portfolio alpha estimation | No page yet |
| OpenAI GPT-4 Turbo/GPT-4/GPT-4o API | LLM text classification of meeting notes for soft/hard information content, MNPI likelihood, and consensus assessment | No page yet |

Sample: January 2007 to December 2015 (nine years). Universe: FTSE All-Share Index (353-703 firms at any given time). Meetings cover firms held at any point across 40-50 UK Equities Desk funds.

This study extends an earlier study of the same asset manager, Becht, Franks, Mayer, and Rossi (2009), which examined Hermes UK fund activism; here the analysis uses the SLI/abrdn proprietary meeting notes.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13495) if you are: studying the regulatory economics of private meetings between institutional investors and portfolio firms; building LLM-based classification of investment research or meeting content; examining whether active asset manager stewardship generates trading advantages; or investigating the empirical distinction between soft and hard information in financial markets. The Internet Appendix contains all LLM prompts, detailed data construction steps, and a timeline of individual firm meetings (Figure IA.2).

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 81(2). This distillation was extracted by an LLM on 2026-06-01 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Becht, Marco, Julian Franks, and Hannes F. Wagner.
> "The Benefits of Access: Evidence from Private Meetings with Portfolio Firms."
> *The Journal of Finance* 81, no. 2 (April 2026): 739-789.
> DOI: 10.1111/jofi.13495. (C) 2026 The Author(s).
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Adverse Selection in Corporate Loan Markets: Beyhaghi, Fracassi & Weitzner (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2026/beyhaghi-adverse-selection-corporate-loans-2026/
# Distilled: using confidential Federal Reserve Y-14Q supervisory data, this paper shows that more banks in a local market raises interest rates, borrower risk, and loan volume, consistent with adverse selection dominating competition effects; it also constructs a novel risk-orthogonalized markup measure and shows repeat-relationship markups and GSIB-shock evidence support the channel. J. Finance 2026, paywalled. Nine core results with source locators, datasets used, the theory tested, and the estimating specifications.
# Tags: paper-summary, adverse-selection, banking, corporate-loans, market-structure, information-asymmetry, panel-regression, difference-in-differences, peer-reviewed, unreplicated, data:fr-y14q, data:fdic-summary-of-deposits, data:bls, data:census
==============================================================================

**What this is.** The paper's core results, datasets, and theory: enough to
know what it found without reading all 46 pages. To replicate or extend it,
read the full source at the
[original DOI](https://doi.org/10.1111/jofi.70011) (paywalled).

## TL;DR

Using the Federal Reserve's confidential Y-14Q supervisory data on corporate
loans (21,924 loans, 2014Q4–2019Q4, 23 large bank holding companies), the
paper documents that markets with more banks have higher interest rates, riskier
borrowers (higher bank-assessed probability of default), and larger loan volume.
These patterns contradict standard competition models but are consistent with
adverse selection: riskier borrowers find it easier to obtain financing when
more banks compete, creating a winner's curse for lenders. The paper constructs
a novel markup measure by orthogonalizing the interest rate to the bank's
internal risk assessment, finds markups also rise with the number of banks,
and shows that firms staying with their existing bank pay 9 bp higher markups,
consistent with information holdup. A GSIB capital-surcharge shock confirms the
channel: counties with more GSIBs in 2015 saw falling interest rates and
borrower risk after surcharges reduced adverse selection.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%.
Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Bank risk assessments strongly predict loan interest rates** and loan performance; after controlling for PD and LGD, interest rates are orthogonal to future performance | Table III, p. 251; Table IV, p. 254 | PD coef 0.077\*\*\* (t=4.66), LGD coef 0.003\*\*\* (t=4.37); adj. R² rises from 0.52 to 0.55; interest-rate coef on nonperformance drops from 0.527\*\*\* to 0.101 (insig) when PD/LGD added |
| R2 | **More banks in a county raises interest rates** (adverse selection reverses standard competition effect) | Table V, p. 255 | coef 0.013\*\*\* (t=6.32) in col.(2); 1-SD increase (~6 banks) ≈ +7 bp vs. average credit spread ~150 bp; nonmonotonic U-shape (Fig. 2, p. 256): rates first fall then rise with bank count |
| R3 | **More banks raises borrower risk** (higher bank-assessed PD, conditional on observables) | Table VI, p. 257 | coef 0.011\*\*\* (t=4.97) in col.(2); PD increases monotonically with bank count (Fig. 3, p. 258); average PD = 134 bp, effect ≈ +1.1 bp per additional bank |
| R4 | **More banks raises total lending volume** (consistent with adverse selection and competition alike, but combined with R2–R3 supports adverse selection) | Table VII, p. 258 | coef 0.140\*\*\* (t=14.66) on log volume; 1 additional bank ≈ +14% volume; monotonically increasing (Fig. 4, p. 259) |
| R5 | **Novel markup measure also rises with number of banks**, consistent with adverse-selection-driven market power | Table IX, p. 265 | coef 0.012\*\*\* (t=5.84) in col.(2); markups U-shaped in bank count (Fig. 8, p. 266) |
| R6 | **Adverse selection effects are concentrated in low-tangibility firms**, for which private information problems are most severe | Table X, pp. 267–268 | Interest rate coef: high-tangibility 0.003 (insig), low-tangibility 0.013\*\*\* (t=3.77); PD coef: high-tangibility 0.000 (insig), low-tangibility 0.014\*\* (t=2.47); markup: high-tangibility insig, low-tangibility 0.011\*\*\* |
| R7 | **Firms rarely switch banks** (~75% stay with existing bank, roughly constant across market structures), consistent with information holdup | Fig. 6, p. 262 | Stay Bank rate ~75% across all bank-count bins; small economically insignificant reduction (Fig. 7, p. 263) |
| R8 | **Firms staying with their existing bank pay higher markups** (information rent extraction); holdup attenuated by prior multi-bank borrowing | Table XI, p. 271 | Stay Bank coef 0.090\*\*\* (t=3.64) in col.(2); interaction Stay Bank × N. of Prior Lenders: −0.052\*\*\* (t=3.50) |
| R9 | **GSIB capital surcharges (2016) reduced the number of banks, interest rates, and borrower risk** in high-GSIB counties; markup direction consistent but not significant | Table XII (reduced-form DiD), p. 273; Table XIII (2SLS), p. 278 | Reduced form: Post × N. of GSIBs(2015) coefs: N. banks −0.201\*\* (t=2.44), log volume −0.084\*\*\* (t=3.16), interest rate −0.031\* (t=1.96), PD −0.069\* (t=1.68); 2SLS: interest rate +0.179\* (t=1.85), PD +0.396\* (t=1.66) per additional bank |

**Overall (paper's conclusion).** Standard competition models fail to capture
the corporate loan market: more banks lead to riskier borrowers and higher
rates, not lower. Adverse selection gives informed incumbent banks market power,
which they exploit via information rents on repeat borrowers. Antitrust
policies that increase the number of banks in local loan markets may raise
rates and borrower risk as an unintended consequence.

## Theory / model

The paper has no original structural model. It tests predictions derived from
Broecker (1990) and related adverse-selection models (Riordan 1995, Marquez
2002, Dell'Ariccia and Marquez 2006). In these models, when lenders cannot
observe whether a borrower has been previously rejected by another bank, they
face a winner's curse: the pool of applicants worsens as the number of banks
in the market increases, because lower-quality borrowers have more chances to
find an approving lender. Two related mechanisms reinforce each other (p. 246):

1. As the number of banks rises, individual banks become more concerned that
   applicants have been screened and rejected elsewhere, raising the expected
   risk of any given applicant and thus the required interest rate.
2. More banks makes it harder for any one bank to infer whether a borrower
   has been rejected (Marquez 2002), further worsening adverse selection.

Both mechanisms predict that more banks leads to (i) higher average borrower
risk conditional on observables, (ii) higher interest rates, and (iii) higher
lending volume (pp. 246-247).

A second class of predictions concerns market power arising from information
advantages. Because incumbent banks learn about borrowers through lending
relationships, they hold private information that competitors lack (Sharpe
1990, Rajan 1992, Dell'Ariccia and Marquez 2006). This advantage lets them
charge above-marginal-cost rates to repeat borrowers who would be pooled with
riskier applicants if they tried to switch banks. More banks in the market
worsens adverse selection further and thus strengthens the incumbent's
information advantage and holdup capacity (p. 246).

**Tested predictions:**

1. More banks raises average bank-assessed PD, conditional on observables.
2. More banks raises interest rates (adverse-selection effect dominates competition).
3. More banks raises total lending volume.
4. Markups (interest rates net of risk) rise with bank count.
5. Repeat borrowers pay higher markups; the effect attenuates when borrowers
   have prior multi-bank relationships that reduce holdup.

The paper provides no welfare ranking and is fully empirical.

## Method

The paper applies three estimation strategies. All use OLS (or 2SLS) at the
loan level with multi-way fixed effects; no structural model is estimated.
The approach builds on `panel-regression`, `difference-in-differences`, and
`instrumental-variables`.

**Risk-assessment validation (Section II, pp. 250-254).** Before testing
market-structure predictions, the paper verifies that banks' internal PD and
LGD reports contain genuine private information. This is critical because the
markup measure relies entirely on the credibility of these assessments. The
validation runs two regressions (equations 1 and 2 in the PDF).

**Markup construction (Section IV, p. 264).** Markup is not an externally
observed price-cost margin; it is constructed residually. Equation (5) regresses
the interest rate on PD, LGD, their interaction, loan controls, bank x quarter
FEs, and industry x quarter FEs. The residual after partialling out the risk
components measures the portion of the interest rate unexplained by the
bank's own assessment of risk. This makes the measure internally consistent:
by construction, Table IV shows that interest rates do not predict loan
performance once PD and LGD are controlled for, so the residual markup
arguably reflects market power rather than unpriced risk.

**GSIB DiD/IV (Section VII, pp. 272-278).** To address endogeneity of the
number of banks, the paper uses a Bartik-style design (Goldsmith-Pinkham,
Sorkin, and Swift 2020): the pre-determined number of GSIBs in a county in
2015, interacted with a post-surcharge dummy, instruments for changes in the
number of lending banks. Following Favara, Ivanov, and Rezende (2021), the
paper treats the GSIB surcharge as a shock to large-bank lending costs and
uses it to build the DiD/IV instrument (p. 272). The first stage (column (1) of Table XII) shows that
more GSIBs in 2015 predicts a significant drop in the total number of banks
after 2016. The IV strategy identifies the adverse-selection channel by
exploiting a supply-side cost shock that affected GSIBs differentially across
counties.

## Empirical specifications

All regressions use loan $$l$$ in industry $$i$$ originated by bank $$b$$ in quarter $$t$$ as
the unit of observation unless noted. Standard errors are clustered by county
throughout.

**Spec (1): Risk assessment and interest rates (Table III, p. 251)**

$$
\text{IR}_l = \beta_0 \cdot \text{PD}_l + \beta_1 \cdot \text{LGD}_l + \beta_2 \cdot (\text{PD}_l \times \text{LGD}_l)
       + \Gamma X_l + \delta_{b,t} + \alpha_{i,t} + u_l \tag{1}
$$

- $$\text{IR}_l$$ is loan interest rate (%);
- $$\text{PD}_l$$ is bank-assessed probability of default (%);
- $$\text{LGD}_l$$ is loss given default (%);
- $$X_l$$ is a vector of loan controls (log maturity, log amount, guarantee dummy, loan purpose FE, loan type FE);
- $$\delta_{b,t}$$ is bank x quarter FE;
- $$\alpha_{i,t}$$ is industry x quarter FE.

Produces results R1 (Table III, p. 251).

**Spec (2): Risk assessments and loan performance (Table IV, p. 254)**

$$
y_l = \beta_0 \cdot \text{IR}_l + \beta_1 \cdot \text{PD}_l + \beta_2 \cdot \text{LGD}_l + \beta_3 \cdot (\text{PD}_l \times \text{LGD}_l)
      + \Gamma X_l + \delta_{b,t} + \alpha_{i,t} + u_l \tag{2}
$$

- $$y_l$$ is either nonperformance (dummy) or realized default (dummy);
- same controls and FEs as spec (1).

Produces results R1 (Table IV, p. 254).

**Spec (3): Market structure and interest rates (Table V, p. 255)**

$$
\text{IR}_l = \beta \cdot \text{NOB}_c + \Gamma_0 X_l + \Gamma_1 Z_{f,t}
       + \delta_{b,t} + \alpha_{i,t} + u_l \tag{3}
$$

- $$\text{NOB}_c$$ is the number of banks in county $$c$$;
- $$Z_{f,t}$$ is firm characteristics (log assets, leverage, tangibility, profitability).

Produces results R2 (Table V, p. 255). The same specification is used with $$\text{PD}$$ as the dependent variable for R3 (Table VI, p. 257).

**Spec (4): Market structure and loan volume (Table VII, p. 258)**

$$
\text{Volume}_{c,t} = \beta_0 \cdot \text{NOB}_c + \Gamma_0 X_{c,t} + \delta_t + u_{c,t} \tag{4}
$$

- $$\text{Volume}_{c,t}$$ is log total dollar loan volume (or log total number of loans) in county $$c$$ in quarter $$t$$;
- $$X_{c,t}$$ is county-level controls (log population density, log wages, log financial industry wages, log population);
- $$\delta_t$$ is year-quarter FE.

Produces results R4 (Table VII, p. 258). The same specification with blanket lien as the outcome produces collateral results (Table VIII, p. 261).

**Spec (5): Market structure and markups (Table IX, p. 265)**

$$
\text{IR}_l = \beta_0 \cdot \text{NOB}_c + \beta_1 \cdot \text{PD}_l + \beta_2 \cdot \text{LGD}_l + \beta_3 \cdot (\text{PD}_l \times \text{LGD}_l)
       + \Gamma_0 X_l + \Gamma_1 Z_{f,t} + \delta_{b,t} + \alpha_{i,t} + u_l \tag{5}
$$

This is spec (3) augmented with the risk measures as controls; the coefficient on $$\text{NOB}_c$$ now captures the markup effect (interest rate variation beyond risk). Produces results R5 (Table IX, p. 265).

**Spec (6): Switching banks and markups (Table XI, p. 271)**

$$
\text{IR}_l = \beta_0 \cdot \text{Stay Bank}_l + \beta_1 \cdot \text{PD}_l + \beta_2 \cdot \text{LGD}_l + \beta_3 \cdot (\text{PD}_l \times \text{LGD}_l)
       + \Gamma_0 X_l + \Gamma_1 Z_{f,t} + \delta_{b,t} + \alpha_{i,t} + \lambda_{c,t} + u_l \tag{6}
$$

- $$\text{Stay Bank}_l$$ equals one when the borrower stays with an existing bank;
- $$\lambda_{c,t}$$ is county x quarter FE (additional to bank x quarter FE);
- Sample restricted to firms with more than one loan.

Produces results R8 (Table XI, p. 271).

**Spec (7): GSIB surcharges - DiD reduced form and 2SLS (Tables XII-XIII, pp. 273-278)**

$$
y_{t,c} = \beta_0 + \beta_1 \cdot (\text{NOG2015}_c \times \text{Post}_t)
          + \Gamma_0 X_l + \Gamma_2 Z_{f,t} + \delta_{b,t} + \gamma_{b,c} + \alpha_{i,t} + u_{t,c} \tag{7}
$$

- $$\text{NOG2015}_c$$ is the number of GSIBs in county $$c$$ in 2015;
- $$\text{Post}_t$$ equals one for 2016 and later;
- $$\gamma_{b,c}$$ is bank x county FE;
- $$y_{t,c}$$ is in turns: number of banks, log loan volume, interest rate, PD, or markup.
- Sample period: 2014Q4-2019Q4 except column (1) of Table XII which uses 2015Q1-2019Q4 for the annual bank-count series.
- In the 2SLS version (Table XIII, p. 278), $$\text{NOG2015}_c \times \text{Post}_t$$ is used as an instrument for the annual number of banks in the county.

Produces results R9.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Federal Reserve Y-14Q (Schedule H.1) | Confidential supervisory loan-level data: interest rates, PD, LGD, loan characteristics, firm financials, 2014Q4-2019Q4, 21,924 loans from 23 BHCs | [FR Y-14Q](/wiki/confidential/fr-y14q/) (confidential) |
| FDIC Summary of Deposits | Alternative county-level bank count (branches), corroboration of main measure | [FDIC Summary of Deposits](/wiki/datasets/fdic-summary-of-deposits/) |
| Bureau of Labor Statistics (BLS) | County-level wage and financial-industry wages | [BLS](/wiki/datasets/bls/) |
| U.S. Census Bureau | County-level population estimates | [Census](/wiki/datasets/census/) |
| Zillow | County-level residential rent (robustness) | no page yet |

Sample: 21,924 new corporate loans (private, nonsyndicated borrowers) originated
2014Q4-2019Q4 by 23 U.S. bank holding companies. Median firm assets $23.6 mm,
median revenue $46 mm; median interest rate 3.66% (~150 bp credit spread).

## When to read the full paper

Read the source at [doi.org/10.1111/jofi.70011](https://doi.org/10.1111/jofi.70011)
if you are: replicating or extending the markup methodology (orthogonalizing
interest rates to bank internal risk assessments); studying adverse selection
identification in loan markets; using the GSIB surcharge as an instrument for
bank presence; or auditing a specific coefficient. The locators above point to
the exact table or figure.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 81(1), February 2026,
pp. 239–284. DOI: 10.1111/jofi.70011. © 2025 American Finance Association.
Paywalled; Wiley VOR licence (no CC grant). The Federal Reserve co-author's
contribution carries a U.S. Government public-domain notice under 17 U.S.C.
§105, which does not extend to the full article.

This distillation was extracted by an LLM on 2026-05-31 and is **not
human-verified or independently reproduced**. Extract-only: no PDF is hosted
here.

> Beyhaghi, Mehdi, Cesare Fracassi, and Gregory Weitzner. "Adverse Selection
> in Corporate Loan Markets." *The Journal of Finance* 81, no. 1 (February
> 2026): 239–284. DOI: 10.1111/jofi.70011.


==============================================================================
# Paying Too Much: Bhutta, Fuster & Hizmo (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2026/bhutta-mortgage-overpayment-borrower-sophistication-2026/
# Distilled: many U.S. mortgage borrowers significantly overpay relative to rates available in their market on the same day; overpayment is largest for FHA and low-FICO borrowers and rises when market interest rates are low; borrower sophistication (shopping and knowledge) strongly predicts lower rates and competition benefits sophisticated borrowers most. J. Finance 2026, paywalled. Eight core results with source locators, datasets used, the EGain model, and the key estimating specifications.
# Tags: paper-summary, mortgage-markets, consumer-finance, price-dispersion, borrower-sophistication, household-finance, panel-regression, peer-reviewed, unreplicated, data:optimal-blue, data:nsmo, data:mcr-nmls, data:hmda
==============================================================================

**What this is.** The paper's core results, datasets, and theory: enough
to know what it found without reading all 42 pages. To replicate or extend
it, access the full source at the
[original](https://doi.org/10.1111/jofi.70001) (paywalled).

## TL;DR

Using a proprietary platform dataset of 3.6 million rate-lock agreements
(Optimal Blue, Jan 2015 to Dec 2019) linked to real-time lender offer
distributions, the paper constructs an *Expected Gain from Additional
Search (EGain)* metric: the rate reduction a borrower could expect by
obtaining one more quote from a randomly drawn lender. FHA borrowers
average 28 bp of EGain; jumbo borrowers average 4 bp. Overpayment rises
when market interest rates are low, consistent with behavioral factors
reducing shopping effort. Substantial rate dispersion persists even within
the same lender, branch, and loan officer on the same day. Expensive
lenders earn higher profits: a 1 pp rate premium translates to $4.05 more
gross income per $100 originated, with no evidence of better service
quality. NSMO survey data show that a composite borrower sophistication
index (shopping and knowledge) predicts rates 23 bp lower for the most
vs. least sophisticated, and the benefit of lower market concentration
accrues primarily to sophisticated borrowers.

## Core results

Magnitudes and significance are as reported; `*`/`**`/`***` = 10%/5%/1%.
Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **FHA borrowers have the largest expected gain from additional search** | Table II, p. 66; Figure 2, p. 65 | Mean EGain = 28 bp for FHA vs. 17 bp conforming, 10 bp super-conforming, 4 bp jumbo; 25% of FHA borrowers have EGain above 40 bp |
| R2 | **Low-FICO and high-LTV borrowers overpay more, conditional on loan amount** | Table III, p. 67 | FICO>=740 vs. [640,660]: coef -0.097\*\*\* (col 1); adding lender-branch FEs reduces but does not eliminate the FICO gap |
| R3 | **EGain falls when market interest rates are higher** | Table IV, p. 70; Figure 3, p. 69 | A 1 pp rise in the 10-yr Treasury yield reduces average EGain by about 5 bp (coef -0.051\*\*\*, col 1); effect persists after lender-branch FEs (coef -0.026\*\*, col 3); stronger for unconstrained borrowers (DTI <= 36: interaction +0.012\*\*\*) |
| R4 | **Substantial residual rate dispersion remains after controlling for all observable characteristics, lender, branch, and LO** | Table V, p. 73; Table VI, p. 75 | 90-10 percentile gap: 55 bp raw (col 3); 26 bp after full lender, branch, and LO FEs (col 10); FHA 90-10 gap = 31 bp, jumbo = 24 bp in most restrictive spec; gap for same-branch same-day identical borrowers = 31 bp (col 8) |
| R5 | **Larger lenders and those with high FHA share are more expensive and more profitable** | Table VII, p. 79; Table VIII, p. 81 | Size quartile 4 lender FE: +0.127\*\*\* pp (col 1); FHA share coef: +0.385\*\*\* (col 2); 1 pp higher rate associated with $4.05\*\*\* more gross income per $100 originated (Table VIII) |
| R6 | **Expensive lenders have higher costs but higher net income; higher costs mostly reflect personnel, not service quality** | Table VIII, p. 81 | Gross expenses rise $3.50\*\*\* per $100 per 1 pp rate; net income rises $0.45\*\* (residential) and $0.24\*\* (all lines); technology and occupancy effects are small |
| R7 | **More expensive mortgages are not associated with better service; borrowers paying more are less satisfied** | Table IX, p. 84 | A 100 bp more expensive mortgage: satisfied with interest rate -13\*\*\* pp, satisfied with lender -2.3\*\*\* pp, application process -1.9\*\* pp; overall satisfaction coefficient -0.035\*\*\* |
| R8 | **Borrower sophistication (shopping and knowledge) strongly predicts lower mortgage rates; competition benefits sophisticated borrowers most** | Table X, p. 86 | Sophistication Index coef: -0.226\*\*\* (col 2); applied to 2+ lenders for better terms: -0.075\*\*\*; knows their interest rate: -0.060\*\*\*; interaction Sophistication x County HHI: +0.050\*\* (col 3), meaning lower concentration helps sophisticated borrowers more |

**Overall (paper's conclusion).** A large fraction of U.S. borrowers,
especially FHA and low-FICO borrowers targeted by government homeownership
programs, overpay for mortgages. Limited borrower sophistication provides
lenders with market power even in low-concentration markets. Overpayment
also impedes pass-through of monetary policy easing to mortgage rates.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Optimal Blue rate-lock and pricing-insight data (Jan 2015 to Dec 2019; ~3.6M locks; 20 MSAs for offer data) | Primary dataset: locked mortgage rates, lender offer distributions, loan/borrower characteristics; basis for EGain construction | No page yet |
| Mortgage Call Reports (MCR/NMLS), 2015:Q1 to 2019:Q4; 162 unique lenders | Nonbank lender income, expenses, and profitability; merged with Optimal Blue to study how rate premiums translate into lender margins and how competition moderates borrower overpayment | [MCR (NMLS)](/wiki/confidential/mcr-nmls/) |
| National Survey of Mortgage Originations (NSMO), waves 1-26, 2013-2019; 22,567 mortgages | Borrower sophistication, shopping behavior, knowledge, satisfaction; merged with administrative credit/servicing data | [NSMO](/wiki/datasets/nsmo/) |
| HMDA (Home Mortgage Disclosure Act) | Lender classification (bank vs. nonbank); market concentration (HHI) | No page yet |

Sample sizes: lock data 3.6M (full) / 67,537 (matched to offer data for EGain); dispersion analysis 2,996,149; MCR-Optimal Blue merge 162 lenders, 1,897 lender-quarters; NSMO 22,567 mortgages.

## Theory / model

The paper has no original structural model. It is an empirical paper that
tests predictions from the consumer-search and price-dispersion literature
(Carlson and McAfee (1983), Baye et al. (2006)) and from the literature on
limited financial sophistication in consumer finance (Woodward and Hall (2012),
Agarwal et al. (2015, 2017)).

The central identification object is the **Expected Gain from Additional
Search (EGain)**. Given $$n$$ lenders posting rates $$r_1 \leq r_2 \leq \cdots \leq r_n$$
for an identical borrower type on a given day, and a borrower who has already
found rate $$r_k$$, the expected gain from one more search is (eq. 1, p. 63):

$$
\text{EGain}_k = \sum_{i=1}^{k-1} (r_k - r_i) \times \frac{1}{n-1}
= \left[ r_k - \frac{\sum_{i=1}^{k-1} r_i}{k-1} \right] \times \frac{k-1}{n-1}
$$

Intuitively, the bracketed term is the locked rate minus the expected rate
among the $$k-1$$ cheaper lenders; this is scaled by $$\frac{k-1}{n-1}$$, the share
of remaining lenders who offer cheaper rates. The measure accounts for both
the borrower's position in the distribution and the width of the distribution.
Lenders that do not offer a given loan type are treated as $$r_i = \infty$$.

The main hypotheses tested:
- (H1) Borrowers with lower financial sophistication (less shopping, less
  knowledge) pay higher rates relative to available offers.
- (H2) Overpayment varies systematically with borrower type (FICO, LTV,
  loan program) even after controlling for risk-based pricing in offers.
- (H3) Market concentration interacts with borrower sophistication: lower
  concentration (HHI) benefits sophisticated borrowers more.
- (H4) Expensive lenders are more profitable but do not provide better service,
  implying pure market power rather than quality differentiation.

Identification for the EGain analysis rests on comparing each locked rate
to the real-time distribution of lender offers for an identical loan (same
MSA, same day, same FICO/LTV/program/amount/points), which is observed in
the Optimal Blue Pricing Insight data. For the dispersion and sophistication
analyses, identification uses within-market OLS with exhaustive fixed effects
(lender, branch, loan officer, MSA x month) and the NSMO survey linked to
administrative loan records.

## Method

The estimating strategy has three parts: (1) constructing EGain from matched
lock and offer data; (2) OLS regressions of EGain and locked rates on
borrower/lender characteristics with layered fixed effects; (3) NSMO-based
OLS regressions of contracted rates on borrower sophistication with
administrative controls.

**EGain construction.** For each rate-lock, the authors match to the
offer distribution on the same day in the same MSA for a loan with nearly
identical FICO, LTV, program, purpose, and amount. They adjust locked rates
for points paid using the empirical points-rate relationship. The EGain is
then computed from equation (1) above. The matched sub-sample is 67,537 loans
from the 20 MSAs with offer data.

**Rate-dispersion regressions (Table V, p. 73).** Locked rates are regressed
on increasingly rich sets of fixed effects. The outcome of interest is the
residual variance (measured as standard deviation and 90th-to-10th percentile
gap) after absorbing each specification:

$$
\text{Rate}_{ilt} = \beta \, X_{ilt} + \text{FE\_set} + \varepsilon_{ilt}
$$

- $$\text{FE\_set}$$ escalates across columns:
- col (1): Lock Date x MSA F.E.
- col (2): + FICO x LTV x Program x Lock Month F.E., ZIP Code F.E., Discount Points x Program x Lock Month F.E.
- col (3): + Discount Points x Program x Lock Month F.E. (finer)
- col (4): + Lender F.E.
- col (5): + Lender x Lock-Day x Program x LTV x Loan Month F.E.
- col (6): + Lender x FICO x LTV x Program x Lock Month F.E.
- col (7): + Branch F.E.
- col (8): + Branch x Lock Month F.E.
- col (9): + Loan Officer F.E.
- col (10): + Loan Officer x Lock Year F.E. x Program
- SE: two-way clustered by month and lender.

Standard errors are two-way clustered by month and lender. Sample: 2,996,149
loans locked 2015 to 2019, 30-year fixed-rate, fully documented, owner-
occupied, single-unit purchase mortgages in 277 MSAs.

**EGain-on-Treasury-yield regression (Table IV, p. 70).** The time-series
variation in overpayment is estimated as:

$$
\text{EGain}_{it} = \beta \cdot \text{TreasuryYield}_t + \gamma \, Z_{it}
+ \text{MSA F.E.} + [\text{MSA} \times \text{Month F.E.}] + [\text{Lender-Branch F.E.}]
+ \varepsilon_{it}
$$

- $$\text{TreasuryYield}_t$$: daily 10-year Treasury yield at the lock date.
- $$Z_{it}$$: FICO, LTV, loan amount controls and a DTI-below-36 dummy.
- $$\text{DTI} \leq 36$$ is interacted with Treasury yield to test whether affordability-constrained borrowers drive the relationship.
- SE: two-way clustered by month and lender.
- Sample: 67,241 matched loans.

## Empirical specifications

**EGain cross-section (R1, R2; Table II and III).** Cross-sectional
differences in EGain by loan program, FICO, and LTV are first shown in
summary statistics (Table II, p. 66), then confirmed in regressions:

$$
\text{EGain}_{it} = \beta_1 \, I_{\text{FICO\_bin}} + \beta_2 \, I_{\text{LTV\_bin}}
+ \gamma \cdot \text{LoanOfficerComp}_{it}
+ \text{Loan Amount F.E. (\$10k bins)}
+ \text{MSA} \times \text{Month F.E.}
+ [\text{Lender-Branch F.E.}]
+ \varepsilon_{it}
$$

- $$I_{\text{FICO\_bin}}$$: indicator dummies for FICO bins; omitted category [640, 660).
- $$I_{\text{LTV\_bin}}$$: indicator dummies for LTV bins; omitted category [60, 80].
- $$\text{LoanOfficerComp}_{it}$$: loan officer compensation controls.
- SE: two-way clustered by month and lender.
- Sample: 67,637 matched loans, 30-year fixed-rate purchase, 20 MSAs, 2016-2019.
- Key coefficients: FICO >= 740 vs [640,660) is -0.097\*\*\* (col 1); adding lender-branch FEs reduces but does not eliminate the FICO gradient (Table III).

**EGain and market rates (R3; Table IV, p. 70).** The core specification is:

$$
\text{EGain}_{it} = \beta \cdot \text{TreasuryYield}_t + \gamma \, Z_{it} + \text{MSA F.E.} + \varepsilon_{it}
$$

with columns progressively adding MSA x Month F.E. and Lender-Branch F.E.
Key result: a 1 pp rise in Treasury yield reduces EGain by about 5 bp
(col 1: -0.051\*\*\*); with lender-branch FEs: -0.026\*\* (col 3). This
builds on Fuster, Lo, and Willen (2024) and their time-varying
price-of-intermediation result: lenders also make worse offers when rates
are low, and borrowers are less likely to shop (p. 71).

**Rate dispersion (R4; Table V, p. 73).** The 10-spec rate dispersion
regression described in Method above. Key results: 90-10 gap = 55 bp in
spec (3) (pure observable controls); = 26 bp in spec (10) after full lender,
branch, and LO FEs. Same-branch same-day gap (spec 8) = 31 bp. This
contrasts with Alexandrov and Koulayev (2017), who find that negotiation
plays little role; the substantial within-lender within-branch-day
dispersion here is consistent with negotiation (p. 52).

**Lender expensiveness and profits (R5, R6; Tables VII-VIII).** Lender FEs
from spec (4) of Table V are the dependent variable in lender-level
regressions on size quartiles, nonbank indicator, and FHA share (Table VII,
p. 79). Lender expensiveness is then regressed on income/expense line items
from MCR filings in median regressions with year-quarter FEs (Table VIII, p. 81):

$$
\text{FinancialOutcome}_l = \beta \cdot \text{LenderExpensiveness}_l + \text{year-quarter F.E.} + \varepsilon_l
$$

Key: 1 pp higher rate corresponds to $4.05\*\*\* extra gross income and
$3.50\*\*\* extra gross expenses per $100 originated; net income rises
$0.45\*\* (residential originations, precorporate).

**Service quality (R7; eq. 2, Table IX, p. 84).** NSMO borrower survey
outcomes are regressed on contracted rate, with rich controls:

$$
Y_{ijtw} = \beta \cdot \text{Rate}_i + \Gamma \, Z_{ij} + \alpha_t + \delta_w + \varepsilon_{ijtw}
\tag{2}
$$

- $$Y_{ijtw}$$: binary satisfaction/delay indicator.
- $$\text{Rate}_i$$: contracted mortgage rate.
- $$Z_{ij}$$: credit-score and LTV flexible controls, county FEs, program indicators, income/employment/wealth/race/ethnicity controls, and likelihood-of-moving controls.
- $$\alpha_t$$: origination-month FEs.
- $$\delta_w$$: survey-wave FEs.
- SE: robust.
- Sample: 22,567 NSMO mortgages, 2013-2019.

**Borrower sophistication and rates (R8; eq. 3, Table X, p. 86).** NSMO
contracted rates are regressed on a sophistication index and market
concentration:

$$
\text{Rate}_{ijtw} = \beta \, X_i + \Gamma \, Z_{ij} + \alpha_t + \delta_w + \varepsilon_{ijtw}
\tag{3}
$$

- $$X_i$$: either (col 1) individual shopping/knowledge binary indicators or (col 2) a composite Sophistication Index (sum of six shopping/knowledge dummies divided by 6, range 0-1), or (col 3) Sophistication Index plus County HHI (last year) and their interaction.
- $$Z_{ij}$$, $$\alpha_t$$, $$\delta_w$$: the same rich controls as eq. (2).
- SE: robust.
- Sample: 22,567 (cols 1-2), 22,563 (col 3).
- Key results: Sophistication Index coef = -0.226\*\*\* (col 2); HHI x Sophistication interaction = +0.050\*\* (col 3), meaning lower concentration primarily benefits sophisticated borrowers.

## When to read the full paper

Access the [original](https://doi.org/10.1111/jofi.70001) if you are:
replicating (code available in the journal's Supporting Information);
studying the EGain measure construction in detail (Section III and Internet
Appendix Sections IV-V); reviewing robustness across loan programs and
lender types; or tracing the monetary policy transmission implications
(Section VI). The locators above point to the exact tables and figures.
For "what did this paper find," the table above is sufficient.

## Attribution & rights

Source: peer-reviewed, *The Journal of Finance* 81(1), February 2026,
pp. 49–90. © 2025 the American Finance Association. This distillation
was extracted by an LLM on 2026-05-31 and is **not human-verified or
independently reproduced**. The paper is paywalled; redistribution is
extract-only. Cite as:

> Bhutta, Neil, Andreas Fuster, and Aurel Hizmo. "Paying Too Much?
> Borrower Sophistication and Overpayment in the U.S. Mortgage Market."
> *The Journal of Finance* 81, no. 1 (February 2026): 49–90.
> DOI: 10.1111/jofi.70001.


==============================================================================
# Going Public and the Internal Organization of the Firm: Bias, Lochner, Obernberger & Sevilir (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2026/bias-going-public-internal-organization-2026/
# Distilled: German IPO firms become more hierarchical and standardized organizations in the two years before and during the IPO, adding management layers, narrowing control spans, expanding administrative functions, and standardizing job profiles. Hierarchy growth is more pronounced in firms with greater human capital risk. J. Finance 2026, CC BY 4.0. Eight core results with source locators, datasets used, the theory tested, and the estimating equations.
# Tags: paper-summary, ipo, organizational-economics, firm-structure, human-capital, labor-economics, difference-in-differences, panel-regression, peer-reviewed, unreplicated, open-access, cc-by, data:ieb-germany, data:orbis-bvd, data:sdc-platinum, data:iab-establishment-panel
==============================================================================

**What this is.** The paper's core results, datasets, theory, and estimating
equations: enough to know what it found and how without reading all 47 pages.
To replicate or extend it, read the original:
[DOI 10.1111/jofi.70012](https://doi.org/10.1111/jofi.70012).
The CC BY 4.0 licence permits mirroring; this batch does not host the PDF.

## TL;DR

Using administrative employment data matched to 312 German IPOs (1986–2015),
the paper tracks how firms change their internal hierarchies across an
eight-year window (five years pre-IPO to two years post-IPO) relative to
matched private-firm controls. IPO firms become more hierarchical in
preparation for listing: they add management layers (especially a new middle
management layer), narrow control spans, shift employment toward managerial
and administrative functions, hire finance, legal, and public-firm experts,
and standardize job profiles to align with industry job ladders. Approximately
40% of the hierarchical change is not explained by firm growth. Firms with
greater human capital risk undergo the largest hierarchical changes, consistent
with the theory that going public requires reducing dependence on key
individuals. Firms that withdraw their IPOs reverse these changes;
private-equity-backed firms show no analogous hierarchical transformation.

## Core results

Magnitudes and significance are as reported; `**`/`***` = 5%/1%. Locators
point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | IPO firms become **more hierarchical starting two years pre-IPO**, with changes concentrated in the anticipation and IPO year | Table III, p. 479; Figure 2, p. 478 | DiD coef on layers: +0.119\*\*\* at t-2, +0.202\*\*\* at t-1, +0.388\*\*\* at t (all vs. matched private controls; t-stats 3.1–7.3); 11.8% more hierarchical by IPO year (0.388/3.28 mean layers) |
| R2 | About **60% of hierarchical changes cannot be explained by firm growth**; the residual reflects going public per se | Table III col. (1) vs. (2), p. 479 | Without growth controls: +0.394 at t+2; with controls: +0.244 (7.4% more hierarchical); growth explains ~40% of the raw change |
| R3 | **Middle and top management (layers 3 and 4) grow disproportionately**; control spans narrow significantly | Table IV, p. 482 | Layer 3 employment +23.9% (p=0.002); layer 4 employment +18.9% (p=0.009) relative to controls; control span of layer 3 falls 23.3% (p=0.003), layer 4 falls 17.0% (p=0.015) |
| R4 | **Production share falls; administration and management shares rise** in IPO firms | Table V, p. 484 | Production share: -4.971pp\*\*\* at t+2 (10.3% relative decline vs. 48.1% mean share); administration: +2.708pp\*\*\* (p=0.003); management: +1.986pp\*\*\* (p=0.001); management share rises 31.5% relative to its t-3 base |
| R5 | **Finance, accounting, legal, and public-firm expert shares all rise** around the IPO | Table V cols. (4)–(6), p. 484 | Finance/accounting: +1.376pp\*\* at t+2 (53.1% relative increase, p=0.018); legal experts: +0.254pp\*\*\* (158.8% relative); public-firm experts: +2.779pp\*\* at t+2, peak +7.570pp\*\*\* at IPO year |
| R6 | **Firms with higher human capital risk experience larger hierarchical changes**, consistent with Rajan (2012) | Table VI Panel A, p. 486 | Triple-DiD: human-capital-intensive industry: interaction 0.202\* (p=0.054); highly skilled labor force: 0.261\*\*\* (p=0.009); R&D-intensive industry: 0.207\* (p=0.062) in post-IPO period |
| R7 | IPO firms **add job roles that align with industry-specific job ladders** (standardization); hierarchical changes correlate with more formalized internal processes | Table VII, p. 491; Figure 4, p. 489 | Promotion levels DiD: +0.196\*\*\* to +0.423\*\*\* across t+1/t+2 (3.8–8.3% increase relative to mean of 5.11; col. (1) without controls); with controls col. (2): +0.196\*\*\* at t+2, +0.256\*\*\* at t+1; more layers linked to written HR plans, job descriptions, performance reviews (all p<0.10) |
| R8 | **Withdrawn IPO firms build layers before withdrawal then reverse them**; PE-backed firms show no comparable hierarchical change beyond scale | Table VIII, p. 492; Table IX, p. 494 | Withdrawn firms: +0.183 at t-1 (p=0.123), -0.089 at t+2 (p=0.609); difference significant (F-stat 3.699\*). PE firms: all DiD estimates statistically indistinguishable from zero except borderline at t+2 (0.254\*, p~0.10) and fully explained by scale |

**Overall (paper's conclusion).** Firms reorganize to reduce dependence on key
individuals' human capital when transitioning to public markets. Most
organizational changes precede the listing date, occur in the two years before
the IPO, and are not fully explained by growth. The internal organization of a
firm is linked to its financing choices.

## Theory / model

The paper has no structural model and no formal estimation of structural
parameters. It derives testable predictions from two bodies of organizational-
economics theory and tests them empirically.

**Bolton and Dewatripont (1994) / Garicano (2000): firms as communication
networks (pp. 464–465).** In Bolton and Dewatripont (1994), firms are
information-processing networks that balance specialization and communication
costs. In Garicano (2000), employees in lower layers handle routine problems
while complex tasks escalate to specialized problem-solvers in upper layers.
Going public raises operational complexity (regulations, disclosure
requirements, investor scrutiny), making problems less predictable and
increasing the information load. The model predicts:

- BG.1: IPO firms increase the number of hierarchical layers.
- BG.2: IPO firms allocate more of the workforce to upper layers (more top-heavy),
  narrowing control spans.
- BG.3: IPO firms increase the proportion of specialized-role employees
  (finance, accounting, legal experts).
- BG.4: Hierarchical changes are greater in firms facing stricter listing
  standards (higher regulatory complexity).

**Rajan (2012): standardization to reduce key-person dependence
(pp. 465–466).** Early-stage firms rely on founders and key employees whose
human capital creates bargaining power and a risk to outside shareholders.
Going public requires transferring control to professional managers and
creating conventional job profiles that exist across firms in the same
industry and can be staffed with external recruits. The predictions are:

- R.1: IPO firms expand management capacity (top-heavy hierarchy, replaceable
  managers instead of founders/early employees).
- R.2: IPO firms grow administrative functions (personnel management, etc.).
- R.3: Hierarchical changes are more pronounced in firms with more valuable
  human capital (harder to replace).
- R.4: IPO firms align their hierarchies with industry-specific job ladders
  (jobs that exist at other firms in the same industry).

The two frameworks are treated as complementary (p. 467): both predict more
hierarchy; they differ in mechanism (complexity vs. key-person dependence).
The paper's goal is not to pit them against each other but to measure
organizational changes and assess which channel explains more variation.

**Identification logic (pp. 475–476).** The design is a stacked
difference-in-differences with matched never-treated private controls. The
identifying assumptions are (i) no treatment anticipation more than two years
before the IPO (preparation for a German IPO typically takes 12–24 months,
so $$t{-}3$$ and earlier are assumed clean); and (ii) conditional parallel trends:
matched controls would have followed the same hierarchy trajectory as IPO
firms in the absence of the IPO. Parallel trends cannot be tested directly
but is supported by near-zero, statistically insignificant pre-IPO period
coefficients at $$t{-}5$$ and $$t{-}4$$ (Table III).

## Method

The estimator is a stacked difference-in-differences (DiD) that builds on
`difference-in-differences` and `matching` as its technique primitives
(pp. 475–476). It is applied-method, not method-proposing; the stacked DiD
design follows Gormley and Matsa (2011) and Cengiz et al. (2019).

**Hierarchy measurement.** Three complementary measures of a firm's
hierarchical structure are constructed from German administrative occupation
codes (KldB1988):

1. `layers`: the number of hierarchical levels a firm has, where each
   occupation is mapped to one of four layers (layer 1 = production/blue-collar;
   layer 2 = supervisors/experts; layer 3 = middle management; layer 4 = top
   management/directors). Follows Caliendo, Monte, and Rossi-Hansberg (2015)
   and Gumpert, Steimer, and Antoni (2021).

2. `refined layers` (log): sublayers within each of the four layers, formed
   by clustering employees by wages within each layer (Bonhomme, Lamadon, and
   Manresa (2022)). Captures within-layer heterogeneity. Mean of 5.89 for IPO
   firms and 5.68 for controls at $$t{-}3$$ (Table II, p. 474).

3. `promotion levels`: industry-specific job-ladder positions derived from
   within-firm occupational transitions associated with wage increases, building
   on Huitfeldt et al. (2023). Captures the alignment of a firm's hierarchy
   with the industry job ladder.

**Sample construction (pp. 472–473).** Starting from 888 German IPOs (1984–2016,
compiled from SDC, Deutsche Boerse website, Bloomberg, and a manual dataset
from Christoph Kaserer at TU Munich), the paper links each IPO to employment
data via Orbis (BvD) firm identifiers and the ADIAB linking table to the
Betriebs-Historik-Panel. The matching algorithm has two steps: (1) match
each IPO to up to 20 private firms in the same industry with the most similar
size, age, employment growth, and mean wage three years before the IPO; then
(2) restrict to controls with the same layer structure as the IPO firm at
$$t{-}3$$. This yields 312 matched pairs (312 IPO firms and 312 matched private
firms, 4,992 firm-years in the main panel).

## Empirical specifications

The main specification is the dynamic stacked DiD (Equation 1, p. 476):

$$
y_{f,t,c} = \alpha
  + \sum_{k=-5}^{-4} \beta_k \cdot \mathbf{1}(\text{IPO})_f \cdot \mathbf{1}(t{+}k)_{f,t}
  + \sum_{k=-2}^{0}  \beta_k \cdot \mathbf{1}(\text{IPO})_f \cdot \mathbf{1}(t{+}k)_{f,t}
  + \sum_{k=1}^{2}   \beta_k \cdot \mathbf{1}(\text{IPO})_f \cdot \mathbf{1}(t{+}k)_{f,t}
  + \phi_f + \psi_{t,c} + \Pi \cdot X_{f,t} + \epsilon_{f,t}
$$

- $$f$$ = firm, $$t$$ = year, $$c$$ = cohort of IPO firms going public in the focal year.
- $$\mathbf{1}(\text{IPO})_f$$ is an indicator for the IPO (treated) firm.
- $$\mathbf{1}(t{+}k)_{f,t}$$ is an indicator for calendar year $$t$$ being $$k$$ years relative to the firm's IPO year.
- The omitted reference period is $$t{-}3$$ (the matching year).
- $$\phi_f$$ = firm fixed effects (equivalent to cohort-by-firm fixed effects).
- $$\psi_{t,c}$$ = cohort-by-year fixed effects.
- $$X_{f,t}$$ = controls: log number of layer-1 employees and log number of establishments.
- Standard errors are clustered at the firm level.

The dependent variable $$y_{f,t,c}$$ varies by table:
- **Table III**: `layers` (cols. 1-2) and $$\log(\text{refined layers})$$ (cols. 3-4),
  the extensive margin of hierarchy. Columns (2) and (4) add firm-growth
  controls.
- **Table IV**: $$\log(\text{employment in layer } L)$$ for layers 2, 3, 4 and
  $$\log(\text{control span of layer } L)$$ for layers 2, 3, 4, the intensive margin.
- **Table V**: employment share in production/service, administration,
  management, finance/accounting, legal experts, and public-firm experts.
- **Table VII**: `promotion levels` (industry-specific job-ladder alignment).
- **Tables VIII and IX**: the same specification estimated on withdrawn-IPO
  and PE-investment samples respectively.

**Triple-DiD specification for heterogeneity (Table VI, p. 486).**
To test whether human capital risk or regulatory complexity moderates the
main effect, the paper adds a triple interaction:

$$
y_{f,t,c} = \alpha
  + \sum_{\text{period}} \beta_k \cdot \mathbf{1}(\text{IPO})_f \cdot \mathbf{1}(\text{period})_{f,t}
  + \sum_{\text{period}} \gamma_k \cdot \mathbf{1}(\text{IPO})_f \cdot \mathbf{1}(\text{period})_{f,t} \cdot \mathbf{1}(\text{Split})_f
  + \delta \cdot \mathbf{1}(\text{Split})_f \cdot \mathbf{1}(\text{period})_{f,t}
  + \phi_f + \psi_{t,c} + \Pi \cdot X_{f,t} + \epsilon_{f,t}
$$

- $$\mathbf{1}(\text{Split})_f$$ is an indicator for higher human capital risk (human-capital-intensive industry, highly skilled labor force, R&D-intensive industry) or for higher IPO proceeds or listing in a more regulated market segment.
- The coefficient of interest is $$\gamma_k$$ in the post-IPO period.

**Formalization cross-section (Figure 4, p. 489, footnote p. 490).**
Using the IAB Establishment Panel survey, the paper regresses 10 standardization/
reorganization indicators on number-of-layers dummies, controlling for
establishment and year fixed effects:

$$
\mathbf{1}(\text{standardization/reorganization})_{j,t}
  = \alpha
  + \beta_1 \cdot \mathbf{1}(\text{firm has two layers})_{j,t}
  + \beta_2 \cdot \mathbf{1}(\text{firm has three layers})_{j,t}
  + \beta_3 \cdot \mathbf{1}(\text{firm has four layers})_{j,t}
  + \phi_j + \psi_t + \epsilon_{j,t}
$$

- $$j$$ = establishment.
- $$\phi_j$$ = establishment fixed effects.
- $$\psi_t$$ = year fixed effects.
- The omitted category is firms with one layer.
- Vertical bars in Figure 4 are 95% confidence intervals from heteroskedasticity-consistent standard errors clustered at the establishment level.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Integrated Employment Biographies (IEB), German Institute for Employment Research (IAB) | Administrative employee-level data: occupational codes, wages, layering, functions, tenure; main source for all hierarchy measures | No page yet |
| Orbis (Bureau van Dijk) | Firm identifier linking IPO list to employment data; firm characteristics | [Orbis (BvD)](/wiki/commercial/orbis-bvd/) (licensed) |
| Thomson Reuters Securities Data Corporation (SDC) | German IPO list (888 IPOs, 1984–2016) | No page yet |
| Bloomberg database | IPO list compilation (supplementary to SDC) | [Bloomberg](/wiki/commercial/bloomberg/) (licensed) |
| Deutsche Boerse AG website | IPO list compilation | No page yet |
| Manually collected data (Christoph Kaserer, TU Munich) | German IPO identifiers | No page yet |
| IAB Establishment Panel | Representative survey of German establishments; used to test whether hierarchical changes correlate with formalization of internal processes | No page yet |
| Bureau van Dijk Orbis / VentureSource | PE growth investment sample (71 firms) | [Orbis (BvD)](/wiki/commercial/orbis-bvd/) (licensed) |

Sample: 312 IPO firms and 312 matched private-firm controls; 4,992 firm-years in the main panel (eight-year window, $$t{-}5$$ to $$t{+}2$$). IPOs span 1986–2015; primarily Manufacturing (34.9%), Information and Communication (21.8%), Wholesale and Retail Trade (15.3%), and Professional, Scientific, and Technical Activities (14.4%).

## When to read the full paper

Read the original if you are: studying organizational economics of IPOs or
corporate governance; extending the hierarchy-measurement methodology
(layers, refined layers, promotion levels); using the German IPO sample or
IAB/IEB data; examining the standardization-via-job-ladders mechanism; or
auditing a specific coefficient. The locators above point to the exact tables.
For "what did this paper find," the table above is the intended default.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 81(1), February 2026. This
distillation was extracted by an LLM on 2026-05-31 and augmented on 2026-06-01;
it is **not human-verified or independently reproduced**. The article is
published under CC BY 4.0, which permits mirroring; the PDF is not hosted in
this batch.

> **Attribution (CC BY 4.0).** Bias, Daniel, Benjamin Lochner, Stefan
> Obernberger, and Merih Sevilir. "Going Public and the Internal Organization
> of the Firm." *The Journal of Finance* 81, no. 1 (February 2026): 459–505.
> DOI: 10.1111/jofi.70012. © 2025 The Author(s). Licensed under
> [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). Open access
> funding enabled and organized by Projekt DEAL.
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Monetary Policy and Wealth Effects: Caramp & Silva (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2026/caramp-monetary-policy-wealth-effects-2026/
# Distilled: In a heterogeneous-agent New Keynesian model with rare disasters and heterogeneous beliefs (D-HANK), monetary contractions raise risk premia and redistribute wealth from optimists to pessimists; the resulting time-varying precautionary motive accounts for roughly 60% of the aggregate consumption response, the wealth effect for 30%, and the standard intertemporal-substitution channel for less than 10%. J. Finance 2026, CC BY 4.0. Eight core results with source locators, datasets used, the model, and the method with its defining equations.
# Tags: paper-summary, monetary-policy, asset-pricing, macro, heterogeneous-agents
==============================================================================

**What this is.** The paper's core results, the D-HANK model, and the analytical
method (aggregation via a market-implied disaster probability) with the defining
equations: enough to understand what was found and how, without reading all
42 pages. To replicate or extend, read the full source at the
[original](https://doi.org/10.1111/jofi.70021).

## TL;DR

Caramp and Silva build D-HANK: an analytical heterogeneous-agent New Keynesian
model that adds rare aggregate disasters and heterogeneous beliefs about disaster
risk to the HANK setting of Kaplan, Moll, and Violante (2018), who find only a
minor role for the standard ISE. A contractionary monetary shock redistributes wealth from
optimistic to pessimistic savers, raising the market-implied disaster
probability and risk premia on stocks and bonds. This time-varying precautionary
motive accounts for roughly 60% of the aggregate consumption response on impact;
the aggregate wealth effect (via government bond revaluation) accounts for 30%;
and the standard intertemporal-substitution effect (ISE) accounts for less than
10%. The model also matches the forward-curve dynamics documented by Hanson and
Stein (2015), the equity premium, and the response of corporate spreads, all
without requiring a high elasticity of intertemporal substitution. The paper
derives conditions (Proposition 4) under which risk premia have no real effects
("risk-premium neutrality"), and shows that belief heterogeneity is necessary to
break this neutrality.

## Core results

Magnitudes and significance are as reported. Locators point into the source PDF
(pages are printed page numbers, beginning at 1011).

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Time-varying precautionary motive (TVP) accounts for roughly 60% of the initial output response to a 100 bps monetary shock (estimated fiscal backing solution) | Figure 5, Panel A, p. 1039 | Output drops 1.15% on impact; TVP component accounts for ~60%; aggregate wealth effect (GE factor) accounts for ~30%; ISE accounts for less than 10% |
| R2 | Heterogeneous beliefs amplify the consumption response by more than 3x relative to the homogeneous-beliefs economy with disaster risk | Figure 6, Panel B, p. 1040 | Homogeneous-beliefs output drop: ~0.35%; heterogeneous-beliefs output drop: ~1.15% (more than three times larger) |
| R3 | Model generates a 100 bps monetary shock that raises the five-year yield by 32 bps and matches the full forward curve estimated by Hanson and Stein (2015) | Figure 3, Panel B, p. 1037 | Calibrated psi_lambda = 0.57 (half-life ~4 months); eps_lambda = 315 (33 bps change in disaster probability per 25 bps shock) |
| R4 | Model produces a 4.0% drop in equity prices in response to a 100 bps increase in interest rates, driven primarily by the risk premium rather than the discount-rate channel | Figure 4, Panel C, p. 1038 | Stock price decline: 4.0%; consistent with Bernanke and Kuttner (2005) point estimate |
| R5 | Corporate spread rises by 11 bps in response to a 100 bps shock; VAR estimate is 6.5 bps (SE 3.1) | Figure 4, Panel B, p. 1038 | Model spread response: 11 bps; excess bond premium (EBP) from VAR: 6.5 bps, SE 3.1 |
| R6 | Introducing long-term risky household debt raises the output drop by 48 bps (from ~1.15% to ~1.6%) relative to D-HANK without household debt; the TVP channel accounts for roughly half, the wealth effect for ~40% | Figure 7, Panels A-B, p. 1043 | Output drop with HH debt (estimated fiscal): ~1.6%; baseline without HH debt: ~1.15%; additional drop: ~48 bps |
| R7 | Standard RANK model with high EIS (sigma=1) generates a 1.1% output drop via wealth effect amplified by GE, but requires counterfactually large implied fiscal backing (more than 20x the empirical estimate) | Figure 8, Panels A-B, p. 1044 | RANK MSV sigma=1: output drops 1.1%; RANK MSV sigma=4: output drops 0.1% (about 11x smaller); D-HANK with estimated fiscal: ~1.15% at sigma=4 |
| R8 | Risk-premium neutrality (Proposition 4) holds exactly when fiscal backing is independent of risk premia; heterogeneous beliefs break this condition since belief differences create differential wealth effects even at the same real interest rate | Proposition 4, p. 1027; Figure 1, p. 1028 | Two economies with lambda_o < lambda_p: identical [y_t, pi_t, i_t] paths despite larger asset-price drop in the heterogeneous-beliefs economy; output difference = 0 by construction |

**Overall (paper's conclusion).** In the D-HANK model, the standard
intertemporal-substitution channel of textbook New Keynesian models plays only a
minor role. Aggregate risk combined with heterogeneous beliefs generates large
time-varying precautionary motives that dominate the transmission of monetary
shocks. The model can simultaneously match the equity premium, the term premium,
the forward-curve dynamics, and the real effects of monetary shocks at a low EIS,
a combination that RANK and standard HANK models cannot achieve.

## Theory / model

The D-HANK model is a continuous-time heterogeneous-agent New Keynesian model.
The economy is populated by three types: workers $$w$$, optimistic savers $$o$$,
and pessimistic savers $$p$$, with masses $$\mu_w$$, $$\mu_o$$, $$\mu_p$$ summing to 1.
Savers invest in short-term bonds, long-term government bonds, and corporate
equity. They have heterogeneous subjective beliefs $$\lambda_j$$ ($$j \in \{o,p\}$$)
about the Poisson arrival rate of aggregate disasters, with $$\lambda_o \leq \lambda_p$$.
The aggregation uses the heterogeneous-beliefs framework of the risk-centric model
of demand recessions in Caballero and Simsek (2020).

**Savers' problem.** Each saver $$j$$ maximizes (p. 1017):

$$
V_{j,t}(B_{j,t}) = \max E_j \!\left[ \int_t^{t^*} e^{-\int_t^z \rho_{j,u}\,du}
                    \frac{C_{j,z}^{1-\sigma}}{1-\sigma}\, dz
                    + e^{-\int_t^{t^*} \rho_{j,u}\,du} V^*_{j,t^*}(B^*_{j,t^*}) \right]
$$

subject to the flow budget constraint:

$$
dB_{j,t} = \left[(i_t - \pi_t)B_{j,t} + r_{L,t}B^L_{j,t} + r_{E,t}B^E_{j,t}
             + T_{j,t} - C_{j,t}\right]dt + \left[B^*_{j,t} - B_{j,t}\right]dN_t
$$

where $$i_t$$ is the nominal rate, $$\pi_t$$ is inflation, $$r_{L,t}$$ and $$r_{E,t}$$
are excess returns on long-term bonds and equities conditional on no disaster,
$$T_{j,t}$$ are transfers, and $$N_t$$ is the Poisson disaster process with
arrival rate $$\bar{\lambda}$$.

**Euler equations.** The Euler equation for short-term bonds (eq. 1, p. 1018):

$$
\dot{C}_{j,t} / C_{j,t} = \sigma^{-1}(i_t - \pi_t - \rho_{j,t})
                          + (\lambda_j / \sigma) \left[(C_{j,t}/C^*_{j,t})^{\sigma} - 1\right]
$$

The first term is the standard ISE; the second captures the precautionary
savings motive from disaster risk. The Euler equations for long-term bonds
and equities (eqs. 2-3, p. 1018):

$$
r_{L,t} = \lambda_j (C_{j,t}/C^*_{j,t})^{\sigma} \cdot (Q_{L,t} - Q^*_{L,t})/Q_{L,t}
$$

$$
r_{E,t} = \lambda_j (C_{j,t}/C^*_{j,t})^{\sigma} \cdot (Q_{E,t} - Q^*_{E,t})/Q_{E,t}
$$

where the risk premium equals the price of disaster risk times the quantity of
risk (the relative loss in asset value in the disaster state).

**Market-implied disaster probability (Proposition 1, p. 1019).** With
heterogeneous beliefs, the economy aggregates as if a representative saver holds
a CES-weighted belief:

$$
\lambda_t = \left[ \frac{\mu_o C_{o,t}}{\mu_o C_{o,t} + \mu_p C_{p,t}} \lambda_o^{1/\sigma}
           + \frac{\mu_p C_{p,t}}{\mu_o C_{o,t} + \mu_p C_{p,t}} \lambda_p^{1/\sigma} \right]^{\sigma}
$$

and $$\eta_t = e^{-\int_0^t \rho_{s,z}\,dz} C_{s,t}^{-\sigma}$$ is a valid SDF. This
aggregation result is the key: the heterogeneous economy behaves as a
representative-agent model with an endogenous, time-varying disaster probability
that responds to monetary shocks via wealth redistribution. This extends the
redistribution-via-risk-premia channel of Kekre and Lenel (2022) into a full New
Keynesian model with analytical aggregation.

**New Keynesian Phillips Curve and interest rate rule (eq. 11, p. 1023; eq. 6, p. 1020):**

$$
\dot{\pi}_t = (\rho_s + \lambda) \pi_t - \kappa y_t, \qquad \kappa = \phi^{-1}(\varepsilon-1)\phi Y
$$

$$
i_t = r_n + \phi_{\pi} \pi_t + u_t
$$

**Aggregate Euler equation (Proposition 2, eq. 10, p. 1023):**

$$
\dot{y}_t = -\bar{\sigma}^{-1}(i_t - \pi_t - r_n) + \chi_{p_d} p_{d,t}
$$

$$
\bar{\sigma}^{-1} = \frac{1 - \mu_w}{1 - \mu_w \chi_y} \sigma^{-1}
$$

$$
\chi_{p_d} = (\lambda/\bar{\sigma}) (C_s/C^*_s)^{\sigma}
$$

$$
p_{d,t} = \sigma(c_{s,t} - c^*_{s,t}) + \hat{\lambda}_t \qquad \text{(price of disaster risk)}
$$

The aggregate EIS is amplified by the cyclicality of income inequality $$\chi_y$$.
The extra term $$\chi_{p_d} p_{d,t}$$ connects aggregate risk and asset prices
to real output.

**Wealth effect decomposition (Proposition 6, eq. 26, p. 1033).** The prior paper
by the same authors, Caramp and Silva (2023), decomposes output into the ISE and
the wealth effect; the D-HANK adds aggregate risk and heterogeneity to that
framework. Output decomposes into three components:

$$
y_t = \bar{\sigma}^{-1} \hat{y}_{m,t}                                 \quad \text{(ISE)}
    + \chi_{\lambda} \hat{y}_{\lambda,t}                               \quad \text{(time-varying precautionary motive)}
    + (\rho - \omega) e^{\omega t} \Omega_0                            \quad \text{(GE factor} \times \text{aggregate wealth effect)}
$$

where $$\Omega_0$$ is the aggregate wealth effect at impact, $$\chi_{\lambda} =
\chi_{p_d} \varepsilon_{\lambda}$$, and the GE factor $$(\rho - \omega)$$ ensures that $$\int_0^{\infty}
e^{-\rho t}(\rho-\omega)e^{\omega t}\,dt = 1$$, so the wealth effect shifts output in
all periods by $$\rho \Omega_0$$, amplified in general equilibrium.

## Method

The paper contributes two methodological innovations that allow analytical
aggregation in a setting with heterogeneous portfolios.

**Approximate block-recursivity (Proposition 3, p. 1025).** The market-implied
disaster probability $$\hat{\lambda}_t$$ and relative net worth $$b_{p,t} - b_{o,t}$$
can be solved independently of output and inflation if the effect of savers'
aggregate consumption $$c_{s,t}$$ on risk premia is second-order (the term
$$r_k \sigma c_{s,t}$$ is $$O(\|i_t - r_n\|^2)$$). Under this approximation:

$$
\hat{\lambda}_t = e^{-\psi_{\lambda} t} \hat{\lambda}_0
$$

$$
\hat{\lambda}_0 = \varepsilon_{\lambda} (i_0 - r_n)
$$

where $$\varepsilon_{\lambda} \geq 0$$ and is strictly positive if and only if $$\lambda_p > \lambda_o$$.
The persistence parameter $$\psi_{\lambda} = \xi$$ (the speed of reversion in Uzawa
preferences). The initial price of risk depends linearly on the initial monetary
shock, with coefficient $$\varepsilon_{\lambda}$$ that captures the pass-through of nominal
rates to the disaster probability via wealth redistribution.

**Four-equation system.** Combining the aggregate Euler equation (10), the
NKPC (11), the Taylor rule (6), and the price of risk equation (19):

$$
p_{d,t} = \bar{\sigma} y_t + e^{-\psi_{\lambda} t} \hat{\lambda}_0
$$

the system has the same structure as the textbook three-equation NK model but
with an additional term connecting asset prices to aggregate dynamics. The
system is solved analogously to `buildsFrom: value-function-iteration` (matrix
eigendecomposition of the $$2 \times 2$$ system in $$[y_t, \pi_t]$$), with the unstable root
solved forward and the stable root backward.

**Wealth effect formula (eq. 29, p. 1035).** The aggregate wealth effect can
be written entirely in terms of policy variables:

$$
\Omega_0 = \frac{\rho - \omega}{(\rho - \omega)\chi_{\tau} + \bar{d}_G \kappa}
           \left[ \int_0^{\infty} e^{-\rho t} \Delta B^L_t (i_t - r_n + r_L \hat{\lambda}_t)\,dt
               - \bar{d}_G \int_0^{\infty} e^{-\rho t} \hat{\pi}_t\, dt
               - \int_0^{\infty} e^{-\rho t} \tau_t\, dt \right]
$$

where $$\Delta B^L_t = (1 - e^{-\psi_L t})\bar{d}_G$$ is the portfolio exposure to
long-term bonds, and $$\tau_t$$ is the fiscal backing (taxes on savers). This
expression shows that equity revaluations do not affect $$\Omega_0$$ (the term
$$r_E$$ drops out), only government bond revaluations do.

**Calibration method.** Parameters are disciplined by four moments: (i) natural
interest rate $$r_n = 1\%$$; (ii) equity premium of 7.0% in stationary equilibrium
(implies $$\sigma = 4$$); (iii) initial five-year yield response of 32 bps per 100
bps monetary shock (from a four-lag VAR on 1962-2007 US data); (iv) the entire
forward curve estimated by Hanson and Stein (2015), used to pin $$\varepsilon_{\lambda}$$ and
$$\psi_{\lambda}$$. The fiscal backing $$\tau_t$$ is estimated from the VAR impulse
responses for government revenues and expenditures.

## Empirical specifications

The paper is primarily a structural theory paper; the empirical content consists
of VAR-based calibration moments and model-vs-data comparisons, not standalone
reduced-form regressions.

**VAR for fiscal and interest-rate dynamics (Section III.A, p. 1036).** A
four-lag VAR is estimated on quarterly US data 1962:Q1 to 2007:Q3. Variables
(in order): real GDP per capita, CPI inflation, real consumption per capita,
real investment per capita, capacity utilization, hours worked per capita, real
wages, tax revenues/GDP, government expenditures/GDP, federal funds rate,
five-year constant maturity rate, real value of government debt/GDP. The
recursiveness assumption identifies the monetary shock: the federal funds rate
is ordered third to last, five-year rate and government debt last. Bootstrapped
68% confidence bands are reported (Figure 2). The VAR is used to estimate the
fiscal backing $$\tau_t$$ and the interest-rate impulse response that the model
then matches.

**Forward-curve calibration (Figures 3-4, pp. 1037-1038).** The response of
forward rates to a 25 bps change in the two-year yield around FOMC meetings is
taken from Hanson and Stein (2015). The model forward curve is derived by
solving the partial differential equation (PDE) for bond prices (Appendix S1,
Internet Appendix Section III) and compared to the data-based forward rates.
The model matches the long-horizon forward-rate response that standard models
cannot explain.

**Model-vs-data checks (Figures 4-7, pp. 1038-1043):**

- Corporate spread: model predicts 11 bps rise per 100 bps shock, matching the
  VAR-based estimate of Gertler and Karadi (2015) of 6.5 bps (SE 3.1 bps) for the
  excess bond premium. Untargeted.
- Equity price: model predicts 4.0% drop; point estimate from Bernanke and
  Kuttner (2005) is the comparison benchmark. Untargeted.
- Output decomposition: model output drop (~1.15%, estimated fiscal) is
  compared against Miranda-Agrippino and Ricco (2021) estimate of ~1.15%.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| US macroeconomic VAR data (FRED/BEA/BLS) | Calibration of fiscal backing and interest-rate dynamics via VAR; variables include real GDP, CPI, consumption, investment, capacity utilization, hours, wages, tax revenues, govt expenditures, federal funds rate, five-year CMT rate, real govt debt (1962:Q1-2007:Q3) | [FRED](/wiki/datasets/fred/) |
| Hanson & Stein (2015) forward-rate estimates | Calibration targets for the forward curve and persistence of risk premia | No page yet |
| Bernanke & Kuttner (2005) equity-price estimates | Untargeted model comparison for equity price response | No page yet |
| Gilchrist & Zakrajsek (2012) excess bond premium | Untargeted model comparison for corporate spread response | No page yet |

Sample for VAR: quarterly, 1962:Q1 to 2007:Q3. Model calibration uses
stationary equilibrium moments (equity premium 7.0%, credit spread 200 bps,
debt-to-income ratio 10%, duration 5 years).

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.70021) if you are:
building or extending HANK models with aggregate risk; studying the
transmission mechanism of monetary policy through asset prices; seeking the
proofs of the eight propositions and lemma (Appendix, pp. 1045-1049);
calibrating the term premium or credit spread to monetary shocks; or extending
the model to richer capital structures or full quantitative HANK settings.
The Internet Appendix contains the forward-curve PDE, additional robustness
(sticky wages, investment, wealthy hand-to-mouth households), and the mapping
between $$\varepsilon_{\lambda}$$ and underlying belief parameters.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 81(2), April 2026.
This distillation was extracted by an LLM on 2026-06-01 and is **not
human-verified or independently reproduced**. The CC BY 4.0 licence permits
mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Caramp, Nicolas, and Dejanir H. Silva.
> "Monetary Policy and Wealth Effects: The Role of Risk and Heterogeneity."
> *The Journal of Finance* 81, no. 2 (April 2026): 1011-1052.
> DOI: 10.1111/jofi.70021. (c) 2026 The Author(s).
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# What Drives Investors' Portfolio Choices: Choukhmane & de Silva (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2026/choukhmane-portfolio-choices-risk-preferences-2026/
# Distilled: using 401(k) default-fund quasi-experiments, Choukhmane and de Silva show that 94% of retirement investors prefer stock market participation absent frictions and estimate relative risk aversion of 2.54, EIS of 0.25, and a $156 portfolio adjustment cost. J. Finance 2026, CC BY 4.0. Eight core results with source locators, datasets used, the model, and the method.
# Tags: paper-summary, household-finance, portfolio-choice, stock-market-participation, life-cycle, frictions, default-effects, risk-preferences, structural, panel-regression, peer-reviewed, unreplicated, open-access, cc-by, data:401k-admin, data:wrds, data:sipp, data:scf
==============================================================================

**What this is.** The paper's core results, datasets, and theory: enough
to know what it found without reading all 44 pages. To replicate or extend
it, use the [original](https://doi.org/10.1111/jofi.70013).

## TL;DR

Using quasi-experimental variation in 401(k) default asset allocations
(money-market fund vs. target-date fund) across 4 million employees at
hundreds of thousands of plans (Dec 2006-Dec 2017), the paper separates
investors' underlying risk preferences from participation frictions. Absent
frictions, 94% of retirement investors prefer stock market participation
and the average preferred equity share is 76%, declining with age: patterns
broadly consistent with standard life-cycle portfolio choice models. These
preferences differ markedly from observed allocations, where participation
and equity shares are lower and hump-shaped. A life-cycle model estimated
via SMM recovers moderate risk aversion ($$\gamma = 2.54$$), EIS = 0.25, and a
$156 portfolio adjustment cost. Low stock market participation in retirement
accounts is driven by one-time frictions, not nonstandard preferences.

## Core results

Magnitudes and significance are as reported. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Lower bound on fraction preferring stock market participation** is 42%; upper bound is 95% | §II.C.1, Fig. 4 (p. 22), Table IAII | At tenure 2 years: bounds 78%–95% (money-market-to-TDF sample); lower bound rises as more investors reveal preferences over tenure |
| R2 | **Point estimate: 94% of investors prefer stock market participation** in their retirement accounts (under Assumption 6) | §II.C.3, Fig. IA9 | At tenure = 3: average preferred participation = 94%; average preferred stock share = 76%; stable over life cycle |
| R3 | **Preferred equity share is high (>60% at all ages) and declining with age** (opposite of the observed hump-shaped profile) | §II.C.4, Fig. 6 (p. 26), right panel | Preferred share ~80% at age 25, declining to ~60% at age 60; observed share hump-shaped and strictly below preferences at all ages |
| R4 | **Observed participation and equity shares diverge from preferences**; TDF-auto-enrolled investors' choices most closely approximate friction-free preferences | Fig. 7 (p. 27) | SCF 2007-16 stock share: 27%; not-auto-enrolled 401(k): 40%; auto-enrolled TDF: 80%; friction-free preference estimate: 76% |
| R5 | **Baseline structural estimate: relative risk aversion $$\gamma = 2.54$$** (EZW model, SMM on 38 moments) | Table III col. (1), p. 41 | $$\gamma = 2.54$$ (SE 0.09); discount factor $$\beta = 0.94$$ (SE 0.001); EIS $$\sigma^{-1} = 0.253$$ (SE 0.018) |
| R6 | **Portfolio adjustment cost = $156; contribution adjustment cost = $488** | Table III col. (1), p. 41 | $$k_\theta$$ = $156 (SE $6.01); $$k_s$$ = $488 (SE $16.60); contribution cost larger, consistent with frictions in DC plan enrollment as additional driver of nonparticipation |
| R7 | **Without frictions, risk-aversion estimates are implausibly heterogeneous**: $$\gamma = 18.94$$ using money-market-default data alone vs. $$\gamma = 2.25$$ using TDF-default data alone | Table III cols. (3) and (4), p. 41 | Same population (employees hired within 12 months of the same policy change), frictionless model produces $$\gamma$$ 18.94 vs. 2.25 depending on which half of the data is used; baseline model reconciles both with $$\gamma = 2.54$$ |
| R8 | **Treatment group (TDF default) maintains ~95% stock market participation and ~80% equity share** throughout tenure; control group (money market default) starts near 0% and converges over years | Fig. 2 (p. 16), Table IAII | Treatment-control gap: 19-25 pp in participation rate, 20-23 pp in stock share of retirement wealth; convergence is gradual, inconsistent with pure time-dependent (Calvo) frictions |

**Overall (paper's conclusion).** Participation frictions, not nonstandard
risk preferences such as loss aversion or ambiguity aversion, are the
primary driver of limited stock market participation in retirement accounts.
Investors' true preferences align with standard life-cycle models once
one-time adjustment costs are accounted for.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| 401(k) administrative records (large U.S. record-keeper, anonymized), Dec 2006–Dec 2017, ~4 million employees, >600k plans | Main data: portfolio allocations, participation, contribution rates, plan defaults | No page yet (proprietary; no public access) |
| CRSP Value-Weighted Index (1925–2006) | Equity premium and return volatility calibration (6.4% premium, 20% vol) | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| Survey of Income and Program Participation (SIPP) | Labor income process estimation, employment transition probabilities | [SIPP](/wiki/datasets/sipp/) |
| Survey of Consumer Finances (SCF), 2007 and 2016 waves | External validation of financial wealth, stock market participation benchmarks | No page yet |

Sample: ~4 million employees, more than 600,000 401(k) plans, ages 25-64,
2006-2017. Quasi-experiment #1 (money-market-to-TDF): 1,086 control + 1,321
treatment investors at 6 firms. Quasi-experiment #2 (opt-in-to-TDF): 40,337
control + 52,400 treatment investors at 191 firms.

## Theory / model

The paper combines two frameworks: a nonparametric revealed-preference
approach and a structural life-cycle model. The structural side builds on
the survey of life-cycle portfolio choice models in Gomes (2020) as the
benchmark framework.

**Revealed-preference setup (Section II, pp. 19-24).** Individual $$i$$ has
unobserved preferred participation $$Y^*_{it} \in \{0,1\}$$ and preferred equity
share $$\theta^*_{it} \in [0,1]$$ at tenure $$t$$. Observed allocations $$Y_{it}$$,
$$\theta_{it}$$ may differ because of inertia or frictions. The 401(k) plan
carries a default $$D_i \in \{0,1\}$$ ($$D=1$$ means TDF default; $$D=0$$ means money
market or opt-in). Consistency indicators are (p. 19):

$$
C^Y_{it} = 1 \text{ if } Y_{it}(0) = Y_{it}(1), \quad 0 \text{ otherwise.}
$$

$$
C^\theta_{it} = 1 \text{ if } \theta_{it}(0) = \theta_{it}(1), \quad 0 \text{ otherwise.}
$$

Four identifying assumptions (pp. 20-22):

$$
\text{Assumption 1 (Frame Separability):} \quad (Y^*_{it}, \theta^*_{it}) \text{ independent of } D_i.
$$

$$
\text{Assumption 2 (Frame Exogeneity):} \quad D_i \text{ independent of } (Y_{it}(0), Y_{it}(1), \theta_{it}(0), \theta_{it}(1)).
$$

$$
\text{Assumption 3 (Frame Monotonicity):} \quad Y_{it}(1) \geq Y_{it}(0), \quad \theta_{it}(1) \geq \theta_{it}(0).
$$

$$
\text{Assumption 4 (Consistency):} \quad C^Y_{it} = 1 \Rightarrow Y_{it} = Y^*_{it}; \quad C^\theta_{it} = 1 \Rightarrow \theta_{it} = \theta^*_{it}.
$$

Proposition 1 (p. 21) shows the average preference for participation is
partially identified under Assumptions 1-4:

$$
\mathbb{E}_\tau(Y^*_{it}) \in [\mathbb{E}_\tau(Y_{it} \mid D_i=0),\ \mathbb{E}_\tau(Y_{it} \mid D_i=1)]. \tag{1}
$$

For the average equity share (continuous variable), a fifth assumption is
needed. Assumption 5 (p. 22): an investor who deviates from the default
chooses her preferred share (consistent with fixed-cost models). Proposition 2 (p. 22):

$$
\mathbb{E}_\tau(\theta^*_{it}) \geq \mathbb{E}_\tau(\theta_{it} \mid D_i=0).
$$

For point identification, Assumption 6 (p. 23): preferences of consistent
(active) and inconsistent (passive) investors are uncorrelated at any tenure:

$$
\text{cov}_\tau(Y^*_{it},\, C^Y_{it}) = \text{cov}_\tau(\theta^*_{it},\, C^\theta_{it}) = 0.
$$

Under Assumption 6, Proposition 3 (p. 23) gives:

$$
\mathbb{E}_\tau(Y^*_{it}) = \mathbb{E}_\tau(Y^*_{it} \mid C^Y_{it} = 1) - \frac{1}{\mathbb{E}_\tau(C^Y_{it})} \cdot \text{cov}_\tau(Y^*_{it},\, C^Y_{it}) \tag{2}
$$

$$
\mathbb{E}_\tau(\theta^*_{it}) = \mathbb{E}_\tau(\theta^*_{it} \mid C^\theta_{it} = 1) - \frac{1}{\mathbb{E}_\tau(C^\theta_{it})} \cdot \text{cov}_\tau(\theta^*_{it},\, C^\theta_{it}) \tag{3}
$$

where the first term is the preferences of consistent investors (identified
as their observed active choices, eq. 4) and the second is a selection bias
term. Under Assumption 6 the bias is zero and preferences equal the active
investors' average choices.

**Life-cycle model (Section III, pp. 29-36).** Investors have Epstein-Zin-Weil
recursive preferences. The estimated preferences are broadly consistent with
the CRRA life-cycle model of Merton (1969) and its prediction of an equity
share that declines with age. The value function for a retired investor (state
vector $$X_t$$) satisfies (p. 35):

$$
V_t = \max_{d^{dc}_t,\, s^l_t,\, \Xi_t} \left\{ (1-\beta) n_t \left[ \frac{c_t - k_\theta \mathbf{1}\{\Xi_t \neq \Xi_{d,t}\}}{n_t} \right]^{1-\sigma} + \beta \left[ m_t \mathbb{E}_t V^{1-\gamma}_{t+1} \right]^{(1-\sigma)/(1-\gamma)} \right\}^{1/(1-\sigma)}
$$

subject to: (10), (11), (12), (14), (15), (17), and budget constraint

- $$\gamma$$ = relative risk aversion
- $$\sigma^{-1}$$ = EIS
- $$\beta$$ = discount factor
- $$n_t$$ = equivalence scale
- $$k_\theta$$ = portfolio adjustment cost (utility units)
- $$m_t$$ = survival probability

For the working life (employment states $$E$$ or $$JJ$$), an additional contribution adjustment cost
$$k_s$$ is incurred when $$s^{dc}_t \neq s_{d,t}$$ (p. 36):

$$
V_t = \max_{s^{dc}_t,\, s^l_t,\, \Xi_t} \left\{ (1-\beta) n_t \left[ \frac{c_t - k_\theta \mathbf{1}\{\Xi_t \neq \Xi_{d,t}\} - k_s \mathbf{1}\{s^{dc}_t \neq s_{d,t}\}}{n_t} \right]^{1-\sigma} + \beta \left[ m_t \mathbb{E}_t V^{1-\gamma}_{t+1} \right]^{(1-\sigma)/(1-\gamma)} \right\}^{1/(1-\sigma)}
$$

subject to: (7), (8), (10), (11), (12), (14), (15), (17), and

$$
s^{dc}_t \cdot w_t + s^l_t = w_t - c_t - \text{tax}_i(y^{\text{tax}}_t).
$$

Three financial assets: risk-free bond (gross return $$R_f$$), risky stock
(log return process, p. 31):

$$
\ln R^S_t = \ln R_f + \mu_s + \epsilon_t, \quad \epsilon_t \sim N(0,\, \sigma^2_s). \tag{10}
$$

Liquid savings account (p. 31):

$$
L_{t+1} = (L_t + s^l_t)[1 + r(1 - \tau_c)], \quad L_0 = 0. \tag{11}
$$

Labor income follows an AR(1) process with a deterministic cubic-in-age
component (pp. 30-31):

$$
\ln w_t = \delta_0 + \delta_1 a_t + \delta_2 a^2_t + \delta_3 a^3_t + \eta_t, \tag{7}
$$

$$
\eta_t = \rho \eta_{t-1} + \xi^E_t, \quad \xi^E_0 \sim N(0,\, \sigma^2_{\xi_0}), \quad \xi^E_t \sim N(0,\, \sigma^2_\xi) \text{ for } t>0.
$$

Default options follow: at a new job, defaults are the employer's settings
$$\theta^j_{e}$$ (portfolio) and $$s^{dc}_{e}$$ (contribution rate); in
subsequent periods the default is the prior period's choice (eqs. 14-16,
pp. 33-34).

## Method

The estimation has two stages. The first stage sets demographics,
income-process parameters, asset returns, and tax/benefit rules outside the
model using auxiliary data and calibration (pp. 36-38). The second stage
estimates five preference parameters by SMM (Simulated Method of Moments):
$$\beta$$, $$\gamma$$, $$\sigma^{-1}$$, $$k_\theta$$, $$k_s$$ (pp. 36, 39-40).

The life-cycle model is solved by standard numerical discrete-time
dynamic programming. The state vector has 10 dimensions (age, labor
productivity, employment status, employer identity, tenure, average lifetime
earnings, DC retirement savings, liquid savings, default portfolio shares,
default contribution rate). Controls are: consumption, portfolio shares for
new and accumulated DC assets, DC contribution rate, DC withdrawal, and
liquid savings.

The SMM objective minimizes the weighted squared distance between
model-simulated and empirical moments (pp. 39-40):

$$
\min_{\theta} (\hat{m} - m(\theta))' W (\hat{m} - m(\theta))
$$

- $$\hat{m}$$ are 38 empirical moments
- $$m(\theta)$$ are their model counterparts simulated on 7,500 investors (approximately five times the estimation sample size)
- $$W$$ is the inverse covariance matrix of the empirical moments (i.e., the optimal SMM weight matrix), estimated via Erickson-Whited (2002) to avoid bootstrapping-weight-matrix bias

The method builds on Epstein-Zin-Weil preferences (separating $$\gamma$$
from $$\sigma^{-1}$$, following Epstein and Zin (1989) and Weil (1990)) and
on revealed-preference bounds (Goldin and Reck (2020) extended to
continuous shares here).

**Key identifying variation for each parameter:**
- $$k_\theta$$ and $$k_s$$: identified by bunching at default options at various
  tenure levels; the degree of bunching pins down adjustment cost size.
- $$\gamma$$: identified primarily by asset allocation decisions of consistent
  (active) investors who deviate from the default.
- $$\sigma^{-1}$$: identified by bunching at the employer match threshold (6%
  of salary cap), following Best et al. (2020) and Choukhmane (2025).
- $$\beta$$: identified by the overall level of retirement contributions.

## Empirical specifications

The paper does not use OLS regressions with fixed effects as the primary
estimating procedure. Instead, two reduced-form exercises and one SMM
structural estimation generate the results.

**Quasi-experimental comparison (R8, R1, pp. 13-17).** For each sample
(money-market-to-TDF and opt-in-to-TDF), the estimating object is the
difference in observed portfolio outcomes between investors hired within 12
months before versus after the 401(k) default asset allocation change at
the same set of firms:

$$
\text{Outcome}_{it} = f(\text{tenure}_t,\, D_i), \quad D_i = \mathbf{1}[\text{hired after default change at firm } e]
$$

No regression equation is reported; the result is the raw time-path of
stock market participation $$Y_t$$ and equity share $$\theta_t$$ by years of
tenure for treatment vs. control (Figure 2, p. 16). Standard errors are
clustered by investor (quasi-experiment #1) or by firm (quasi-experiment #2).
Sample: money-market-to-TDF: 1,086 control + 1,321 treatment at 6 firms;
opt-in-to-TDF: 40,337 control + 52,400 treatment at 191 firms.

**Nonparametric preference bounds and point estimates (R1-R4, pp. 21-27).**
The estimating objects are the tenure-specific conditional means from
Propositions 1-3 applied to the quasi-experimental data (p. 25):

$$
\mathbb{E}_\tau(Y^*_{it}) = \mathbb{E}_\tau(Y_{it} \mid Y_{it} \neq D_i,\, \text{age}=A) \tag{5}
$$

$$
\mathbb{E}_\tau(\theta^*_{it}) = \mathbb{E}_\tau(\theta_{it} \mid \theta_{it} \neq \theta^d_i(D_i),\, \text{age}=A) \tag{6}
$$

Standard errors are clustered by investor (quasi-experiment #1) and by
firm (quasi-experiment #2). Life-cycle preference profiles by age are
displayed in Figures 5 and 6 (pp. 25-26). The point estimates (Assumption 6)
are displayed in Figure IA9: average preferred participation = 94%, average
preferred equity share = 76%, at tenure = 3.

**SMM estimation (R5-R7, pp. 39-41).** Targeting 38 moments: 14 stock
market participation rates by tenure from the money-market-to-TDF
quasi-experiment; 16 average stock shares by age for each default group at
end of first tenure year; and 8 contribution rate distribution moments
from the opt-in-to-TDF quasi-experiment. Table III (p. 41) reports the
five preference parameters with standard errors:

$$
\beta = 0.940\ (\text{SE}\ 0.001), \quad \gamma = 2.54\ (\text{SE}\ 0.09), \quad \sigma^{-1} = 0.253\ (\text{SE}\ 0.018),
$$

$$
k_\theta = \$156\ (\text{SE}\ \$6.01), \quad k_s = \$488\ (\text{SE}\ \$16.60).
$$

Robustness: column (2) imposes CRRA ($$\sigma = \gamma$$); column (3) zeros
adjustment costs and uses only control-group moments (yields $$\gamma = 18.94$$);
column (4) zeros adjustment costs and uses only TDF-default moments (yields
$$\gamma = 2.25$$). The stark contrast between columns (3) and (4) is the
paper's key identification claim for the role of frictions (R7).

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.70013) if you are:
replicating the nonparametric bounds or the SMM estimation; extending the
life-cycle model to brokerage accounts or international settings; reviewing
the Internet Appendix robustness (peer effects, passive rebalancing,
survivorship, cohort/year effects, income conditioning); or auditing a
specific coefficient from Table III. The locators above point to the exact
table or figure.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 81(1). This distillation
was extracted by an LLM on 2026-05-31 and is **not human-verified or
independently reproduced**. The paper is CC BY 4.0 and mirroring is
permitted; no PDF mirror has been set up in this batch.

> **Attribution (CC BY 4.0).** Choukhmane, Taha, and Tim de Silva.
> "What Drives Investors' Portfolio Choices? Separating Risk Preferences
> from Frictions." *The Journal of Finance* 81, no. 1 (February 2026):
> 5–48. DOI: 10.1111/jofi.70013. © 2025 The Author(s). Licensed under
> [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Asset Pricing and Risk-Sharing under DB vs DC Pensions: Coimbra, Gomes, Michaelides & Shen (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2026/coimbra-pension-plans-asset-pricing-2026/
# Distilled: a general equilibrium model with an explicit defined-benefit pension fund matches the historical equity premium and riskless rate better than a standard PPG model; a shift to defined-contribution plans raises the riskless rate, lowers the Sharpe ratio, increases retiree consumption volatility and decreases worker consumption volatility. J. Finance 2026, CC BY 4.0. Eight core results with source locators, datasets used, and the model equations and method.
# Tags: paper-summary, asset-pricing, pensions, risk-sharing, equity-premium, open-access, cc-by, peer-reviewed, unreplicated, data:wrds, data:fred, data:flow-of-funds, data:nipa, data:scf, data:nber-cycles
==============================================================================

**What this is.** The paper's core results, model equations, solution method, and datasets: enough to know what it found and how without reading all 46 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1111/jofi.13507).

## TL;DR

The paper builds a general equilibrium incomplete-markets model with an
explicit defined-benefit (DB) pension fund. Calibrated to U.S. data, the
model matches the historical equity premium (7.46% vs. 7.55% in data), the
riskless rate (1.16% vs. 0.86%), and the Sharpe ratio (0.39 vs. 0.36)
better than a standard pass-through model that ignores the fund's endowment
and asset demands. The DB fund's relatively conservative portfolio lowers the
riskless rate and raises the equity premium; stochastic contribution rates
create a new risk channel that raises consumption volatility for workers
and firms. A shift to a DC-only economy produces a higher riskless rate
(3.34% vs. 1.16%), a lower equity premium (4.96% vs. 7.46%), a lower Sharpe
ratio (0.27 vs. 0.39), higher consumption volatility for retirees, and lower
consumption volatility for workers.

## Core results

Magnitudes and significance are as reported. Locators point into the source
PDF (page numbers match the journal pagination printed on each page).

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Baseline DB model closely matches key asset pricing moments | Table II, p. 164 | Equity premium 7.46% (data 7.55%); riskless rate 1.16% (data 0.86%); Sharpe ratio 0.39 (data 0.36); stock market participation 59.1% (data 51.1%) |
| R2 | DB model dominates pure-pass-through (PPG) model at same calibration: much higher equity premium and Sharpe ratio | Table III, p. 167 | Baseline: equity premium 7.46%, Sharpe 0.39 vs. PPG: 5.53%, 0.31; PPG riskless rate 4.04% vs. baseline 1.16% |
| R3 | DB pension fund's conservative bond demand is the first pricing channel: raising equity premium and lowering riskless rate | Table V, p. 171 | Varying risky share from 42% to 72% moves equity premium from 10.60% to 5.48% and Sharpe ratio from 0.55 to 0.28 |
| R4 | Stochastic DB contribution rates create a new risk channel: higher cross-sectional consumption volatility for workers | Table IV, p. 170 | SD consumption growth ages 20-35: 10.7% (baseline) vs. 10.0% (rPPG1); ages 36-65: 8.4% vs. 7.6%; retirees (66+): 2.2% vs. 2.6% |
| R5 | Asset pricing results are robust across alternative DB fund portfolio allocation rules | Table II cols (1)-(3), p. 164 | Fixed vs. elastic vs. reaching-for-yield: equity premium 7.46%, 7.58%, 7.39%; Sharpe ratio 0.39, 0.39, 0.38 |
| R6 | DC-only economy (DB phased out) has a substantially higher riskless rate and lower equity premium | Table IX col (2), p. 182 | Riskless rate 3.34% vs. 1.16%; equity premium 4.96% vs. 7.46%; Sharpe ratio 0.27 vs. 0.39 |
| R7 | In the DC-only economy, retiree consumption volatility increases while worker consumption volatility decreases | Table IX col (2), p. 182 | SD cons. growth ages 20-35: 10.1% vs. 10.7%; ages 36-65: 7.4% vs. 8.4%; retirees (66+): 2.8% vs. 2.2% |
| R8 | Lower stock market participation costs in DC economy raise participation but have modest aggregate price effects | Table IX col (3), p. 182 | Participation rises to 79.7% (vs. 57.2% at baseline DC costs); equity premium 4.60% and riskless rate 3.44%, both close to base DC scenario |

**Overall (paper's conclusion).** The endowment and asset demands of DB
pension funds matter for asset pricing and risk sharing in ways that purely
PPG models miss. The shift toward DC plans is characterized by a higher
riskless rate, a lower equity premium and Sharpe ratio, and a redistribution
of consumption risk from retirees to workers.

## Theory / model

The model is an incomplete-markets overlapping-generations (OLG) production
economy, using the framework of Storesletten,
Telmer, and Yaron (2007) as a benchmark. Households live from age 20 (adult age 1) to 100 (adult age 81),
working until age 65 and retiring thereafter. Two household types (*A* and
*B*) have heterogeneous discount factors and EIS but the same risk aversion.

**Production technology (eq. 1, p. 149).**

$$
Y_t = Z_t K_t^{\alpha} L_t^{1-\alpha}
$$

where $$Z_t = G_t U_t$$ is aggregate productivity ($$G_t = (1+g)^t$$ is
deterministic growth; $$U_t$$ is a two-state Markov business-cycle shock),
$$K$$ is the beginning-of-period capital stock, and $$L$$ is labor supply (eq.
2, p. 149).

**Stochastic depreciation (eq. 3, p. 149).**

$$
\delta_t = \bar{\delta}(U_t) + \sigma^{\delta}(U_t) \cdot \eta_t
$$

where $$\eta_t$$ is i.i.d. standard normal and both the conditional mean and
standard deviation of depreciation are correlated with $$U_t$$. This device
avoids explicit adjustment costs while generating return volatility
in the incomplete-markets setting, sharing the stochastic depreciation device
for tractability with Favilukis, Ludvigson, and Van Nieuwerburgh (2017).

**Household preferences: Epstein-Zin-Weil (eq. 9, p. 151).**

$$
V_{a,t} = \left\{ (1-\beta)(C^i_{a,t})^{1-1/\psi}
             + \beta\left( \mathbb{E}_t\!\left(p_a V^{1-\gamma}_{a+1,t+1}\right) \right)^{(1-1/\psi)/(1-\gamma)} \right\}^{1/(1-1/\psi)}
$$

where $$\beta$$ is the discount factor, $$\gamma$$ is relative risk aversion, and
$$\psi$$ is the elasticity of intertemporal substitution (EIS). Type-*B*
households: $$\beta^B = 0.965$$, $$\psi^B = 0.65$$; type-*A*: $$\beta^A = 0.83$$,
$$\psi^A = 0.25$$; both have $$\gamma = 6$$ (Table I, p. 160).

**Labor income (eqs. 13-16, pp. 152-153).** Individual labor income is
$$H^i_{a,t} = W_t L^i_{a,t}$$, with individual productivity $$L^i_{a,t} = P^i_{a,t} \epsilon^i_t$$
(permanent component times transitory shock). The permanent component follows:

$$
P^i_{a,t} = \exp(f(a))\, P^i_{a-1,t-1}\, \xi^i_t
$$

where $$f(a)$$ is a deterministic age profile (hump-shaped). Following Guvenen,
Ozkan, and Song (2014), $$\ln \xi^i_t$$ is a mixture of two normals conditional on
the aggregate state $$U_t$$, capturing countercyclical earnings risk.

**Retirement income (eq. 17, p. 153).** Retired households receive:

$$
H^i_{a,t} = (\lambda^{ss} + \lambda^{db})\, P^i_{a^R, t^R}\, W_t, \qquad a > a^R
$$

where $$\lambda^{ss}$$ and $$\lambda^{db}$$ are the social security and DB pension
replacement ratios, calibrated to 0.4596 and 0.2225 respectively (Table I).

**Pension fund endowment and return (eq. 19, p. 155).**

$$
R^P_t = \alpha^P R^K_t + (1 - \alpha^P) R^B_t
$$

where $$\alpha^P$$ is the risky (equity) share of the pension fund portfolio,
calibrated to 52% to match the historical Flow of Funds average.

**Pension fund budget constraint and contribution rates.** The fund keeps
endowment $$\omega^P$$ constant and adjusts contribution rates each year. In the
general case (eq. 22-24, pp. 156-157), the shortfall before adjustments is:

$$
\tilde{\omega}^P_t = (1 + R^P_t)\, \omega^P
                  + \sum_{a=20}^{65} \int_{i \in I^a} \bar{\tau}^{db} L^i_{a,t}\, w_t\, di
                  - \sum_{a=66}^{100} \int_{i \in I^a} \left[\lambda^{db} \exp(f(a^R))\, w_t\, P^i_{a^R, t^R}\right] di
$$

The fraction $$\theta^P$$ of the shortfall is absorbed by employer contributions:

$$
\tau^{kdb}_t = \theta^P \cdot \frac{\omega^P - \tilde{\omega}^P_t}{k_t - \alpha^P \omega^P}
$$

and the remainder $$(1 - \theta^P)$$ by employee contributions:

$$
\tau^{db}_t - \bar{\tau}^{db} = (1 - \theta^P) \cdot \frac{\omega^P - \tilde{\omega}^P_t}{\displaystyle\sum_{a=20}^{65} \int_{i \in I^a} L^i_{a,t}\, w_t\, di}
$$

Baseline calibration: $$\theta^P = 0.5$$ (equal split). This stochastic
adjustment is the new risk channel: return shocks feed into net wages and firm
profits, raising cross-sectional consumption volatility. Building on
Constantinides and Duffie (1996), stochastic contribution rates raise
idiosyncratic income risk for workers and firms, which raises the equity premium.

**Government budget constraint (eq. 8, p. 151).**

$$
C^G_t + (1 + R^B_t)\, B_t = B_{t+1} + T_t
$$

Bond supply is calibrated to a debt-to-GDP ratio of 42% (average U.S.
Treasury holdings by the public). Interest payments are financed by taxes on
capital income (rate $$\tau^K = 40\%$$), bond interest ($$\tau^B = 20\%$$), wages
($$\tau^W$$), and bequests.

**Equilibrium conditions (eqs. 34-36, p. 160).** Markets clear in capital,
bonds, and the consumption good:

$$
k_{t+1} = \int \int P^i_{a,t}\, k^i_{a,t+1}\, da\, di
$$

$$
b_{t+1} = \int \int P^i_{a,t}\, b^i_{a,t+1}\, da\, di
$$

$$
U_t k_t^{\alpha} L_t^{1-\alpha} = \frac{C^G_t}{G_t^{1/(1-\alpha)}}
                              + (1+g)^{1/(1-\alpha)} k_{t+1} - (1-\delta_t) k_t
                              + \int \int P^i_{a,t}\, c^i_{a,t}\, da\, di
$$

## Method

The paper contributes a calibrated structural model, not a new econometric
method. The solution follows the `krusell-smith` approximate-aggregation
approach, building on `overlapping-generations`, `epstein-zin-weil`
preferences, and `incomplete-markets-olg` techniques.

**Household optimization (eq. 30, pp. 158-159).** Households solve the
Bellman equation:

$$
V_a(x^i_{a,t}, E^i_a, k_t, U_t, \eta_t, P^B_t) =
  \max_{k^i_{a+1,t+1},\, b^i_{a+1,t+1}} \left\{
    (1-\beta) (c^i_{a,t})^{1-1/\psi}
    + \beta \left( \mathbb{E}_t \left[ \left(\frac{P^i_{a+1,t+1}}{P^i_{a,t}} (1+g)\right)^{1-\rho}
             p_a V^{1-\rho}(x^i_{a+1,t+1}, E^i_{a+1}; k_{t+1}, U_{t+1}, \eta_{t+1}, P^B_{t+1}) \right] \right)^{(1-1/\psi)/(1-\rho)} \right\}^{1/(1-1/\psi)}
$$

subject to: $$k^i \geq 0$$, $$b^i \geq 0$$, budget constraint $$c + b' + k' = x$$ (eq. 32),
and the wealth transition (eqs. 26-27) that includes after-tax capital and
bond income, net labor income (less social security and DB contribution taxes
for workers), and retirement income (for retirees).

State variables: age $$a$$, normalized cash-on-hand $$x^i_{a,t}$$, entry-cost
dummy $$E^i_a$$, plus four aggregate variables $$(k_t, U_t, \eta_t, P^B_t)$$.

**Aggregate forecasting rules (eqs. 28-29, p. 158).**

$$
k_{t+1} = \Gamma_K(k_t, U_t, \eta_t)
$$

$$
P^B_{t+1} = \Gamma_P(P^B_t, k_t, U_t, \eta_t)
$$

These log-linear rules are estimated on simulated data and iterated to
convergence (following Krusell and Smith 1998, and Gomes and Michaelides 2008).

**Calibration procedure.** Aggregate parameters are calibrated to NBER
business cycle frequencies (Markov chain $$\pi_r = 16/37$$), capital share
$$\alpha = 34\%$$, depreciation $$\text{mean}(\delta) = 10\%$$, $$\text{vol}(\delta) = 10\%$$.
Household parameters are chosen to jointly match: the standard deviation of
consumption growth, the riskless rate level, and limited stock market
participation. The pension fund risky share ($$\alpha^P = 52\%$$) is calibrated
to the 1970-2023 Flow of Funds average (Table I, p. 160; Section I.H.3, p. 162).

**Alternative DB pension fund portfolio rules (eqs. 37-38, p. 166).**

The baseline uses a constant $$\alpha^P$$. Two alternatives are also studied:
elastic allocation proportional to the equity premium:

$$
\alpha^P_t = a^P + c^P \bigl(\mathbb{E}[R^K_t] - R^B_t\bigr), \qquad c^P = 0.25
$$

and reaching-for-yield allocation varying with the riskless rate:

$$
\alpha^P_t = a^P + b^P R^B_t, \qquad b^P = -2
$$

Both deliver nearly identical asset pricing moments (Table II, cols 1-3), so
the results are insensitive to the precise specification.

**DC-only counterfactual (eq. 39, p. 178).** Setting $$\lambda^{db} = 0$$ and
$$\omega^P = 0$$ removes the DB pension fund entirely; households finance
retirement from private savings and social security only. The DC economy
incorporates tax benefits via a reduced capital gains tax rate scaled to the
increase in private household wealth, and a 10% early-withdrawal penalty
(Sections V.B.1-V.B.2, pp. 179-180). The numerical solution adds an outer
loop to find the fixed point for the implied capital gains tax adjustment.

## Empirical specifications

The paper does not estimate regression equations. All quantitative results
are steady-state moments from the calibrated structural model, compared
against empirical counterparts. The "specifications" are the alternative
calibrated economies:

**Baseline DB economy (R1, R5):** $$\alpha^P = 52\%$$, $$\theta^P = 0.5$$, two household
types (A and B), calibrated to match riskless rate SD, participation rate, and
equity premium. Key moments computed at the stationary distribution of the
model (Table II, p. 164). Asset pricing data from CRSP; real risk-free rate
from Croce et al. (2012); participation from SCF; consumption/GDP from NIPA
1929-2023.

**Comparison with PPG model (R2):** same parameter values, but $$\omega^P = 0$$
so the fund is a pure pass-through with no endowment and constant
contribution rates (Table III, p. 167). Two recalibrations (rPPG1, rPPG2)
additionally match the riskless rate or consumption growth SD of the baseline.

**Risk channel decomposition (R3):** the pension fund risky share $$\alpha^P$$
is varied from 42% to 72% (Table V, p. 171) to isolate the bond-demand
channel; all other parameters are held at the baseline.

**Consumption risk sharing by age group (R4):** cross-sectional standard
deviation of consumption growth reported by cohort (ages 20-35, 36-65, 66+)
for the baseline and rPPG1 economy (Table IV, p. 170).

**Adjustment rule robustness (R5):** $$\theta^P$$ is varied from 0.2 (mostly
employee adjustment) to 0.8 (mostly employer adjustment) with the same
aggregate parameters (Table VIII, p. 177).

**DC-only counterfactual (R6-R8):** the DB fund is shut down ($$\lambda^{db} = 0$$,
$$\omega^P = 0$$) and the model is solved for the new stationary equilibrium.
Three scenarios: (1) same participation costs, (2) lower participation costs
($$F^0 = 3\%$$, $$F^1 = 0.1\%$$), (3) higher debt-to-GDP (0.6). Asset pricing and
macro moments from Table IX (p. 182) are compared to the baseline.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CRSP (Center for Research in Security Prices) | Asset pricing moments: equity return mean and SD, riskless rate (via Croce et al. 2012 for real rate) | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| NIPA tables (BEA / Federal Reserve Bank of St. Louis) | Real consumption growth mean and SD; capital-output ratio (tables 1.1.3 and 1.1.5, 1929-2023) | [FRED](/wiki/datasets/fred/), free |
| Flow of Funds (Federal Reserve) | DB pension fund total financial assets and endowment-to-GDP ratio (1970-2023) | [FRED](/wiki/datasets/fred/), free |
| Survey of Consumer Finances (SCF, Federal Reserve) | Historical stock market participation rate (used as calibration target) | no page yet |
| NBER business cycle dates | Calibration of productivity shock Markov chain (recession/expansion probabilities) | [NBER cycles](/wiki/datasets/nber-cycles/) |
| Public Plans Data / Social Security Administration data | Decomposition of DB replacement ratio vs. social security replacement ratio | no page yet |

Sample: U.S. aggregate, 1929-2023 for returns and consumption; 1970-2023 for pension fund data.

## When to read the full paper

Use the [original article](https://doi.org/10.1111/jofi.13507) if you are:
replicating or extending the quantitative model; examining Internet Appendix
robustness checks (bequest motives, alternative bond supply, transition
dynamics); doing a literature review of intermediary asset pricing or pension
finance; or auditing a specific parameter value from the calibration (Table I,
p. 160). The locators in the table above point to the exact figures and tables.
For "what did this paper find," the table above is the intended default.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 81(1), February 2026.
This distillation was extracted by an LLM on 2026-05-31 and augmented on
2026-06-01; it is **not human-verified or independently reproduced**.

> **Attribution (CC BY 4.0).** Coimbra, Nuno, Francisco Gomes, Alexander
> Michaelides, and Jialu Shen. "Asset Pricing and Risk-Sharing Implications
> of Alternative Pension Plan Systems." *The Journal of Finance* 81, no. 1
> (February 2026): 143-188. DOI: 10.1111/jofi.13507. (c) 2025 The Author(s).
> Published by Wiley Periodicals LLC on behalf of the American Finance
> Association. Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**. CC BY 4.0
> permits mirroring; the PDF is not hosted in this batch.


==============================================================================
# Can Social Media Inform Corporate Decisions: Cookson, Niessner & Schiller (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2026/cookson-social-media-merger-withdrawals-2026/
# Distilled: abnormal StockTwits sentiment after a merger announcement predicts a 0.64 percentage point higher withdrawal probability (16.6% of the baseline rate), robust to market reactions, news, and analyst signals; the effect strengthens after firms register corporate Twitter accounts and is driven by fundamental rather than technical or meme tweets. J. Finance 2026, paywalled. Eight core results with source locators, datasets used, the model, and the method with formal equations.
# Tags: paper-summary, mergers-and-acquisitions, social-media, text-as-data, corporate-decisions, fintech, panel-regression, event-study, peer-reviewed, unreplicated, data:wrds, data:stocktwits, data:sdc-platinum, data:ravenpack, data:ibes, data:ken-french
==============================================================================

**What this is.** The paper's core results, datasets, and formal specifications:
enough to know what it found and how without reading all 52 pages. To replicate
or extend it, access the original via the DOI:
[10.1111/jofi.13508](https://doi.org/10.1111/jofi.13508) (paywalled).

## TL;DR

Using 260 million StockTwits posts (2010–2021) matched to 6,438 U.S.
M&A deals from SDC Platinum, the paper measures firm-specific abnormal
social media sentiment in the four days after a merger announcement. A
one-standard-deviation decrease in abnormal sentiment predicts a 0.64
percentage point higher probability of deal withdrawal (16.6% of the
unconditional rate). This effect survives controls for acquirer CARs,
traditional news sentiment (RavenPack), analyst recommendation changes
(IBES), and deal characteristics. The effect is absent for deals withdrawn
by regulators or target boards (only one-third the magnitude), is
concentrated after firms register corporate Twitter accounts (especially
high-follower or verified accounts), and is driven by fundamental investor
tweets and M&A-relevant tweet topics, not meme or technical tweets.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%.
Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Negative abnormal social media sentiment predicts merger withdrawal**, controlling for market reactions, news, and analyst signals | Table II Panel A, pp. 107–108 | beta = -0.643\*\*\* (col. 1, no controls) to -0.711\*\*\* (col. 6, full controls); 1-SD decrease in AbnSent associated with 0.64 pp higher withdrawal probability (16.6% of baseline 3.89%) |
| R2 | **StockTwits and Twitter sentiment each predict withdrawal independently** | Table II Panel B, pp. 107–108 | StockTwits col. (1): -0.706\*\*; Twitter (SMA) col. (2): -1.075\*\*\*; both significant when included jointly in col. (3) (StockTwits: -0.593\*\*, Twitter: -0.998\*\*\*) |
| R3 | **Effect is not driven by governance channel**: coefficients are similar for positive-CAR and negative-CAR deals | Figure 5, p. 111 | Estimated AbnSent coefficient indistinguishable across CAR < 0 and CAR >= 0 subsamples; governance-metric splits (ISS, board independence, antitakeover provisions) also yield no significant differences (Table IA.VI) |
| R4 | **Social media sentiment predicts market-favorable deal outcomes**: initial negative sentiment followed by withdrawal yields higher post-announcement BHARs | Table IV, p. 119 | Interaction AbnSent x 1(Deal Withdrawn): -0.101\*\* to -0.153\*\*\* across columns; withdrawal predicted by negative AbnSent associated with 10.09%–15.28% higher BHAR from day 11 to deal conclusion |
| R5 | **Effect strengthens after acquirer registers corporate Twitter account**, especially for high-follower or verified accounts | Table V, pp. 122–123 | AbnSent x Post(Twitter, HighFollow) col. (4): -0.956\*\*; AbnSent x Post(Twitter, Verified) col. (5): -1.301\*\*; coefficients on AbnSent alone (pre-Twitter) near zero and insignificant |
| R6 | **Effect is driven by fundamental investor tweets, not technical investor tweets** | Table VI Panel A, pp. 126–127 | Fundamental (non-technical) AbnSent: -1.105\*\*\* (col. 1); technical AbnSent: -0.350 (insignificant); result holds after adding media controls (col. 2): fundamental -1.080\*\*\*, technical -0.274 (insig.) |
| R7 | **Effect is driven by longer tweets and by M&A-relevant topics (Company/Business, Deal Terms, Disclosure), not by meme or trading tweets** | Table VI Panels A–B, pp. 126–127 | Long tweets: -1.214\*\*\* to -1.222\*\*\*; short tweets: -0.053 to +0.014 (insig.); Company/Business: -1.202\*\*\*; Deal-Terms: -1.139\*\*\*; Disclosure: -0.868\*\*; Meme: -0.280 (insig.); Technical: -0.487 (insig.) |
| R8 | **Effect is stronger when AbnSent disagrees with market and news signals, and when social media volume is higher** | Table VII, pp. 129, 130 | High N Tweets vs. low: -2.030\*\*\* vs. -0.516\* (coef. diff. t = 2.28, p = 0.023); split by social media vs. news disagreement drives the result; news-article-count split shows no statistically significant difference |

**Overall (paper's conclusion).** Building on the parallel test of traditional
news sentiment for M&A decisions in Liu and McConnell (2013), the paper adds
social media as a distinct channel. Social media sentiment contains
information about M&A deal outcomes that is not subsumed by market prices,
traditional media, or analyst signals. The evidence is consistent with a
revelatory channel: managers learn from social media, particularly after
engaging with the platform through a corporate Twitter account.

## Theory / model

The paper has no original structural model. It tests a conceptual distinction
introduced by Bond, Edmans, and Goldstein (2012) and Bai, Philippon, and Savov
(2016) between two channels of informativeness:

- **Revelatory**: social media reveals genuinely new information to the firm
  manager, analogous to revelatory price efficiency (RPE), where prices
  aggregate dispersed information that managers subsequently learn from.
- **Forecasting**: social media merely correlates with information the manager
  already has (market prices, analyst signals, news), providing no incremental
  input.

The key empirical test distinguishes these channels: under the revelatory channel
the predictive coefficient on $$\text{AbnSent}_i$$ is large and robust to controlling for
all other signals managers are known to rely on; it grows stronger after the firm
engages actively with social media (registering a corporate Twitter account); and
it weakens for deals where withdrawal cannot reflect manager learning (regulator-
or target-rejected deals). Under a pure forecasting channel these patterns would
not arise. Under a governance channel, predictability would be concentrated in
negative-CAR mergers and in firms with weaker governance, which is not observed.
Whereas the Chinese message board study of Ang et al. (2021) emphasizes the
governance channel, this US-sample paper finds no governance-channel pattern and
instead supports revelatory learning.

**Identification strategy.** The baseline design of using market reactions to
predict M&A withdrawal follows Luo (2005), which the paper extends by adding
social media signals. The paper exploits deal-level cross-sectional
variation in $$\text{AbnSent}_i$$ (eq. 1, p. 103), which is constructed as the
difference between announcement-window and pre-announcement-window sentiment,
removing firm-level baseline sentiment differences. The main specification
(eq. 2, p. 106) further controls for the acquirer CAR, traditional news
sentiment, analyst recommendations, deal characteristics, year-by-quarter time
fixed effects, and acquirer-industry fixed effects. Coefficient stability as
controls are added ($$R^2$$ rising from 0.001 to 0.216 with a stable beta, following
Oster (2019)) is the primary guard against omitted-variable concerns. The
Twitter-account timing test (eq. 4, p. 121) uses within-acquirer before/after
variation as an additional quasi-experimental design.

## Method

**Step 1: Constructing abnormal sentiment (equation 1, p. 103).**

$$
\text{AbnSent}_i = \left( \frac{1}{|J_{i,[0,3]}|} \sum_{j \in J_{i,[0,3]}} \text{Sentiment}_{i,j(t)} \right)
- \left( \frac{1}{|J_{i,[-13,-7]}|} \sum_{j \in J_{i,[-13,-7]}} \text{Sentiment}_{i,j(t)} \right)
\tag{1}
$$

- $$\text{Sentiment}_{i,j(t)}$$ is the sentiment of tweet $$j$$ about acquiring firm $$i$$ occurring $$t$$ days after the merger announcement date (day 0).
- $$J_{i,[t_1,t_2]}$$ is the set of tweets about firm $$i$$ between day $$t_1$$ and $$t_2$$.
- The first term averages sentiment over the four-day announcement window $$[0, 3]$$; the second term averages over a reference period $$[-13, -7]$$ (7 to 13 days prior to announcement), omitting the 6 days immediately before the announcement to address information leakage.
- The primary sentiment scores are from StockTwits' proprietary MarketLex classifier, bounded in $$[-1, 1]$$; robustness uses maximum entropy and naive Bayes classifiers trained on user-labelled bullish/bearish tags, and SMA Twitter sentiment bounded in $$[0, 1]$$.

**Step 2: Estimating abnormal sentiment in a text classifier.** To construct
alternative sentiment measures, a maximum entropy (MaxEnt) classifier and a
naive Bayes (Bayes) classifier are trained on StockTwits posts with user-provided
sentiment tags (bullish/bearish), following Antweiler and Frank (2004) and
Cookson and Niessner (2020). The classifiers assign a continuous sentiment
score to each tweet and are validated against held-out samples (Section I.A of
the Internet Appendix; cross-sample correlations 0.85-0.90, Table IA.I, p. 99
of the Internet Appendix).

**Step 3: Topic classification via Biterm Topic Model (BTM, p. 128).** To
decompose the social media signal into fundamental vs. meme/technical content,
the paper trains a BTM with eight topics on all tweets in the $$[0, 3]$$ window
around merger announcements, following Yan et al. (2013). Six topics are
retained: Company/Business, Disclosure, Deal Terms, Trading, Technical, and
Memes. Separate $$\text{AbnSent}$$ measures are constructed for each topic subset.

The method builds on `panel-regression` for the estimating equation, `event-study`
for the CAR controls, and `text-classification` for the sentiment and topic scores.

## Empirical specifications

**Baseline withdrawal regression (equation 2, p. 106, produces R1–R3).**

$$
\text{Deal\_Withdrawn}_i = \beta_1 \cdot \text{AbnSent}_i + \beta_2 \cdot \text{CAR}_i + \Gamma \cdot X_i
+ \alpha_t + \gamma_j + \epsilon_i
\tag{2}
$$

- $$\text{Deal\_Withdrawn}_i$$ is an indicator equal to 1 if M&A deal $$i$$ was subsequently withdrawn, multiplied by 100.
- $$\text{AbnSent}_i$$ is eq. (1) above, standardized to mean 0, SD 1.
- $$\text{CAR}_i$$ is the acquirer CAR[-1, 10] from the Fama-French three-factor model (100-day pre-event window, 10-day gap), also standardized.
- $$X_i$$ includes CAR[-5, -1], news sentiment (RavenPack ESS), analyst recommendation changes (IBES), deal value, pct. shares held, white-knight/competing-bidder/rumored/hostile deal indicators, termination fee, N tweets, N news articles, and acquirer firm size, leverage, cash.
- $$\alpha_t$$ = year-by-quarter fixed effects; $$\gamma_j$$ = acquirer industry (GIC 2-digit) fixed effects.
- Standard errors clustered at the year-by-quarter level. Sample: 5,932-6,306 deal-level observations (2010-2021).

**BHAR regression (equation 3, p. 118, produces R4).**

$$
\text{BHAR}_{i,[11, T_{\text{conclusion}}]} = \beta_1 \cdot \mathbf{1}(\text{Deal Withdrawn})_i + \beta_2 \cdot \text{AbnSent}_i
+ \beta_3 \cdot \mathbf{1}(\text{Deal Withdrawn})_i \times \text{AbnSent}_i
+ \Gamma \cdot X_i + \epsilon_i
\tag{3}
$$

- $$\text{BHAR}_{i,[11, T_{\text{conclusion}}]}$$ is the buy-and-hold abnormal return from day 11 after the merger announcement until deal conclusion (withdrawal or completion), using the Fama-French three-factor model.
- $$\mathbf{1}(\text{Deal Withdrawn})_i$$ equals 1 for withdrawn deals.
- The coefficient of interest is $$\beta_3$$ on the interaction: a negative $$\beta_3$$ means that initial negative social media reaction predicts the market eventually responds positively to a merger withdrawal.
- Industry (GIC2) and year-by-quarter fixed effects included. Standard errors clustered at year-by-quarter level.
- Sample restricted to deals with interim period > 25 days (and > 75 days for robustness) to avoid overlap with the announcement window; target-rejected deals excluded. N = 1,784-3,343 (Table IV, p. 119).

**Twitter-account timing test (equation 4, p. 121, produces R5).**

$$
\text{Deal\_Withdrawn}_i = \beta_1 \cdot \text{AbnSent}_i + \beta_2 \cdot \text{AbnSent}_i \times \text{Post}_{i,t}
+ \beta_3 \cdot \text{Post}_{i,t} + \Gamma \cdot X_i + \alpha_t + \gamma_j + \epsilon_i
\tag{4}
$$

- $$\text{Post}_{i,t}$$ equals 1 for acquisitions by firm $$i$$ announced after firm $$i$$ registered its corporate Twitter account.
- Columns vary whether any Twitter account counts (cols 1-2), whether the account must have an above-median follower count (cols 3-4), or whether it must be a verified account (cols 5-6).
- $$X_i$$ includes the same controls as eq. (2) including $$\text{CAR}_i$$.
- The identification relies on within-acquirer before/after variation in Twitter engagement.
- Sample: 5,932 observations with year-by-quarter and industry FE; acquirer FE added in cols 2, 4, 6 (Table V, pp. 122-123).

**Content heterogeneity regressions (Table VI, pp. 126-127, produces R6-R7).**
Variants of eq. (2) replacing the single $$\text{AbnSent}_i$$ with two simultaneous
signals: $$\text{AbnSent}_i(\text{Technical}=\text{N})$$ and $$\text{AbnSent}_i(\text{Technical}=\text{Y})$$ in Panel A
columns 1-2 (technical vs. fundamental traders); $$\text{AbnSent}_i(\text{Long}=\text{Y})$$ and
$$\text{AbnSent}_i(\text{Long}=\text{N})$$ in columns 3-4 (above-/below-median word count); and one
topic-specific $$\text{AbnSent}$$ per topic in Panel B (Company/Business, Deal Terms,
Disclosure, Meme, Technical, Trading). Same FE structure and clustering.

**Information-source heterogeneity regressions (Table VII, p. 129, produces R8).**
Variants of eq. (2) with sample split at median N Tweets (cols 1-2), N News
Articles (cols 3-4), and $$|\text{CAR}|$$ (cols 5-6); coefficient difference tested with
a Wald t-statistic.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| StockTwits (260M posts, Jan 2010–Dec 2021, proprietary via Social Market Analytics / direct) | Primary social media sentiment measure (AbnSent); 6,438 M&A deals matched | [StockTwits](/wiki/commercial/stocktwits/) (licensed) |
| SDC Platinum (Thomson Reuters) | M&A deal universe (announcement dates, deal values, completion/withdrawal status, deal characteristics) | [SDC Platinum](/wiki/commercial/sdc-platinum/) (licensed) |
| CRSP | Stock returns for acquirer and target CARs; Fama-French three-factor model inputs | [WRDS / CRSP / Compustat](/wiki/commercial/wrds/) (licensed) |
| Compustat North America | Acquirer firm controls (market cap, leverage, cash holdings, M/B ratio) | [WRDS / CRSP / Compustat](/wiki/commercial/wrds/) (licensed) |
| RavenPack News Analytics (v. RPA 1.0) | Traditional news media sentiment (Event Sentiment Score) for M&A-related articles | [RavenPack](/wiki/commercial/ravenpack/) (licensed) |
| IBES (via Refinitiv) | Analyst recommendation changes as external signal control | [WRDS / CRSP / Compustat](/wiki/commercial/wrds/) (licensed via WRDS) |
| Ken French Data Library | Fama-French three-factor returns for CAR estimation | [Ken French Library](/wiki/datasets/ken-french/) |
| Social Market Analytics (SMA) Twitter data (2012–2021) | Robustness alternative sentiment measure from Twitter | No page yet |
| Refinitiv Streetevents (M&A conference call transcripts) | Conference call textual analysis (% constrained/negative words in presentation vs. Q&A) | No page yet |
| BoardEx / LinkedIn | CEO professional network proxies (education, employment, digital connections) | No page yet |

Sample: 6,438 M&A announcements (6,187 completed, 251 withdrawn); acquirers are U.S. public firms; deal values >= $25M; 2010–2021.

## When to read the full paper

Use the original DOI link if you are: examining the Internet Appendix
robustness tables (Tables IA.I–IA.XIV); studying the conference call
presentation vs. Q&A split in detail (Table VIII); using the BTM tweet-topic
methodology; or replicating the BHAR analysis (Table IV). The locators
above point to the main tables. For "what did this paper find," the table
above covers the core results.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 81(1), February 2026.
© 2025 the American Finance Association. Published by Wiley; paywalled.
Licence confirmed via Crossref (Wiley VOR terms, no Creative Commons entry).

This distillation was extracted by an LLM on 2026-05-31 and augmented on
2026-06-01; **not human-verified or independently reproduced**. Extract-only:
no PDF mirror is hosted here.

> Cookson, J. Anthony, Marina Niessner, and Christoph Schiller.
> "Can Social Media Inform Corporate Decisions? Evidence from Merger
> Withdrawals." *The Journal of Finance* 81, no. 1 (February 2026): 91–142.
> DOI: 10.1111/jofi.13508. © 2025 the American Finance Association.
> All rights reserved. This page contains extracted findings only; no
> reproduction of the original text.


==============================================================================
# Dynamic Trading with Realization Utility: Dai, Qin & Wang (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2026/dai-dynamic-trading-realization-utility-2026/
# Distilled: a jump-diffusion model with two-layered mental accounts shows that investors can optimally sell stocks at deep losses when savings are sufficient, and sell losing stocks after a price rebound when savings are low; leverage strengthens the disposition effect while leverage constraints mitigate it. J. Finance 2026, paywalled. Seven core results with source locators, the structural model with its equations, and the solution method.
# Tags: paper-summary, asset-pricing, disposition-effect, behavioral-finance, realization-utility, prospect-theory, portfolio-choice, structural-model, peer-reviewed, unreplicated
==============================================================================

**What this is.** The paper's core theoretical results, model structure, and
predictions: enough to know what it found without reading all 50 pages. To
replicate or extend it, read the full source at the
[canonical DOI](https://doi.org/10.1111/jofi.13472) (paywalled).

## TL;DR

Dai, Qin, and Wang build a continuous-time jump-diffusion model in which an
investor receives utility bursts from realizing stock gains and losses at the
individual stock level, while also managing a dynamic mental trading budget
shared across all investment episodes. The key departure from prior
realization-utility models (Barberis-Xiong 2012, Ingersoll-Jin 2013,
He-Yang 2019) is that the investor is not forced to invest his entire budget
in a single stock: he can save a fraction in the risk-free asset or use
leverage. This intensive margin, combined with downward jumps in stock prices,
generates two new predictions: (i) an investor with sufficient savings
voluntarily sells a stock at an arbitrarily deep loss to reset his reference
level, and (ii) an investor with low savings will not sell a deep loser but
will sell it after its price rebounds just enough. Leverage amplifies the
disposition effect; leverage constraints dampen it.

## Core results

Magnitudes and thresholds are as reported. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | With sufficient savings ($$w^* > 0$$), the investor optimally saves 63.5% of his budget and allocates only 36.1% to the stock each trade | §III.A, Figure 2, p. 205-206 | Baseline: $$w^* = 1.76$$; stock share $$= 1/(1 + \theta_p + w^*) = 36.1\%$$; value of saving option = 21% of mental trading budget ($$\Delta(w^*) = 21\%$$, Figure 2 Panel B) |
| R2 | Savings cause the investor to realize losses sooner than in IJ (2013), reducing the disposition effect | Figure 3, p. 207 | Loss-realization boundary $$x^* = 0.69$$ in baseline model vs 0.55 in IJ (2013); downside loss in dollars is one-third of the IJ (2013) model because savings absorb the hit at the trading-account level |
| R3 | Leverage ($$w^* < 0$$) strengthens the disposition effect: loss-realization threshold falls and gain-realization threshold rises | Figure 4, p. 208 | At $$\sigma = 20\%$$: $$w^* = -0.36$$; $$x^*$$ falls from 0.6 (IJ 2013) to 0.47; $$\bar{x}$$ rises from 1.03 to 1.04; the option to use leverage is worth 31% of the investor's trading budget |
| R4 | Binding leverage constraints mitigate the disposition effect by forcing earlier loss realization | Figure 5, p. 209 | Tightening $$\kappa$$ from 0.79 to 0.59 raises loss-realization threshold $$x^*$$ from 0.47 to 0.52; gain-realization threshold $$\bar{x}$$ unchanged at 1.04 |
| R5 | Investors prefer stocks with high or low volatility, not intermediate volatility; leverage users prefer low-volatility stocks, savers prefer high-volatility stocks | Figure 6, p. 210 | Scaled value $$\hat{v}$$ is U-shaped in $$\sigma$$ with minimum at $$\sigma = 25\%$$; investors use leverage when $$\sigma < 25\%$$ and save when $$\sigma > 25\%$$; $$\hat{v}$$ at $$\sigma = 25\%$$ equals 7.95 (same as IJ 2013, where no saving/leverage is available) |
| R6 | With downward jumps and sufficient savings (Case A, $$\sigma = 30\%$$), the investor voluntarily realizes deep losses for all $$x \in (0, 0.38)$$; saving $$w^* = 0.24$$ means 19.2% of budget is in the risk-free asset, making deep-loss realization optimal | Figure 7, §IV.A, p. 213-214 | Three-region solution: gain-realization at $$x \geq 1.03$$; holding for $$x \in (0.38, 1.03)$$; voluntary loss realization for all $$x \in (0, 0.38)$$ including $$x$$ near 0; payoff function $$f(w^*, 0) = 2.3 > 0$$ |
| R7 | With low savings (Case B, $$\sigma = 24\%$$), the investor holds a deep-loss stock (deep-loss holding region $$x \in (0, 0.04)$$) but sells after the price rebounds just enough to exit the deep-loss region | Figure 8, §IV.B, p. 214-215 | Four-region solution: deep-loss holding for $$x \in (0, 0.04)$$; loss-realization for $$x \in (0.04, 0.34)$$; normal holding for $$x \in (0.34, 1.03)$$; gain-realization for $$x \geq 1.03$$; $$w^* = 0.02$$ (only 1.9% in savings) |

**Overall (paper's conclusion).** The two-layered mental account generates
qualitatively new loss-realization predictions that diffusion-only models
cannot produce. The sell-at-deep-loss (Case A) and sell-after-rebound
(Case B) predictions arise from the interaction between the dynamic trading
budget and downward jumps. Both predictions are consistent with observed
retail investor behavior. Quantitatively, the option to save in the risk-free
asset is worth over 20% of the investor's total trading budget in calibrated
diffusion models.

## Theory / model

**Two-layered mental accounts.** The investor has a trading account with budget
$$\Pi_t > 0$$ at time $$t$$, used solely for realization-utility optimization.
At each trade he allocates a fraction to a risky stock and saves the rest in
the risk-free asset (the intensive margin $$w$$). Prior models (BX 2012, IJ 2013,
HY 2019) force $$w = 0$$ at all times; here $$w$$ is a choice variable.

**State variables and dynamics.** Three state variables: risk-free wealth
$$W_t$$, risky wealth $$X_t$$, and reference level $$B_t$$ (eq. 1, p. 198):

$$
\frac{dP_{n,t}}{P_{n,t}} = \mu \, dt + \sigma \, dZ_{n,t}, \qquad t > 0 \tag{1}
$$

Between two consecutive trading times $$(\tau_i, \tau_{i+1})$$, risky wealth
follows the same GBM (eq. 2, p. 198):

$$
dX_t = \mu X_t \, dt + \sigma X_t \, dZ_{n,t}, \qquad t \in (\tau_i, \tau_{i+1}) \tag{2}
$$

$$
dW_t = r W_t \, dt, \qquad t \in (\tau_i, \tau_{i+1}) \tag{3}
$$

The mental budget at $$t$$ is (eq. 4, p. 198):

$$
\Pi_t = W_t + (1 - \theta_s) X_t \tag{4}
$$

where $$\theta_s$$ is the proportional sale cost. Post-purchase wealth satisfies
(eq. 5, p. 199):

$$
W_{\tau_i^+} = \Pi_{\tau_i} - (1 + \theta_p) X_{\tau_i^+} \tag{5}
$$

The leverage constraint (eq. 6, p. 199):

$$
X_t \geq -W_t / \kappa, \qquad \text{where } 0 < \kappa < 1 - \theta_s \tag{6}
$$

**Reference level and realization utility.** The reference level grows at the
risk-free rate (eq. 7-8, p. 199):

$$
dB_t = r B_t \, dt \qquad \text{for } t \in (\tau_i, \tau_{i+1}) \tag{7}
$$

$$
B_{\tau_i^+} = X_{\tau_i^+} \tag{8}
$$

Realized gain (loss) at $$\tau_i$$ is (eq. 9-10, p. 199-200):

$$
G_{\tau_i} = (1 - \theta_s) X_{\tau_i} - B_{\tau_i} \tag{9}
$$

$$
g_{\tau_i} = G_{\tau_i} / B_{\tau_i} \tag{10}
$$

Utility burst (eq. 11, p. 200):

$$
U(G, B) = B^{\beta} u(G/B) = B^{\beta} u(g) \tag{11}
$$

The scaled utility function is S-shaped CPT (eq. 12, p. 200):

$$
u(g) = \begin{cases} g^{\alpha_+} & \text{if } g \geq 0 \\ -\lambda (-g)^{\alpha_-} & \text{if } g < 0 \end{cases} \tag{12}
$$

with $$\lambda \geq 1$$ (loss aversion), $$\alpha_+, \alpha_- \in (0, 1]$$
(diminishing sensitivity), and $$\beta \leq \min\{\alpha_+, \alpha_-\}$$ (eq. 13,
p. 200) to ensure $$|U(G,B)|$$ decreases in $$B$$ for fixed $$G$$.

**Optimization problem.** The investor chooses trading times
$$\{\tau_i \geq t\}$$ and stock allocations $$X_{\tau_i^+}$$ to maximize
(eq. 14, p. 201):

$$
\max \; \mathbb{E}_t \!\left[ \sum_{i=1}^{\infty} e^{-\delta(\tau_i - t)} U(G_{\tau_i}, B_{\tau_i}) \mathbf{1}_{\tau_i < \tau_L} + e^{-\delta(\tau_L - t)} U(G_{\tau_L}, B_{\tau_L}) \right] \tag{14}
$$

subject to the leverage constraint (6) and dynamics (2), (3), (5), (7), (8),
where $$\delta > 0$$ is the subjective discount rate and $$\tau_L$$ is the
liquidity-shock arrival time (exogenous Poisson with rate $$\xi$$).

**Baseline calibration** (Table II, p. 205): $$\alpha_+ = \alpha_- = 0.5$$,
$$\lambda = 1.5$$, $$\beta = 0.3$$, $$r = 3\%$$, $$\delta = 5\%$$, $$\mu = 9\%$$,
$$\sigma = 30\%$$, $$\theta_s = \theta_p = 1\%$$, $$\kappa = 0.79$$, $$\xi = 0$$.

**Jump-diffusion extension** (Section IV, p. 211-216): stock prices follow
(eq. 29, p. 212):

$$
\frac{dP_{n,t}}{P_{n,t^-}} = \mu \, dt + \sigma \, dZ_{n,t} - (1 - Y) \, dJ_{n,t}, \qquad P_0 > 0 \tag{29}
$$

where $$J_n$$ is a Poisson process with arrival rate $$\rho = 0.73/\text{year}$$ and
jump size $$Y \in [0,1]$$ drawn from cdf $$\Omega(Y) = Y^{\psi}$$ with $$\psi = 6.3$$,
implying expected price drop $$\mathbb{E}[1-Y] = 1/(\psi+1) = 14\%$$ per jump. In
scaled variables (eq. 30, p. 212):

$$
\frac{dx_t}{x_{t^-}} = (\mu - r) \, dt + \sigma \, dZ_{n,t} - (1 - Y) \, dJ_{n,t} \tag{30}
$$

## Method

**Homogeneity reduction.** Using the homogeneity of the value function
$$V(W,X,B) = B^{\beta} v(w,x)$$ and payoff function $$F(W,X,B) = B^{\beta} f(w,x)$$
(p. 202), the three-state problem reduces to a two-state scaled problem
with scaled variables (eq. 19, p. 203):

$$
w_t = W_t / B_t \qquad \text{and} \qquad x_t = X_t / B_t \tag{19}
$$

Since $$w_t$$ is constant between trades ($$dw_t = 0$$, eq. 21, p. 203), the
investor optimally picks a constant target ratio $$w^*$$ at each trade. The
scaled value with budget one is (eq. 24-25, p. 203):

$$
\hat{v} = \max_{w \geq -\kappa} m(w) \tag{24}
$$

$$
m(w) = \left(\frac{1}{w + 1 + \theta_p}\right)^{\!\beta} v(w, 1) \tag{25}
$$

The simplified scaled optimization problem is (eq. 22, p. 203):

$$
v(w_t, x_t) = \max_{\tau} \; \mathbb{E}_t \!\left[ e^{-\delta_e(\tau - t)} f(w_{\tau}, x_{\tau}) \mathbf{1}_{\tau < \tau_L} + e^{-\delta_e(\tau_L - t)} u\!\left((1-\theta_s)x_{\tau_L} - 1\right) \right] \tag{22}
$$

where $$\delta_e = \delta - \beta r$$ is the effective discount rate and
$$f(w, x) = u((1-\theta_s)x - 1) + [(1-\theta_s)x + w]^{\beta} \hat{v}$$ (eq. 23,
p. 203).

**HJB / variational inequality.** In the holding domain the scaled value
satisfies (eq. 26, p. 204 for the diffusion model; eq. 31, p. 212 for
jump-diffusion):

$$
\delta_e v(w,x) = \tfrac{1}{2} \sigma^2 x^2 v_{xx} + (\mu - r) x v_x + \xi \!\left[u\!\left((1-\theta_s)x - 1\right) - v(w,x)\right] \tag{26}
$$

$$
\delta_e v(w,x) = \frac{\sigma^2 x^2}{2} v_{xx} + (\mu - r) x v_x + \rho \!\left(\mathbb{E}[v(w, Yx)] - v(w,x)\right) \tag{31}
$$

The full variational inequality (Appendix A, pp. 218-219). The unscaled form is (A.1); after applying the homogeneity reduction the scaled variational inequality is (A.4):

$$
\max \!\left\{ \mathcal{L} v(w,x),\; f(w,x) - v(w,x) \right\} = 0 \tag{A.4}
$$

$$
\mathcal{L} v = \tfrac{1}{2} \sigma^2 x^2 v_{xx} + (\mu - r) x v_x - \delta_e v + \xi \!\left[u\!\left((1-\theta_s)x - 1\right) - v\right] \tag{A.5}
$$

When the leverage constraint (6) binds ($$w = -\kappa x$$):
$$v(w,x) = f(w,x)$$ (eq. A.3, p. 219).

**Closed-form solution (diffusion, no liquidity shocks).** In the holding
region, the value function has the form (Appendix B, eq. B.1, p. 222):

$$
v(w,x) = C_1(w)\, x^{\eta_1} + C_2(w)\, x^{\eta_2} \tag{B.1}
$$

where $$\eta_1 > 0$$ and $$\eta_2 < 0$$ are the two roots of the fundamental
quadratic (eq. B.2, p. 222):

$$
h(\eta) = \frac{\sigma^2}{2} \eta(\eta - 1) + (\mu - r)\eta - \delta_e = 0 \tag{B.2}
$$

The optimal $$w^*$$ is found by (eq. B.4, p. 222):

$$
w^* = \operatorname*{argmax}_{w \geq -\kappa} \frac{C_1(w) + C_2(w)}{[w + (1 + \theta_p)]^{\beta}} \tag{B.4}
$$

Value-matching and smooth-pasting conditions at the two boundaries
$$\bar{x}(w)$$ (gain) and $$x^*(w)$$ (loss) give a system of four equations
(B.5)-(B.8) jointly with the FOC (B.9), p. 223.

For the **jump-diffusion model**, the variational inequality (B.10, p. 225):

$$
\max \!\left\{ \mathcal{L}^J v(w,x),\; f(w,x) - v(w,x) \right\} = 0 \qquad \text{for } x \geq 0,\; w \geq 0 \tag{B.10}
$$

is solved numerically via a penalty method (Appendix B.3).

## Empirical specifications

This is a purely theoretical paper. There are no regression equations,
no portfolio sorts, and no empirical datasets used. All quantitative results
derive from numerical solution of the variational inequality or its
closed-form analogue under the baseline and jump-diffusion calibrations.

**Calibration targets** (Table II, p. 205 and §IV parameter choices,
p. 213):

- $$\alpha_+ = \alpha_- = 0.5$$, $$\lambda = 1.5$$, $$\beta = 0.3$$ match the CPT
  parameter values in IJ (2013), enabling direct comparison.
- $$\mu = 9\%$$, $$r = 3\%$$ target a 6% risk premium consistent with U.S. equity
  estimates (Hansen and Singleton 1982; Mehra and Prescott 1985).
- $$\kappa = 0.79$$ targets an 80% maximum loan-to-value ratio.
- Jump parameters $$\rho = 0.73/\text{year}$$, $$\psi = 6.3$$ (implying 14% expected
  price drop per jump) follow Barro and Jin (2011) and the rare-disaster
  literature.

**Comparative statics** are conducted by varying one parameter at a time
(Figures 2-9, pp. 205-217): $$\sigma$$ from 10% to 50% (Fig. 6); $$\kappa$$ from
0.79 to 0.59 (Fig. 5); $$\sigma$$ from 30% (Case A) to 24% (Case B) in the
jump-diffusion extension (Figs. 7-8).

Model predictions are discussed qualitatively against empirical findings in
Barber et al. (2019), Heimer and Imas (2022), An et al. (2024), and
Hartzmark (2015) but the paper does not run statistical tests against data.

## Datasets used

This paper is purely theoretical. No empirical dataset is used; all results
are derived analytically or via numerical solution of the model. No data tags
apply.

| Dataset | Role in paper | Wiki page |
|---|---|---|
| None | Theory and calibration only; parameter values ($$\lambda$$, $$\mu$$, $$\sigma$$, $$r$$, etc.) are taken from prior literature (Andersen et al. 2022, standard equity-premium estimates) | N/A |

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.13472) if you are: extending
the realization-utility framework to allow saving or leverage; studying the
disposition effect under jump risk; looking for the closed-form solution
procedure (Appendix B) or the variational-inequality proofs (Appendix A);
or evaluating the model's quantitative calibration against empirical
disposition-effect magnitudes. The locators above point to the exact figures.
For "what did this paper find," the table above is sufficient.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 81(1), February 2026,
pp. 189–238. DOI: 10.1111/jofi.13472. © 2026 the American Finance
Association. This distillation was extracted by an LLM on 2026-05-31 and is
**not human-verified or independently reproduced**. The article is paywalled;
no open-access or CC licence was found in Crossref metadata. This page
contains only extracted findings (extract-only).

> Dai, Min, Cong Qin, and Neng Wang. "Dynamic Trading with Realization
> Utility." *The Journal of Finance* 81, no. 1 (February 2026): 189–238.
> DOI: 10.1111/jofi.13472.


==============================================================================
# Default Risk and Sovereign Bond Pricing: Dittmar, Hsu, Roussellet & Simasek (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2026/dittmar-default-risk-sovereign-bonds-2026/
# Distilled: U.S. Treasury default risk is significantly positively related to the spread between inflation-linked swap rates and breakeven inflation (ILSBEI); the channel operates primarily through inflation dynamics upon default, not differential recovery rates; a no-arbitrage affine term structure model shows credit risk explains most of the ILSBEI differential at longer maturities outside the financial crisis. J. Finance 2026, paywalled. Eight core results with source locators, datasets used, the model equations, the estimation method, and the empirical specifications.
# Tags: paper-summary, fixed-income, sovereign-debt, default-risk, inflation, term-structure, tips, breakeven-inflation, panel-regression, instrumental-variables, affine-term-structure, peer-reviewed, unreplicated, data:gsw-yields, data:bloomberg, data:fred, data:bls
==============================================================================

**What this is.** The paper's core results, the model it builds on (an affine
no-arbitrage term structure with sovereign default), the estimation method
(Extended Kalman Filter maximum likelihood), and the empirical specifications
(IV regression, GMM): enough to know what it found and how, without reading
all 42 pages. To replicate or extend it, read the full source at
[https://doi.org/10.1111/jofi.70014](https://doi.org/10.1111/jofi.70014) (paywalled).

## TL;DR

The ILSBEI differential is a version of the TIPS-Treasury no-arbitrage
mispricing studied by Fleckenstein, Longstaff, and Lustig (2014), and this
paper extends their liquidity and slow-moving-capital account by adding a
credit risk channel. Treating U.S. default risk as nontrivial follows the
macrofinance view of U.S. sovereign CDS premia in Chernov, Schmid, and
Schneider (2020), applied here to relative bond pricing.

Using monthly U.S. data from June 2005 to December 2020, the paper
documents that the spread between the inflation-linked swap (ILS) rate and
the breakeven inflation rate (BEI), the ILSBEI differential, is
significantly positively related to two measures of U.S. sovereign default
risk: growth in Treasury debt held by the public and Euro-denominated
five-year CDS spreads. Controlling for liquidity (VIX, LIBOR-OIS, OTR
spread, HPW noise), debt growth remains a robust predictor. Treasury debt
growth is used as an instrument for CDS to address endogeneity, and
instrumented CDS loads significantly on the ILSBEI spread across tenors of
two through seven years. A new affine no-arbitrage term structure model
estimated by Extended Kalman Filter decomposes the ILSBEI spread into
credit and liquidity components: at the 10-year maturity, most of the
ILSBEI spread is explained by the credit component outside the financial
crisis. The dominant transmission channel is the correlation between
inflation dynamics and default probability (hyperinflation upon default),
not differential loss-given-default between nominal Treasuries and TIPS.

## Core results

Magnitudes and significance are as reported; `*`/`**`/`***` = 10%/5%/1%.
Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **ILSBEI differential is positively related to Treasury debt growth** (G) | Table II, p. 837 | G coef 0.013\*\*\* (level), 0.025\*\*\* (first diff); R^2 = 0.217 (level), 0.113 (FD); debt growth captures ~22% of level variation |
| R2 | **Debt growth relationship is robust to liquidity and slow-moving-capital controls** (HPW noise, LIBOR-OIS, VIX) | Table III, p. 839 | G coef 0.004\*\* (level), 0.018\*\*\* (FD) with full liquidity controls; HPW noise also significantly positive (0.053\*\*\* level); overall R^2 = 0.808 |
| R3 | **Postcrisis (2010-2020) results are qualitatively similar**; CDS is a stronger predictor than in the full sample | Table IV, p. 840 | CDS coef 0.352\*\*\* (spec 2); with both G and CDS, CDS coef 0.287\*\*\* and G coef 0.002\* |
| R4 | **Treasury debt growth is a strong instrument for CDS**; instrumented CDS is significantly positively related to ILSBEI | Table V Panel B, p. 842 | 2SLS: instrumented CDS coef on ILSBEI = 1.205\*\* (SE 0.475); on ILS = -2.174\*\*\*, on nominal TSY = -3.643\*\*; TIPS coefficient insignificant (-0.265) |
| R5 | **Effect is robust across maturities 2-7 years**; 10-year loading positive but not statistically significant | Table VI, p. 843 | Instrumented CDS on ILSBEI: 2y = 1.097\*\* (SE 0.522), 3y = 1.324\*\*\* (SE 0.511), 7y = 0.868\*\* (SE 0.341), 10y = 0.385 (SE 0.250) |
| R6 | **Term structure model fits the data well** (R^2 93-99% for ILS; 99%+ for nominals; ILSBEI R^2 81-91% for maturities 2-7y) | Table VIII, p. 859 | ILS RMSE 6-13 bps; Nominal RMSE 6-10 bps; ILSBEI RMSE 8-11 bps (range 7.77-10.92 bps across maturities); model-implied default probabilities peak at ~3% (10y) during 2008-2009 crisis |
| R7 | **Credit component of ILSBEI is small at short maturities but large at long maturities**; outside the crisis, more than half of the 10-year spread is credit | Table IX / Figure 7, pp. 859-862 | Mean credit component (Cred.): 7.57 bps (2y), 10.07 bps (3y), 14.46 bps (5y), 18.16 bps (7y), 23.05 bps (10y); credit share rises from ~46% (2y) to ~78% (10y) on average |
| R8 | **Dominant channel is inflation jump upon default** (hyperinflation), not differential LGD or default priced in SDF alone | Figure 10 / Section IV.F, pp. 865-866 | Comparative statics: inflation-upon-default channel (kappa_delta^pi) contributes 1-10 bps to ILSBEI (95% CI, orange curve); inflation/PD correlation channel contributes ~1 bp at 10y median (~3.5 bps at 97.5th percentile); default-in-SDF channel contributes near zero |

**Overall (paper's conclusion).** Credit risk can drive persistent deviations
between ILS rates and BEI rates that are often attributed solely to
liquidity. Where Pflueger and Viceira (2016) attribute the ILSBEI differential
to a liquidity premium, this paper finds credit risk is also a nontrivial
driver, especially at long maturities. The interaction between inflation
dynamics and default is the primary source of differential pricing between
nominal Treasuries and TIPS.

## Theory / model

The paper builds intuition through a one-period two-equation toy model
(Section II, p. 844) before embedding these mechanisms in the full
multivariate term structure model (Section III, p. 852).

**Toy model (Section II, eq. 1, p. 844).** The real pricing kernel and
inflation rate are:

$$
\log M^*_{t+1} = \bar{M} + \Lambda_\delta \cdot \mathbf{1}\{\delta^{(c)}_{t+1} > 0\}
$$

$$
\pi_{t+1} = \kappa_0 + \kappa_y \lambda_t + \kappa_\delta \cdot \mathbf{1}\{\delta^{(c)}_{t+1} > 0\}
$$

where $$\delta^{(c)}_{t+1}$$ is a nonnegative default process with conditional
jump probability $$\lambda_t$$, $$\Lambda_\delta > 0$$ is the price of default
risk, $$\kappa_0$$ governs average inflation, $$\kappa_y$$ is the loading of
inflation dynamics on default probability, and $$\kappa_\delta$$ is the
inflation jump upon default. Under independence (default event is transitory),
all bond prices are in closed form.

**Bond yield decomposition (eq. 2, p. 845).** Any nominal Treasury yield can
be decomposed into four parts:

$$
\begin{aligned}
R^{(n)}_t \;=\; & r^{(n)*}_t & &\text{(a) real risk-free rate} \\
               + & \bigl(r^{(n)}_t - r^{(n)*}_t\bigr) & &\text{(b) ILS} \\
               + & \bigl(R^{(n)*}_t - r^{(n)*}_t\bigr) & &\text{(c) real credit spread} \\
               + & \bigl(R^{(n)}_t - r^{(n)}_t\bigr) - \bigl(R^{(n)*}_t - r^{(n)*}_t\bigr) & &\text{(d) -ILSBEI}
\end{aligned}
$$

where $$*$$ denotes real yields and capitalization denotes defaultable bond
yields. The ILSBEI spread in component (d) is negative when the nominal
credit spread exceeds the real credit spread, that is, when nominal
Treasuries are more exposed to default than TIPS. Expanding around small
default probability $$\lambda_t$$, the nominal Treasury yield decomposes as
(eqs. 3.a-3.d, pp. 845-846):

$$
\begin{aligned}
R^{(n)}_t \;\approx\; & -\bar{M} + \lambda_t(1 - e^{\Lambda_\delta}) & &\text{(3.a: risk-free real yield)} \\
                    + & \kappa_0 + \kappa_y \lambda_t + \lambda_t e^{\Lambda_\delta}(1 - e^{-\kappa_\delta}) & &\text{(3.b: ILS)} \\
                    + & \lambda_t e^{\Lambda_\delta} \cdot \text{LGD}^* & &\text{(3.c: real credit spread)} \\
                    + & \lambda_t e^{\Lambda_\delta}(e^{-\kappa_\delta} \cdot \text{LGD} - \text{LGD}^*) & &\text{(3.d: -ILSBEI)}
\end{aligned}
$$

where $$\text{LGD}$$ ($$\text{LGD}^*$$) is the loss given default of a nominal (real) Treasury.
Equation (3.d) shows that even equal LGDs produce a nonzero ILSBEI if
$$\kappa_\delta > 0$$, that is, if there is hyperinflation upon default.

**Three channels.** The model identifies three channels through which
default affects ILSBEI (pp. 844, 865):

1. Default priced in the SDF ($$\Lambda_\delta$$): a price of default risk that
   lowers riskless real bond yields through the pricing kernel.
2. Inflation jump upon default ($$\kappa_\delta^{(\pi)}$$): hyperinflation upon
   default raises the value of inflation swaps relative to nominal
   Treasuries, widening ILSBEI.
3. Negative correlation between default probability and inflation
   ($$\kappa_y^{(\pi)} < 0$$): higher default probability predicts lower current
   inflation, lowering the ILS rate.

**Lucas tree motivation (Section II.D, p. 850).** The inflation/default
correlations arise naturally in a CRRA representative-agent Lucas tree
economy with long-run risk and a central bank following a Taylor rule with
coefficient on inflation less than one (passive monetary policy). The
passive central bank stance means the Taylor principle is violated: inflation
must fall when default probability rises, consistent with the negative
$$\kappa_y$$ in the regression results.

## Method

**Full term structure model (Section III, pp. 852-855).** The paper builds a
dynamic discrete-time affine no-arbitrage term structure model following
Monfort et al. (2020), using `affine-term-structure` and `gamma-zero-processes`
as building blocks.

**Risk factors (eqs. 5-8, pp. 852-853).** Three blocks of state variables:

Riskless real short rate driven by three Gaussian factors $$x_t$$ (eq. 5):

$$
r^{(1)*}_t = \kappa_0^{(r)} + \kappa_x^{(r)\prime} x_t
$$

with VAR(1) dynamics (eq. 6):

$$
x_t = \Phi_x x_{t-1} + \epsilon_{x,t}, \qquad \epsilon_{x,t} \sim \text{iid}\; N(0, I_3)
$$

Credit and liquidity event processes $$\delta_t = (\delta^{(c)}_t, \delta^{(l)}_t)$$
modelled as gamma-zero processes (eq. 7): for $$i \in \{c, l\}$$,

$$
\delta^{(i)}_t = \sum_{j=1}^{P^{(i)}_t} \xi^{(i)}_{j,t},
\quad \text{where } P^{(i)}_t \mid \lambda^{(i)}_t \sim \text{Poisson}(\lambda^{(i)}_t)
\text{ and } \xi^{(i)}_{j,t} \sim \text{Exp}(1/c^{(i)}_\delta)
$$

Event intensities driven by nonnegative factors $$y_t = (y^{(c)}_{1,t}, y^{(c)}_{2,t}, y^{(l)}_t)$$ following VARG dynamics (eq. 8):

$$
\begin{aligned}
\lambda^{(c)}_t &= \beta^{(c)}_{\lambda,1} y^{(c)}_{1,t} + \beta^{(c)}_{\lambda,2} y^{(c)}_{2,t} \\
\lambda^{(l)}_t &= \beta^{(l)}_\lambda y^{(l)}_t
\end{aligned}
$$

**Pricing kernel (eq. 9, p. 853):**

$$
\log(M^*_{t+1}) = -r^{(1)*}_t + \Lambda_{x,t}' x_{t+1} + \Lambda_y' y_{t+1}
                  + \Lambda_\delta \delta^{(c)}_{t+1} - \zeta_t
$$

where $$\Lambda_{x,t} = \Lambda_{0,x} + \Lambda_{1,x} x_t$$ are the prices of
riskless factor risk, $$\Lambda_y$$ are the prices of credit/liquidity factor
risk, and $$\Lambda_\delta$$ is the price of credit event risk. The structure-
preserving property of the pricing kernel ensures the model belongs to the
affine class, so all pricing formulas are in closed form.

**Inflation dynamics (eq. 10, p. 854):**

$$
\pi_t = \kappa_0^{(\pi)} + \kappa_x^{(\pi)\prime} x_t + \kappa_y^{(\pi)\prime} y_t + \kappa_\delta^{(\pi)} \delta^{(c)}_t
$$

The expected positive sign on $$\kappa_\delta^{(\pi)}$$ (hyperinflation upon
default) and expected negative sign on $$\kappa_y^{(\pi)}$$ (lower inflation when
default probability is high) are both key model predictions tested against
the data.

**Estimation (eq. 11, p. 856).** The model is cast in state-space form with
23 observable variables $$\mathbf{y}_t \in \mathbb{R}^{23}$$ (ILS, nominal yields, TIPS, CDS at
multiple maturities, OIS, CPI, TIPS liquidity proxy):

$$
\mathbf{y}_t = F(x_t, y_t, \delta_t, \theta^Q) + \eta_t, \qquad \eta_t \sim N(0, \Sigma_\eta)
$$

where $$F(\cdot)$$ is a nonlinear closed-form function summarizing the affine
pricing equations and $$\theta^Q$$ is the set of risk-neutral parameters.
Since $$F(\cdot)$$ is nonlinear (due to the gamma-zero processes), the model is
estimated by Extended Kalman Filter approximate maximum likelihood (`extended-kalman-filter`). CDS measurement error standard deviations are allowed to
depend on a CDS liquidity proxy (Grischenko and Huang (2013)) to account
for mismeasurement.

## Empirical specifications

**Baseline regression (Tables II-IV, pp. 837-840).** OLS in levels and first
differences of the five-year ILSBEI spread on measures of default risk:

$$
\text{ILSBEI}_t = \alpha + \beta G_t + \epsilon_t \qquad \text{(level)}
$$

$$
\Delta\text{ILSBEI}_t = \alpha + \beta\, \Delta G_t + \epsilon_t \qquad \text{(first diff)}
$$

- $$\text{ILSBEI}_t$$: five-year inflation-linked swap rate minus breakeven inflation rate at time $$t$$
- $$G_t$$: year-over-year log growth in Treasury debt held by the public (level regressions) or monthly log variation (first-difference)
- $$\alpha$$: intercept
- $$\beta$$: slope coefficient on default risk proxy
- $$\epsilon_t$$: error term
- Standard errors: Newey-West with three lags
- Sample: June 2005 to December 2020, monthly

Extended specification adds liquidity controls:

$$
\text{ILSBEI}_t = \alpha + \beta_G G_t + \beta_{\text{CDS}} \text{CDS}_t
                 + \beta_{\text{VIX}} \text{VIX}_t + \beta_{\text{OIS}} (L\text{-}OIS)_t
                 + \beta_{\text{OTR}} \text{OTR}_t + \beta_{\text{HPW}} \text{HPW}_t + \epsilon_t
$$

- $$G_t$$: Treasury debt growth (default risk proxy)
- $$\text{CDS}_t$$: Euro-denominated five-year U.S. sovereign CDS spread
- $$\text{VIX}_t$$: CBOE Volatility Index (liquidity/risk-aversion control)
- $$(L\text{-}OIS)_t$$: LIBOR-OIS spread (counterparty risk / liquidity control)
- $$\text{OTR}_t$$: off-the-run/on-the-run 10-year Treasury spread (liquidity control)
- $$\text{HPW}_t$$: Hu-Pan-Wang noise measure (slow-moving capital proxy)

**Instrumental variables (Tables V-VI, pp. 841-843).** Single-stage GMM
using $$G_t$$ (Treasury debt growth) as an instrument for CDS spreads.
First stage: regress each liquidity/slow-moving-capital control on $$G_t$$
to obtain orthogonalized residuals $$\text{VIX}^\perp$$, $$(L\text{-}OIS)^\perp$$, $$\text{OTR}^\perp$$,
$$\text{HPW}^\perp$$; then regress $$\text{CDS}_t$$ on $$G_t$$ and these residuals. Second stage:
regress $$\text{ILSBEI}_t$$ (and individually: ILS, nominal TSY, TIPS) on the
first-stage predicted $$\widehat{\text{CDS}}_t$$ and the orthogonalized liquidity controls.
Standard errors are Newey-West with three lags, estimated simultaneously
via single-stage GMM. Specification repeated at 2y, 3y, 5y, 7y, 10y tenors.
Identifying assumption: Treasury debt issuance decisions are exogenous to
financial market pricing frictions (p. 841).

**Term structure model estimation (Section IV, pp. 855-856).** Extended
Kalman Filter maximum likelihood on 23 observable monthly series from
November 2004 to December 2019 (data availability for TIPS liquidity proxy).
Observable variables are the term structures of ILS (2-10y), nominal
Treasuries (1-10y), TIPS (implied from BEI), sovereign CDS (5y, 10y), OIS
(6m), monthly CPI inflation, and the Grischenko-Huang TIPS liquidity index.
The riskless real one-period yield and all bond prices are derived in closed
form under the affine structure. Model parameters are estimated jointly
including $$\kappa_\delta^{(\pi)}$$, $$\kappa_y^{(\pi)}$$, $$\Lambda_\delta$$, recovery fractions,
and the TIPS disindexation rate $$\rho^*$$.

**Comparative statics (Figure 10, Section IV.F, pp. 865-866).** Sequential
counterfactuals setting $$\Lambda_\delta = 0$$, $$\kappa_\delta^{(\pi)} = 0$$, or
$$\kappa_y^{(\pi)} = 0$$ in turn, computing the deviation between fitted and
counterfactual yield curves. The median contribution of each channel to
ILSBEI is measured in basis points at each maturity, with 95% confidence
intervals across sample dates.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Gurkaynak, Sack, and Wright (GSW) zero-coupon yields (nominal Treasury and TIPS) | BEI construction; nominal yield curve estimation | no page yet (Federal Reserve Board data, publicly available) |
| Bloomberg ILS rates (zero-coupon, 2-10y maturities) | ILSBEI spread construction; model estimation target | [Bloomberg](/wiki/commercial/bloomberg/) (licensed) |
| U.S. Treasury debt held by public (monthly, Federal Reserve/Treasury) | Default risk proxy (G); instrument for CDS | [FRED](/wiki/datasets/fred/), series available via FRED |
| Euro-denominated 5-year U.S. Treasury CDS spreads | Alternative default risk measure; model estimation target | [Bloomberg](/wiki/commercial/bloomberg/) / [Markit CDS](/wiki/commercial/markit-cds/) (licensed) |
| BLS CPI-U (monthly) | Inflation measure for model | [BLS](/wiki/datasets/bls/) (CPI-U), pulled via [FRED](/wiki/datasets/fred/) series `CPIAUCSL` |
| VIX (CBOE) | Liquidity/slow-moving capital control | [FRED](/wiki/datasets/fred/), series `VIXCLS` |
| LIBOR-OIS spread | Liquidity control (counterparty risk) | no page yet |
| Off-the-run / on-the-run 10y Treasury spread (OTR) | Liquidity control | no page yet (derived from GSW and Bloomberg) |
| HPW noise measure (Hu, Pan & Wang 2013) | Slow-moving capital proxy; TIPS liquidity intensity proxy | no page yet (academic dataset, Jun Pan's website) |
| OIS 6-month rate | Short-term riskless nominal yield for model | [Bloomberg](/wiki/commercial/bloomberg/) (licensed) |

Sample: monthly, Nov 2004 to Dec 2019 (model estimation); Jun 2005 to Dec 2020 (regressions).

## When to read the full paper

Consult the original via [https://doi.org/10.1111/jofi.70014](https://doi.org/10.1111/jofi.70014)
if you are: replicating (code in journal Supporting Information); extending
the affine model to other sovereign issuers; studying how monetary policy
passivity interacts with default to generate inflation dynamics; auditing a
specific coefficient; or tracing the closed-form bond pricing derivations
(Internet Appendix Sections I-XII). The locators above point to the exact
tables and figures. For "what did this paper find," the table above is
sufficient and is the intended default.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 81(2). (c) 2026 the
American Finance Association. This distillation was extracted by an LLM
on 2026-05-31 and augmented on 2026-06-01; it is **not human-verified or
independently reproduced**. The paper is paywalled; no verbatim content is
reproduced here.

> Dittmar, Robert F., Alex Hsu, Guillaume Roussellet, and Peter Simasek.
> "Default Risk and the Pricing of U.S. Sovereign Bonds." *The Journal of
> Finance* 81, no. 2 (April 2026): 829-869. DOI: 10.1111/jofi.70014.
> (c) 2026 the American Finance Association. All rights reserved.
> This page is an **extract-only distillation** by the Institute for
> Automated Research: core results summarised; no verbatim text reproduced.


==============================================================================
# FinTech Lending and Cashless Payments: Ghosh, Vallee & Zeng (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2026/ghosh-fintech-lending-cashless-payments-2026/
# Distilled: Borrowers' use of cashless payments improves access to capital from FinTech lenders and predicts lower default probability, with outflows and information-intensive payment records showing the strongest effects. J. Finance 2026, CC BY-NC 4.0. Ten core results with source locators, datasets used, the signaling model, and empirical specifications.
# Tags: paper-summary, fintech, lending, cashless-payments, credit-markets
==============================================================================

**What this is.** The paper's core results, the signaling model it contributes,
and the empirical specifications behind each finding: enough to know what it
found and how, without reading all 49 pages. To replicate or extend it, read
the full source at the [original](https://doi.org/10.1111/jofi.70003).

## TL;DR

Using a dataset of 316,719 loan applications to Indifi, an Indian
FinTech lender (2015-2022), the paper shows that borrowers who conduct more of
their business through cashless payment technologies obtain better financing
outcomes: higher loan approval rates, lower interest rates, and larger loan
amounts. They also default less. These patterns are stronger for payment
outflows than inflows, for information-intensive than information-light payment
records, and for applicants with higher credit scores (an accuracy effect).
A within-applicant specification with applicant fixed effects and an
instrumental variable strategy exploiting the 2016 Indian Demonetization
support a causal interpretation. The research question is motivated by the
FinTech lending landscape surveyed in Berg, Fuster, and Puri (2022), and the
result that digital footprints complement credit bureau data, documented in
Berg et al. (2020), here confirmed for cashless payment records specifically. The authors rationalize the findings with a
signaling model in which cashless payment records serve as "digital
collateral": bad borrowers face higher expected costs from posting their
records (because discrepancies are more easily detected upon default), so only
good borrowers endogenously choose cashless payments.

## Core results

Magnitudes and significance are as reported; `**`/`***` = 5%/1%. Locators
point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | A one-SD increase in cashless payment share raises loan approval probability by **2 percentage points** (full controls) | Table II, col. (2), p. 1067 | $$\beta = 0.017^{***}$$ (SE 0.001); baseline approval ~21%; represents ~8% of baseline |
| R2 | A one-SD increase in cashless payments **reduces offered interest rate by 44 bps** (full controls) | Table II, col. (4), p. 1067 | $$\beta = -0.440^{***}$$ (SE 0.024); mean rate 25.4% |
| R3 | A one-SD increase in cashless payments **raises offered loan amount by 13%** (log) | Table II, col. (6), p. 1067 | $$\beta = 0.134^{***}$$ (SE 0.006) |
| R4 | A one-SD increase in cashless payments **reduces default probability by 2 pp** (11% of baseline) | Table III, col. (2), p. 1069 | $$\beta = -0.023^{***}$$ (SE 0.003); baseline default rate ~18% of matured loans |
| R5 | **Payment outflows** have 2x larger effects on loan approval and amount than inflows; interest-rate effects are not significantly different by direction | Table V, Panel A, p. 1073 | Outflow approval $$\beta = 0.022^{***}$$ vs inflow $$\beta = 0.008^{***}$$; outflow default $$\beta = -0.021^{***}$$ vs inflow $$\beta = -0.009^{***}$$ |
| R6 | **Information-intensive** cashless payments (individually identifiable counterparty) have larger effects than information-light payments | Table V, Panel B, p. 1073 | Info-intensive approval $$\beta = 0.050^{***}$$ (vs Table II baseline $$0.017^{***}$$); info-intensive default $$\beta = -0.017^{***}$$ |
| R7 | **Within-applicant** (repeat borrowers, applicant FE): a firm-level increase in cashless payments raises approval probability; applicant fixed effects absorbed | Table VI, col. (2), p. 1075 | $$\beta = 0.009^{***}$$ (SE 0.003); spending-only $$\beta = 0.010^{***}$$, info-intensive $$\beta = 0.021^{***}$$ |
| R8 | Cashless payment benefit is **complementary, not substitute**, with credit score (accuracy effect): interaction term positive and significant for approval and amount | Table VII, Panel A, p. 1077 | Cashless $$\times$$ Cibil Score $$\beta$$ on approval $$= 0.012^{***}$$ (SE 0.001); on log amount $$= 0.039^{***}$$ (SE 0.005) |
| R9 | **2SLS IV estimate** (Demonetization $$\times$$ currency chest instrument): a one-SD increase in cashless use raises approval by **12 pp** (larger than OLS, consistent with attenuation bias) | Table VIII, col. (4), p. 1082 | IV $$\beta = 0.117^{*}$$ (SE 0.064), second-stage col. (4); full-FE first-stage F-stat = 26.90 (col. 2); banking at chest branch in 2018-19 reduces cashless share by ~10 pp |
| R10 | **Online marketplace transactions** (BigTech-style digital records) show similar effects to cashless payments and their effects are cumulative | Table IV, p. 1071 | Marketplace approval $$\beta = 0.061^{***}$$ (SE 0.002); conditional on cashless share, marketplace $$\beta$$ remains $$0.016^{***}$$; default $$\beta = -0.006^{***}$$ |

**Overall (paper's conclusion).** The use of cashless payments provides an
informational synergy for FinTech lending that operates through a selection
mechanism (payment technology choice signals creditworthiness) and an accuracy
effect (the signal is most useful for otherwise good borrowers). This provides
a rationale for the joint rise of cashless payments and FinTech lending, and
for open banking policies that expand access to historical payment records.
This complements the open banking and FinTech-bank competition analysis of
He, Huang, and Zhou (2023) by showing the informational value of payment
records.

## Theory / model

The paper develops a three-date model of signaling via payment technology
choice, building on Parlour, Rajan, and Zhu (2022). There is a continuum of
risk-neutral entrepreneurs (fraction $$\alpha$$ good, $$1-\alpha$$ bad) and a competitive
lender. Entrepreneurs operate a production technology $$n$$ times during $$t=0$$ to
$$t=1$$ (the "production stage"). Output $$y_i \in \{0, \theta\}$$ with failure probability
$$\pi_j$$ for type $$j \in \{G, B\}$$ (p. 1083).

At t=0, the entrepreneur chooses a payment method: cash (production outcomes
$$Y$$ are unverifiable) or cashless (outcomes become verifiable payment records $$X$$).
The cashless payment technology generates a record $$x_i$$ for each production
outcome $$y_i$$ with informational precision $$q \in [1/2, 1]$$ (p. 1084, eq. 5):

$$
\Pr(x_i = \theta \mid y_i = \theta) = \Pr(x_i = 0 \mid y_i = 0) = q
$$

$$
\Pr(x_i = 0 \mid y_i = \theta) = \Pr(x_i = \theta \mid y_i = 0) = 1 - q
$$

If a borrower with cashless payment records defaults at t=2, she incurs a cost
$$C$$ proportional to the discrepancies between production capacity $$\theta$$ and
realized records $$X$$ (p. 1084, eq. 6):

$$
C = \phi \sum_{i=1}^{n} (\theta - x_i), \quad \phi > 0
$$

This cost $$C$$ rises with the number of cashless records $$n$$, failure probability
$$\pi_j$$ (bad types have more deviations), and informational precision $$q$$. In
equilibrium the lender offers two contracts: $$r_{\text{cashless}}$$ (for borrowers who
committed to cashless at t=0) and $$r_{\text{cash}}$$ (for those who did not). The
borrower type $$j$$'s expected utility under each contract is (pp. 1085-1086,
eqs. 7-8):

$$
w_j(r_{\text{cashless}}, C \mid \pi_j) = (1 - \pi_j)(\theta - r_{\text{cashless}}) - \pi_j E_j[C]
$$

$$
w_j(r_{\text{cash}} \mid \pi_j) = (1 - \pi_j)(\theta - r_{\text{cash}})
$$

**Proposition 1** (p. 1086): A separating equilibrium with $$r_{\text{cashless}} < r_{\text{cash}}$$
exists when two conditions hold:

(i) The cost $$\phi n \theta$$ is intermediate so that good types prefer cashless but
bad types do not (incentive compatibility, eq. 14):

$$
\frac{\pi_B - \pi_G}{\pi_B (1 - \pi_G)(1 - q - \pi_B + 2q\pi_B)}
  \leq \phi n \theta \leq
\frac{\pi_B - \pi_G}{(1 - \pi_B)\pi_G (1 - q - \pi_G + 2q\pi_G)}
$$

(ii) The average borrower quality $$(1 - \bar{\pi})$$ is not so high that a pooling
contract dominates (eq. 15):

$$
1 - \bar{\pi} \leq \frac{1 - \pi_G}{1 + \phi n \theta \pi_G (1 - q - \pi_G + 2q\pi_G)}
$$

The selection effect is stronger when $$n$$ (number of cashless records) is larger
and when $$q$$ (informational precision) is higher, which the model maps directly
to the empirical patterns: outflows have higher $$q$$ than inflows, and
information-intensive records have higher $$q$$ than information-light records.

**Proposition 2** (p. 1088, eq. 16): The accuracy effect: in the separating
equilibrium, the benefit from cashless payments $$(r_{\text{cash}} - r_{\text{cashless}})$$ is larger
for firms with lower relative default probability:

$$
\frac{d(r_{\text{cash}} - r_{\text{cashless}})}{d(\pi_B - \pi_G)} > 0
$$

This implies better entrepreneurs benefit more, consistent with the empirical
complementarity between cashless payment usage and credit score (Table VII).

## Method

The paper is primarily an applied empirical paper (`applies-method`), but also
contributes a signaling model. The econometric approach uses four distinct
estimators:

**Cross-sectional OLS (baseline, `panel-regression`):** The financing outcome
specifications (eqs. 1 and 2) are cross-sectional OLS over the full applicant
sample, with a comprehensive set of fixed effects to nonlinearly absorb
selection on observable dimensions. Standard errors are clustered at the
application-month level throughout.

**Within-applicant panel regression (eq. 3, `panel-regression`):** To address
time-invariant unobserved heterogeneity, repeat borrowers are used with
applicant fixed effects and application-year fixed effects, exploiting
within-firm time-series variation in cashless payment usage.

**Text classification (`text-classification`):** Payment records from bank
statements are classified into cash vs. cashless, and further into
information-intensive vs. information-light, using text analysis of payment
labels (Appendix A). This covers 75% of all payment records; unclassified
records show no meaningful predictive power.

**2SLS instrumental variables (`instrumental-variables`):** The share of
cashless payments is instrumented by the interaction of an indicator for
banking at a currency chest branch and an indicator for the 2018-2019 period
when Demonetization cash shortages were most severe (Table VIII). Currency
chest branches had better access to new banknotes post-Demonetization, so
their clients reverted to cash more than non-chest clients during 2018-2019,
creating plausibly exogenous within-district variation in cashless payment use.
The cross-district currency chest variation from the 2016 Demonetization used
here as identification follows Chodorow-Reich et al. (2020), and the design is
adapted to within-district variation in the spirit of Crouzet, Gupta, and
Mezzanotti (2023), who also leverage the same Demonetization variation.

## Empirical specifications

### Baseline financing outcomes (R1-R3, R5-R6, R8, R10)

Specification (1), p. 1065:

$$
\text{FinancingOutcome}_i = \beta \cdot \text{CashlessShare}_i + \gamma X_i
                   + \sum_k \theta_{F_k(i)} + \epsilon_i
$$

- $$\text{FinancingOutcome}_i$$ is (i) an indicator for loan approval, (ii) offered interest rate, or (iii) log offered loan amount.
- $$\text{CashlessShare}_i$$ is the standardized amount-weighted share of transactions in cashless technology over six months prior to application (average of inflow and outflow shares).
- $$X_i$$ includes log number of payments, credit history length, business vintage, log owner age, missing credit score indicator, and top-up loan indicator.
- Fixed effects $$F_k(i)$$ include industry (67), application month, Cibil score group (10-point bands), 3-digit zip code, and revenue decile.
- Sample: 311,942-314,538 observations (cols. 1-6 of Table II). Standard errors: clustered at application month level.

### Loan default (R4)

Specification (2), p. 1068:

$$
\text{Default}_i = \beta \cdot \text{CashlessShare}_i + \gamma X_i
           + \sum_k \theta_{F_k(i)} + \epsilon_i
$$

- Same controls and fixed effects as specification (1).
- Sample: restricted to matured/outstanding loans as of November 2022 (41,227 observations, Table III cols. 1-4).
- Additional columns (3-4) add log offered interest rate and log disbursed amount as controls.
- Columns (5-8) use time-to-first-delinquency and time-to-full-repayment as dependent variables (OLS on the loans that reached the respective event; note repayment regression run internally by Indifi due to data-privacy regulation).

### Complementarity with credit quality (R8)

Specification (4), p. 1075:

$$
\text{FinancingOutcomes}_i = \beta_1 \cdot \text{CashlessShare}_i + \beta_2 \cdot \text{FirmQuality}_i
                    + \beta_3 \cdot \text{CashlessShare}_i \times \text{FirmQuality}_i
                    + \gamma X_i + \sum_k \theta_{F_k(i)} + \epsilon_i
$$

- $$\text{FirmQuality}$$ proxied by (i) Cibil score (Table VII Panel A, N=277,323) and (ii) weekly outflow volatility constructed from payment records (Table VII Panel B, N=311,938).
- Full set of fixed effects including credit score band FE.

### Instrumental variable (R9)

First stage of specification (Table VIII), p. 1082:

$$
\text{CashlessShare}_i = \delta \cdot (\text{ChestBank}_i \times \text{Year2018-19}_i) + \lambda \cdot \text{ChestBank}_i
               + \gamma X_i + \sum_k \theta_{F_k(i)} + v_i
$$

- Table VIII layout (p. 1082): cols. (1) and (2) report the first stage (year FE only, then full fixed effects); cols. (3), (4), and (5) report the second stage (approved loan indicator on instrumented cashless share).
- The effective F-statistic is reported only for the first-stage columns: 37.41 (col. 1) / 26.90 (col. 2), above Montiel Olea-Pflueger critical values.
- Column (5) additionally instruments the interaction of cashless share $$\times$$ Cibil Score.
- Sample: 316,407 (year FE) / 311,941 (full FE) observations.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Indifi loan application data (proprietary) | Primary dataset: 316,719 complete applications with payment records, applicant characteristics, credit bureau data, and loan outcomes (September 2015 to November 2022) | No page yet |
| Indian Cibil (credit bureau) scores | Borrower credit score control variable; credit history length and past loans | No page yet |
| Indian Demonetization / currency chest branch designations | Identification of IV; Reserve Bank of India currency chest branch list | No page yet |

Sample: 316,719 complete loan applications from micro and small businesses,
with 152 million transactions; approximately half are repeat applicants
traceable via unique applicant identifier. Loan outcomes available for 66,017
approved loans; default outcome for 41,123 matured loans as of November 2022.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.70003) if you are:
studying the informational role of payment technology in credit markets;
designing open banking policies (the model and discussion in Sections VI-VII
give a formal rationale); working on FinTech or BigTech lending mechanisms;
interested in the Indian Demonetization as a natural experiment for financial
digitization; or building on the digital-collateral concept vs. traditional
collateral. The Internet Appendix (Tables IA.I-IA.XIV) provides extensive
robustness by sector, subperiod, clustering level, and specification.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 81(2), April 2026 (published
online December 18, 2025). This distillation was extracted by an LLM on
2026-06-01 and is **not human-verified or independently reproduced**.
The CC BY-NC 4.0 licence permits reproduction for non-commercial use with
attribution; the verbatim PDF is not hosted in this batch.

> Ghosh, Pulak, Boris Vallee, and Yao Zeng.
> "FinTech Lending and Cashless Payments."
> *The Journal of Finance* 81, no. 2 (April 2026): 1053-1101.
> DOI: 10.1111/jofi.70003. (C) 2025 The Author(s).
> Licensed under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.
> Non-commercial use only.


==============================================================================
# Losing Control: Griffin, Nini & Smith (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2026/griffin-loan-covenant-violations-decline-2026/
# Distilled: the annual share of U.S. public firms reporting a financial covenant violation fell roughly 70% from 1997 to 2019; a structural decomposition shows the drop is driven mainly by fewer false-positive violations and a lower corporate distress rate, not a deterioration in lender monitoring ability. J. Finance 2026, paywalled. Nine core results with source locators, datasets used, and the theory tested.
# Tags: paper-summary, corporate-finance, debt-covenants, credit-markets, lender-control, panel-regression, peer-reviewed, unreplicated, data:wrds, data:edgar, data:dealscan, data:moodys-urd, data:audit-analytics, data:lopucki-brd
==============================================================================

**What this is.** The paper's core results, the model it builds on, and the
method it contributes with the defining equations: enough to know what it
found and how, without reading all 42 pages. To replicate or extend it,
read the full source at the
[original](https://doi.org/10.1111/jofi.70005) (paywalled).

## TL;DR

Using hand-collected SEC filings (10-K/10-Q, 1997-2019) merged with
Dealscan and Compustat, the paper extends the violation sample of Nini,
Smith, and Sufi (2012) from 1997-2008 through 2019 and adds a structural
decomposition framework. It documents that the annual share of U.S.
public firms reporting a financial covenant violation fell roughly 70%,
from about 18% at the 2001 peak to around 5% by 2012 and below that
thereafter. A structural model of optimal covenant design, cast as a
medical-diagnostic analogy (true positives, false positives, true
negatives, false negatives), decomposes the trend. The dominant driver is
a collapse in false-positive violations: lenders set looser covenants,
reducing nuisance violations for nondistressed borrowers. The corporate
distress rate also fell. The true-positive rate (distressed firms that
violate) stayed near 75% through 2011 and only declined modestly after the
global financial crisis. Observable loan-market changes (larger borrowers,
more investment-grade firms, universal-bank arrangers, the shift from
balance-sheet to cash-flow covenants) explain about two-thirds of the
overall drop; the remaining third reflects an unexplained shift in lender
preferences post-2012 consistent with heightened investor sentiment.

## Core results

Magnitudes and significance are as reported; `*`/`**`/`***` = 10%/5%/1%.
Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Covenant **violation rate fell roughly 70%** from 1997 to 2019 | Figure 1, p. 378 | Peak ~18% in 2001; ~5% by 2012; stable at ~5% through 2019; new-violation rate fell from ~9% to ~2% over same period |
| R2 | Loan **covenants became fewer and looser** over 1997-2019 | Figure 2, p. 380 | Average covenants per package: ~2.75 (late 1990s) to ~1.5 (2016); tightest-covenant slack ~3x larger at end of sample vs. late 1990s; ex ante covenant strictness (Murfin measure) fell roughly half |
| R3 | **More than half of violations are false positives** (waived with no consequential change) | Figure 3, p. 385 | True positives 2.5%; false positives 2.9%; false negatives 1.1%; true negatives 93.6% of firm-year observations |
| R4 | **False negatives cost lenders 7-11 pps in creditor recovery rates** at bankruptcy | Table I, p. 387 | No-Violation coefficient: -11.4\*\*\* (col. 1), -10.1\*\*\* (col. 2), -9.8\*\*\* (col. 3), -6.7\*\* (col. 4, with Bank Debt Share); N = 403 corporate defaults from Moody's URD |
| R5 | **FPR fell from ~6% to under 0.5%** (a 90% drop), explaining 61% of total violation decline | Table II, Figure 5, pp. 389-391 | Period 1 (1997-2003) FPR 5.3%; Period 3 (2012-2019) FPR 0.5%; FPR drop explains 4.6 pps of total 7.6 pp decline |
| R6 | **Distress rate decline explains 25%** and **TPR decline explains 14%** of the total drop | Table II Panel B, p. 390-391 | Distress rate contributes -1.9 pps; TPR decline (from ~72% to ~52%) contributes -1.1 pps over full period |
| R7 | **Preference parameter R nearly doubled then nearly quintupled** across periods in structural estimates (roughly tenfold overall), while covenant technology improved modestly | Table III, p. 393 | R: 3.10 (1997-2003) to 6.06 (2004-2011) to 30.29 (2012-2019); technology parameter $$\mu_D$$: 2.21 to 2.68 to 2.65 |
| R8 | **Observable loan-market changes** (borrower size, rating, lender type, covenant type) explain ~52% of FPR decline and ~69% of TPR decline | Table V, p. 402; Table VI, p. 404 | Observable characteristics explain 5.1 pp of the 7.6 pp total violation decline (~two-thirds); borrower size and speculative-grade rating are the largest contributors |
| R9 | **Unexplained post-2012 drop in TPR** is consistent with heightened investor sentiment, not deterioration in covenant technology | Table VII, Figure 10, pp. 405-406 | Absent characteristic changes, R rises ~80% (from 3.10 to 5.69) vs. ~880% raw; counterfactual violation rate 4.0% vs. realized 1.4% in Period 3 |

**Overall (paper's conclusion).** The dramatic decline in covenant
violations is best attributed to a compositional shift in the public
borrower population toward larger, rated firms with transactional lenders
who prefer fewer costly renegotiations, with a modest but real residual
reflecting looser lender preferences post-GFC. The decline does not
primarily signal a deterioration in lenders' ability to monitor distressed
borrowers.

## Theory / model

The paper develops a model of optimal financial covenant thresholds as a
binary-classification problem (Section II, pp. 381-383). Borrowers are
either distressed ($$D = 1$$) or nondistressed ($$D = 0$$). The lender observes
a financial metric $$r$$ (e.g., a coverage or leverage ratio) correlated with
but not perfectly identifying $$D$$. A covenant sets a contractual threshold
$$t$$; a violation occurs when $$r > t$$ (for covenants written so that a higher
ratio signals distress). Distressed and nondistressed firms have
distributions over $$r$$:

$$
\Pr(r < t \mid D = 1) = F_D(t) \qquad \text{(distressed CDF)}
$$

$$
\Pr(r < t \mid D = 0) = F_{ND}(t) \qquad \text{(nondistressed CDF)}
$$

The false negative rate (FNR) and false positive rate (FPR) follow:

$$
\text{FNR}(t) = \Pr(r < t \mid D = 1) = F_D(t)
$$

$$
\text{FPR}(t) = \Pr(r > t \mid D = 0) = 1 - F_{ND}(t)
$$

Because $$\text{FNR}(t)$$ is increasing in $$t$$ and $$\text{FPR}(t)$$ is decreasing in $$t$$, there
is a trade-off: tighter thresholds catch more truly distressed firms
(lower FNR, higher $$\text{TPR} = 1 - \text{FNR}$$) at the cost of more nuisance violations
for healthy firms (higher FPR). The ROC curve traces this trade-off;
better covenant "technology" shifts the ROC curve further from the
45-degree diagonal, allowing a lower FPR at any given TPR (Figure 4, p. 385).

The optimal threshold minimizes total expected costs of both error types
(equation (1), p. 382):

$$
\min_t \; (1 - \rho) \cdot \text{FPR}(t) \cdot C_{\text{FP}} \;+\; \rho \cdot \text{FNR}(t) \cdot C_{\text{FN}}
$$

where $$\rho$$ is the unconditional probability of distress and $$C_{\text{FP}}$$, $$C_{\text{FN}}$$
are the costs of false positives and false negatives, respectively. The
first-order condition identifies the optimal threshold $$t^*$$:

$$
\frac{1 - \rho}{\rho} \cdot \frac{C_{\text{FP}}}{C_{\text{FN}}} = \frac{f_D(t^*)}{f_{ND}(t^*)} \tag{1}
$$

The left side is the preference parameter $$R = \frac{1-\rho}{\rho} \cdot \frac{C_{\text{FP}}}{C_{\text{FN}}}$$,
the ratio of expected costs of false positives to false negatives. The right
side is the likelihood ratio for the relative probability of violation for a
distressed vs. nondistressed borrower at $$t^*$$. Thus:
- A higher $$R$$ (more costly false positives relative to false negatives)
  raises $$t^*$$, loosening the covenant and reducing both FPR and TPR.
- Better technology (ROC curve farther from 45 degrees) allows a lower FPR
  without reducing the TPR.
- Lower distress prevalence $$\rho$$ also raises $$R$$, loosening covenants.

One covenant-technology channel is the shift from balance-sheet to
cash-flow covenants in loan agreements documented by Demerjian (2011),
which the paper treats as a change in how well the metric separates
distressed from nondistressed borrowers.

The paper estimates parameters by assuming $$r \sim N(\mu_D, 1)$$ for distressed
firms and $$r \sim N(0, 1)$$ for nondistressed firms, so $$\mu_D$$ indexes covenant
technology (how well the metric separates the two populations), and $$R$$ is
estimated from the observed FPR and TPR via the ROC curve slope (p. 393).
As an ex ante measure of covenant tightness the paper adopts the covenant
strictness measure of Murfin (2012) together with its Demerjian-Owens
update.

## Method

The paper builds on `probit-regression`, `roc-curve-analysis`, and
`blinder-oaxaca-decomposition` to estimate and decompose structural
parameters from realized violation rates.

**Step 1: Structural parameter estimation (Section IV, pp. 392-393).**
Under the normality assumption, the optimal threshold satisfies
$$\text{FPR} = 1 - \Phi(t^*)$$, so:

$$
t^* = \Phi^{-1}(1 - \text{FPR}_t)
$$

where $$\Phi(\cdot)$$ is the standard normal CDF. Given $$t^*$$, the technology
parameter $$\mu_D$$ is recovered from the TPR:

$$
\text{TPR} = 1 - \Phi(t^* - \mu_D) \implies \mu_D = t^* - \Phi^{-1}(1 - \text{TPR})
$$

The preference parameter $$R$$ is then recovered from the likelihood ratio
condition of equation (1):

$$
R = e^{(t^* - \mu_D / 2) \cdot \mu_D}
$$

Standard errors use the delta method, since the parameters are functions
of sample proportions with known sampling variances (Table III, p. 393).

**Step 2: Violation-rate decomposition (Section III, p. 388).** The annual
violation rate is decomposed using the identity (equation (2), p. 388):

$$
V_t = \rho_t \cdot \text{TPR}_t \;+\; (1 - \rho_t) \cdot \text{FPR}_t \tag{2}
$$

- $$\rho_t = (\text{FN}_t + \text{TP}_t) / N_t$$ is the realized distress rate.
- $$\text{TPR}_t = \text{TP}_t / (\text{TP}_t + \text{FN}_t)$$.
- $$\text{FPR}_t = \text{FP}_t / (\text{FP}_t + \text{TN}_t)$$.

The change in violation rate from period $$s$$ to period $$t$$ decomposes as:

$$
\Delta V_{s,t} = \underbrace{\rho_s \cdot (\Delta\text{TPR})}_{\text{"TPR"}} \;+\; \underbrace{(1 - \rho_s) \cdot (\Delta\text{FPR})}_{\text{"FPR"}} \;+\; \underbrace{(\text{TPR}_t - \text{FPR}_t) \cdot (\Delta\rho)}_{\text{"Distress"}}
$$

**Step 3: Blinder-Oaxaca decomposition (Section V, pp. 399-402).** To
separate the portion of the FPR and TPR trends explained by observable
characteristics from the unexplained portion, the paper employs the
decomposition of Blinder (1973) and Oaxaca (1973). It
estimates two probit regressions: one for the FPR (nondistressed firms
only) and one for the TPR (distressed firms only). For each, the marginal
effects of borrower size, credit rating, lender type, and covenant type
are estimated. The Blinder-Oaxaca decomposition of the change in violation
rates across periods takes the form (equation (3), p. 402):

$$
\bar{V}^t - \bar{V}^s = \underbrace{\bigl[\Phi(\bar{X}^t \hat{\beta}) - \Phi(\bar{X}^s \hat{\beta})\bigr]}_{\text{"Explained"}} \;+\; \underbrace{U}_{\text{"Unexplained"}} \tag{3}
$$

- $$\Phi$$ is the normal CDF.
- $$\hat{\beta}$$ is the estimated probit coefficient vector from Table IV.
- $$X$$ is the set of explanatory variables (borrower size, rating, lender type, loan type, covenant type).
- $$U$$ captures changes in the mapping from $$X$$ to violation status beyond observable composition shifts. The Yun (2004) method attributes the explained portion to individual characteristics.

## Empirical specifications

All main specifications use annual firm-year observations from the
Compustat-EDGAR panel (85,876 firm-years) merged with the Dealscan loan
sample (17,724 packages) for the FPR/TPR analysis. The Dealscan-to-Compustat
merge uses the link file of Chava and Roberts (2008), whose imputed-violation
measure the paper also uses as a robustness check.

**Creditor recovery rate regression (R4, Table I, p. 387).**

$$
\text{RecoveryRate}_i = \alpha + \beta \cdot \text{NoViolation}_i + \gamma' X_i + \epsilon_i
$$

- $$\text{RecoveryRate}_i$$: firm-level par value-weighted average recovery rate (Moody's URD).
- $$\text{NoViolation}_i$$: indicator equal to 1 if the firm did not report a covenant violation in the year before bankruptcy.
- $$X_i$$: controls including operating cash flow/assets, debt/assets, interest expense/assets, net worth/assets, current ratio, market-to-book, cash/assets, and (in column 4) bank debt share.
- Specification: OLS, year and industry fixed effects (columns 2-4), standard errors in parentheses.
- Sample: 403 corporate bankruptcies (Moody's URD, 1997-2020). Winsorized at 1/99%.

**Probit regressions for FPR and TPR determinants (R8, Table IV, p. 401).**

$$
V_i = \Phi\!\left(\alpha + \beta_1 \cdot \text{BorrowerSize}_i + \beta_2 \cdot \text{SpeculativeRating}_i + \beta_3 \cdot \text{InvestmentRating}_i + \beta_4 \cdot \text{UniversalBank}_i + \beta_5 \cdot \text{InstitutionalLoan}_i + \beta_6 \cdot \text{BalanceSheetCovenant}_i + \gamma' \cdot \text{IndustryFE}_i + \epsilon_i\right)
$$

- $$V_i$$: covenant violation indicator (1 = violation).
- Two separate regressions: (i) nondistressed firms only (FPR regression, N = 29,719 full sample; period sub-samples 5,792 / 6,715 / 7,500) and (ii) distressed firms only (TPR regression, N = 1,102 full sample; period sub-samples 968 / ... / 763).
- Specification: probit, Fama-French 12-industry fixed effects, standard errors via delta-method for marginal effects. Coefficients reported as estimated marginal effects.
- $$\text{BorrowerSize}$$: log(total assets, real 2000 dollars). Ratings and lender/covenant-type indicators are as described in Section V.

**Blinder-Oaxaca decomposition (R8, Table V, p. 402).** Based on probit
coefficients from Table IV and period-specific sample means. The explained
portion equals the predicted change due to shifts in the distribution of
$$X$$ using a constant coefficient vector; the unexplained portion is the
residual. Standard errors via delta method.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Compustat (WRDS) | Firm financials, firm-year sample construction (85,876 firm-years, 9,618 firms, 1997-2019) | [WRDS](/wiki/commercial/wrds/) (licensed) |
| SEC EDGAR (10-K and 10-Q filings) | Hand-collected covenant violation disclosures; loan amendment exhibits | [EDGAR](/wiki/datasets/edgar/) |
| Dealscan (WRDS) | Loan package and covenant data; covenant strictness measure; lender characteristics (17,724 packages, 5,258 firms) | [DealScan](/wiki/commercial/dealscan/) (licensed) |
| CRSP (WRDS) | Stock price and shares outstanding for sample filters and controls | [WRDS](/wiki/commercial/wrds/) (licensed) |
| Moody's Ultimate Recovery Database (URD) | Creditor recovery rates for 403 corporate defaults, 1997-2020 | [Moody's URD](/wiki/commercial/moodys-urd/) (licensed) |
| Audit Analytics | Bankruptcy filings cross-check for false-negative identification | no page yet |
| UCLA-LoPucki Bankruptcy Research Database | Bankruptcy filing dates for false-negative classification | [LoPucki BRD](/wiki/datasets/lopucki-brd/) |

Sample: 85,876 firm-year observations (Compustat-EDGAR); 17,724 loan packages
(Dealscan); 403 bankruptcies with recovery data (Moody's URD).

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.70005) if you are:
replicating the violation-coding procedure (the text-search and manual
inspection process from SEC filings); extending the structural
covenant-design model; analyzing the Blinder-Oaxaca decomposition in
detail; or checking the Internet Appendix robustness results (alternative
distress measures, ratio-manipulation tests, loan amendment trends).
The locators above point to the exact table. For "what did this paper
find," the table above is sufficient.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 81(1), February 2026,
pp. 371-412. (c) 2025 the American Finance Association. This
distillation was extracted by an LLM on 2026-05-31 and is **not
human-verified or independently reproduced**. The paper is paywalled;
only the extract (core results, datasets, theory) is reproduced here
under fair-use principles.

> Griffin, Thomas P., Greg Nini, and David C. Smith. "Losing Control?
> The Two-Decade Decline in Loan Covenant Violations." *The Journal of
> Finance* 81, no. 1 (February 2026): 371-412.
> DOI: [10.1111/jofi.70005](https://doi.org/10.1111/jofi.70005).
> Extract-only; no PDF hosted here.


==============================================================================
# Investment under Upstream and Downstream Uncertainty: Grigoris & Segal (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2026/grigoris-investment-upstream-downstream-uncertainty-2026/
# Distilled: upstream (supplier-level) uncertainty reduces firm investment, hiring, and working capital while downstream (customer-level) uncertainty has a weaker and often positive effect; the asymmetry is amplified for long time-to-build firms and scales to the macro level. J. Finance 2026, paywalled. Nine core results with source locators, datasets used, and the theory tested.
# Tags: paper-summary, corporate-investment, uncertainty, supply-chains, real-options, production-networks, macro, panel-regression, peer-reviewed, unreplicated, data:wrds, data:factset-revere, data:bea-io, data:compustat-segments, data:nber-ces, data:fred
==============================================================================

**What this is.** The paper's core results, datasets, and theory: enough
to know what it found without reading all 45 pages. To replicate or extend
it, use the [original](https://doi.org/10.1111/jofi.70010) (paywalled) or
the replication code available in the journal's Supporting Information.

## TL;DR

Using granular supplier-customer link data from Compustat Segments and
FactSet Revere (1976-2019), the paper measures each firm's upstream
(supplier-level) and downstream (customer-level) uncertainty as the
realized stock return volatility of its trading partners. Upstream
uncertainty robustly suppresses investment, hiring, and working capital.
Downstream uncertainty has a weaker and often positive effect, flipping
sign for firms with long time-to-build periods. A production-based
real-option model with time-to-build generates this asymmetry: upstream
uncertainty raises the option value of waiting via the bad news principle,
while downstream uncertainty raises the opportunity cost of waiting via the
good news principle (convex future cash flows). The asymmetry scales to the
macro level: macrolevel upstream (downstream) uncertainty negatively
(positively) predicts GDP growth, consumption, investment, and
price-dividend ratios.

## Core results

Magnitudes and significance are as reported; `**`/`***` = 5%/1%. All
firm-level independent variables are scaled by their unconditional
standard deviation. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Upstream uncertainty negatively predicts firm investment** | Table III col. (4), p. 438 | 1-SD increase: investment rate -0.03 (t = -2.80)\*\*\*; firm + year FE, controls |
| R2 | **Downstream uncertainty positively predicts firm investment** | Table III col. (6), p. 438 | 1-SD increase: investment rate +0.03 (t = 3.15)\*\*\*; firm + year FE, controls |
| R3 | **Asymmetry extends to working capital, employment, COGS, and intangibles** | Table IV, p. 439 | Upstream: significant negative in all four outcomes; downstream: positive (working capital t=2.18-3.32) or insignificant (employment, intangibles); upstream effect weakly larger in absolute value |
| R4 | **Downstream uncertainty effect on investment is stronger for long time-to-build firms** | Table V, p. 442 | Long vs. short interaction: coef 0.03-0.06 (t=3.15-4.60)\*\*\* across three proxies (depreciation, sector, R&D); Wald test rejects equality at 10% in all specs |
| R5 | **Upstream uncertainty effect on investment is amplified for low-reversibility (hard-to-abandon) firms** | Table VI col. (2), p. 443 | LowReverse: -0.05 (t=-3.31)\*\*\*; HighReverse: -0.03 (t=-1.42, insignificant) |
| R6 | **Downstream uncertainty effect on investment is stronger for high-reversibility firms** | Table VI col. (4), p. 443 | HighReverse: +0.05 (t=3.60)\*\*\*; LowReverse: +0.02 (t=1.32, insignificant) |
| R7 | **Macrolevel upstream uncertainty shock leads to economic contraction** | Figure 6, p. 448 | 1-SD shock: industrial production and GDP fall ~0.15 SD, consumption and investment fall ~0.10 SD; P/D ratio falls ~0.10-0.15 SD; effects persist ~4 quarters (90% CI excludes zero) |
| R8 | **Macrolevel downstream uncertainty shock leads to economic expansion** | Figure 7, p. 449 | 1-SD shock: industrial production, consumption, investment, and GDP rise ~0.10 SD for at least 4 quarters; P/D ratio rises ~0.10 SD for ~12 quarters; upstream impacts up to 50% larger in absolute magnitude |
| R9 | **COVID-19 onset was driven by downstream uncertainty spike, consistent with fast recovery** | Figure 9 / §IV.C, pp. 451-452 | Orthogonal downstream uncertainty spiked in March 2020 (while upstream uncertainty also rose); downstream dominance consistent with the recession being short-lived relative to upstream-driven recessions |

**Overall (paper's conclusion).** Uncertainty is not uniformly contractionary:
downstream uncertainty may have an expansionary impact. The asymmetry
arises from the time-to-build mechanism and the real-option structure of
investment, not from the magnitude of uncertainty.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CRSP / Compustat (via WRDS), 1976-2019 | Investment rates, firm characteristics (size, leverage, tangibility, Tobin's q, profitability, past returns), stock return volatility | [WRDS / CRSP / Compustat](/wiki/commercial/wrds/) (licensed) |
| Compustat Segments database, 1976-2002 | Supplier-customer links for early subsample (pre-FactSet) | no page yet |
| FactSet Revere Relationship database, 2003-2019 | Supplier-customer links (primary source for post-2003 period; more comprehensive than Segments) | no page yet |
| NBER-CES Manufacturing Industry database | Validates link between input price uncertainty and supplier return volatility; upstream and downstream price correlation check | [NBER-CES](/wiki/datasets/nber-ces/) |
| BEA Input-Output (Make and Use) tables, 1977-2012 | Industry upstreamness scores for macrolevel analysis; published every 5 years | no page yet |
| FRED (VIX, industrial production index) | VIX used in COVID-19 episode illustration; IP index as macro outcome variable | [FRED](/wiki/datasets/fred/) |

Sample: firm-year panel 1976-2019; ~17,000-50,786 observations depending on
specification (Table III). Macrolevel analysis: 1976Q1-2019Q4.

## Theory / model

The paper builds a production-based real-option model with time-to-build
(Section II, pp. 421-434). It extends the time-to-build real-option model of
Majd and Pindyck (1987) to stochastic volatility and to the supply-chain
location of uncertainty. The focal firm has assets-in-place (installed
capacity $$k_t$$, depreciating at rate $$\delta$$) and a growth option to expand.
Operating cash flow per period (eq. 1, p. 421):

$$
\pi_t = P_t^{\text{Out}} \cdot k_t^{\alpha} - \omega \cdot k_t - P_t^{\text{In}} \cdot \delta \cdot k_t
$$

- $$P_t^{\text{Out}}$$ is the stochastic output price
- $$P_t^{\text{In}}$$ is the stochastic input price
- $$\alpha \in (0,1)$$ is returns to scale
- $$\omega$$ is a proportional operating cost

The last term captures maintenance (replacing
depreciated inputs purchased at the current input price).

Both input and output log-prices follow mean-reverting stochastic volatility
processes (eqs. 2-3, p. 422). For $$j \in \{\text{In}, \text{Out}\}$$:

$$
\begin{aligned}
p^j_{t+1} &= \rho_p \, p^j_t + \sigma_p \exp\!\left(\sigma^j_t / 2\right) \epsilon^j_{t+1} \\
\sigma^j_{t+1} &= \rho_\sigma \, \sigma^j_t + \sigma_w \, \eta^j_{t+1}
\end{aligned}
$$

- $$p^j_t = \log(P^j_t)$$; innovations $$\epsilon$$ and $$\eta$$ are i.i.d. standard normal
- $$\rho_p$$ governs price persistence
- $$\rho_\sigma$$ governs volatility persistence
- $$\sigma_w$$ governs the volatility of volatility

The firm's recursive Bellman equation (eq. 4, p. 422), choosing future
capacity $$k'$$ to maximize cum-dividend value $$V(k, \Gamma)$$, where
$$\Gamma = [p^{\text{In}}, \sigma^{\text{In}}, p^{\text{Out}}, \sigma^{\text{Out}}]$$:

$$
V(k, \Gamma) = \max_{k'} \left\{ \pi(k, \Gamma) + \Phi(k, k') +
  \max\!\begin{cases}
    P^{\text{In}}(k - k') & \text{if } k' \leq k \text{ (Contraction)} \\
    -f \cdot k - P^{\text{In}} \cdot w_1(k' - k) + \beta \, \mathbb{E}[V^{\text{Build}}(k, k', \Gamma, H-1)] & \text{if } k' > k \text{ (Expansion)}
  \end{cases}
\right\}
$$

- $$f$$ is the fixed cost of expansion
- $$w_1$$ is the fraction of excess capacity purchased in period 1 of time-to-build
- $$V^{\text{Build}}$$ (eq. 5, p. 423) is the firm's continuation value during the build-up stage

The price of the focal firm's input equals the output price of its supplier
$$s$$, and its output price equals the input price of its customer $$c$$
(eq. 8, p. 424):

$$
P_t^{\text{In}} = P_t^{s,\text{Out}} \qquad \text{and} \qquad P_t^{\text{Out}} = P_t^{c,\text{In}}
$$

This links the focal firm's input and output price uncertainty to its
trading partners' fundamentals. The observable proxy for each uncertainty
type is the realized stock return volatility of the supplier (customer)
over a rolling window (eq. 9, p. 424):

$$
\begin{aligned}
\sigma_t^{\text{Upstream}}   &= \operatorname{Std}(R^s_{t-W}, \ldots, R^s_t) \\
\sigma_t^{\text{Downstream}} &= \operatorname{Std}(R^c_{t-W}, \ldots, R^c_t)
\end{aligned}
$$

**Key asymmetry (pp. 428-431).** The paper builds on the canonical
bad-news-principle channel of Bloom (2009) by decomposing total uncertainty
into upstream and downstream components. Both uncertainties increase the option
value of waiting (bad news principle, Bernanke 1983). Only downstream
uncertainty also raises the opportunity cost of waiting: during
time-to-build, forgone revenues are a convex function of the future output
price (the firm can disinvest if the price falls), so higher downstream
uncertainty raises the cost of delay. Upstream uncertainty is unaffected
because all input purchases are made up front. Net result: upstream
uncertainty unambiguously suppresses investment; downstream uncertainty can
hasten investment when the time-to-build period is sufficiently long.

Four testable hypotheses (§II.C.4, p. 434):

1. Upstream-investment association is unambiguously negative.
2. Downstream-investment association is weaker in absolute value, can be positive.
3. Downstream effect is more positive for firms with longer time-to-build.
4. Harder-to-abandon firms show a more negative (less positive) upstream
   (downstream) effect.

## Method

**Model solution.** The model is solved numerically by value function
iteration (Section II.B, p. 425). Gaussian autoregressive processes are
discretized using a Tauchen (1986) variant that allows time-varying
conditional volatility, similar to Alfaro et al. (2024). The state space
uses a refined, endogenous grid for capital centered around the stochastic
steady state, with a dense grid near the free boundaries where the growth
option is exercised. The model is calibrated at the quarterly frequency
(Table I, p. 425); key parameters: $$\alpha = 0.40$$, $$\beta = 0.997$$,
$$\delta = 0.025$$, $$f = 0.020$$, $$\rho_p = 0.950$$, $$\sigma_p = 0.200$$.
Model-implied moments (Table II, p. 426) match $$\sigma(I/K) = 0.165$$ and
skewness $$= 0.626$$ in the data within the 95% confidence interval.

This builds on `real-options` and `value-function-iteration`; the macrolevel
evidence builds on `smooth-local-projections`.

**Uncertainty measures.** Upstream (downstream) uncertainty is the
equal-weighted average realized daily stock return volatility of
the firm's suppliers (customers), computed over the prior calendar year
using CRSP daily data. Firm-level supplier-customer networks are identified
from Compustat Segments (1976-2002) and FactSet Revere (2003-2019), merged
to maximize coverage.

## Empirical specifications

All firm-level regressions (Section III, pp. 434-443) are estimated on a
firm-year panel of CRSP/Compustat firms (NYSE, AMEX, NASDAQ, excl.
financials SIC 6000-6999 and utilities SIC 4900-4999), 1976-2019.
Standard errors are clustered at the firm level. Each independent variable
is scaled by its unconditional standard deviation.

**Baseline investment regression (eq. 10, p. 437; R1-R3):**

$$
y_{i,t} = \alpha_i + \delta_t + \beta_1 \cdot \sigma(\text{Own})_{i,t}
          + \beta_2 \cdot \sigma(\text{SupplyChain})_{i,t} + \gamma' Z_{i,t} + \epsilon_{i,t}
$$

- $$\sigma(\text{SupplyChain}) \in \{ \sigma(\text{Upstream}),\, \sigma(\text{Downstream}) \}$$
- $$y_{i,t}$$ is the investment rate (I/K) of firm $$i$$ at time $$t$$, measured from the most recent annual report as of June $$t$$
- $$\alpha_i$$ = firm fixed effects; $$\delta_t$$ = year fixed effects
- $$\sigma(\text{Own})_{i,t}$$ is the firm's own stock return volatility
- $$Z_{i,t}$$ includes firm size, leverage, tangibility, Tobin's q, profitability, and past returns (Leary and Roberts 2014)
- Sample: OLS, firm + year FE; ~17,456-50,786 observations (Table III, p. 438)

The same equation with $$y$$ replaced by working capital growth,
employment growth, COGS growth, or intangibles growth gives Table IV results
(R3).

**Time-to-build heterogeneity regression (eq. 11, p. 440; R4):**

$$
y_{i,t} = \alpha_i + \delta_t + \beta_1 \cdot \sigma(\text{Own})_{i,t}
          + \beta_2 \cdot \sigma(\text{Downstream})_{i,t} \times I[\text{Long}]_{i,t}
          + \beta_3 \cdot \sigma(\text{Downstream})_{i,t} \times I[\text{Short}]_{i,t}
          + \gamma' Z_{i,t} + \epsilon_{i,t}
$$

- $$I[\text{Long}]$$ and $$I[\text{Short}]$$ are indicator variables for long and short time-to-build firms
- Three proxies: (i) inverse depreciation rate, (ii) sector (nondurables/services = short; investment goods/durables = long, Gomes et al. 2009), (iii) R&D intensity
- The null $$H_0: \beta_2 = \beta_3$$ (Wald test) is rejected at 10% in all specifications (Table V, p. 442)

The same interaction structure is used to test reversibility heterogeneity
(Table VI, p. 443), replacing $$I[\text{Long}]$$ with $$\text{HighReverse}$$ / $$\text{LowReverse}$$
(Kim and Kung 2017 capital redeployability measure).

Where Acemoglu, Akcigit, and Kerr (2016) study the production-network
propagation of shocks, the paper tests analogous channels for second-moment
(uncertainty) shocks at the macro level.

**Macrolevel impulse responses (eq. 13, p. 447; R7-R8):** Smooth local
projections (SLPs, Barnichon and Brownlees 2019) estimated for forecast
horizons $$h \in \{1, \ldots, H\}$$ quarters:

$$
y_{t+h} = \beta_{0(h)} + \beta_{1(h)} y_t + \beta_{2(h)} \sigma_{U,t}
           + \beta_{3(h)} \sigma_{D,t} + \sum_{p=1}^{P} \gamma'_{p(h)} \Gamma_{t-p}
           + \epsilon_{t+h}
$$

- $$y_{t+h}$$ is one of: quarterly real growth rates of industrial production, consumption, private investment, GDP, and the level of market price-dividend ratio and risk-free rate
- $$\sigma_{U,t}$$ ($$\sigma_{D,t}$$) is macrolevel upstream (downstream) uncertainty, constructed as the value-weighted average realized volatility of firms classified in the top (bottom) 10th percentile of the industry upstreamness score (eq. 12), built on the upstreamness measure from BEA I-O tables of Antras and Chor (2018) to form the macrolevel upstream-downstream industry classification
- $$\Gamma_{t-p}$$ includes the dependent variable, both macrolevel uncertainties, excess market return, term spread, default spread, and inflation
- $$P = 4$$ lags; 1976Q1-2019Q4 quarterly data; all variables standardized
- SE/CIs: IRFs plotted with 90% confidence intervals (Figures 6-7, pp. 448-449)

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.70010) (institutional
access required) if you are: replicating (code in Supporting Information);
extending the supply-chain uncertainty measures or time-to-build
heterogeneity tests; auditing the IV strategy or the macrolevel SLP
estimates; or reviewing the COVID-19 application in §IV.C. The locators
above point to the exact table or figure.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 81(1), February 2026.
© 2025 the American Finance Association. This distillation was
extracted by an LLM on 2026-05-31 and is **not human-verified or
independently reproduced**. The paper is paywalled; no CC licence is
present in Crossref metadata or on the artifact. Reproduction of the
verbatim text requires a licence from the publisher.

> Grigoris, Fotis, and Gill Segal. "Investment under Upstream and Downstream
> Uncertainty." *The Journal of Finance* 81, no. 1 (February 2026): 413–457.
> DOI: 10.1111/jofi.70010. Extract-only; all rights reserved by the
> American Finance Association / Wiley.


==============================================================================
# Bank Monitoring with On-Site Inspections: Heitz, Martin & Ufier (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2026/heitz-bank-monitoring-onsite-inspections-2026/
# Distilled: Using proprietary transaction-level data on nearly 30,000 construction loans from a failed bank, this paper provides empirical evidence that banks trade off monitoring intensity with loan origination terms, use inspection report text to inform draw decisions, and that increased on-site inspections causally reduce loan default. J. Finance 2026, CC BY 4.0. Eight core results with source locators, datasets, the identification strategy, and the regression specifications.
# Tags: paper-summary, banking, bank-monitoring, credit-risk, moral-hazard
==============================================================================

**What this is.** The paper's core results, the empirical design, and the
regression specifications: enough to know what it found and how, without
reading all 51 pages. To replicate or extend, read the full source at
[https://doi.org/10.1111/jofi.70026](https://doi.org/10.1111/jofi.70026).

## TL;DR

Using proprietary transaction-level data on nearly 30,000 construction loans
from a large bank that failed during the financial crisis, Heitz, Martin, and
Ufier provide direct empirical evidence on the determinants and consequences
of bank monitoring via on-site inspections. They find that banks trade off
monitoring intensity with loan terms (more monitoring pairs with lower spreads,
higher amounts, shorter maturities), that monitoring escalates when borrower
credit quality declines or the bank approaches failure, and that inspection
report text predicts draw denials: negative language raises denial probability
while positive language lowers it. Using three independent instrumental
variables (the draw schedule, time to first inspection, and inspector fixed
effects), they establish that increased inspection frequency causally reduces
loan default by 3.8 to 6.1 percentage points per one-percentage-point increase
in monitoring frequency, approximately double the OLS estimate. The gains come
primarily through the threat of inspections inducing borrowers to complete
projects rather than through the bank catching and stopping failing projects
early.

## Core results

Magnitudes and significance are as reported; `*`/`**`/`***` = 10%/5%/1%.
Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Larger loans receive more and more frequent inspections; lower spreads and lower fees associate with more monitoring, consistent with banks trading off monitoring intensity for favorable terms | Table III, p. 709; §III.A, p. 712 | LOG(LOANAMT) coeff on ALLINSPECTIONS = 0.168\*\*\* (t=37.83); ORIGSPREAD coeff = -0.0140\*\*\* (t=-5.08); ORIGFEES coeff = 0.00190\*\*\* (t=9.86) |
| R2 | Shorter-maturity loans are monitored more frequently per 100 days; longer-maturity loans have more inspections in total | Table III cols (1)-(2), p. 709 | TERM: ALLINSPECTIONS = +0.0172\*\*\* (t=39.51); ALLTOTERMINAL = -0.0417\*\*\* (t=-4.06) |
| R3 | Riskier borrowers (lower FICO, higher CLTV, speculative loans, owner-builders) receive significantly more intensive monitoring | Table III, pp. 709-710 | FICO: ALLTOTERMINAL = -0.00326\*\*\* (t=-18.35); SPECULATING = 0.344\*\*\* (t not reported for col 2); OWNERBUILDER: ALLINSPECTIONS = +0.149\*\*\* (t=29.85) |
| R4 | Draw denials increase and inspection probability increases as the bank approaches failure, inconsistent with gambling-for-resurrection | Table V, p. 713 | YEARBEFOREFAILURE: DRAWDENIED = +0.0645\*\*\* (t=13.60); INSPECTIONDATE = +0.00291\*\*\* (t=6.96) |
| R5 | Negative language in inspection reports increases draw denial probability; positive language decreases it; results hold with inspector, loan, and day fixed effects | Table VI, p. 714 | NEGATIVEWORDS: DRAWDENIED = +0.00164\*\* (t=2.21) to +0.00372\*\*\* (t=4.54); POSITIVEWORDS: DRAWDENIED = -0.00207\* (t=-1.71) to -0.00234\*\*\* (t=-2.27) |
| R6 | OLS: more frequent monitoring (ALLTOTERMINAL) associates with lower eventual default; IV estimates are approximately twice as large as OLS | Table VIII Panel A col (1) vs col (3), pp. 727-728 | IV coeff = -0.0380\*\*\* (t=-23.70); OLS coeff = -0.0195\*\*\* (t=-19.46); IV: 1 s.d. increase in ALLTOTERMINAL -> -5.85 pp default probability |
| R7 | All three IV specifications (draw schedule, time to first inspection, inspector fixed effects) produce consistent causal estimates of monitoring on default | Table VIII Panels A-C, pp. 727-729 | IV 2nd-stage coeffs: -0.0380 (draw schedule), -0.0396 (TIMETOFIRST), -0.0261 (inspector FE); all significant at 1%, Cragg-Donald F > 29.8 across panels |
| R8 | Decomposition shows monitoring reduces defaults primarily through the threat channel (63-77% of total benefit); the direct early-stopping effect is smaller but also present | §V.A, pp. 723-724 | 1 extra inspection per 100 days: -1.14 pp maturity default, -0.81 pp term default net (total -1.95 pp); threat share = 63-77% of total benefit |

**Overall (paper's conclusion).** Bank on-site inspections provide direct
support for three theoretical monitoring channels: adverse selection screening
(riskier loans receive more monitoring and better terms, deterring high-risk
applicants), moral hazard mitigation (monitoring intensifies when collateral
values fall and economic conditions worsen), and early intervention (monitoring
causally reduces default, primarily through the credible threat of stopping
failing projects). Both the occurrence and the threat of inspections matter.

## Theory / model

The paper has no formal structural model. The tested hypotheses derive from a
set of theoretical predictions in the bank monitoring literature, and the
results provide empirical support for the delegated monitoring theory of
Diamond (1984), under which banks reduce default through monitoring. The
analysis extends the syndicated-loan monitoring study of Gustafson, Ivanov, and
Meisenzahl (2021) to single-lender construction loans, adding draw-denial
outcomes and a causal IV framework:

**H1 (Adverse selection / screening, Diamond 1991, Rajan 1992).** Banks that
commit to monitoring can offer better terms to borrowers they screen, because
monitoring reduces the residual risk borne by the lender. Testable implication:
monitoring intensity trades off negatively with loan spreads and fees and
positively with loan amounts.

**H2 (Moral hazard, Calomiris and Kahn 1991, Rajan and Winton 1995).** Moral
hazard is greater when project returns are lower. As collateral value declines
or foreclosure risk rises, banks intensify monitoring to discipline borrowers.
Building on the prior direct evidence in Cerqueiro, Ongena, and Roszbach (2016)
that collateral value affects monitoring frequency, this paper extends the link
to loan outcomes and causal identification. Testable implication: monitoring
and draw denials increase when local housing price growth is negative or local
foreclosure rates rise.

**H3 (Early intervention, Diamond and Rajan 2001, Acharya, Hasan and Saunders
2006).** Monitoring allows banks to detect default risk early and cut credit
extension to failing projects. Testable implication: monitoring reduces loan
default; the effect operates both through directly stopping failing loans
(maturity-to-term default conversion) and through the threat of stopping them
(discouraging borrowers from deviating).

**Identification challenge.** Banks endogenously assign more inspections to
higher-risk loans, so OLS confounds the treatment with selection. The IV
strategy is described in the Method section. The paper also exploits the
within-loan time-series structure to include loan fixed effects for the
draw-denial and inspection-date analyses, absorbing all time-invariant
loan-level heterogeneity.

## Method

The paper applies three standard empirical methods. No novel method is
proposed.

**Cross-sectional monitoring determinants (equation 1, p. 707).** OLS and
Poisson regressions of monitoring measures on loan origination characteristics:

$$\text{Monitoring}_l = \gamma \cdot \text{Origination}_l + \beta \cdot X_l + \epsilon_l$$

- $$\text{Monitoring}_l$$ is ALLINSPECTIONS, ALLTOTERMINAL, or TIMETOFIRST.
- $$\text{Origination}_l$$ are loan terms (log amount, spread, fees, CLTV, TERM).
- $$X_l$$ are fixed effects: property zip (3-digit), borrower zip, loan origination day. Poisson regression is used for count variables (ALLINSPECTIONS, columns 1 and 4); OLS elsewhere.

**Panel regressions for draw denials and inspections (equations 2-3, pp. 708,
712).** Daily loan-day or draw-request panel regressions:

$$
\text{DrawDenied}_{dtl} = \gamma \cdot \text{Covariate}_{dtl} + \zeta \cdot Z_l + \epsilon_{dtl} \tag{2}
$$

$$
\text{InspectionDate}_{lt} = \gamma \cdot \text{Covariate}_{lt} + \zeta \cdot Z_l + \epsilon_{lt} \tag{3}
$$

- $$Z_l$$ are loan fixed effects (absorbing all time-invariant loan characteristics).
- Covariates include HOUSING PRICE INDEX and FORECLOSURE RATE (time-varying, at the five-digit zip code level) and YEARBEFOREFAILURE / STARTOFYEARBEFOREFAILURE (bank distress).
- Standard errors clustered at the three-digit property zip code level.

**Instrumental variable estimation (Table VIII).** Three separate 2SLS
specifications for the causal effect of monitoring on eventual default:

$$\text{First stage:} \quad \text{ALLTOTERMINAL}_l = \pi \cdot \text{Instrument}_l + \delta \cdot \text{Controls}_l + \nu_l$$

$$\text{Second stage:} \quad \text{EVENTUALDEFAULT}_l = \alpha \cdot \widehat{\text{ALLTOTERMINAL}}_l + \beta \cdot \text{Controls}_l + \epsilon_l$$

Three instruments are used in three independent panels:
- Panel A: DRAWTOTERMINAL (the draw schedule frequency, set at origination).
- Panel B: TIMETOFIRST (time to first inspection, proxy for project complexity).
- Panel C: Inspector indicator variables (inspector fixed effects, akin to
  judge-IV design from Kling (2006) and Frandsen, Lefgren, and Leslie (2023)).

Controls throughout include log loan amount, origination fee, FICO, CLTV, loan
term, a speculating indicator, owner-builder indicator, and fixed effects for
property zip, borrower zip, and loan origination day. Standard errors clustered
by three-digit zip code and loan origination day.

**Text sentiment scoring (Loughran-McDonald, §IV, pp. 719-720).** From each
on-site inspection report, the paper calculates:

$$\text{POSITIVEWORDS} = \frac{\text{count of positive words}}{\text{comment length in chars}} \times 100$$

$$\text{NEGATIVEWORDS} = \frac{\text{count of negative words}}{\text{comment length in chars}} \times 100$$

using the Loughran-McDonald (2011) financial sentiment dictionary. These
sentiment scores are then entered as covariates in a draw-denial panel
regression (equation 2 format) with inspector, loan, day, and zip fixed
effects.

## Empirical specifications

**Determinants of monitoring (Table III, p. 709).** Cross-sectional
Poisson (cols 1, 4) and OLS (cols 2, 3) regressions. LHS:
ALLINSPECTIONS (total lifetime inspections), ALLTOTERMINAL (inspections per
100 active days), TIMETOFIRST (months to first inspection). RHS: LOG(LOANAMT),
ORIGSPREAD, TERM, ORIGFEES, CLTV, FICO, SPECULATING, OWNERBUILDER, BUDGETITEM.
FEs: property zip (3-digit), borrower zip, loan origination day. N = 28,939
(27,567 for TIMETOFIRST). Sample: full construction loan portfolio.

**Bank actions and macroeconomy (Table IV, p. 711).** Panel OLS at the draw
request (DRAWDENIED) or loan-day (INSPECTIONDATE) level. LHS: DRAWDENIED or
INSPECTIONDATE. RHS: HOUSING PRICE INDEX (annualized price change, 5-digit zip,
from FHFA), FORECLOSURE RATE (monthly foreclosure rate, 5-digit zip, from
CoreLogic). FEs: loan-level. SE: clustered at three-digit zip code.
N (DRAWDENIED) = 330,579 draw requests; N (INSPECTIONDATE) = 10,806,815
loan-days.

**Bank actions approaching failure (Table V, p. 713).** Same panel framework
as Table IV with loan fixed effects. LHS: DRAWDENIED or INSPECTIONDATE. RHS:
YEARBEFOREFAILURE (indicator, last 365 days before bank failure) and
STARTOFYEARBEFOREFAILURE (indicator for start of that calendar year).
SE: clustered at three-digit property zip.

**Draw decisions based on inspection comments (Table VI, p. 714).** Panel OLS.
LHS: DRAWDENIED. RHS: POSITIVEWORDS, NEGATIVEWORDS, COMMENTLENGTH. FEs vary
across columns: Day, Loan, Inspector, Property Zip, Borrower Zip, Loan
Origination Day. N = 143,074 (matched inspection reports to draw requests).
SE: clustered by three-digit property zip code and day.

**Determinants of default (Table VII, pp. 722-723).** Cross-sectional OLS
(equation 4, p. 721). LHS: EVENTUALDEFAULT (col 1), MATURITYDEFAULT (col 2),
TERMDEFAULT (col 3), and variants adding ALLTOTERMINAL and DRAWTOTERMINAL
(cols 4-8). RHS: ALLTOTERMINAL, LOG(LOANAMT), ORIGSPREAD, TERM, FEES, CLTV,
FICO, SPECULATING, OWNERBUILDER, BUDGETITEM. FEs: property zip, borrower zip,
loan origination day. SE: clustered by three-digit zip code and day.
N = 28,939.

**IV estimation (Table VIII, pp. 727-729).** Three panels, each with 2SLS
(col 1), first-stage OLS (col 2), OLS (col 3), and reduced form (col 4).
LHS (second stage): EVENTUALDEFAULT. Endogenous: ALLTOTERMINAL. Instruments:
Panel A: DRAWTOTERMINAL. Panel B: TIMETOFIRST. Panel C: inspector indicators.
Controls: log loan amount, origination fees, FICO, CLTV, TERM, SPECULATING,
OWNERBUILDER, BUDGETITEM. FEs: property zip, borrower zip, loan origination
day. SE: clustered by zip and day. Weak instrument tests: Cragg-Donald F
exceeds 29 in all panels (Panel A: 13,839; Panel B: 1,558; Panel C: 29.8).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| FDIC proprietary construction loan servicing data (single failed bank) | Primary: loan terms, draw requests, inspection reports, inspection dates, borrower identifiers, default outcomes; ~11.6 M loan-day obs, 28,939 loans, ~10 years | No page yet |
| FHFA Housing Price Index (five-digit zip, monthly) | Time-varying collateral value proxy for moral hazard analysis (Tables IV, V) | [FHFA HPI](/wiki/datasets/fhfa-hpi/) |
| CoreLogic foreclosure rate data (five-digit zip, monthly) | Foreclosure rate as measure of local economic stress (Tables IV, V) | [CoreLogic](/wiki/commercial/corelogic/) (licensed) |

Sample: approximately 10 years of transaction-level data from a single large
bank (over $1 billion in assets) that failed during the financial crisis.
Construction loans primarily for residential single-family properties across
the continental US. Daily frequency; 11,586,385 loan-day observations; 28,939
loans; 143,074 matched inspection-to-draw-request observations for text
analysis.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.70026) if you need: the
full set of robustness checks (Internet Appendix, including exclusion of
non-single-family loans); the univariate correlations for the IV first stages;
the detailed construction of the inspector strictness measure via leave-one-out;
or the back-of-the-envelope optimality calculation on monitoring costs versus
benefits (pp. 732-733). The paper is also the primary source for institutional
background on construction loan draw schedules and the two-part default
taxonomy (maturity versus term default) that is specific to this lending market.
The locators above point to the exact tables and figures.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 81(2). This distillation
was extracted by an LLM on 2026-06-01 and is **not human-verified or
independently reproduced**. The CC BY 4.0 licence permits mirroring; the
verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Heitz, Amanda Rae, Christopher Martin, and Alexander Ufier.
> "Bank Monitoring with On-Site Inspections."
> *The Journal of Finance* 81, no. 2 (April 2026): 687-737.
> DOI: 10.1111/jofi.70026. (c) 2026 The Author(s).
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Monetary Policy, Inflation, and Crises: Jimenez, Kuvshinov, Peydro & Richter (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2026/jimenez-monetary-policy-inflation-crises-2026/
# Distilled: A U-shaped monetary policy rate path (prolonged cuts followed by hikes) substantially raises banking crisis risk across 17 countries from 1870 to 2020, via credit and asset price boom-bust cycles, with loan-level evidence from Spain confirming the credit supply channel. J. Finance 2026, CC BY 4.0. Ten core results with source locators, datasets used, the empirical specifications, and identification strategy.
# Tags: paper-summary, macro, banking, monetary-policy, financial-crises
==============================================================================

**What this is.** The paper's core results, the identification strategy (trilemma IV),
and the key empirical specifications with equations: enough to know what it found and how,
without reading all 48 pages. To replicate or extend, read the full source at the
[original](https://doi.org/10.1111/jofi.70023).

## TL;DR

The paper shows that what matters for banking crisis risk is not the level of
monetary policy rates, but the full path: a U-shaped path of prolonged cuts
followed by hikes is associated with roughly double the unconditional crisis
probability. Using long-run data for 17 advanced economies back to 1870 and
Spanish loan-level administrative data (1995 to 2008), the paper finds that
prolonged rate cuts fuel credit supply expansions and asset price booms (the
financial red zone), and that subsequent rate hikes crystallize these
vulnerabilities into crises, primarily through realized credit risk rather
than interest rate risk. Neither cuts alone nor hikes alone are strongly
linked to crises; it is the combination that matters.

## Core results

Magnitudes and significance are as reported; `*`/`**`/`***` = 10%/5%/1%.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | U-shaped monetary rate path is more than twice as frequent before banking crises as unconditionally; 100% of deep post-WWII crises preceded by U shape | Table I, p. 935 | Crisis conditional frequency: U shape 55% (all crises) vs 27% unconditional; 100% for post-WWII deep crises |
| R2 | U-shaped rate path is associated with 18% three-year crisis frequency, roughly double the 10% unconditional probability; deep and post-WWII crises show even larger gaps | Table II, p. 936 | U shape: 18%`***` crisis frequency vs 6-9% for other rate paths; deep crisis: 12%`***` vs 1-4% |
| R3 | OLS regression: the interaction of rate hikes with previous cuts (U-shape) significantly raises crisis probability; 1 ppt hike after cuts raises three-year crisis probability by 9 ppts (sum of coefficients) | Table III col. (2), p. 939 | $$\Delta_3\text{Rate} \times \text{Cut}$$ interaction = 0.03`**` (s.e. 0.01); sum of first three coefficients approx 0.09 |
| R4 | IV result (trilemma instrument): 1 ppt rate increase over three years, after rates were cut for five years, raises three-year crisis probability by approximately 10 to 12 ppts | Table III col. (4), p. 939-940 | $$\Delta_3\text{Rate} \times \text{Cut}$$ (IV) = 0.07`**` (s.e. 0.03); sum approx 0.10-0.12; Kleibergen-Paap weak ID = 27.48 |
| R5 | U-shaped rates are not associated with nonfinancial recession risk; for recessions the interaction term is small and insignificant | Table IV, p. 942 | $$\Delta_3\text{Rate} \times \text{Cut}$$ in recession regression = 0.02 (s.e. 0.01), insignificant; rate level alone raises recession risk |
| R6 | A residual (above-and-beyond-systematic) U-shaped monetary path raises the three-year crisis frequency to 26%; combining the residual U shape with the financial red zone raises it to 45% (all crises), versus 36% for any U-shaped path with a red zone and 22% for a systematic U shape with a red zone | Tables V and VIII, pp. 943, 951 | Strong residual U crisis frequency: 26% (Table V, all crises). Table VIII: any U-shape + red zone = 36% (Panel A, 18/50); residual U-shape + red zone = 45% (Panel B, 14/30 all crises; 48% post-WWII, 11/22); systematic U-shape + red zone = 22% (3/15) |
| R7 | Red zones (high credit and asset price growth) are strongly associated with future crises only if preceded by a U-shaped monetary rate path; monetary rate hikes while in the red zone raise crisis risk (R-zone x rate hike interaction = 0.18`***` to 0.38`***`) | Tables VII and IX, pp. 950, 953 | $$\text{R-zone} \times I(\Delta_3\text{Rate} \geq 0)$$ = 0.18`***` (OLS, s.e. 0.05), 0.38`***` (IV, s.e. 0.15); R-zones pre-raised: interaction 0.22`***` (OLS, s.e. 0.08) |
| R8 | Spain loan level: monetary rate cuts increase credit growth, especially from weaker banks to riskier firms; 1 ppt cut raises credit growth by 4.8 ppt at the bank-firm level, rising 2.8 ppt further per interquartile increase in bank NPL ratio | Table XI Panel A, p. 962 | $$\text{Cut}$$ = 4.80`**` (col. 2); $$\text{Cut} \times \text{Bank NPL ratio}$$ = 2.62`**` (col. 3); triple interaction with real estate firms: up to 7.9 ppt additional (col. 6) |
| R9 | Spain loan level: monetary rate cuts reduce firm cost of debt by 20 bps on average, with larger reductions for firms borrowing from weaker (high-NPL) banks, consistent with credit supply and mispricing | Table XI Panel B, p. 962 | $$\text{Cut}$$ = -0.20`***` (col. 1); $$\text{Cut} \times \text{Bank NPL ratio}$$ = -0.13`***` to -0.32`***` (cols. 2-5) |
| R10 | Spain loan level: U-shaped monetary path raises loan default probability; 1 ppt rate hike after cuts raises three-year delinquency probability by 11.2% relative, with effects stronger for loans by weaker banks to real estate firms | Table XII, pp. 964-965 | $$\Delta_3\text{Rate}$$ (col. 3) = 0.002`***`; $$\text{Cut}$$ = 0.011`***`; $$\Delta_3\text{Rate} \times \text{Cut}$$ (col. 3) = 0.005`***`; quadruple interaction ($$\Delta_3\text{Rate} \times \text{Cut} \times \text{Bank NPL} \times \text{Real estate}$$, col. 6) = 0.005`***` |

**Overall (paper's conclusion).** The dynamic path of monetary policy rates is
crucial for financial stability. Prolonged rate cuts fuel credit and asset price
booms through credit supply (including bank risk-taking and mispricing), and
subsequent rate hikes crystallize these vulnerabilities into banking crises
through realized credit risk. Neither the red zone alone nor U-shaped monetary
rates alone are sufficient to produce a high crisis probability; their
combination is what generates the largest crisis risks historically (p. 965-966).
This differs from Grimm et al. (2023) on mechanism: where they emphasize loose
policy, this paper uses nominal rates to show the full U-shaped path (not just the
easing leg) matters, and adds administrative loan-level evidence.

## Theory / model

The paper has no structural model. It tests a path-dependency hypothesis:
crisis risk depends not on the current level of monetary rates but on the
sequence of cuts and subsequent hikes. The economic mechanism it proposes is
consistent with the theoretical framework of Boissay et al. (2023), in which
a long period of monetary loosening triggers an investment and credit boom
and a search for yield/risk-taking; the subsequent tightening then collapses
credit markets through the fear of loan defaults.

**Identification strategy.** The key endogeneity concern is that central banks
raise rates when the economy (and the financial sector) is overheating,
so a positive correlation between rate hikes and crisis risk could reflect
omitted financial-sector vulnerabilities rather than a causal effect. The
paper addresses this by:

1. Controlling for contemporaneous and eight lags of country-level and global
   GDP growth and inflation in all specifications.
2. Residualizing monetary rate changes with respect to the main business-cycle
   variables (GDP growth, inflation, investment, consumption, current account,
   short- and long-term rates, decade fixed effects) to separate the systematic
   from the discretionary component.
3. Using the Mundell trilemma instrumental variable (see Method section), which
   exploits variation in base-country monetary policy transmitted through
   fixed-exchange-rate pegs and open capital accounts (Jorda, Schularick, and
   Taylor, 2020).

The paper shows the U-shape result is present for both raw and residualized
rate changes, and that the effect is larger for the residual (discretionary)
component, ruling out the possibility that the U shape merely reflects
mechanical policy responses to business-cycle conditions.

## Method

**Crisis-window regressions (equation 1, p. 932).**

$$
y_{i,t+h} - y_{i,t} = \alpha_{i,h} + \alpha_{d,h} + \beta_h \cdot 1[\text{Crisis}_{i,t}=1] + \epsilon_{i,t+h}
$$

- $$h = -7, \ldots, 0, \ldots, 7$$: years relative to crisis onset
- $$y$$: monetary policy rate level
- $$\alpha_i$$: country FE; $$\alpha_d$$: decade FE

This plots the average path of monetary rates around historical crisis events,
with 90% confidence intervals, for different crisis definitions and subsamples.

**Linear probability model for crisis risk (equation 2, p. 937-938).**

$$
\text{Crisis}_{i,t\text{ to }t+2} = \alpha_i + \beta_1 \cdot \Delta_3\text{Rate}_{i,t} + \beta_2 \cdot \text{Cut}_{i,t-8,t-3}
  + \beta_3 \cdot \Delta_3\text{Rate}_{i,t} \times \text{Cut}_{i,t-8,t-3}
  + \gamma \cdot X_{i,t} + u_{i,t}
$$

- $$\Delta_3\text{Rate}$$: three-year change in monetary policy rate (ppts)
- $$\text{Cut}$$: 1 if monetary rates were cumulatively cut (t-8 to t-3)
- $$X$$: contemporaneous values and eight lags of local and global inflation and GDP growth
- SE: Driscoll-Kraay (five lags) to account for cross-country, cross-time correlation

The coefficient $$\beta_3$$ is the U-shape test: it captures whether rate hikes
are especially crisis-inducing when preceded by prolonged cuts.

**Trilemma IV (equation 3, p. 938).**

$$
\text{Trilemma IV}_{i,t} = \Delta\text{Rate}^{\text{Residual}}_{b(i),t} \times \text{PEG}_{i,t} \times \text{PEG}_{i,t-1} \times \text{KOPEN}_{i,t}
$$

- $$\Delta\text{Rate}^{\text{Residual}}_{b(i),t}$$: residualized monetary rate change of the base country $$b(i)$$ (e.g. Germany for ERM members)
- $$\text{PEG}$$: 1 if fixed exchange rate regime
- $$\text{KOPEN}$$: degree of capital account openness (Quinn-Schindler-Toyoda rescaled)

The IV strategy instruments $$\Delta_3\text{Rate}$$ with the three-year change in the
residualized trilemma variable, and the interaction with $$\text{Cut}$$ with the
trilemma variable interacted with the cut dummy. Standard errors remain
Driscoll-Kraay. First-stage Kleibergen-Paap weak ID statistics are well above
conventional thresholds (27.48 to 65.68 across columns; Table III, p. 939).

**Local projections for red zone interaction (equation 4, p. 954).**

$$
\Delta_h y_{i,t} = \alpha_{i,h} + \beta_{1,h} \cdot \Delta\text{Rate}_{i,t} + \beta_{2,h} \cdot I(\Delta_3 y_{i,t} \geq \text{Rz})
  + \beta_{3,h} \cdot \Delta\text{Rate}_{i,t} \times I(\Delta_3 y_{i,t} \geq \text{Rz})
  + \gamma \cdot X + \epsilon_{i,t+h}, \quad h = 1, \ldots, 10
$$

- $$y$$: household credit, house prices, business credit, or equity prices
- $$\text{Rz}$$: red zone threshold (80th pctile for credit, 66.7th pctile for asset prices, following Greenwood et al. 2022)
- $$\beta_{3,h}$$: main coefficient: does a rate hike reverse vulnerabilities more strongly when the financial variable is already elevated?
- SE: Driscoll-Kraay with $$1.5 \times h$$ lags; 10% confidence intervals

**Spain loan-level credit supply regression (equation 5, p. 960).**

$$
\Delta y_{i,j,t} = \beta_1 \cdot \text{Cut}_{t-5,t} + \beta_2 \cdot \text{Cut}_{t-5,t} \times \text{Bank risk}_{i,t-1}
  + \beta_3 \cdot \text{Cut}_{t-5,t} \times \text{Bank risk}_{i,t-1} \times \text{Firm risk}_{j,t-1}
  + \gamma_1 \cdot F_{j,t-1} + \gamma_2 \cdot B_{i,t-1} + \gamma_3 \cdot M_t + u_{i,j,t}
$$

- $$\Delta y$$: log change in credit granted by bank $$i$$ to firm $$j$$
- $$\text{Cut}$$: 1 if overnight rates were below their average between t-5 and t
- $$\text{Bank risk}$$: bank NPL ratio (proxy for ex ante bank risk)
- $$\text{Firm risk}$$: 1 if firm is in construction/real estate sector
- $$F$$: firm-level controls (industry, location) and FE
- $$B$$: bank-level controls and FE
- $$M$$: macro controls and time FE; also firm*bank and firm*time FE variants
- SE: clustered at time and bank levels

**Spain loan-level default regression (equation 6, p. 963).**

$$
\text{Loan Default}_{i,j,t,t+3} = \beta_1 \cdot \Delta_3\text{Rate}_{t,t+3} + \beta_2 \cdot \text{Cut}_{t-5,t}
  + \beta_3 \cdot \Delta_3\text{Rate}_{t,t+3} \times \text{Cut}_{t-5,t}
  + \gamma_1 \cdot F_{j,t-1} + \gamma_2 \cdot B_{i,t-1} + \gamma_3 \cdot M_t + u_{i,j,t,t+3}
$$

- $$\text{Loan Default}$$: 1 if loan becomes delinquent (>90 days overdue) in t+1 to t+3
- $$\Delta_3\text{Rate}$$: ppt change in monetary rate between t and t+3
- $$\beta_3 > 0$$: hikes are more crisis-inducing when preceded by cuts (U-shape test)
- SE: clustered at time and bank levels

## Empirical specifications

The headline results tie to the following specification choices:

- **Macro panel (R1-R7).** Sample: 17 advanced economies, 1870 to 2020, 87
  banking crises (Jorda, Schularick, and Taylor 2016a chronology), annual
  frequency. Baseline uses the narrative crisis definition of Schularick and
  Taylor (2012). Robustness: Baron, Verner, and Xiong (2021) crisis dates;
  probit vs linear probability models; one-year vs three-year crisis windows;
  alternative rate-path window lengths; global credit-growth controls; decade
  fixed effects (Table III; Internet Appendix Tables IA.VI-IA.XI).

- **U-shape path classification.** An eight-year window is classified into four
  shapes based on the direction of the cumulative change in t-8 to t-3 and
  in t-3 to t. U shape = cumulative cut in the first five years followed by
  a raise in the last three years. This classification is used in frequency
  comparisons (Tables I, II, V, VIII) and interacted with rate changes in
  regression (equation 2).

- **Financial red zone (R7, R8 mechanism).** Defined following Greenwood et al.
  (2022) as periods when both credit growth (above the 80th percentile of the
  three-year change in the credit-to-GDP ratio) and asset price growth (above
  the 66.7th percentile of three-year real asset price changes) are
  simultaneously elevated. Computed separately for household and business
  sectors. Used as a mechanism variable to test whether the U-shape crisis
  effect runs through financial booms.

- **Spain micro panel (R8-R10).** The loan-level analysis follows the Spain CIR
  approach to bank risk-taking and monetary policy of Jimenez et al. (2014) and
  the CIR loan-level methodology for separating credit supply from credit demand
  of Jimenez et al. (2012). Sample: 10% random sample of Spanish
  nonfinancial corporate loans from the Central de Informacion de Riesgos
  (CIR), quarterly 1995 Q1 to 2008 Q3, matched to bank supervisory data and
  firm Mercantile Register data. Credit growth regressions: 1.9 million
  bank-firm-quarter observations. Cost of debt regressions: 1.2 million
  firm-year observations. Default regressions: 1.1 million loan observations
  (sample ends at 2011 Q3 to allow three-year default look-ahead). Fixed-effect
  saturation reaches firm-time and bank-time FE in the most demanding
  specifications (Table XI col. 6, Table XII col. 5-6).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Jorda-Schularick-Taylor (JST) Macrohistory Database | Monetary policy rates, banking crisis chronology, macro controls (GDP, inflation, credit, investment) for 17 advanced economies 1870-2020 | no page yet |
| Greenwood et al. (2022) financial red zone data | Credit-to-GDP and asset price growth series to define the financial red zone mechanism variable | no page yet |
| Baron, Verner & Xiong (2021) BVX crisis chronology | Alternative crisis dates using bank equity returns for robustness | no page yet |
| Spain Central de Informacion de Riesgos (CIR) | Loan-level monthly data on all corporate loans by Spanish banks 1984 to 2008 Q3 (10% random sample used); credit volumes, maturities, defaults | no page yet |
| Spain Mercantile Register (Registro Mercantil) | Annual firm balance sheet and income statement data; financial expenses over liabilities as cost-of-debt proxy | no page yet |
| Banco de Espana bank supervisory data | Bank-level balance sheet characteristics (total assets, capital ratio, liquidity, ROA, NPL ratio) matched to CIR | no page yet |
| Quinn-Schindler-Toyoda (2011) KOPEN index | Rescaled capital account openness measure for trilemma IV construction | no page yet |

Sample scope: macro panel covers 17 advanced economies (Australia, Belgium, Canada, Denmark, Finland, France, Germany, Italy, Japan, Netherlands, Norway, Portugal, Spain, Sweden, Switzerland, UK, US), annually 1870 to 2020 (77 crisis observations after data availability filters). Spain panel covers quarterly 1995 to 2008 Q3 (boom period) and defaults through 2011 Q3.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.70023) if you are:
studying banking crisis predictors and want the full robustness battery (30+
Internet Appendix tables); extending the trilemma IV to other contexts;
building on the financial red zone mechanism to study credit supply dynamics;
analyzing the 2022-2025 rate-hiking cycle as a potential U-shape episode;
or using the Spanish CIR administrative data methodology for loan-level
identification of credit supply. Tables III and XII contain the key
specifications; Figures 2 and 3 show the event-study path estimates.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 81(2). This distillation
was extracted by an LLM on 2026-06-01 and is **not human-verified or
independently reproduced**. The CC BY 4.0 licence permits mirroring; the
verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Jimenez, Gabriel, Dmitry Kuvshinov,
> Jose-Luis Peydro, and Bjorn Richter.
> "Monetary Policy, Inflation, and Crises: Evidence from History and
> Administrative Data."
> *The Journal of Finance* 81, no. 2 (April 2026): 923-970.
> DOI: 10.1111/jofi.70023. (C) 2026 The Author(s).
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Institutional Investor Attention: Kwan, Liu & Matthies (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2026/kwan-liu-matthies-2026/
# Distilled: institutional funds shift attention to macro news when volatility rises; attention tracks holdings; attention to a stock predicts that position's value-add; attention by buying hedge funds predicts stock returns. J. Finance 2026, CC BY 4.0. Eight core results with source locators, datasets used, the theory tested, and empirical specifications.
# Tags: paper-summary, limited-attention, institutional-investors, return-predictability, fund-performance, panel-regression, fama-macbeth, portfolio-sort, open-access, cc-by, peer-reviewed, unreplicated, data:fred, data:edgar, data:wrds, data:ravenpack, data:factset-lionshares, data:revelio, data:form-adv
==============================================================================

**What this is.** The paper's core results, datasets, and theory: enough
to know what it found without reading all 38 pages. To replicate or extend
it, read the full source: the
[verbatim PDF](/library/kwan-liu-matthies-2026-institutional-investor-attention.pdf)
(machine-accessible) or the
[original](https://doi.org/10.1111/jofi.70009).

## TL;DR

Using a proprietary dataset of institutional investors' Internet news
reading (Nov 2017 to Jun 2022; ~482M fund-firm-quarters, 4,075 funds), the
paper measures fund attention to *macro* vs *firm-specific* news. Funds
reallocate attention to macro news when aggregate volatility rises; funds
that reallocate more strongly earn higher future returns. Firm-specific
attention tracks holdings ("attention habitats"), and attention to a stock
predicts that position's value-add, most so for value-relevant news and for
buying hedge funds, whose attention predicts stock returns.

## Core results

Magnitudes and significance are as reported; `**`/`***` = 5%/1%. Locators
point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Funds shift attention toward **macro news when aggregate volatility is high** | Table III, p. 804 | β = 0.25\*\* on VIX²ₜ₋₁; robust to VIX and realized vol; ≈ 5% of sample-SD in macro-attention share per 1-SD VIX² |
| R2 | Funds with **higher attention-reallocation sensitivity (β^VIX²) earn higher future returns** | Table IV, p. 806 | coef 0.31→0.36 (sig 5–1%); ≈ +0.36%/qtr (~1.4%/yr) per 1-SD; ~2× stronger in top VIX quartile (interaction 0.73\*\*) |
| R3 | High-β^VIX² funds look **more efficient** | §III.C | 0.78–1.74% less attention-weighted salience (sig 5%); +17% advanced-degree staff; trait persistent (62–66% stay vs 25% random); hedge funds >2× mutual funds' β^VIX² |
| R4 | Firm-specific **attention strongly tracks portfolio holdings** ("attention habitats") | Table V, p. 809 | held read 5–6× more than non-held (t sig 1%); with firm×time FE, 1-SD holdings ≈ 1.02-SD attention; fund×firm FE dominate the variance |
| R5 | **Attention to a stock predicts that position's value-add** | Table VI, p. 813 | 1-SD attention ≈ +3.3% SD position value-add; trade-based coef 0.074\*\*; ×trade-size 0.58\*\*\* (bigger trades, more value) |
| R6 | Value-add is **concentrated in value-relevant news** (business/financial newswires) | Table VIII, p. 817 | biz/fin-newswire attention×holdings 0.107\*\* / 0.123\*\*\*; retail and general news insignificant |
| R7 | Funds **attend more to buys than sells; attentive buys outperform** | Table IX, p. 819 | residualized attention: buy ≈ 1.6 vs sell ≈ 0.6 vs hold ≈ 0 (buy>sell sig 1%); attentive buys add value, sells mixed/insignificant |
| R8 | **Attention by buying hedge funds predicts future stock returns** | Tables X–XI, pp. 821–823 | Fama-MacBeth: buying-HF attention 0.51\*/0.56\*\*\* (MF negative/insignificant, other-fund negative/significant at 5%); ≈ +0.35%/mo (~4%/yr) per 1-SD; HF long-short 0.53%/mo EW (t=2.75), 0.80%/mo VW; FF5 α ≈ 0.45%\*\*\* EW; no predictability for held/sold |

**Overall (paper's conclusion).** Attention is a resource that funds
allocate, and the allocation contributes to performance. Funds that
reallocate attention to macro news in volatile times, and that attend to
value-relevant firm news, do better; the strongest stock-return signal is
attention by *buying hedge funds*.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Proprietary Internet news-reading data ("Data Partner", anonymized analytics firm), Nov 2017–Jun 2022 | The attention measure itself | Proprietary; not public or redistributable; no page |
| RavenPack 1.0 | News topic / subject / sentiment; stock-ticker mapping | [RavenPack](/wiki/commercial/ravenpack/) (licensed) |
| FactSet LionShares | Institutional holdings (13-F), institution classification | [FactSet LionShares](/wiki/commercial/factset-lionshares/) (licensed) |
| CRSP & Compustat | Returns, fundamentals, stock characteristics | [WRDS / CRSP / Compustat](/wiki/commercial/wrds/) (licensed) |
| VIX (CBOE) | Aggregate-volatility measure (VIX²) | [FRED](/wiki/datasets/fred/), free, series `VIXCLS` |
| SEC Form ADV, Form N-1A | Fund descriptions for classification | [Form ADV](/wiki/datasets/form-adv/) (via SEC IAPD, *not* EDGAR); [N-1A](/wiki/datasets/edgar/#form-n-1a-open-end-fund-registration) via SEC EDGAR |
| LinkedIn / Revelio Labs | Fund human capital (advanced-degree share) for R3 | [Revelio Labs](/wiki/commercial/revelio/) (licensed) |

Sample: 481,820,400 fund-firm-quarters; 4,075 distinct funds.

## Theory / model

The paper has no original structural model. It is purely empirical,
testing predictions from the limited- and rational-inattention literature.

The primary theoretical framework is Peng and Xiong (2006) and
Kacperczyk, Van Nieuwerburgh and Veldkamp (2016), whose models predict
that attention-constrained investors allocate more attention to
macroeconomic news when aggregate uncertainty is high. In those models,
information about macro conditions becomes more valuable during high-VIX
periods, so a rational fund reallocates limited attention toward it. A
competing interpretation (also from Peng and Xiong (2006), Kacperczyk
et al. (2014)) is that funds with more binding constraints react more
to macro news when volatility rises, which would predict lower, not
higher, future returns. The paper tests both predictions against data.

For firm-specific attention, the theoretical baseline is Van Nieuwerburgh
and Veldkamp (2010): in equilibrium, investors prefer to hold assets they
are more informed about and have incentives to acquire information about
assets they expect to hold. This generates a positive feedback between
attention and holdings. Further, a fund's attention to a stock should be
positively associated with the value the position contributes to the
fund's performance (position-level value-add), because learning about a
stock becomes more valuable when the fund expects to hold a larger
position.

There is no identification via a natural experiment. The paper controls
for potential confounds through rich fixed-effect structures (see Method
and Empirical specifications).

## Method

The paper applies standard panel-regression, Fama-MacBeth, and portfolio-sort
methods (`panel-regression`, `fama-macbeth`, `portfolio-sort`); it proposes
no new estimator.

**Attention measures.** The core attention measure is the fraction of a
fund's reading of news articles covering either macroeconomic conditions
($$\text{InstAttn}^{\text{macro}}_{it}$$) or a specific firm ($$\text{InstAttn}_{ist}$$), derived
from event-level Internet news-reading data matched to fund identities via
the Data Partner (p. 796-799). Article topics and firm-stock mappings come
from RavenPack 1.0 (p. 797).

**Macroeconomic attention beta.** For each fund-quarter, the macroeconomic
attention sensitivity to aggregate volatility ($$\beta^{\text{VIX}^2}_{it}$$) is
estimated by regressing weekly attention to macro news on contemporaneous
$$\text{VIX}^2$$ over a trailing 52-week window (p. 804). This fund-quarter-level
beta is then used as a predictor in quarterly fund-return regressions.

**Position-level value-add.** The outcome $$h_{ist-1} \times R_{st}$$ weights
the stock return by the fund's prior-period dollar holding share, following
Berk and Van Binsbergen (2015) in spirit (p. 811).

**Trade-based value-add.** The outcome $$\Delta\text{Position}_{ist-1} \times R_{st}$$
multiplies the future stock return by the dollar change in holdings from
$$t-2$$ to $$t-1$$, isolating performance attributable to trading (p. 814).
These trade-based tests build on Akepanidtaworn et al. (2023), who find
that institutional buying adds value while selling does not (p. 815 of this
paper).

**Stock-level return predictability.** Fama-MacBeth cross-sectional
regressions at the monthly frequency, with controls for size,
book-to-market, profitability, investment, and news coverage (p. 821).
Portfolio sorts rank stocks into quintiles by attention from buying funds
within each investor-type category (p. 822). Newey-West standard errors
with two lags are used in both (pp. 821, 822).

## Empirical specifications

### Specification 1: Macro attention and aggregate volatility (eq. 1, p. 803)

$$
\text{InstAttn}^{\text{macro}}_{it} = \beta \cdot \text{VIX}^2_{t-1} + \mu_i + \text{Controls} + \epsilon_{it}
$$

- LHS: fraction of fund $$i$$'s reading in month $$t$$ about macro news.
- RHS: $$\text{VIX}^2_{t-1}$$, the average VIX-squared in the prior month (normalized to mean 0, SD 1); $$\mu_i$$ = fund fixed effects; controls include market return.
- SE: clustered by fund and time (p. 804).
- Sample: 234,934 fund-month observations (Table III, p. 804).
- Robustness: replaces $$\text{VIX}^2$$ with VIX and with realized S&P 500 daily volatility (columns 3-4 of Table III).

### Specification 2: Fund-level return predictability (Table IV, p. 806)

$$
\text{Fund Ret}_{it} = \gamma \cdot \beta^{\text{VIX}^2}_{it-1} + \lambda \cdot [\beta^{\text{VIX}^2}_{it-1} \times \text{VIX}_{t-1}]
                + \text{FundFE} + \text{TimeFE} + \text{Controls} + \epsilon_{it}
$$

- LHS: quarterly fund return weighted by holdings at $$t-1$$.
- RHS: $$\beta^{\text{VIX}^2}_{it-1}$$ (attention-reallocation sensitivity, normalized); interaction with VIX level; controls include log AUM and log articles read.
- SE: clustered by fund and time (p. 806).
- Sample: 51,207 fund-quarter observations (Table IV, p. 806).

### Specification 3: Attention and holdings - intensive margin (eq. 2, p. 810)

$$
\text{InstAttn}_{ist} = \alpha + \beta \cdot h_{ist-1} + \epsilon_{ist}
$$

- LHS: share of fund $$i$$'s reading devoted to firm $$s$$ in quarter $$t$$.
- RHS: $$h_{ist-1}$$ = dollar share of firm $$s$$ in fund $$i$$'s portfolio at end of quarter $$t-1$$; $$\alpha$$ denotes fixed effects (fund $$\times$$ time, firm $$\times$$ time in various columns).
- SE: clustered by fund, firm, and time (p. 809).
- Sample: held stocks only ($$h_{ist-1} > 0$$); 11,910,288 fund-stock-quarter observations (Table V, p. 809).

### Specification 4: Position-level value-add (eq. 3, p. 812)

$$
h_{ist-1} \times R_{st} = \alpha + \beta \cdot \text{InstAttn}_{ist-1} + \delta \cdot h_{ist-2} + \epsilon_{ist}
$$

- LHS: position-level value-add (holding weight times next-quarter return, scaled by 100).
- RHS: $$\text{InstAttn}_{ist-1}$$ = fraction of reading about firm $$s$$ in $$t-1$$ relative to total fund reading; $$h_{ist-2}$$ controls for prior-period holdings.
- FE: fund $$\times$$ time; or fund $$\times$$ time + firm $$\times$$ time (columns 1-3, Table VI, p. 813).
- SE: clustered by fund, firm, and time.
- Sample: held stocks ($$h_{ist-1} > 0$$); ~11.9M fund-stock-quarter obs.

### Specification 5: Trade-based value-add (eq. 4, p. 814)

$$
\Delta\text{Position}_{ist-1} \times R_{st} = \alpha + \beta \cdot \text{InstAttn}_{ist-1} + \delta \cdot h_{ist-2} + \epsilon_{ist}
$$

- LHS: trade-based value-add (dollar change in holdings from $$t-2$$ to $$t-1$$ times return at $$t$$, scaled by 100).
- RHS: same regressors as specification 4; also interacted with trade size.
- FE: firm $$\times$$ time; fund $$\times$$ time (Table VI, Panel B, p. 813).
- SE: clustered by fund, firm, and time.
- Sample: traded positions only; ~11.1M fund-stock-quarter observations.

### Specification 6: Stock-level Fama-MacBeth return predictability (Table X, p. 821)

$$
R_{st+1} \times 100 = a_t + b_t \cdot \text{AttnBuying}_{st} + \text{controls}_t + u_{st}
$$

- LHS: monthly stock return times 100.
- RHS: attention by buying (or holding/selling) hedge funds, mutual funds, or other funds; controls include size, book-to-market, gross profitability, investment, log news coverage, and log number of funds in the action category.
- Coefficients: time-series average of cross-sectional OLS slopes.
- SE: Newey-West with two lags (p. 821).
- Sample: 54 monthly periods (Nov 2017 to Jun 2022); each period ~500-3000 stocks with adequate news coverage.

### Specification 7: Portfolio sorts by buying-fund attention (Table XI, p. 822-823)

Stocks are ranked into five quintiles each quarter by the attention-to-buys
measure from the Fama-MacBeth specification, separately for hedge funds,
mutual funds, and other funds. Equal-weighted (EW) and value-weighted (VW)
returns to each quintile and the long-short (H-L) spread are computed over
the next three months. Factor alphas (CAPM, FF3, FF3+MOM, FF5) are estimated
with Newey-West two-lag adjustment (Table XI, Panel B, p. 823).

## When to read the full paper

Use the [mirrored PDF](/library/kwan-liu-matthies-2026-institutional-investor-attention.pdf)
if you are: replicating (code in the journal's Supporting Information);
extending the attention measure or the value-add tests; doing a literature
review where the Internet Appendix robustness matters; or auditing a
specific coefficient. The locators above point you to the exact table. For
"what did this paper find," the table above is sufficient and is the
intended default.

## Attribution & rights

Source: peer-reviewed, *The Journal of Finance* 81(2). This distillation
was extracted by an LLM on 2026-05-17 and is **not human-verified or
independently reproduced**. Licence, verification trail, and takedown
policy: [Open Library](/library).

> **Attribution (CC BY 4.0).** Kwan, Alan, Yukun Liu, and Ben Matthies.
> "Institutional Investor Attention." *The Journal of Finance* 81, no. 2
> (April 2026): 791–827. DOI: 10.1111/jofi.70009. © 2026 The Author(s).
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**. The
> verbatim, unmodified PDF is mirrored in the [Open Library](/library).


==============================================================================
# Communism and Financial Markets: Laudenbach, Malmendier & Niessen-Ruenzi (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2026/laudenbach-communism-attitudes-2026/
# Distilled: East Germans invest less in stocks and hold more negative attitudes toward capital markets decades after reunification, with the gap explained by lasting adherence to anti-capitalist ideology shaped by personal experiences under communism. J. Finance 2026, paywalled. Ten core results with source locators, datasets used, the identification strategy, and the empirical specifications.
# Tags: paper-summary, household-finance, stock-market-participation, ideology
==============================================================================

**What this is.** The paper's core results, the identification strategy, and
the empirical specifications with enough detail to know what it found and how,
without reading all 43 pages. To replicate or extend, read the full source at
the [original](https://doi.org/10.1111/jofi.70006).

## TL;DR

Using three independent data sets (a representative survey of 9,695 Germans,
proprietary bank records for 326,437 customers, and brokerage data for 230,229
retail investors), the paper shows that East Germans are 25-28% less likely to
participate in the stock market than West Germans, and that a significant gap of
roughly 10% persists after controlling exhaustively for wealth, income,
financial literacy, trust, social capital, and risk aversion. Where prior work
such as Fuchs-Schundeln and Haliassos (2021) documents the East-West
participation gap, this paper shows that a residual gap survives richer controls
and traces it to ideology. The paper argues
that the gap is explained by lasting adherence to the GDR's anti-capitalist
ideology: East Germans with stronger positive experiences of life under
communism (proxied by geography-based variation and by survey memories) show
greater stock-market aversion, while those with more negative experiences (e.g.,
living in heavily polluted areas or areas without Western TV access) invest more.
The financial cost is real: East German investors earn 7-11 basis points per
month less, hold fewer assets, pay higher fund fees, and hold less diversified
portfolios than comparable West Germans.

## Core results

Magnitudes and significance are as reported; `**`/`***` = 5%/1%. Locators
point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Raw East-West gap in stock market participation is large and consistent across all three data sets | Table III, pp. 1119-1120; Table II, p. 1114 | Survey: 27.6% gap (East 26.9% vs. West 35.5%); bank: 25.2% gap; broker: 27.7% gap |
| R2 | A significant residual gap of ~10% remains after controlling for the full set of demographic and financial variables | Table III col (2) all panels, pp. 1119-1120 | Survey: -0.026\*\*\* (SE 0.009), effect size 8.4%; bank: -0.007\*\*\* (SE 0.001), 7.7%; broker data: -0.156\*\*\* (SE 0.002), 19.1% |
| R3 | East Germans who moved to West Germany after Reunification and now live in the same economic environment as West Germans still invest significantly less | Table IV col (1)-(2), p. 1125 | Survey movers: -0.072\*\*\* (SE 0.026) on all Germans, -0.075\*\* (SE 0.029) restricting to West Germans only |
| R4 | East Germans are less willing to buy "capitalist" stocks (financial firms, U.S. firms) and more willing to buy stocks of formerly communist countries | Table V, p. 1130 | Survey: -1.9 pp for financial firms (4.7% relative gap), -1.9 pp for U.S. firms (5.5% gap); +5.9 pp for Chinese firms (23.5% gap); broker data: -4.9 pp financial firms\*\*\*, -1.9 pp U.S. firms\*\*\*, +0.4 pp East European firms\*\*\* |
| R5 | Within East Germany, anti-capitalist and anti-stock-market attitudes directly predict lower stock market participation | Table VI, p. 1132 | "I generally reject stocks": -0.103\*\*\* (SE 0.004); "Investing in stocks is immoral": -0.070\*\*\* (SE 0.006); "Capitalism should be abolished": -0.020\*\*\* (SE 0.005) |
| R6 | Negative GDR experiences (pollution, no Western TV access) predict higher stock market participation; positive experiences (renamed showcase cities, Olympic gold medal wins) predict lower participation | Table VII, p. 1136 | Pollution: +0.058\*\*\* (SE 0.009); No West-TV: +0.090\*\*\* (SE 0.016); Renamed city: -0.169\*\*\* (SE 0.010); Olympic gold (population-weighted): -0.044\*\*\* (SE 0.006) |
| R7 | Survey-based GDR memories directly link to stock market participation in the expected direction | Table VIII, p. 1138 | High GDR life standard: -0.032\*\*\* (SE 0.008); Wishing GDR back: -0.062\*\*\* (SE 0.016); Positive GDR memories: -0.050\*\*\* (SE 0.019) |
| R8 | East German investors earn lower risk-adjusted portfolio returns than West Germans | Table IX Panel A, p. 1140 | CAPM alpha (equal-weighted long East, short West): -0.082\*\* (SE 0.040); FF3 alpha: -0.070\* (SE 0.040); FF4 alpha: -0.099\*\* (SE 0.041) per month |
| R9 | East German investors hold significantly fewer passive investments, fewer assets, pay higher fund fees, and hold less diversified portfolios | Table IX Panel B, p. 1141 | Passive investments: -0.017\*\*\* (SE 0.002), 44.7% lower relative to baseline; fund fees: +3.78% higher (col 3); number of assets: -33.1% (col 2) |
| R10 | The stock market participation gap did not exist before the separation of Germany (1920-1924 bank data), ruling out pre-existing persistent differences | §II.A, p. 1116 | Historical stock-market participation: East 66.7% vs. West 68.2%; difference not statistically significant (t-statistic: 0.34) |

**Overall (paper's conclusion).** Exposure to anti-capitalist ideology can
exert lasting influence on individual investment behavior for years and even
decades. Individuals who remember life in the GDR positively are more likely
to continue holding anti-capitalist views and to refrain from stock market
investment, with adverse financial consequences. Negative personal experiences
under communism reverse the effect. These findings suggest that ideology, not
only financial experience, shapes long-run investment behavior, and offer a
micro-level foundation for macroeconomic growth differentials between formerly
communist and capitalist countries. The broader context for this East-West
comparison is the review of the long-term effects of communism in Eastern Europe
by Fuchs-Schundeln and Schundeln (2020).

## Theory / model

The paper has no formal model. The theoretical framework is the
**emotional-tagging hypothesis** from cognitive science, applied to ideological
formation. The core logic is as follows.

Emotionally charged stimuli are encoded more strongly in long-term memory. The
valence (positive vs. negative) of an emotional context during an experience
shapes the memory trace: positive emotions create favorable associations with
the context, and negative emotions create unfavorable ones (Richter-Levin and
Akirav, 2003; Dolan, 2002). This extends the experience-effects framework of
Malmendier and Nagel (2011), in which lifetime returns shape stock market
participation, to ideological rather than financial experiences. Applied to the GDR setting, individuals who lived
under communism with positive experiences are predicted to (i) form stronger
positive associations with the communist ideology and its anti-capitalist
stance, (ii) retain these associations in long-term memory, and (iii) carry
them into financial behavior decades later. Those with negative experiences are
predicted to reject the ideology and, as a result, embrace capitalist financial
markets more readily.

The tested hypothesis is:

- **H1 (positive tagging):** East Germans with more positive experiences under
  communism show greater adherence to anti-capitalist views and lower stock
  market participation.
- **H2 (negative tagging):** East Germans with more negative experiences under
  communism show less adherence to anti-capitalist views and higher stock
  market participation.
- **H3 (mechanism):** The link from experience to investment runs through
  ideological attitudes: anti-capitalist beliefs (toward stocks specifically
  and capitalism generally) mediate the effect on stock market participation.

**Identification strategy.** The paper uses three complementary identification
approaches:

1. **East-West comparison** (Sections II-III): the quasi-natural experiment of
   Germany's post-WWII division and 1990 reunification. Pre-trends are
   addressed using historical 1920-1924 bank data showing no pre-existing gap.
   Movers (East Germans who relocated to West Germany after Reunification) are
   used to rule out contemporaneous environmental confounders, addressing the
   concern raised by Becker, Mergele, and Woessmann (2020) that reunification
   may not be a clean natural experiment.

2. **Within-East survey variation** (Section III.B): variation in anti-capitalist
   attitudes among East Germans, related directly to stock market participation
   via logit regressions.

3. **Within-East geographic variation** (Section IV): four geography-based,
   predetermined (pre-reunification) proxies for positive or negative GDR
   experiences - pollution levels (negative), access to West German TV (negative
   for lack thereof, following Bursztyn and Cantoni (2016), who use Western TV
   access in East Germany as a quasi-exogenous proxy), living in GDR renamed
   showcase cities (positive), and
   proximity to Olympic gold medal winners (positive) - are used as
   quasi-exogenous instruments for emotional tagging. These proxies are
   orthogonal to current economic conditions and to each other (Internet
   Appendix Table IA.XIV).

## Method

The paper applies logit regression as the primary estimator for all stock
market participation outcomes. For portfolio characteristics and returns, OLS
is used. The estimating framework is described in Section II.B (p. 1118).

**Main logit specification (equation 1, p. 1118):**

$$
P(y_{it} = 1 \mid \text{East}_i, x_{it}, z_{c(i),t}, v_t) = \Phi(\alpha + \beta \cdot \text{East}_i + \gamma' x_{it} + \delta' z_{c(i),t} + v_t)
$$

where:
- $$y_{it}$$ = 1 if investor $$i$$ participates in the stock market in year $$t$$
- $$\text{East}_i$$ = 1 if investor lives in East Germany (former GDR)
- $$x_{it}$$ = individual-level controls (gender, age, marital status, risk
  tolerance, wealth, income, financial literacy, trust, social capital,
  return expectations, peer effects; data-set specific)
- $$z_{c(i),t}$$ = municipality-level controls (in broker data: number of bank
  branches, population, real estate wealth, share with high-school degree,
  county GDP, number of local firms, Facebook social connectedness index)
- $$v_t$$ = year fixed effects (broker data only)
- $$\Phi(\cdot)$$ = logistic CDF (the paper's notation for the logit link function)

Coefficients are reported as **average marginal effects**. Standard errors are
clustered by municipality in the survey and bank data; by broker customer in
the broker data.

The `builds-on` technique primitives used are `panel-regression` (OLS for
returns and portfolio characteristics), `logit-regression` (the logit AME
specification for participation), `difference-in-differences` (East-West
comparison before/after, with matched-city and Berlin sub-samples), and
`matching` (characteristics-matched cities: Eisenach vs. Bad Hersfeld, broker
data col (4) Table III Panel C).

**Alternative estimation.** The paper also uses Conley (1999, 2008) spatial
HAC standard errors with a 50 km distance cutoff and a two-year linear
Bartlett window (§II.C) to address spatial and serial autocorrelation in the
broker data. Results are qualitatively unchanged.

## Empirical specifications

### Baseline East-West participation gap (R1, R2)

Three separate logit regressions are run on each of the three data sets (Table
III, pp. 1119-1120):

$$
P(\text{stock market participant} = 1) = \Phi(\alpha + \beta \cdot \text{East}_i + \text{controls})
$$

- **Survey data (Panel A):** N = 9,695; cross-section, 2023; SE clustered by
  municipality. Controls include gender, age, marital status, wealth (8-level),
  income (6-level), education (4-level), trust, risk tolerance (1-7), financial
  literacy (0-3), familiarity with stocks, peer effects, social capital (two
  measures), return expectations. Berlin excluded.

- **Bank data (Panel B):** N = 326,437; cross-section, 2019; SE clustered by
  municipality. Controls include gender, age, marital status, employment,
  wealth, income, risk tolerance, product ownership (consumer credit, retirement
  savings plans, credit card, mortgage, savings plans), number of consultations.

- **Broker data (Panel C):** N = 839,292 investor-years, June 2004-December
  2012; SE clustered by broker customer. Controls include investor age, marital
  status, portfolio value, time account open, municipality controls (see above);
  year FE.

Robustness variants: (i) Berlin-only subsample (col 3, Panel C), (ii) matched
cities Eisenach-Bad Hersfeld (col 4, Panel C, N = 574), (iii) restricting to
active bank accounts (col 3, Panel B), (iv) single-stock holding as dependent
variable (col 4, Panel B), (v) HAC standard errors.

### Movers specification (R3)

$$
P(\text{stock market participant} = 1) = \Phi(\alpha + \beta_1 \cdot \text{Mover}_i + \beta_2 \cdot \text{East}_i + \text{controls})
$$

where $$\text{Mover}_i$$ = 1 if respondent moved from East to West Germany after 1989
and lived in GDR for at least 10 years. The $$\text{East}_i$$ coefficient is set to
zero in columns (2) and (4) (West-Germans-only subsample). Estimated on survey
data (col 1-2, N = 9,695/4,409) and a bank survey subsample (col 3-4, N =
241/198). Table IV, p. 1125.

### Stock type preference: "communist" vs. "capitalist" stocks (R4)

$$
P(\text{hold stock type } k = 1) = \Phi(\alpha + \beta \cdot \text{East}_i + \text{controls})
$$

run separately for financial-industry stocks, U.S.-company stocks, Chinese
stocks, and East European stocks. Survey: willingness to buy (Table V Panel A,
p. 1130); bank data: actual holdings conditional on participating (N = 29,768,
Table V Panel B); broker data: portfolio share (N = 611,410, Table V Panel C).
Year FE included in broker data. SE clustered by municipality (survey, bank) or
broker customer (broker data).

### Anti-capitalist attitudes and participation (R5)

$$
P(\text{stock market participant} = 1) = \Phi(\alpha + \beta \cdot \text{Attitude}_{qi} + \text{controls})
$$

estimated separately for each of nine attitude survey questions on four-point
Likert scales, restricted to the East German survey subsample (N = 5,286). Questions
span (i) anti-stock-market attitudes (Panel A), (ii) anti-capitalist attitudes
(Panel B), and (iii) pro-capitalist attitudes (Panel C). Table VI, p. 1132. SE
clustered by municipality.

### Geographic experience proxies and participation (R6)

$$
P(\text{stock market participant} = 1) = \Phi(\alpha + \beta_k \cdot \text{Proxy}_k + \text{controls} + \text{year FE})
$$

run separately for each of four proxy variables on the East German broker
subsample (N = 171,343 investor-years):

- $$\text{Pollution}_c$$ = 1 if investor lives in a municipality on the 1990 GDR
  environmental emergency list (negative experience proxy)
- $$\text{NoWestTV}_c$$ = 1 if municipality did not receive West German TV signals
  (negative experience proxy)
- $$\text{RenamedCity}_c$$ = 1 if municipality was renamed under the GDR communist
  regime (positive experience proxy)
- $$\text{OlympicGold}_c$$ = indicator scaled by inverse population rank for whether
  an Olympic gold medal winner was born in the same municipality (positive
  experience proxy)

Table VII, p. 1136. SE clustered by broker customer.

### GDR memories and participation (R7)

Same specification as R6 but using survey self-reports for five GDR memory
questions (Likert scale): living standard, wishing GDR back, disappointed in
FRG, positive GDR experience, positive GDR memories. East German survey
subsample only (N = 1,661-4,874). Table VIII, p. 1138. SE clustered by
municipality.

### Portfolio return and characteristic regressions (R8, R9)

$$
\alpha_{E\text{-}W,\, t} = \text{Long East portfolio} - \text{Short West portfolio}
$$

Monthly portfolio returns (including dividends, from Thomson Reuters
Datastream) are regressed on CAPM, Fama-French three-factor, and Carhart
four-factor models using German risk factors (CFR Cologne). Both equal- and
value-weighted portfolio constructions used. N = 92 monthly observations. Table
IX Panel A, p. 1140.

Portfolio characteristics (passive investment indicator, number of assets, fund
fees, Herfindahl index, bank-owned product share) regressed on East dummy plus
the same broker-data controls. N = 515,600-839,292. Table IX Panel B,
p. 1141.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Bilendi online survey (2023) | Representative survey of 9,695 Germans (5,286 East, 4,409 West); attitudes, stock market participation, demographics, trust, risk tolerance | no page yet |
| Proprietary bank data (2019) | 326,437 randomly selected bank customers; financial product holdings, wealth, income, stock market participation | no page yet |
| Online broker data (2004-2012) | 839,292 investor-year observations for 230,229 retail investors; security holdings, portfolio characteristics, returns | no page yet |
| Thomson Reuters Datastream | Monthly stock returns (including dividends) for portfolio return calculations; merged into broker data | no page yet |
| SAVE survey (Germany) | Municipality-level real estate wealth panel, merged as control into broker data | no page yet |
| German Census / Federal Statistical Office | Education variables (share with high-school degree) and economic indicators at municipality level | no page yet |
| GDR 1990 Environmental Emergency Report | Municipality-level air-pollution indicator (18 GDR municipalities requiring immediate action) | no page yet |
| Braggion, von Meyerinck & Schaub (2023) bank data (1920-1924) | Historical baseline: stock market participation for 2,000+ East and West German customers before the GDR, to rule out pre-existing differences | no page yet |

Sample: survey 2023 (cross-section); bank 2019 (cross-section); broker June
2004-December 2012 (panel). Geographic scope: Germany (East former GDR vs.
West FRG). The broker data cover 171,343 East German investor-year observations
used in the within-East experience analysis.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.70006) if you are:
studying the long-run behavioral effects of political or ideological exposure
on financial markets; replicating the geography-based experience proxies
(pollution, West-TV access, renamed cities, Olympic victories); designing
surveys to measure ideology and financial behavior jointly; or studying the
financial welfare costs of ideological aversion to capital markets. The
Internet Appendix contains all robustness tables (IA.III-IA.XV), variable
definitions (Table IA.I), and the exact survey question wording (Table IA.X).

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 81(2), April 2026, pp.
1103-1145. DOI: [10.1111/jofi.70006](https://doi.org/10.1111/jofi.70006).
Copyright 2025 the American Finance Association. This article is paywalled; no
Creative Commons licence is in effect.

This distillation was extracted by an LLM (claude-sonnet-4-6) on 2026-06-01
and is **not human-verified or independently reproduced**. Access to the full
text requires a subscription to *The Journal of Finance* or institutional
access via Wiley Online Library.

> Laudenbach, Christine, Ulrike Malmendier, and Alexandra Niessen-Ruenzi.
> "The Long-Lasting Effects of Experiencing Communism on Attitudes toward
> Financial Markets." *The Journal of Finance* 81, no. 2 (April 2026):
> 1103-1145. DOI: 10.1111/jofi.70006.


==============================================================================
# Investor Composition and Liquidity Component: Li & Yu (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2026/li-investor-composition-corporate-bond-liquidity-2026/
# Distilled: The loading of U.S. corporate bond credit spreads on bid-ask spreads more than doubled from 2005 to 2019 as mutual funds and ETFs grew, raising the liquidity component from roughly 10% to 30% of credit spreads. A directed-search model with heterogeneous investors and bonds shows that the inflow of short-term investors (mutual funds, ETFs) amplifies secondary-market frictions on prices via both a direct trading-frequency channel and an indirect trade-delay channel. J. Finance 2026, paywalled. Eight core results with source locators, datasets used, the model, and the method with its key equations.
# Tags: paper-summary, fixed-income, liquidity, corporate-bonds, investor-composition
==============================================================================

**What this is.** The paper's core results, the directed-search model with two-sided
heterogeneity, and the key equations that connect investor composition to the
liquidity component of corporate bond credit spreads: enough to understand what
it found and how, without reading all 52 pages. To replicate or extend it, read
the full source at the [original](https://doi.org/10.1111/jofi.70024).

## TL;DR

Building on Bao, Pan, and Wang (2011), who find that illiquidity explains a
significant part of common credit spread variation, the paper documents that
from 2005 to 2019 the loading of U.S. corporate bond credit spreads on bid-ask
spreads more than doubled, driven by the rapid entry of mutual funds and ETFs
into the bond market. The authors document this trend and use a 10-year
time-to-maturity discontinuity in fund mandates as an instrument to show
causally that bonds held by more short-term investors have higher trading
activity and a stronger credit-spread sensitivity to secondary-market frictions.
A directed-search model with heterogeneous investors (differing in liquidity
shock frequency) and heterogeneous bonds (differing in maturity and default
probability) shows that declining risk-free rates induce more short-term
investors to reach for yield in the illiquid bond market, amplifying the effect
of secondary-market frictions on prices through two channels: a direct channel
(more frequent trading means each unit of transaction cost is incurred more
often) and an indirect channel (bid-ask spreads are correlated with
seller-buyer ratios, so trade delays are priced into credit spreads). The
calibrated model matches the observed 2.5 to 2.8x growth in the liquidity
component and shows the indirect channel accounts for more than half of
the total sensitivity by the end of the sample.

## Core results

Magnitudes and significance are as reported; `\*`, `\*\*`, `\*\*\*` = 5%, 1%, 0.1%.
Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | The loading of credit spreads on bid-ask spreads (beta) has increased significantly for both investment-grade and high-yield bonds since 2005 | Figure 1, p. 880; Figure A3, p. 910 | IG: loading rose from ~10-20 bp/100 bp BA to ~60 bp/100 bp BA by 2019; HY: from ~60 bp/100 bp BA to ~250 bp/100 bp BA; for the full sample, beta rose from 0.54 to 1.2 |
| R2 | The median liquidity component (loading x BA / CS) grew from roughly 10% to 30% of credit spreads, a ~2.8x increase | Figure 2, p. 882; §I.B, p. 881-882 | Median liq. component grew ~2.8x from pre-GFC to 2019 (11.3% in first three years of sample; 31.4% in last three years); model predicts 6.77% in 2005 and 16.62% in 2019, a 2.5x model increase |
| R3 | Bonds held by more short-term investors have significantly higher investor composition measures and significantly more trades (IV first stage and second stage) | Table I, cols. (1)-(6), p. 886 | First stage: 1_{ttm>10} coefficient on Inv_Comp = -0.00581\*\*\* (t=-5.86, full) and -0.00692\*\*\* (t=-7.24, subsample). Second stage: 1 s.d. increase in Inv_Comp raises log No. of Trades by 75.02\*\*\* (t=8.02) and 61.62\*\*\* (t=9.69) |
| R4 | Higher investor composition (more short-term investors) significantly increases the loading of credit spreads on bid-ask spreads (Inv_Comp x Bid_Ask interaction) | Table II, cols. (2) and (4), p. 887 | Inv_Comp x Bid_Ask: 48.65\*\*\* (t=6.61, full sample) and 55.86\*\*\* (t=3.93, subsample); a 1 s.d. increase in investor composition (0.03) is associated with ~1.44 increase in the loading coefficient |
| R5 | The model generates positive assortative matching: short-term investors endogenously hold short-maturity and high-default-probability bonds (Proposition 1) | §III, p. 895; Figure 3, p. 884 | Theoretical result (Prop. 1): theta'(delta+d) > 0 in equilibrium; empirically confirmed by monotone relationships in Figure 3 across 20 investor-composition bins |
| R6 | As the risk-free rate declines, more short-term investors enter the illiquid bond market, reducing the seller-buyer ratio in all submarkets (Proposition 3) | §III, p. 898; Figure 4, p. 898 | Proposition 3 result; numerically verified: cutoff theta increases from 0.688 in 2005 to 0.766 in 2019 in calibration (Table IV, p. 903) |
| R7 | The calibrated model matches the key empirical moments; the indirect channel (bid-ask spreads correlated with seller-buyer ratios) accounts for more than 50% of the total sensitivity by 2019 | Table III and Table V, pp. 903-905 | beta_exo (direct component): 0.27 in 2005, 0.35 in 2019; beta_endo (indirect component): 0.65 in 2005, 1.7 in 2019; total empirical beta: 0.54 in 2005, 1.2 in 2019 |
| R8 | The change in investor composition amplifies the effect of a dealer regulation change for short-term bonds but alleviates it for long-term bonds, highlighting the dual liquidity-provision role | §IV, p. 907; Appendix A | Qualitative result from calibration: for short-term bonds, more short-term investors amplify dealer regulation frictions; for long-term bonds, investor inflows provide liquidity that dampens the regulation effect |

**Overall (paper's conclusion).** Where Wu (2020) attributes similar trends to
dealer regulation changes, this paper shows that investor composition is an
independent quantitative explanation even for investment-grade bonds where
dealer regulation may matter less. The massive growth of mutual funds and ETFs in
the corporate bond market is quantitatively important in explaining the rising
sensitivity of credit spreads to secondary-market frictions. The model shows
this operates through two amplification channels (direct trading frequency,
indirect trade-delay pricing) and that investor composition can interact with
dealer regulatory changes in complex ways depending on bond maturity.

## Theory / model

The paper builds a directed-search model with two-sided heterogeneity: investors
differ in their liquidity shock frequency $$\theta$$ and bonds differ in their
maturity intensity $$\delta$$ and default intensity $$d$$. Time is continuous. All agents
are infinitely lived and risk-neutral.

**Investors.** Each period measure $$m_I$$ of new investors enter with discount rate
$$\rho$$. Investor $$j$$ faces liquidity shocks arriving at Poisson rate $$\theta_j \in
[\bar{\theta}, +\infty)$$, a permanent feature with CDF $$F(\cdot)$$ and PDF $$f(\cdot)$$. Upon a
liquidity shock an investor becomes impatient and values flow coupon at a
discount $$\Delta$$ units less. The effective flow to an impatient investor is thus
$$r_i - \Delta$$ (p. 889, §II.A).

**Bonds.** There is a continuum of bonds indexed by $$i \in I$$. Bond $$i$$ has face value 1,
coupon $$r_i$$ (determined in equilibrium), matures at rate $$\delta_i$$, and defaults at
Poisson intensity $$d_i$$ with recovery $$s_i$$. Bonds are ordered so that $$\delta_i + d_i
\leq \delta_j + d_j$$ for $$i \leq j$$ (p. 889, §II.A.2).

**Secondary market and bid-ask spread.** In each submarket (indexed by bond type
and price), sellers search for buyers via a Cobb-Douglas matching function
(p. 890, II.A.2):

$$
m(\alpha_{s,i},\, \alpha_{b,j}) = \eta \cdot \alpha_{s,i}^{\gamma} \cdot \alpha_{b,j}^{1-\gamma}, \quad \eta > 0,\; \gamma \in (0,1)
$$

Following the reduced-form bid-ask spread specification of Lester, Rocheteau,
and Weill (2015), which micro-founds the bid-ask spread as a fraction of the
trade surplus in directed search, the bid-ask spread decomposes into an
endogenous component $$\xi$$ (proportional to the trade surplus) and an exogenous
component $$\epsilon_{\xi}$$ capturing dealers' balance sheet conditions:

$$
P_{b,i,t} - P_{s,i,t} = \xi_i + \epsilon_{\xi,i} \tag{9}
$$

$$
\text{Assumption 1:} \quad \xi_i = \kappa \bigl(V_{h,i}(\theta) - V_{b,i}(\theta) - V_{s,i}\bigr), \quad \kappa \in (0,1) \tag{14}
$$

**Value functions.** The HJB equation for a patient bondholder of type $$\theta$$
holding bond $$i$$ is (eq. 10, p. 891):

$$
\rho \, V_{h,i}(\theta) = r_i + \delta_i \bigl(1 - V_{h,i}(\theta)\bigr)
                           + d_i \bigl(s_i - V_{h,i}(\theta)\bigr)
                           + \theta \bigl(V_{s,i} - V_{h,i}(\theta)\bigr) \tag{10}
$$

The seller's HJB is (eq. 11, p. 892):

$$
\rho \, V_{s,i} = r_i - \Delta + \delta_i \bigl(1 - V_{s,i}\bigr)
              + d_i \bigl(s_i - V_{s,i}\bigr)
              + \mu_s\!\bigl(\lambda(i, P_s)\bigr) \bigl(P_s - V_{s,i}\bigr) \tag{11}
$$

The buyer's HJB for type $$\theta$$ searching in bond $$i$$'s submarket is (eq. 12,
p. 892):

$$
\rho \, V_{b,i}(\theta) = \mu_b\!\bigl(\lambda(i, P_s)\bigr) \bigl(V_{h,i}(\theta) - V_{b,i}(\theta) - P_{b,i}\bigr) \tag{12}
$$

where $$\mu_b(\lambda) = \eta \lambda^{\gamma}$$ and $$\mu_s(\lambda) = \eta \lambda^{\gamma-1}$$
are the buyer and seller meeting intensities given Cobb-Douglas matching.

**Equilibrium conditions.** In equilibrium buyers are indifferent between
primary and secondary markets:

$$
V_{h,i}(\theta) - 1 = V_{b,i}(\theta) \tag{15}
$$

The investor participation decision features a cutoff strategy: investors with
$$\theta < \bar{\theta}(r_f)$$ hold risk-free assets; those with $$\theta \geq \bar{\theta}$$
participate in the bond market. As $$r_f$$ decreases, the cutoff falls and more
short-term investors enter (Proposition 3, p. 898).

**Equilibrium characterization.** Unlike Amihud and Mendelson (1986), where
investors sort on exogenous bid-ask spreads, here bid-ask spreads are endogenous
and investors sort on maturity and default probability. Under Assumption 1 and
$$\epsilon_{\xi} = 0$$, Proposition 1 (p. 895) establishes positive assortative
matching: $$\theta'(\delta+d) > 0$$.
Proposition 2 (p. 896) characterizes the full equilibrium as a system of ODEs
in $$\bar{\delta} = \delta + d$$ (eq. 24, p. 896):

$$
\begin{cases}
\lambda'(\bar{\delta}) = \dfrac{\lambda \left[\frac{1-\gamma}{\alpha_b}\,\lambda - \frac{\gamma}{\alpha_b}\,\theta'\!\left(\frac{1}{\alpha_b} + (1-\gamma)\!\left(\frac{1}{\rho} + \frac{1}{\alpha_b}\right)\right)\right]}{\gamma(1-\gamma)(\rho + \bar{\delta} + \theta)} \\[10pt]
\theta'(\bar{\delta}) = \dfrac{\bar{\delta} + \rho\,\alpha_s / (\rho + \mu_s + \bar{\delta} + d)}{m_I\, f(\theta)}
\end{cases} \tag{24}
$$

with boundary conditions $$\theta(\bar{\delta} + \bar{d}) = \bar{\theta}$$ and
$$\theta(\bar{\delta}_{\text{end}} + \bar{d}) = \bar{\theta}_{\text{end}}$$.

**Interest rate and credit spread sensitivity.** In equilibrium the interest
rate for bond $$j$$ is (Lemma 2, eq. 26, p. 899):

$$
r = \frac{\rho}{\rho + \delta + \theta} + \frac{d(1-s)}{\rho + \delta + \theta} + \frac{\theta\,\Delta}{\rho + \delta + \theta} + [\text{exo and endo bid-ask spread terms}] \tag{26}
$$

The sensitivity of the interest rate to exogenous bid-ask spread changes is
(Corollary 2, eq. 27, p. 900):

$$
\frac{dr}{d\epsilon_{\xi}} = \frac{\dfrac{\theta}{\rho+\delta+\theta}\cdot\dfrac{\gamma}{\lambda} - \dfrac{\rho+\delta}{\rho}\cdot(1-\gamma)}{\left(\dfrac{1}{\mu_s} + \dfrac{1-\kappa}{\rho+\delta+\theta}\right)\dfrac{\gamma}{\lambda} + \left(\dfrac{1-\kappa}{\rho} + \dfrac{1}{\mu_b}\right)(1-\gamma)} \tag{27}
$$

When $$r_f$$ decreases (more short-term investors enter), $$dr/d\epsilon_{\xi}$$ increases
for all bonds.

## Method

The paper applies two complementary methods: reduced-form panel regressions
with an instrumental variables (IV) design, and a calibrated structural model
solved via ordinary differential equations (ODEs).

**Reduced-form evidence.** The baseline cross-sectional regression is
(eq. 1, p. 880):

$$
\text{CS}_{i,t} = \alpha_t + \beta_t \cdot \text{BA}_{i,t} + \gamma_t^T X_{i,t} + \epsilon_{i,t} \tag{1}
$$

Run quarter by quarter; $$X_{i,t}$$ includes bond characteristics (bond age,
time-to-maturity, coupon, offering amount, rating) and firm characteristics
(leverage, size, profitability, equity volatility, total asset value) and
industry fixed effects. Standard errors are clustered at the firm level.

The liquidity component is defined following Dick-Nielsen, Feldhutter, and
Lando (2012) as (eq. 2, p. 881):

$$
\text{liquidity\_component}_{i,t} = \frac{\beta_t \times \text{BA}_{i,t}}{\text{CS}_{i,t}} \tag{2}
$$

**Investor composition measure.** Fund-level net transaction rate (eq. 3, p. 882):

$$
\text{net\_transaction}_{j,t} = \frac{\left|\sum_i \text{holding}_{i,j,t} - \sum_i \text{holding}_{i,j,t-1}\right|}{\sum_i \text{holding}_{i,j,t-1}} \tag{3}
$$

Smoothed over four quarters (eq. 4, p. 883):

$$
\text{NT}_{j,t} = \frac{1}{4} \sum_{t'=1}^{4} \text{net\_transaction}_{j,t-t'} \tag{4}
$$

Bond-level investor composition (eq. 5, p. 883):

$$
\text{investor\_comp}_{i,t} = \frac{\sum_j \text{holding}_{i,j,t} \times \text{NT}_{j,t}}{\sum_j \text{holding}_{i,j,t}} \tag{5}
$$

**Instrumental variables design.** Exploiting the 10-year time-to-maturity
threshold in intermediate-term bond fund mandates documented by Bai, Li, and
Manela (2022) as a sharp discontinuity, using the two-stage specification
(eqs. 6-7, pp. 885-886):

First stage: $$\text{Inv\_Comp}_{i,t} = \alpha + \beta_1 \cdot \mathbf{1}_{\text{ttm}>10} + \gamma^T X_{i,t} + \epsilon_{i,t}$$

Second stage: $$Y_{i,t} = \alpha + \beta_1 \cdot \widehat{\text{Inv\_Comp}}_{i,t} + \gamma^T X_{i,t} + \epsilon_{i,t}$$

with optimal bandwidth around the 10-year cutoff (Calonico, Cattaneo, and
Titiunik 2014). Also instruments the interaction term $$\text{Inv\_Comp} \times \text{BA}$$ with
$$\mathbf{1}_{\text{ttm}>10} \times \text{BA}$$ for the credit-spread loading regression (eq. 8, p. 886):

$$
\text{CS}_{i,t} = \alpha + \beta_1 \cdot \text{BA}_{i,t} + \beta_2 \cdot \widehat{\text{Inv\_Comp}}_{i,t}
           + \beta_3 \cdot \text{BA}_{i,t} \times \widehat{\text{Inv\_Comp}}_{i,t} + \gamma^T X_{i,t} + \epsilon_{i,t} \tag{8}
$$

Standard errors are clustered by industry and time; estimation uses demeaned
bid-ask spreads for interpretability.

**Calibration.** The structural model is calibrated to annual-level moments
for bonds with 2 to 15 years to maturity ($$\bar{\delta} \in [1/15,\, 0.5]$$).
Parameters are set to match six moments in 2005 and 2019: bond turnover rates,
the loading coefficient on bid-ask spreads, average credit spreads, average
bid-ask spreads, and the ratio of residualized standard deviations (Table III,
p. 903). The model is then simulated for $$N = 1{,}000$$ bonds and the following
regression run on simulated data to obtain the model-implied $$\beta$$:

$$
\text{CS}_i = \beta_0 + \beta \cdot \text{BA}_i + \beta_M \cdot \delta_i + \epsilon_i \tag{32}
$$

The mechanism decomposition separates the loading into an exogenous component
(eq. 34, p. 904):

$$
\text{CS}_i = \beta_0 + \beta_{\text{exo}} \cdot \epsilon_{\xi,i} + \beta_M \cdot \delta_i + \epsilon_i \tag{34}
$$

and an endogenous component (eq. 35, p. 905):

$$
\text{CS}_i = \beta_0 + \beta_{\text{endo}} \cdot \text{BA}_{\text{endo},i} + \beta_M \cdot \delta_i + \epsilon_i \tag{35}
$$

where $$\text{BA}_{\text{endo},i}$$ is the endogenous part of bid-ask spreads with $$\epsilon_{\xi}$$ set
to zero.

## Empirical specifications

All regressions are quarterly, 2005Q2 to 2019Q2, on U.S. corporate debentures
with fixed coupon, nonconvertible, nonputtable, and nonexchangeable. The primary
sample excludes bonds that have more than 95% of days with no trading. Bonds
with ratings below CCC- are excluded.

**Aggregate trend (R1, R2).** Regression (1) run quarter by quarter on all
bonds using WRDS bid-ask spreads; standard errors clustered at the firm level
(Figure 1, p. 880). The aggregate loading $$\beta_t$$ and the liquidity component
median are the key time-series outcomes.

**Cross-sectional analysis, investor composition effects (R3).** Table I
(p. 886) reports the first-stage and second-stage IV regressions using the
10-year maturity threshold as an instrument, with bandwidth 1.699 for the full
sample and 2.282 for the subsample with maturity > 10 years at issuance.
Controls include time-to-maturity, age, coupon, log amount outstanding, firm
total assets, fraction of long-term debt, leverage ratio, profitability, equity
price volatility, slope and level of Treasury yields, and rating x industry x
date fixed effects.

**Credit spreads and bid-ask spreads interaction (R4).** Table II (p. 887)
reports four specifications of regression (8): full sample OLS, full sample IV,
subsample OLS, subsample IV. Investor composition is instrumented with
$$\mathbf{1}_{\text{ttm}>10}$$; the interaction $$\text{Inv\_Comp} \times \text{BA}$$ is instrumented with $$\mathbf{1}_{\text{ttm}>10} \times \text{BA}$$.
Standard errors are clustered by industry and time. The key finding is the
significant positive coefficient on $$\text{Inv\_Comp} \times \text{Bid\_Ask}$$ (48.65\*\*\* OLS full
sample; 55.86\*\*\* IV subsample).

**Structural equilibrium and calibration (R5-R7).** The model is simulated in
2005 and 2019 with calibrated parameters (Table IV, p. 903). The mechanism
decomposition (Table V, p. 905) shows $$\beta_{\text{exo}} = 0.27/0.35$$ (2005/2019) vs
empirical $$\beta = 0.54/1.2$$, confirming the indirect channel (via seller-buyer
ratio correlation) must explain the remainder. The indirect loading
$$\beta_{\text{endo}} = 0.65/1.7$$ matches the total pattern.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| TRACE (Enhanced, FINRA) | Corporate bond transaction prices and volumes for bid-ask spread calculation, number of trades, bond turnover; filtered following Dick-Nielsen (2014) | [TRACE](/wiki/commercial/trace/) (licensed) |
| WRDS Bond Return database + Mergent FISD | Bond characteristics: maturity, coupon, rating, offering amount, issuance date; credit spreads calculated from reported yields minus matched Treasury yield | [WRDS](/wiki/commercial/wrds/) (licensed) |
| CRSP equity returns | Equity price volatility for bond issuers | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| Compustat annual fundamentals | Firm characteristics: leverage, size, profitability, total asset value, fraction of long-term debt | [WRDS / Compustat](/wiki/commercial/wrds/) (licensed) |
| Lipper eMaxx (Thomson Reuters) | Quarterly investor holdings of corporate bonds at the CUSIP level for insurance companies, mutual funds, ETFs, and annuities; covers 40-50% of total bonds outstanding | [no page yet] |
| Flow of Funds (Federal Reserve) | Aggregate corporate and foreign bond holdings by investor type for benchmarking eMaxx coverage | [no page yet] |
| Gurkaynak, Sack & Wright (2007) Treasury yield curve | Used to calculate credit spreads by subtracting matched Treasury yields | [no page yet] |

Sample: 2005Q2 to 2019Q2 (58 quarters, quarterly). Primary analysis covers
U.S. corporate debentures with fixed coupon; 15,256 unique bonds, 3,217 unique
firms (cross-sectional sample with eMaxx coverage > 20%).

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.70024) if you are:
investigating the corporate bond liquidity premium and its time-series variation;
building or calibrating a model of OTC bond markets with heterogeneous investors;
studying how the growth of mutual funds and ETFs affects bond market fragility;
evaluating the interaction between investor-side and dealer-side regulatory changes;
or replicating the regression discontinuity design around the 10-year maturity
threshold. The locators above point to the exact tables and figures.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 81(2), April 2026. This
distillation was extracted by an LLM on 2026-06-01 and is **not human-verified
or independently reproduced**. The article is paywalled; redistribution is
extract-only. Access the original at [https://doi.org/10.1111/jofi.70024](https://doi.org/10.1111/jofi.70024).

> Li, Jian, and Haiyue Yu. "Investor Composition and the Liquidity Component
> in the U.S. Corporate Bond Market." *The Journal of Finance* 81, no. 2
> (April 2026): 871-922. DOI: 10.1111/jofi.70024.
> (c) 2026 the American Finance Association.


==============================================================================
# Second Chance: Di Maggio, Kalda & Yao (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2026/maggio-student-debt-second-chance-2026/
# Distilled: exploiting plausibly random private student debt discharge (National Collegiate paperwork loss), the paper finds debt relief reduces other debt balances and delinquency rates, and raises geographic and job mobility and income for defaulted borrowers. J. Finance 2026, CC BY 4.0. Eight core results with source locators, datasets used, the identification strategy, and the estimating specifications with equations.
# Tags: paper-summary, household-finance, student-debt, debt-relief, consumer-credit, labor-markets, panel-regression, difference-in-differences, open-access, cc-by, peer-reviewed, unreplicated, data:equifax-credit, data:lexisnexis-court
==============================================================================

**What this is.** The paper's core results, datasets, and theory: enough to
know what it found without reading all 44 pages. To replicate or extend it,
read the full source at the
[original](https://doi.org/10.1111/jofi.70002).

## TL;DR

Using hand-collected court-filing data matched to Equifax credit bureau
records, the paper exploits a plausibly random debt discharge shock: National
Collegiate Student Loan Trusts lost paperwork for thousands of borrowers,
causing courts to dismiss collection lawsuits and discharge the debt. Treated
borrowers (debt discharged) are compared with similar defaulted borrowers
whose cases were not dismissed. Debt relief leads to lower balances and
delinquency rates on other accounts, higher geographic and job mobility, and
approximately $3,000 more income over three years. Both the treated group
(via debt relief) and the control group (via wage garnishment and collections)
contribute to these differential outcomes.

The paper builds on Dobbie and Song (2015), a benchmark for debt-relief
effects on credit and labor outcomes via chapter 13 bankruptcy, but examines
private student debt discharge outside bankruptcy. It also builds on Dobbie
and Song (2020), a targeted credit-card debt relief experiment, and finds
faster and broader effects for student debt discharge.

## Core results

Magnitudes and significance are as reported; `**`/`***` = 5%/1%. Locators
point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Debt relief **reduces student loan balances** by approximately $7,400 and credit scores improve by 6.8 points | Table II, p. 524 | SL balance: -$7,404.56\*\*\* (SE 340.5); credit score: +6.81\*\*\* (SE 1.23); N=1,283,639 |
| R2 | **Total non-student debt balances fall** by more than $4,500 for treated borrowers | Table III, p. 528 | -$4,600.88\*\*\* (SE 387.55); credit card -$618.76\*\*\*; mortgage -$1,564.60\*\*\*; auto loan not significant |
| R3 | Treated borrowers **deleverage actively**: credit card utilization falls, credit limits decline, new account opening falls, repayments rise | Table IV, p. 529 | Utilization -0.023\*\*\* (SE 0.007); credit limit -$971.82\*\*; new accounts -0.002\*\*\*; monthly payment +$13.57\*\* |
| R4 | **Delinquency rates on other accounts fall by 3 percentage points** (12% relative decline); bankruptcy and foreclosure also decline | Table V, p. 531 | Total delinquency -0.03\*\*\* (SE 0.003); bankruptcy -0.04\*\*\* pp; foreclosure -0.03\*\*\* pp; medical default -0.1\*\*\* pp |
| R5 | **Geographic mobility rises** 0.3 pp; **job change probability rises** 0.3 pp; **industry switching rises** 0.3 pp | Table VI, p. 532 | Mobility dummy +0.003\*\*\* (SE 0.001); job change +0.003\*\* (SE 0.001); new industry +0.003\* (SE 0.002) |
| R6 | **Monthly income rises by approximately $80** (1 pp higher income growth); cumulative gain over three years approximately $3,000 | Table VI, p. 532 | Income (level): +$79.98\*\*\* (SE 31.99); % change in income: +0.01\*\* (SE 0.004) |
| R7 | **Control group drives credit outcomes** via liquidity constraints (wage garnishment): estimates are larger when control group is more likely to be garnished | Table VII, p. 537 | Panel A (adjudicated controls): total balance -$5,099.56\*\*\* vs baseline -$4,600.88\*\*\*; delinquency -0.029\*\*\* |
| R8 | **Treated group drives labor outcomes** via debt overhang: fraction of variable pay rises +3.1 pp and hours worked rise by 1.20/week for treated borrowers post-discharge | Table VIII, p. 539 | Variable pay share: +0.031\*\*\* (SE 0.015) all jobs; +0.042\*\*\* (SE 0.018) no-job-change sample; hours: +1.20\*\* (SE 0.510) for hourly workers |

**Overall (paper's conclusion).** A $7,400 debt-relief shock translates into
a $4,600 reduction in non-student indebtedness, a $3,000 income gain over
three years, and a $4,700 decline in the amount of debt in delinquency.
Effects persist for at least two years and are not transitory.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| LexisNexis court filings (all U.S. civil courts, 2010–2017) | Hand-collected National Collegiate lawsuit data: borrower identity, court, filing date, outcome | no page yet |
| Equifax credit bureau (matched anonymized panel) | Monthly credit outcomes: balances, delinquency, credit score, account types for treatment/control borrowers | no page yet |
| Equifax employment and income verification (payroll data, 5,000+ U.S. firms) | Labor market outcomes: monthly gross earnings, hours, job type, tenure, employer | no page yet |

Sample: 9,878 treated borrowers; 6,388 control borrowers; 1,283,639 borrower-month observations in the credit data. Lawsuits cover 2010–2017.

## Theory / model

The paper has no structural model; it is a reduced-form causal study. It
tests three hypotheses derived from debt-overhang and liquidity-constraint
models (pp. 519-521):

- **Hypothesis 1** (p. 520): Student debt discharge leads to a relative
  decline in other debt balances. Mechanism: discharge removes a delinquent
  account from the credit report, relaxes credit constraints, and protects
  future wages from garnishment, reducing the incentive to borrow further.
  Building on Herkenhoff, Phillips, and Cohen-Cole (2021), who show credit
  access improving self-employment, the paper tests whether student debt
  discharge relaxes credit constraints.
- **Hypothesis 2** (p. 520): Student debt discharge leads to a relative
  decline in the likelihood of default and experiencing distress. Mechanism:
  improved financial condition reduces delinquency on all accounts and
  provides positive spillovers to other creditors.
- **Hypothesis 3** (p. 521): Student debt discharge leads to a relative
  increase in mobility and income. Mechanism: debt overhang (analogous to
  the corporate finance problem) and liquidity constraints distort labor
  supply decisions; discharge removes these frictions. Melzer (2017) documents
  debt overhang reducing homeowner investment; here the same mechanism is
  tested for student-debt borrowers in the labor market.

The paper formalizes the borrower's budget constraint (p. 517) to clarify
what is and is not observable:

$$
c_{i,t} = (1 - s^u_{i,t}) \cdot \gamma_{i,t-1}
         + (1 - G_i(t_d))(1 - s_{i,t})(Y^o_{i,t} + Y^h_{i,t} + Y^f_{i,t})
         - \sum_{l=1}^{A} (1 + r_{i,l}) \cdot b_{i,t,l}
$$

where $$c_{i,t}$$ is consumption of individual $$i$$ in period $$t$$;
$$\gamma_{i,t-1}$$ is wealth at $$t-1$$; $$G_i(t_d)$$ is a piecewise garnishment
function equal to rate $$g$$ if $$i$$ is in the control group after default date
$$t_d$$, zero otherwise; $$Y^o_{i,t}$$ is observed wage income; $$Y^h_{i,t}$$ is
hidden wage income; $$Y^f_{i,t}$$ is financial income; $$s^u_{i,t}$$ is the
savings rate out of assets; $$s_{i,t}$$ is the savings rate out of income;
$$r_{i,l}$$ and $$b_{i,t,l}$$ are interest rates and balances for loans of type
$$l$$ (p. 517-518). The paper explicitly notes it cannot observe consumption,
savings, or hidden income, which limits the set of testable predictions.

**Identification.** The source of variation is the plausibly random loss of
paperwork by National Collegiate Student Loan Trusts. Courts dismissed
collection lawsuits against borrowers whose chain of title could not be
proved. This documentation loss is argued to be orthogonal to borrower
characteristics (p. 516): the same trust held loans it could and could not
prove, and the distinction was driven by clerical errors, not by borrower
type. The paper verifies balance on pre-treatment observables and the absence
of pre-trends across all outcomes (Figure 2, pp. 525-527).

## Method

The empirical strategy uses a difference-in-differences (DiD) estimator
building on `difference-in-differences` and `panel-fe`.

**Treatment and control groups.** Treated borrowers are those whose
National Collegiate cases were dismissed and debt discharged. Control
borrowers were sued by the same trusts but their debt was not discharged
during the sample period (either the trust did not lose the paperwork or
the case was not adjudicated by end of sample). Both groups defaulted on
loans owned by the same trust and were subject to collection by the same
agency, making the groups likely to be similar on unobservables (p. 518).

**Standard-error treatment.** Standard errors are clustered at the zip code
level throughout, allowing within-neighborhood error correlation across
borrowers (p. 519). Robustness checks use individual-level clustering and
double clustering by zip code and calendar month (Internet Appendix Table
IA.I).

**Treated-only robustness.** To address concerns that the control group
is confounded by wage garnishment, the paper re-estimates the main
specifications using only treated borrowers, exploiting the staggered
timing of discharges as the source of variation (Table IX, p. 541).
Results are qualitatively similar and economically meaningful, supporting
the validity of the treated group's contribution.

## Empirical specifications

**Baseline DiD (equation 1, p. 518).** The main estimating equation is:

$$
\text{Outcome}_{i,j,t} = \alpha + \beta \cdot (\text{DebtRelief}_i \times \text{Post}_t)
                + \mu_i + \gamma_{j \times t} + \epsilon_{i,j,t}
\tag{1}
$$

- $$\text{Outcome}_{i,j,t}$$ is the outcome variable for borrower $$i$$, filing year $$j$$, calendar year-month $$t$$.
- $$\text{DebtRelief}_i$$ = 1 for treated borrowers (debt discharged), 0 for control.
- $$\text{Post}_t$$ = 1 after debt discharge and 0 before.
- $$\mu_i$$ are individual fixed effects.
- $$\gamma_{j \times t}$$ are filing-year by calendar-year-month fixed effects (ensuring treated and control borrowers are compared within the same filing-year cohort and calendar time).
- $$\epsilon_{i,j,t}$$ is the error term.
- Standard errors clustered at zip code level. N = 1,283,639 borrower-month observations (credit data).

**Dynamic event-study (equation 2, p. 519).** To assess pre-trends and
persistence:

$$
\text{Outcome}_{i,j,t} = \alpha + \sum_{\tau=-5}^{9} \beta_\tau \cdot (\text{DebtRelief}_i \times \text{Post}_\tau)
                + \mu_i + \gamma_{j \times t} + \epsilon_{i,j,t}
\tag{2}
$$

- $$\tau$$ indexes event-quarters relative to discharge.
- $$\tau = -5$$ captures all months before five quarters pre-treatment and $$\tau = 9$$ captures all months nine or more quarters post-treatment.
- Coefficients $$\beta_\tau$$ are plotted with confidence intervals (Figure 2, pp. 525-526).
- Pre-trend coefficients are indistinguishable from zero across all outcomes; post-discharge coefficients diverge persistently for at least two years.

**Outcome variable groups and samples:**
- Credit outcomes (R1-R4): full sample, N = 1,283,639; outcomes include
  SL balance, credit score, total non-SL balance, credit card balance,
  mortgage balance, credit card utilization, credit limit, account openings,
  monthly payments, delinquency rates, bankruptcy, foreclosure, medical
  default.
- Labor outcomes (R5-R6): subsample with Equifax employment data;
  N = 211,716 (job change), 197,874 (new industry), 106,580 (income level),
  91,230 (income growth).
- Mechanism tests (R7): subsample with adjudicated control-group cases
  (Table VII Panel A, N = 1,028,559); and control borrowers with prior
  collections on file (Table VII Panel B, N = 838,295).
- Wage composition (R8): subsample of workers in Equifax payroll data
  (Table VIII, N = 39,459 all jobs; 28,653 no-job-change; 22,128 hourly).

The same two-way FE structure (individual + filing-year x YM) applies in
all specifications; the only variation across tables is the outcome variable
and the sample restriction.

## When to read the full paper

Use the [original DOI](https://doi.org/10.1111/jofi.70002) if you are:
replicating (code in Supporting Information); extending the design to other
debt types or populations; examining mechanism tests (wage composition,
above/below median debt relief, liquidity-constraint heterogeneity); or
auditing a specific coefficient. The locators above point you to the exact
table. For "what did this paper find," the table above is the intended
default.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 81(1). This distillation was
extracted by an LLM on 2026-05-31 and is **not human-verified or
independently reproduced**. The CC BY 4.0 licence (confirmed via Crossref)
permits reproduction with attribution. No verbatim PDF is hosted in this
batch; the canonical source is the publisher DOI.

> **Attribution (CC BY 4.0).** Di Maggio, Marco, Ankit Kalda, and Vincent
> Yao. "Second Chance: Life with Less Student Debt." *The Journal of
> Finance* 81, no. 1 (February 2026): 507–550. DOI: 10.1111/jofi.70002.
> © 2025 The Author(s). Licensed under
> [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Deposit Inflows and Outflows in Failing Banks: Martin, Puri & Ufier (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2026/martin-deposit-flows-failing-banks-2026/
# Distilled: Using confidential daily account-level FDIC data from a failing U.S. bank, this paper shows that gross deposit inflows are first-order in a distressed bank's funding dynamics: deposit insurance stabilizes outflows while simultaneously enabling large insured deposit inflows that nearly offset departing uninsured funds. J. Finance 2026, U.S. Government public domain. Ten core results with source locators, datasets used, and the estimating equations.
# Tags: paper-summary, banking, deposit-insurance, bank-runs, financial-stability
==============================================================================

**What this is.** The paper's core results, estimating equations, and datasets:
enough to know what the paper found and how, without reading the 43-page source.
To replicate or extend, read the full article at [doi.org/10.1111/jofi.70007](https://doi.org/10.1111/jofi.70007).

## TL;DR

Using confidential daily, account-level deposit records collected by the FDIC from a
single failing U.S. bank (roughly $2 billion in assets, failed during the
financial crisis), the paper documents that gross deposit *inflows* are as
important as outflows for understanding bank funding under distress. Granja,
Matvos, and Seru (2017) document the failed banks of the Great Recession and
the resulting FDIC Deposit Insurance Fund costs that motivate this question. Deposit
insurance (regular FDIC coverage, the FDIC-extended $250,000 limit, and the
temporary TAG program) significantly reduces outflows from insured accounts.
Simultaneously, the distressed bank raised deposit rates to the 95th percentile
of the industry and attracted nearly $400 million in new insured term deposits,
mostly from out-of-state credit unions and banks via internet listing services,
replacing roughly one-third of its departing deposit base. These inflows are
generalizable: a panel of over 2,000 U.S. banks facing regulatory enforcement
actions shows similar shifts in deposit composition toward small and medium-term
insured deposits and away from brokered and large-uninsured deposits. The
central policy implication is that insured deposit inflows substantially weaken
the depositor discipline otherwise exerted by uninsured outflows.

## Core results

Magnitudes and significance are as reported in the paper. `***`/`**`/`*` = 1%/5%/10%.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | In the Formal (bank-specific distress) period, uninsured transaction accounts liquidate 18 pp faster than insured accounts | Table III, col. (4), p. 656 | Uninsured coeff = 0.183\*\*\* (t = 8.45); compared to near-zero baseline in Placebo period (0.0249\*) |
| R2 | TAG temporary unlimited insurance reduces transaction-account liquidation as much as ordinary deposit insurance: the two coefficients are not statistically different (p-value 0.67 for equality) | Table III, col. (3), p. 656; text p. 659 | TAG/DFA Eligible coeff = -0.0944\*\* (t = -2.00) in Postcrisis; difference from Uninsured coeff not significant |
| R3 | Uninsured term deposits are 14.7 pp more likely to liquidate in the Formal period; brokered/placed term deposits are 55 pp more likely to liquidate | Table VI, col. (4), p. 662 | Uninsured coeff = 0.147\*\*\* (t = 2.92); Brokered/Placed coeff = 0.550\*\*\* (t = 25.76) |
| R4 | Uninsured transaction depositors draw balances well below the insurance limit under stress: in the Formal period, 20.8% liquidate to less than $1, far above the 5-7% in earlier periods | Table VIII, Panel B, p. 665 | Formal period: 20.79% in < $1 bin; Placebo: 4.98% |
| R5 | The failed bank attracted nearly $400 million in new insured term deposits in its last year, roughly one-third of its deposit base, mostly in the 90 days before failure | Figure 3, p. 666; text p. 666 | ~$400M new insured term deposits; nearly all from institutional depositors (81% of 2,600 new accounts in Formal period) |
| R6 | Higher deposit rate spreads are associated with more new deposit inflows: estimated interest elasticity of demand of 0.61, in line with Egan, Hortacsu and Matvos (2017) | Tables X-XI, p. 670-671; text p. 673 | Rate Spread to Market coeff = 5.549\*\*\* (t = 2.71) in Table X col. (3); elasticity of demand calculated at 0.61 |
| R7 | The Formal period is the only period where new deposit inflows are significantly elevated after controlling for macroeconomic conditions and interest rates, confirming bank-specific (not macro) drivers | Table XI, col. (4), p. 671 | Formal period dummy coeff = 5.124\*\* (t = 2.32) on log new deposits; no other period significant after controls |
| R8 | Banks under regulatory enforcement actions in the broad U.S. panel reduce brokered deposits by 1.24 pp and increase listing-service deposits by 0.826 pp; small-term ($<$100k) deposits rise by 0.364 pp, large-term ($>$250k) fall by 0.830 pp | Table XII, p. 679 | Under Reg. Action coeff: Brokered = -1.24\*\*\* (t = -22.72); Listed = 0.826\*\*\* (t = 11.40); Sm. Term = 0.364\*\*\* (t = 5.21); Lg. Term = -0.830\*\*\* (t = -12.10) |
| R9 | The shift toward listing-service deposits begins before the enforcement action and ramps up sharply after; brokered deposits decline sharply at enforcement, consistent with regulatory restrictions | Table XIII / Figure 5, pp. 678-680 | Dynamic coefficients on Listed deposits turn positive at t+2 (0.477**), reaching ~1.2 pp at t>=5 (1.21***); Brokered coefficients: ~1.2 pp before enforcement (tau-4 = 1.17***), -3.85*** by t>=5 |
| R10 | Large banks (assets > $5B) under enforcement action raise 12-month CD spreads by a statistically significant 41 bps, paying on average ~75 bps above the FDIC national average, near the regulatory rate cap | Table XV, p. 683 | Under Reg. Action coeff = 0.4121\*\*\* (t = 4.48); constant = 0.3547\*\*\* |

**Overall (paper's conclusion).** Gross deposit inflows to a failing bank are
first-order in magnitude: deposit insurance both reduces outflows and enables
insured inflows that offset them. Focusing only on net outflows misses this
mechanism and overstates the stabilizing power of depositor discipline.
Temporary crisis-era guarantees such as TAG are as effective as standard
deposit insurance. These results hold for the single failed bank in the micro
data and generalize to over 2,000 U.S. banks under regulatory enforcement
action (pp. 683-684).

## Theory / model

The paper has no formal structural model. It builds on the framework of Egan,
Hortacsu, and Matvos (2017), who model a market for insured and uninsured
deposits where distressed banks raise deposit rates to attract insured funds
while uninsured funds flee. It also builds on Diamond and Dybvig (1983), the
foundational bank-run theory that motivates the role of deposit insurance in
preventing runs. The paper tests this framework empirically with
granular data. The core hypotheses are:

1. **Outflow hypothesis.** Deposit insurance (regular, temporary TAG/DFA) reduces
   the probability that a covered account liquidates, because insured depositors
   bear no principal risk from bank failure.
2. **Inflow hypothesis.** Even as uninsured funds leave, distressed banks raise
   deposit rates above market to attract new insured deposits, especially from
   rate-sensitive institutional depositors using internet listing services. The
   inflows are concentrated just below the insurance limit.
3. **Depositor-discipline hypothesis.** If insured inflows offset uninsured
   outflows, gross outflows overstate discipline; net outflows (as reported in
   Call Reports) understate it. Understanding bank runs requires decomposing
   gross flows.

The evidence on insured inflows informs the design of optimal deposit
insurance, as modeled by Davila and Goldstein (2023).

**Identification.** The paper exploits two quasi-natural experiments:

- The October 2008 increase in FDIC insurance from $100,000 to $250,000 and the
  simultaneous introduction of the unlimited-guarantee TAG program provide
  sharp policy changes. An event study around the insurance limit change (a
  36-day symmetric window) shows the daily liquidation probability for term
  deposits newly insured under the $250,000 limit falls from 0.0806 to 0.0512
  (p. 664).
- For generalization, the paper uses the quasi-random timing of formal
  regulatory enforcement actions against a large panel of U.S. banks (2000-2016),
  comparing treated banks to untreated controls before and after enforcement,
  and a propensity-score-matched specification for robustness.

## Method

The paper applies linear probability models (LPM), probit, Cox proportional
hazard models, OLS time-series regressions, and propensity-score matching. It
does not propose a new method; the contribution is entirely empirical.

The account-level approach extends Iyer and Puri (2012), an earlier
account-level bank-run study, by using finer daily data and separating
inflows from outflows.

**Outflow regressions (Section II).** The baseline is a cross-sectional LPM for
each of four time periods (Placebo, Precrisis, Postcrisis, Formal), separately
for transaction and term accounts. This builds on `panel-regression` and
`linear-probability-model` primitives. Robustness uses `probit-regression` and
Cox proportional hazard models (Tables IV, VII), giving consistent results.

**Inflow regressions (Section III).** The characteristics of new versus extant
depositors are compared via OLS (equation 2, p. 667). The daily volume of new
deposits is explained with OLS and Newey-West standard errors (Tables X-XI,
equations 3, p. 669).

**Generalization panel (Section IV).** The treatment of a formal enforcement
action is identified by two OLS specifications (equations 4-5, pp. 676-677) and
a propensity-score-matched comparison (equation 6 for large banks, p. 682).
These build on `panel-regression`, `event-study`, and `matching` primitives.

## Empirical specifications

### Outflow regressions (Core results R1-R4)

**Baseline LPM (equation 1, p. 653):**

$$
\mathbf{1}(\text{Liquidation}_i) = \alpha + \beta \times \text{Controls}_i + \delta_b + \epsilon_i
$$

Where subscript $$i$$ indexes deposit accounts; $$\delta_b$$ is a set of branch
fixed effects. $$\text{Liquidation}_i$$ = 1 if the account balance falls by 75% or more
relative to start-of-period and stays at or below 25% of the starting balance
for at least 61 days. $$\text{Controls}_i$$ includes: Uninsured dummy, TAG/DFA Eligible
dummy, Checking dummy, Direct Deposit dummy, Log(Age), Prior Transactions,
$$\text{Prior Transactions}^2$$, Institutional-Any dummy, Trust dummy, Brokered/Placed
dummy (term only), Log(Days to Maturity) (term only). Standard errors are
asymptotically normal. Run separately for each of four periods and for
transaction vs. term deposits (Tables III, VI). Alternative specifications use
probit (marginal effects) and Cox proportional hazard models (Tables IV, VII).

Interest rate controls added in Table V (Postcrisis and Formal periods only):
rate spread to market and past-month fees are added to equation 1 for both
transaction and term deposits.

### New-depositor characteristics (equation 2, p. 667)

$$
\mathbf{1}(\text{Characteristic}_{i,t}) = \alpha + \sum_{t=1}^{5} \beta_t \times \text{Time Period Dummy}_t \times \mathbf{1}(\text{Extant Depositor}_{i,t}) + \epsilon_{i,t}
$$

OLS, N = 188,834, six separate regressions (one per account characteristic as
dependent variable). Omitted category is extant depositors in the Placebo
period.

### Time-series inflow regressions (equations 3, p. 669)

$$
y_t = \alpha + \beta \times \text{Time Period Dummy}_t + \gamma \times X_t + \epsilon_t
$$

$$y_t$$ is either the share of deposits that are new that day (Table X) or the
log of dollar volume of new deposits (Table XI). $$X_t$$ includes Log(VIX), GDP
Growth, Housing Starts, Daily S&P 500 Return, AR(1) term, OFR Financial Stress
Index, Rate Spread to Market (dollar-weighted average). OLS with Newey-West
standard errors (lag length 9, Newey-West rule of thumb).

### Generalization: panel with regulatory action dummy (equations 4-5, pp. 676-677)

$$
y_{j,t} = \alpha + \beta \times \text{Under Reg. Action}_{j,t} + \gamma \times X_{j,t} + \delta_j + \zeta_t + \epsilon_{j,t} \tag{4}
$$

$$
y_{j,t} = \alpha + \sum_{i=-4}^{\geq 5} \beta_i \times \text{Under Reg. Action}_{j,t=\tau+i} + \gamma \times X_{j,t} + \delta_j + \zeta_t + \epsilon_{j,t} \tag{5}
$$

$$y_{j,t}$$ is a funding-share outcome for bank $$j$$ at quarter $$t$$. Bank and
quarter fixed effects ($$\delta_j$$, $$\zeta_t$$). $$X_{j,t}$$ includes NPL/Assets,
one-year asset growth rate, log assets, deposits/assets, term deposits/assets.
Panel covers ~10,000 U.S. banks, 2000-2016, quarterly. Equation (4) produces
Table XII; equation (5) produces Table XIII and Figure 5. A propensity-score-
matched specification (Table XIV) and a pooled OLS for large banks (equation 6,
p. 682, Table XV) round out Section IV.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| FDIC confidential account-level deposit microdata | Daily account balances and transactions for the one failed bank (early 2006 to failure); source of all Section II-III results | No page yet |
| FDIC Call Reports (public) | Panel of ~10,000 U.S. banks for the generalization analysis (Section IV); funding-share outcomes and bank-level controls | No page yet |
| FDIC confidential supervisory data | Identifies enforcement actions (C&D orders, less-than-well-capitalized status) and brokered-deposit waivers for the generalization panel | No page yet |
| RateWatch deposit rate surveys | 12-month CD rate spreads for large banks under enforcement action (Section IV.B); branch-level weekly survey data | No page yet |

Sample (single bank): daily, from early 2006 to bank failure (approximately
late 2010-early 2011). Generalization panel: quarterly, 2000-2016 (brokered
deposits); 2011-2016 (listing-service deposits); 2010-2016 (term deposits by
size). Total observations in panel regressions: up to 554,180 bank-quarters.

## When to read the full paper

Read the [original](https://doi.org/10.1111/jofi.70007) if you are:
studying bank-run dynamics and need account-level evidence on the gross
inflow-outflow decomposition; designing deposit insurance policy (insurance
limit levels, temporary guarantees, rate caps); testing or calibrating models
of depositor behavior under distress (the EHM 2017 framework empirically);
or analyzing the stability of different deposit categories (checking vs.
savings vs. CDs, insured vs. uninsured, brokered/placed vs. listing-service).
The Internet Appendix (Appendix S1, referenced p. 685) contains additional
robustness tables cited throughout.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 81(2), April 2026. This article
is a U.S. Government work and is in the public domain in the USA (stated on
PDF p. 643). This distillation was extracted by an LLM on 2026-06-01 and is
**not human-verified or independently reproduced**. The underlying data are
confidential FDIC supervisory records; replication requires FDIC data access.

> Martin, Christopher, Manju Puri, and Alexander Ufier. "Deposit Inflows and
> Outflows in Failing Banks: The Role of Deposit Insurance." *The Journal of
> Finance* 81, no. 2 (April 2026): 643-685. DOI: 10.1111/jofi.70007.
> U.S. Government work; public domain in the USA.
> This page is an extract-only distillation by the Institute for Automated
> Research; it is not a verbatim reproduction of the article.


==============================================================================
# Carbon Pricing versus Green Finance: Pedersen (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2026/pedersen-carbon-pricing-green-finance-2026/
# Distilled: a unified model shows when carbon taxes and green finance (ESG investing, sustainable finance regulation) can substitute for each other and when green finance fails; the sustainable discount rate equals the normal rate plus a firm's carbon burden rate. J. Finance 2026, CC BY 4.0. Eight core results with source locators, datasets used, the model, and the method with its defining equations.
# Tags: paper-summary, climate-finance, esg, carbon-pricing, asset-pricing, sustainable-finance, factors, panel-regression, open-access, cc-by, peer-reviewed, unreplicated, data:trucost, data:wrds, data:eia-electricity
==============================================================================

**What this is.** The paper's core results, the model it builds on (a dynamic
general-equilibrium model with firms, households, and carbon externalities),
and the theory it contributes (sustainable discount rates as a second-best
substitute for carbon taxes) with the defining equations: enough to know what
it found and how, without reading all 42 pages. To replicate or extend it,
read the full source at the [original](https://doi.org/10.1111/jofi.70022)
(CC BY 4.0; open access).

## TL;DR

In a dynamic general-equilibrium model with green and brown firms, carbon
taxes, and ESG investors, Pedersen shows: (i) when the carbon price is
at the social cost, green finance should not be used; (ii) when carbon
prices are too low, green finance can restore the social optimum if each
firm's cost of capital is set to its *sustainable discount rate*, which
equals the normal rate plus the ratio of untaxed carbon externality to
firm value. Calibration with Trucost emissions and CRSP/Compustat values
shows the market-value-weighted average sustainable discount rate
adjustment for U.S. firms is only 0.19 pp at a 43 $/tCO2 social cost,
while the brown electricity sector requires a 3.2 pp (average) to 8 pp
(alternative calibration) increase. Empirical evidence from Eskildsen
et al. (2024) suggests current green finance has an effect equivalent to
only ~4 $/tCO2, far below what a green transition requires.

## Core results

Magnitudes and significance are as reported. Locators point into the
source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Carbon tax at the social cost is sufficient; green finance should not be used** when carbon is efficiently priced | Proposition 1, p. 576 | Social optimum implemented by scope-1 taxes at $$\tau = S$$; discount rates stay at $$r$$ for all firms; green finance distorts the equilibrium if applied on top |
| R2 | **Sustainable discount rate implements the social optimum** when carbon tax is too low, if firms can commit to future emissions | Proposition 4, pp. 578-579, eq. (17) | $$r^*_{it} = r + (S_{t+1} - \tau_{it+1}) X_{it+1} / v_{it}$$; equals normal rate plus firm's carbon burden rate (untaxed externality scaled by firm value) |
| R3 | **Scope 1+2 sustainable discount rates can handle stranded assets** where scope-1-only rates fail | Proposition 5, pp. 579-580, eq. (18) | Discount rate adds both direct (scope 1) and indirect (scope 2) carbon burden; brown electricity firms collapse as in the social optimum; stranded-asset problem resolved |
| R4 | **Green electricity firms can receive subsidized discount rates** as compensation for implicit over-taxation via scope-2 rules | Proposition 6, pp. 580, eqs. (19)-(20) | Green electricity discount rate $$r_{gt} = r - (S_{t+1} - \tau_{t+1})(F_b - F_g) G_{t+1} / v_{gt}$$; lower than normal rate $$r$$; mirrors Proposition 3 green-subsidy result |
| R5 | **Market-value-weighted average sustainable discount rate adjustment is 0.19 pp** for U.S. firms at S = 43 $/tCO2 (scope 1) | Figure 1 / Figure 3, pp. 564, 584; §VI.A | Median scope-1 burden rate 0.01%; market-weighted average 0.19%; scope 1+2 weighted average 0.23% (median 0.04%); most firms near baseline, a minority account for bulk of economy-wide emissions |
| R6 | **Brown electricity sector requires a 3.2 pp average increase** in cost of capital (scope 1); alternative calibration gives 8 pp | Figure 1, p. 564; §VI.C, p. 587, eq. (27) | Brown electricity average 8.2% (3.2 pp above 5% baseline) at S = 43 $/tCO2; alternative calibration (zero-profit, $$F_b = 820 \times 10^{-6}$$ tCO2/kWh, 40% profit margin): $$r_{bt} = 5\% + 8.0\% = 13\%$$ |
| R7 | **Empirical green finance implies an effective carbon tax of only ~4 $/tCO2**, far below the social cost | Figure 1, p. 564-565; Eskildsen et al. (2024) | Slope of 4 on emission-to-value ratio from regressing firms' implied cost of capital; corresponds to implicit $$S - \tau = 4$$ $/tCO2; at least an order of magnitude below what a green transition requires |
| R8 | **Carbon offset markets exhibit low and variable prices**, predicted by the model as a sign of greenwashing | Appendix D, pp. 599-600 | In equilibrium, offset prices must be proportional to quality ($$\phi_q = \bar{\phi} \cdot q$$); any price dispersion implies poor-quality offsets; firms buying low-quality offsets face an effective carbon tax reduced to ~10% of intended level |

**Overall (paper's conclusion).** Green finance is a second-best response
when carbon pricing is inadequate, but implementation challenges (commitment
problems, stranded assets, heterogeneous investors, greenwashing) make it
difficult to deliver the required cost-of-capital adjustments. Regions able
to impose a carbon tax have a clearer path to a green transition.

## Theory / model

The model is a dynamic general-equilibrium model (Section III, pp. 572-575).
There are $$N$$ goods-producing firms, green and brown electricity producers, and
a representative household. The paper extends the equilibrium ESG investing
framework of Pastor, Stambaugh, and Taylor (2021) to show exactly when green
finance can and cannot replicate a carbon tax.

**Goods-producing firms.** Firm $$i$$ at time $$t$$ chooses labor $$L_{it}$$, green
electricity $$G_{it}$$, brown electricity $$B_{it}$$, scope-1 emissions $$X_{it}$$, and
investment $$I_{it}$$ to maximize endogenous firm value $$V_{it}$$. Output is
$$Y_{it}(z_{it})$$ where $$z_{it} = (K_{it}, L_{it}, G_{it}, B_{it}, X_{it})$$.
Capital accumulates as $$K_{it} = (1 - \delta) K_{it-1} + I_{it-1}$$.
Firm profit (eq. 8, p. 573):

$$
\Pi_{it}(z_{it}) = Y_{it} - w_t L_{it} - p_{gt} G_{it} - p_{bt} B_{it}
                  - \tau_{it} X_{it} - \tau_{it}^{(2)} f_b B_{it} - \tau_{it}^{(2)} f_g G_{it}
$$

where $$\tau_{it}$$ is a scope-1 carbon tax on direct emissions $$X_{it}$$;
$$\tau_{it}^{(2)}$$ is a scope-2 carbon tax on indirect emissions via electricity
use; $$f_b$$ ($$f_g$$) is the fossil intensity of brown (green) electricity in tCO2/kWh.

The firm maximizes its endogenous value (eq. 9, p. 573):

$$
V_{it}(K_{it}) = \max_{z_{it+1}} \frac{\Pi_{it+1}(z_{it+1}) + V_{it+1}(K_{it+1})}{1 + r_{it}} - I_{it}
$$

taking the discount rate $$r_{it}$$ as given.

**Stylized two-period model (Section II, pp. 570-572).** A single firm uses
capital $$K$$ and fuel $$X$$, with output (eq. 4, p. 570):

$$
Y = AK - \frac{1}{fK}(fK - X)^2
$$

where $$A > 0$$ is productivity and $$f > 0$$ captures how polluting the firm is.
Given a carbon tax $$\tau$$, profit is (eq. 5, p. 570):

$$
\Pi = AK - \frac{1}{fK}(fK - X)^2 - \tau X
$$

The profit-maximizing emission is $$X = (1 - \tau/2) fK$$.

**Electricity firms.** Green electricity producer profit (eq. 10, p. 574):

$$
\Pi_{gt}(z_{gt}) = (p_{gt} + v_{gt} - \tau_{gt} F_g) G_t(z_{gt}) - \chi_{gt}(z_{gt})
$$

where $$v_{gt}$$ is a proportional government subsidy and $$\chi_{gt}$$ is production
cost. Brown electricity producer profit (eq. 11, p. 574):

$$
\Pi_{bt}(z_{bt}) = (p_{bt} - \tau_{bt} F_b) B_t(z_{bt}) - \chi_{bt}(z_{bt})
$$

where $$\tau_{bt}$$ is the direct carbon tax and $$F_b > F_g$$ is the fossil intensity
of brown electricity.

**Households.** The household owns shares $$\theta_{it}$$ in each firm. Consumption
(eq. 12, p. 574):

$$
C_t = w_t L_t + \sum_{i \in I} \theta_{i,t-1} (\Pi_{it} + V_{it})
      - \sum_{i \in I} \theta_{it} (V_{it} + I_{it}) + G_t
$$

where $$G_t$$ is the government budget (carbon taxes net subsidies). Household
utility (eq. 13, p. 574):

$$
U = \sum_{t=1}^{\infty} \beta^t [ u_t(C_t) - d_t(X_t) ]
$$

where $$\beta$$ is the time-preference rate, $$u_t$$ is consumption utility, $$d_t$$ is the
damage of aggregate emissions $$X_t = \sum_i X_{it} + X_{gt} + X_{bt}$$.

**Social planner's problem.** The planner maximizes $$U$$ subject to resource
constraints (eq. 14, p. 575):

$$
\sum_i G_{it} = G_t, \quad \sum_i B_{it} = B_t, \quad \sum_i L_{it} = L_t, \quad \theta_{it} = 1
$$

The social cost of carbon $$S_t = d'_t(X_t) / u'_t(C_t)$$ is the marginal
utility cost of pollution relative to the marginal utility of consumption.

**Key propositions.**
- Proposition 1 (p. 576): The social optimum is a competitive equilibrium
  with scope-1 carbon taxes $$\tau_{it} = S_t$$ for all firms, no green subsidies,
  and discount rate $$r_{it} = r$$ for all $$i$$.
- Proposition 2 (p. 576): The social optimum can also be implemented via
  scope-2 carbon taxes $$\tau_{it}^{(2)} = S_t$$ with no direct scope-1 taxes on
  goods producers.
- Proposition 3 (p. 577): With scope-2 taxes treating all electricity as
  brown, the social optimum requires a proportional subsidy to green
  electricity producers $$v_{gt} = S_t(F_b - F_g)$$.

Chittaro, Piazzesi, Sena, and Schneider (2025) generalize this framework with
a rich input-output structure and short-sale constraints (p. 565).

## Method

The paper is theoretical; it derives closed-form sustainable discount rates
analytically and then calibrates them with external data. The method builds on
`dynamic-general-equilibrium` and `pigouvian-taxation`.

**Deriving the sustainable discount rate (Section V.A, pp. 577-579).** Green
finance must set each firm's discount rate $$r_{it}$$ such that the firm's
optimization problem under the too-low carbon tax $$\tau_{it}$$ yields the same
choices $$z_{it+1}$$ as the star-equilibrium under the social cost $$S_t$$. The
condition (eq. 15, p. 577) is:

$$
\max_{z_{it+1}} \frac{\Pi_{it+1}(z_{it+1}) + V_{it+1}}{1 + r_{it}} - I_{it}
  = V_{it}
  = \max_{z_{it+1}} \frac{\Pi_{it+1}(z_{it+1}) + V_{it+1} - (S_{t+1} - \tau_{it+1}) X_{it+1}}{1 + r} - I_{it}
$$

where the right side is the star-equilibrium firm problem under the optimal
carbon tax. Solving for $$r_{it}$$ (eq. 16, p. 578; full derivation pp. 578):

$$
r_{it} = \frac{(1+r)(\Pi_{it+1} + V_{it+1})}{\Pi_{it+1} + V_{it+1} - (S_{t+1} - \tau_{it+1}) X_{it+1}} - 1
       = r + \frac{(S_{t+1} - \tau_{it+1}) X_{it+1}}{v_{it}}
$$

where $$v_{it} = (\Pi_{it+1} + V_{it+1} - (S_{t+1} - \tau_{it+1}) X_{it+1}) / (1 + r)$$
is the firm's social value, equal to its market value in equilibrium.
This is the `carbon-burden-rate` term: the missing carbon tax $$(S - \tau)$$ times
emissions $$X$$, scaled by firm value $$v$$. Equations (17)-(20) state the four
Propositions (4, 5, 6) in this notation (pp. 578-580).

**Investor preferences that generate the sustainable discount rate (Section V.B,
p. 581).** When investors experience a disutility proportional to their
carbon-footprint ownership (eq. 21, p. 581):

$$
\sum_i \theta_{it} (\Pi_{it+1} + V_{it+1}) + \left(W_t - \sum_{i \neq 1} \theta_{it} v_{it}\right)(1+r)
  - \sum_i (S_{t+1} - \tau_{it+1}) \theta_{it} X_{it+1}
$$

the FOC w.r.t. $$\theta_{it}$$ yields the required return (eq. 22, p. 581):

$$
r_{it} = r + \frac{(S_{t+1} - \tau_{it+1}) X_{it+1}}{v_{it}}
$$

which is exactly the sustainable discount rate (17). This shows green finance
works precisely when the marginal investor fully internalizes externalities.

**Long-term sustainable discount rate (Section VII.A, pp. 587-588).** Using a
Gordon growth model with dividend growth $$g_{\Pi}$$ and externality growth $$g_X$$
(eq. 30, p. 588):

$$
\bar{r}^x_i = r + \frac{(S - \tau_i) X_i}{v_i} \cdot \frac{r - g_{\Pi}}{r - g_{X_i}}
           = r + \frac{\text{PV}[(S - \tau_i) X_i]}{v_i} \cdot \frac{1}{\text{Dur}_i}
$$

where $$\text{Dur}_i = 1 / (r - g_{\Pi})$$ is the modified duration of dividends. The
long-term sustainable discount rate is smaller than the short-term rate
$$r^*_{it}$$ when emission growth is below dividend growth ($$g_X < g_{\Pi}$$).

**Alternative calibration for brown electricity (Section VI.C, pp. 586-587).**
With a constant-returns-to-scale brown electricity technology $$B_t = a K_{bt}$$,
cost $$\chi_{bt} = \bar{\chi} \cdot a K_{bt}$$, and the zero-profit condition (eq. 24-27):

$$
p_{bt+1} = \tau_{bt+1} F_b + \bar{\chi} + \frac{r_{bt} + \delta}{a} \tag{24}
$$

$$
r_{bt} = r + (S_{t+1} - \tau_{bt+1}) F_b \cdot a \tag{25}
$$

$$
r_{bt} = r + \frac{(S_{t+1} - \tau_{bt+1}) F_b}{(r+\delta)/a} \cdot (r + \delta) \tag{26}
$$

Calibrated at S = 43 $/tCO2, $$F_b = 820 \times 10^{-6}$$ tCO2/kWh, $$(r+\delta)/a = 0.40 \times 0.11$$ $/kWh (eq. 27, p. 587):

$$
r_{bt} = 5\% + \frac{43 \times 820 \times 10^{-6}}{0.40 \times 0.11} \times (5\% + 5\%)
       = 5\% + \frac{0.035}{0.044} \times 10\% = 5\% + 8.0\% = 13\% \tag{27}
$$

## Empirical specifications

The paper is primarily theoretical; there is no econometric estimation. The
empirical content is a calibration exercise (Section VI, pp. 582-587) using
external data, plus a single regression from Eskildsen et al. (2024).

**Calibration of sustainable discount rates (Section VI.A, pp. 582-584).**
For each U.S. publicly listed firm with Trucost scope-1 emissions data (fiscal
year 2021), the scope-1 sustainable discount rate from Proposition 4 is
computed as (eq. 17):

$$
r^x_{it} = r + \frac{(S_{t+1} - \tau_{it+1}) X_{it+1}}{v_{it}}
$$

- $$X_{it+1}$$: Trucost scope-1 CO2 emissions in tCO2 for fiscal year 2021
- $$v_{it}$$: firm market value = CRSP market equity + Compustat book value of debt,
  beginning of calendar year 2021
- $$S_{t+1} - \tau_{it+1}$$: set to $$S = 43$$ $/tCO2 (Nordhaus (2019) baseline; $$\tau = 0$$ for illustration)
- $$r$$: set to 5% for illustration

The social cost of carbon $$S_t$$ is taken as given from the Nordhaus (2019)
calibrations and translated into a cost-of-capital adjustment. The resulting
firm-level rates are aggregated to value-weighted industry
averages using two-digit GICS sectors (Utilities split into Renewable
Electricity, Brown Electricity, and Other Utilities).

**Scope 1+2 calibration (Section VI.B, p. 585).** Same procedure as above but
using Proposition 5, adding scope-2 emissions $$X^{\text{scope2}}_{it+1} = F_b B_{it+1} + F_g G_{it+1}$$
(also from Trucost). Each firm's scope-2 electricity estimate is
multiplied by 1/0.6 = 1.67 to proxy total brown-equivalent consumption
(60% of U.S. electricity from fossil fuels, 40% from renewables/nuclear).

**Empirical regression (Figure 1, pp. 564-565; Eskildsen et al. 2024).**
Eskildsen et al. (2024) regress firms' implied cost of capital on their
emission-to-asset ratio, controlling for risk characteristics:

$$
\text{implied\_COC}_i = \alpha + \beta \cdot \frac{X_i}{v_i} + \gamma \cdot \text{controls}_i + \epsilon_i
$$

The estimated slope $$\beta = 4$$ ($/value per tCO2/value = $/tCO2). The paper
reads this as empirical evidence that the market is pricing carbon externalities
as if the implicit social cost minus explicit carbon tax is $$S - \tau = 4$$ $/tCO2,
far below Nordhaus's 43 $/tCO2 estimate (R7 above).

No panel regressions, fixed effects, or standard-error treatments are applied
by the paper itself; the calibration is a direct plug-in of equation (17) into
data.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Trucost (scope 1 and scope 2 emissions, fiscal year 2021) | Firm-level carbon emission data for calibrating sustainable discount rates; scope 1 emissions in tCO2 per fiscal year; scope 2 electricity-based emissions | [Trucost](/wiki/commercial/trucost/) (licensed) |
| CRSP and Compustat (2021) | Market value of equity and book value of debt for constructing firm value $$v_{it}$$ used in calibration; also used to identify GICS industry sectors | [WRDS / CRSP / Compustat](/wiki/commercial/wrds/) (licensed) |
| Eskildsen et al. (2024) working paper data | Cross-sectional regression of firms' implied cost of capital on emission-to-asset ratio; slope estimate of 4 used as empirical calibration of effective implicit carbon tax | No page yet |
| U.S. Energy Information Administration (electricity price data, 2021) | Average electricity price 0.11 $/kWh used in alternative calibration (§VI.C) | No page yet |

Sample: U.S. publicly listed firms with Trucost emission data, fiscal year 2021;
roughly 3,000+ firms ranked by scope-1 externality rate (Figure 3, p. 584).
The paper's model is calibrated at a single cross-section; no time-series
econometrics are performed.

## When to read the full paper

Use the [original article](https://doi.org/10.1111/jofi.70022) if you are:
extending the model to heterogeneous investors or multiple externalities
(see also Pedersen 2026, *J. Finance: Insights and Perspectives*, forthcoming);
calibrating sustainable discount rates for a specific sector or carbon-tax
scenario using the exact proposition formulas; auditing a specific proposition
or appendix proof (all proofs are in Appendix A); or doing a literature
review on the carbon-pricing versus ESG debate. The Appendix B Cobb-Douglas
model provides an alternative tractable derivation of the sustainable discount
rate for readers preferring that functional form.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 81(2). This distillation
was extracted by an LLM on 2026-05-31 and augmented on 2026-06-01; it is
**not human-verified or independently reproduced**. The article is open access
under CC BY 4.0; this page is an adaptation (core results extracted and
re-expressed; changes were made).

> **Attribution (CC BY 4.0).** Pedersen, Lasse Heje.
> "Carbon Pricing versus Green Finance." *The Journal of Finance* 81, no. 2
> (April 2026): 561–602. DOI: 10.1111/jofi.70022. © 2026 The Author(s).
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Subtle Discrimination: Pikulina & Ferreira (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2026/pikulina-subtle-discrimination-2026/
# Distilled: a theoretical model of "subtle discrimination" (biased promotion decisions with plausible deniability) showing that small biases generate large gaps in skills and promotions; the direction of the skill gap reverses with career stakes. J. Finance 2026, CC BY 4.0. Eight core results with source locators, theory tested, and further applications.
# Tags: paper-summary, discrimination, labor-economics, promotions, human-capital, career-stakes, diversity, contest-theory, open-access, cc-by, peer-reviewed, unreplicated
==============================================================================

**What this is.** The paper's core results, model equations, and theory: enough to
know what it found without reading all 41 pages. To replicate or extend it,
read the full source at the [original](https://doi.org/10.1111/jofi.13506).
CC BY 4.0 permits redistribution; the PDF is not mirrored in this batch.

## TL;DR

The paper introduces and formalizes *subtle discrimination*: biased acts
that cannot be objectively ascertained as discriminatory because the
decision-maker can invoke a plausible nondiscriminatory defense. In a
model of promotion contests between two ex ante identical agents ("Blue"
favored, "Red" unfavored), even an arbitrarily small bias generates large
equilibrium gaps in skill investment and promotion outcomes. The direction
of the skill gap reverses with career stakes: unfavored agents
overcompensate (invest more) in low-stakes careers and underinvest in
high-stakes careers. The model delivers novel predictions for equity
analysts, lending, fund flows, banking, and entrepreneurial finance.

## Core results

Magnitudes and significance are as reported. All results are
model-derived (propositions and corollaries); this paper is purely
theoretical. Locators point into the source PDF.

| # | Result | Locator | Magnitude / Content |
|---|---|---|---|
| R1 | A unique equilibrium investment profile exists for any bias level | Proposition 2, p. 343 | Closed-form: e\*\_b = [σ(0.5 − β) + 2βσ²(0.5 + β)] / (1 + 4β²σ²); e\*\_r symmetric with β negated; corner solution e\*\_b = 1 if σ > σ̄(β) |
| R2 | At low stakes (σ ≤ 1), unfavored Red invests more than favored Blue; at high stakes (σ > 1), Blue invests more | Corollary 1, p. 344; Figure 1, p. 343 | e\*\_r ≥ e\*\_b if and only if σ ≤ 1 (symmetric-cost case); driven by the overcompensation vs discouragement trade-off |
| R3 | The promotion gap is U-shaped in the premium-cost ratio σ | Proposition 3, p. 348; Figure 2, p. 349 | Gap Δp first decreases then increases with σ; at high σ observable achievement differences (achievement gap) dominate over the direct favoritism gap, so promotion gaps are large with little visible bias |
| R4 | Subtle discrimination raises profits for low-productivity firms and lowers them for high-productivity firms | Proposition 7, p. 353; Figures 3–4, pp. 352, 354 | Optimal bias β\*(θ) = 0.5 for low-productivity-cost ratio θ, and 0 for high θ; threshold θ' ≈ 2.62 (numerical) |
| R5 | High-productivity firms optimally choose zero bias; low-productivity firms optimally choose maximum bias | Proposition 7, eq. (15), p. 353 | β(θ) = 0.5 if θ ∈ (0, θ'], 0 if θ ∈ [θ', θ̄]; firms become polarized between progressive (high-θ) and conservative (low-θ) |
| R6 | Subtle discrimination, not overt discrimination, generates the overcompensation effect | Corollary 3, p. 347 | e\*\_r ≥ e\*\_b if and only if σ ≤ 1/(1 − δ) under overt bias δ; excess subtle bias ε ≡ β − δ/2 must be strictly positive for Red to outinvest Blue |
| R7 | In the equity analyst application, subtle bias against female (Red) analysts causes them to overinvest in forecast accuracy | §IV.A, p. 356; eq. (16) | e\*\_r = σ(1 − δ)(1 − a(1 − β)); female analysts invest more in accuracy than male, rationalizing Kumar (2010) evidence without assuming superior innate ability |
| R8 | In the fund-flow application, subtle investor bias does not generate a performance gap at marginal funds, making outcome (Becker) tests unable to reject the null of no discrimination | §IV.C, pp. 358–359 | When investors use a lexicographic rule (performance first, then manager characteristics for ties), biased choice at the margin does not affect returns, so outcome tests have no power against subtle discrimination |

**Overall (paper's conclusion).** Subtle and overt discrimination have
markedly different empirical predictions. Small subtle biases can
generate large skill and promotion gaps; the gaps are amplified through
strategic interactions between competing agents. High-stakes careers
see discouragement of unfavored agents; low-stakes careers see
overcompensation. Observable achievement differences explain most of
the promotion gap in high-stakes settings, making the discrimination
hard to detect. Firm-level diversity is predicted to correlate with
firm productivity and human capital intensity.

## Theory / model

The paper formalizes a concept and solves a model; there is no empirical
identification. The formal object is a two-player promotion contest
embedded in a principal-agent setting.

**Decision-maker bias (Definition 1, p. 337).** Two candidates Blue (b)
and Red (r) have observable skills $$s_b$$ and $$s_r$$. The principal
privately observes a subjective signal $$x_i$$ for each agent; the
decision function $$P(s_b, s_r, \omega)$$ is the probability of choosing Blue.
Bias toward Blue is the excess probability of choosing Blue not justified
by the qualification gap:

$$
b(s_b, s_r, \omega) = P(s_b, s_r, \omega) - F\!\left(\frac{s_b - s_r}{\omega}\right) \geq 0 \tag{1}
$$

where $$F(\cdot)$$ is the CDF of $$\Delta_x = x_r - x_b$$, $$\omega > 0$$ weights
subjective signals. The bias is *subtle* (Definition 1, p. 337) if
$$F\!\left(\frac{s_b - s_r}{\omega}\right) > 0$$, i.e., there exist signals that could
justify choosing Blue without proof of bias. It is *overt* (Definition 2)
if $$F = 0$$, meaning a single act is conclusive evidence of discrimination.

**Promotion model (§III.A, pp. 338-340).** This setup builds on
Prendergast (1993), which incentivizes firm-specific human capital
through promotions; the present model adds biased tie-breaking. A firm
hires two ex ante identical agents for job 1. At Date 1 both simultaneously choose
investments $$e_i \in [0, 1]$$ at cost $$c(e_i)$$ (quadratic: $$c(e_i) = k e_i^2 / 2$$).
Skill $$s_i \in \{0, 1\}$$ is realized; $$\Pr(s_i = 1) = e_i$$. At Date 2 the
principal promotes one agent to job 2 (the top position), yielding a
productivity gain $$H > 0$$ if skilled. Wages are $$(w_1, w_2 = w_1 + W)$$;
the promotion premium $$W$$ is the main incentive instrument.
Agent $$i$$'s utility is:

$$
u_i = w_i - c(e_i)
$$

- $$w_i$$ is the wage received

The principal's bias toward Blue is captured by $$\beta \in (0, 0.5]$$,
interpreted as agents' belief about the principal's tie-breaking probability:
if $$s_b = s_r$$, $$\Delta_s = 0$$, principal chooses Blue with probability
$$0.5 + \beta$$. The principal always promotes the skilled agent when the
two differ in observable skill (no overt discrimination).

**Firm profit (pp. 339-340):**

$$
\Pi = l + H(e_b + e_r - e_b e_r) - 2w_1 - W
$$

- $$l$$ is the base payoff and $$H(e_b + e_r - e_b e_r)$$ is the expected value added by skilled promotion

**First-best benchmark (Proposition 1, p. 341).** The social planner
maximizes total surplus:

$$
\max_{(e_b,\, e_r) \in [0,1]^2} \; l + H(e_b + e_r - e_b e_r) - c(e_b) - c(e_r) \tag{2}
$$

The first-best investment levels are either (i) symmetric $$e_b^{FB} = e_r^{FB} = \tilde{e} < 1$$,
or (ii) $$e_i^{FB} > 0$$ and $$e_{-i}^{FB} = 0$$ for some $$i \in \{b, r\}$$ (corner solution).
With quadratic costs $$c(e_i) = k e_i^2 / 2$$ and $$H \leq k$$: symmetric
$$\tilde{e} = H / (H + k)$$. If $$H > k$$: corner solution.

**Identification of mechanisms.** The model isolates two opposing forces
on the unfavored agent's investment (p. 344):

- *Discouragement effect*: Red's probability of promotion is low, so the
  marginal benefit of investing is low, discouraging investment.
- *Overcompensation effect*: Red wants to separate from Blue to avoid ties
  (where she loses), so Red invests more to minimize tie probability.

Which force dominates depends on the premium-cost ratio $$\sigma = W / k$$.
This extends Coate and Loury (1993): the self-fulfilling stereotype model
is generalized by adding strategic competition between agents, and unlike
Coate and Loury the unfavored group may invest more than the favored group
under low stakes.

## Method

The paper uses `promotion-contest` analysis and `principal-agent` optimal
contracting; no econometric estimator is involved.

**Equilibrium characterization (§III.D.1, p. 342).** Under the limiting
case $$\omega \to 0$$ (subjective information negligible), agent $$i$$'s
expected utility is:

$$
U_i(e, w) = w_1 + W \left[ e_i(1 - e_{-i}) + \left(\tfrac{1}{2} + \beta_i\right)(1 - e_i - e_{-i} + 2 e_i e_{-i}) \right] - \frac{k e_i^2}{2} \tag{3}
$$

- $$\beta_b = -\beta_r = \beta$$

Maximizing over $$e_i$$ taking $$e_{-i}$$ as given, the reaction functions are (eq. 4, p. 342):

$$
e_b = \frac{W}{k}\!\left(\tfrac{1}{2} - \beta + 2\beta e_r\right) \qquad \text{and} \qquad e_r = \frac{W}{k}\!\left(\tfrac{1}{2} + \beta - 2\beta e_b\right) \tag{4}
$$

**Optimal compensation (§III.E, pp. 349-350).** The principal chooses
promotion premium $$\sigma$$ to maximize expected profit net of entry costs:

$$
\Pi(k, \beta, \theta) = \max_{\sigma \in [0,\, \bar{\sigma}(\beta)]} \; k\theta(e_b + e_r - e_b e_r) - k\sigma \tag{13}
$$

- subject to equilibrium conditions (5) and (6), where $$\theta = H / k$$ is the *productivity-cost ratio*

The IC constraint for each agent is:

$$
e_i = \operatorname*{arg\,max}_{e \in [0,1]} \; e W \!\left[\left(\tfrac{1}{2} - \beta_i\right) + 2\beta_i e_{-i}\right] - \frac{k e^2}{2}, \quad i \in \{b, r\} \tag{12}
$$

**Endogenous bias (§III.F, p. 351, eq. 14).** When the firm also chooses
its subtle bias $$\beta$$, the problem becomes:

$$
\Pi(\theta) = \max_{(\sigma,\, \beta) \in [0,\, \bar{\sigma}(\beta)] \times [0,\, 0.5]} \; \theta(e_b + e_r - e_b e_r) - \sigma \tag{14}
$$

subject to (5) and (6). The optimal policy is characterized in
Proposition 7 (eq. 15, p. 353):

$$
\beta(\theta) = 0.5 \quad \text{if } \theta \in (0, \theta'] \tag{15}
$$

$$
\beta(\theta) = 0 \quad \text{if } \theta \in [\theta', \bar{\theta}]
$$

with $$\sigma(\theta) < 1$$ (low stakes) for $$\theta \in (0, \theta']$$ and
$$\sigma(\theta) > 1$$ (high stakes) for $$\theta \in [\theta', \bar{\theta}]$$.

## Empirical specifications

This paper derives theoretical propositions; it has no regression
specifications or econometric estimates. Empirical content comes in two
forms: (a) comparative-statics predictions tied to observable proxies,
and (b) cross-application predictions showing how the framework matches
existing evidence.

**Promotion-gap specification (Proposition 3 and Figure 2, pp. 348-349).**
The equilibrium promotion gap between Blue and Red is:

$$
\Delta_p = p_b - p_r = (e_b - e_r) + \bigl[e_b e_r + (1 - e_b)(1 - e_r)\bigr] \cdot 2\beta \tag{10}
$$

- the first term is the *achievement gap* and the second is the *favoritism gap*

Prediction: $$\Delta_p$$ is U-shaped in $$\sigma$$; at high
$$\sigma$$, the achievement gap dominates, so promotion gaps are large but
little direct evidence of discrimination is observable. This contradicts
Lazear and Rosen (1990), whose model predicts small promotion gaps in
high-stakes jobs; here promotion gaps are larger at higher stakes when
discrimination is subtle.

**Analyst accuracy specification (§IV.A, eq. 16, p. 356).**
An analyst of type $$i$$ earns promotion via composite metric
$$y_i = y_{1i} + y_{2i}$$ (accuracy + optimism). With subtle bias $$\beta$$
and overt bias $$\delta$$, analyst investment levels are:

$$
e_r^* = \sigma(1 - \delta)(1 - a(1 - \beta)) \qquad \text{and} \qquad e_b^* = \sigma(1 + \delta)(1 - a(1 + \beta)) \tag{16}
$$

- $$a \in (0,1)$$ is the accuracy-optimism trade-off and $$\sigma = W/k$$

A subtle bias ($$\delta = 0$$, $$\beta > 0$$) increases Red accuracy relative
to Blue; an overt bias ($$\delta > 0$$) has the opposite effect.

**Overt vs. subtle specification (Corollary 3, p. 347).**
With overt bias $$\delta \geq 0$$ and subtle bias $$\beta \geq \delta/2$$, the
overcompensation effect dominates if and only if:

$$
\sigma \leq \frac{1}{1 - \delta}
$$

The threshold $$1/(1 - \delta) > 1$$ is strictly larger than 1, implying
overt bias attenuates overcompensation: overt discrimination moderates
the overcompensation effect of subtle discrimination. This extends
Drugov and Ryvkin (2017) in the biased-contest literature by
distinguishing subtle from overt bias and showing their effects differ
qualitatively.

## Datasets used

This paper is entirely theoretical; it presents no empirical analysis
and uses no data. Empirical predictions are linked to existing
evidence from the literature (e.g., Bircan, Friebel and Stahl (2023)
on banking; Kumar (2010) on analyst forecasts; Frame et al. (2025) on
mortgage lending) but the paper itself does not run any regressions or
construct any dataset. No `data:` tags apply.

| Dataset | Role in paper | Wiki page |
|---|---|---|
| None (theoretical paper) | N/A | N/A |

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.13506) if you are:
deriving the model's propositions or checking proofs (Appendix, pp.
361-365); working through the Internet Appendix robustness extensions;
applying the framework to a new context (the lending or fund-flow
application sections are self-contained); or auditing how the
discouragement vs overcompensation trade-off interacts with cost
heterogeneity (§III.D.3). For "what did this paper find," the table
above is sufficient and is the intended default.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 81(1). This distillation
was extracted by an LLM on 2026-05-31 and augmented on 2026-06-01; it is
**not human-verified or independently reproduced**.

> **Attribution (CC BY 4.0).** Pikulina, Elena S., and Daniel Ferreira.
> "Subtle Discrimination." *The Journal of Finance* 81, no. 1
> (February 2026): 329–369. DOI: 10.1111/jofi.13506. © 2025 The Author(s).
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**. CC BY 4.0
> permits redistribution; the verbatim PDF is not hosted in this batch.


==============================================================================
# Corporate ESG Profiles and Investor Horizons: Starks, Venkat & Zhu (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2026/starks-esg-profiles-investor-horizons-2026/
# Distilled: Long-term institutional investors systematically tilt their portfolios toward firms with higher ESG scores; this pattern holds at both the investor and firm level across mutual funds and 13f institutions, and survives controls for investment style, ESG rating disagreement, and errors-in-variables. J. Finance 2026, CC BY-NC-ND 4.0. Nine core results with source locators, datasets used, tested hypotheses, and the empirical specifications behind each result.
# Tags: paper-summary, esg, institutional-investors, investor-horizon, mutual-funds
==============================================================================

**What this is.** The paper's core results, the hypotheses it tests, and the
regression specifications behind each finding: enough to know what it found
and how, without reading all 40 pages. To replicate or extend it, read the
full source at the [original](https://doi.org/10.1111/jofi.70008).

## TL;DR

Using a large sample of US mutual funds and 13f institutional investors from
2000 to 2018, the paper documents that longer-horizon investors (lower
portfolio turnover or churn ratios) tilt their portfolios toward firms with
higher ESG scores, and that firms with better ESG profiles attract
shareholder bases with longer investment horizons. Three mechanisms are
examined: (i) an information channel, where long-term investors specialize in
analyzing long-payoff ESG signals; (ii) a limits-to-arbitrage channel, where
lower flow-performance sensitivity enables long-term investors to hold
illiquid ESG positions; and (iii) a clientele-catering channel. Evidence
supports the first two channels. A 2004 SEC regulatory shock to mandatory
portfolio disclosure frequency provides quasi-causal evidence that
horizon-shortening causally reduces ESG tilts.

## Core results

Magnitudes and significance are as reported; `*`/`**`/`***` = 10%/5%/1%.
Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Long-horizon mutual funds (lowest-turnover quintile) hold portfolios with meaningfully higher ESG scores than short-horizon funds (highest-turnover quintile) | Figure 1, Panel A, p. 614 | Weighted-average ESG score of 0.91 (long-horizon quintile, turnover = 20%) vs. 0.42 (short-horizon quintile, turnover = 161%); spread = 28% of a standard deviation |
| R2 | Same ESG-horizon gradient holds for 13f institutions sorted by churn ratio | Figure 1, Panel B, p. 614 | ESG score of 1.55 (long-horizon, churn ratio = 10%) vs. 0.44 (short-horizon, churn ratio = 92%), monotonically decreasing |
| R3 | Investor-level panel regression: fund turnover ratio is significantly negatively associated with portfolio ESG score, controlling for investment style and fund characteristics | Table II, col. (1), p. 616 | Coeff. on Fund Turnover Ratio = -0.0968`***` (s.e. 0.0197); 1-SD increase in turnover (0.49) corresponds to a 0.05-point decrease in fund ESG score |
| R4 | Result is robust to churn ratio, adjusted churn ratio, and 13f institutions; 2SLS using alternative ESG scores as instruments (errors-in-variables approach) leaves results unchanged | Table II, cols. (2)-(5), p. 616; Table VI, p. 625 | Fund Churn Ratio coeff. = -0.216`***` (s.e. 0.0402); Adjusted Churn Ratio = -0.555`***` (s.e. 0.0962); 13f Churn Ratio = -0.321`***` (s.e. 0.0310) |
| R5 | Firm-level: one-SD increase in MSCI ESG score (2.32) is associated with a decrease of 0.51 percentage points in weighted-average mutual fund shareholder turnover ratio | Table IV, col. (1), p. 621 | MSCI ESG Score coeff. = -0.222`**` (s.e. 0.102); results hold for 13f churn ratio and transient-investor share |
| R6 | Information channel: long-term mutual funds (low churn) are less likely to sell following a negative earnings surprise (patience), but more likely to sell following a negative ES incident (responsiveness to ESG news) | Table VII, p. 627; Table VIII, p. 629 | Earnings shortfall: interaction coeff. (LongTermInvestor x NegEarningsSurprise) opposite sign to main effect. ES incident: long-term fund D(sell) interaction = 4.652`***` (s.e. 3.699), D(liq) = 3.244`***` (s.e. 1.348) percentage points higher probability |
| R7 | Limits-to-arbitrage channel: shorter-horizon funds have higher flow-performance sensitivity (FPS); high FPS is negatively and significantly associated with fund portfolio ESG score | Table IX, cols. (1)-(3), p. 631 | Turnover Ratio coeff. on FPS = 1.125`***` (s.e. 0.294); FPS coeff. on Portfolio ESG = -0.337`***` (s.e. 0.0875) |
| R8 | Causal test (SEC 2004 DiD + 2SLS): funds required to switch from semiannual to quarterly portfolio disclosure shortened their horizon and subsequently reduced ESG portfolio tilts | Table X, p. 633 | DiD reduced-form: TreatedFunds x PostDisclosure coeff. = -0.0430`**` (s.e. 0.0185); 2SLS second stage: Fund Turnover Ratio coeff. on Fund ESG = -0.683`**` (s.e. 0.333) |
| R9 | Clientele-catering channel: no evidence that long-term fund investors respond differentially to Morningstar sustainability globe ratings, and ESG gap does not widen during periods of high climate news attention | Table XI, p. 636; Table XII, p. 638 | Interaction (LongTermInvestor x Globe x Post) coefficients are statistically indistinguishable from zero; Horizon x CCNI interaction is near zero and insignificant in most specifications |

**Overall (paper's conclusion).** Long-horizon investors consistently prefer
firms with better ESG profiles, supporting the view that investors have
heterogeneous preferences regarding corporate ESG profiles and that this
heterogeneity depends on investor horizons, related to work linking investor
horizon to corporate policies (Derrien, Kecskés, and Thesmar 2013). The
information channel and
limits-to-arbitrage channel both receive empirical support; the
clientele-catering channel does not.

## Theory / model

The paper has no structural model. It tests three theoretical mechanisms
from prior literature:

**Information channel (Froot, Perold, and Stein 1992; Van Nieuwerburgh and
Veldkamp 2009).** Long-term investors specialize in collecting and
interpreting long-payoff information such as ESG profiles; short-term
investors specialize in short-frequency signals such as quarterly earnings.
In equilibrium, long-term investors overweight firms with ESG-related
projects because of an information advantage. The testable implication is
that long-term funds are patient toward negative earnings surprises but
reactive to ES incidents (the two carry different horizon implications).

**Limits-to-arbitrage channel (Stein 2005; Giannetti and Kahraman 2018;
Pedersen, Fitzgibbons, and Pomorski 2021).** If ESG investing is a
long-horizon arbitrage opportunity, open-end fund managers with high
flow-performance sensitivity (FPS) are deterred from maintaining ESG tilts
because short-run underperformance triggers outflows before the mispricing
corrects. Low-FPS (long-horizon) funds face weaker redemption risk and can
hold ESG positions longer. The testable implication is that FPS negatively
mediates the horizon-ESG relationship.

**Clientele-catering channel (Heinkel, Kraus, and Zechner 2001; Pastor,
Stambaugh, and Taylor 2021).** If end-investors of long-term funds have
stronger nonpecuniary ESG preferences, managers cater by tilting toward
high-ESG stocks. The testable implication is that ESG preferences of
end-investors of long-term funds differ from those of short-term funds
(tested via Morningstar globe flow responses and the CCNI interaction).
The globe DiD design follows Hartzmark and Sussman (2019).

**Identification.** The primary cross-sectional tests control for investment
objective by time, fund size, portfolio characteristics, and past return rank
with two-way clustering of standard errors (fund and quarter). The key
quasi-causal test (R8) exploits the 2004 SEC rule requiring mutual funds to
disclose holdings quarterly rather than semiannually, using treated funds
(those newly required to switch frequency) vs. control funds (those already
disclosing quarterly) in both a DiD and a 2SLS framework where the
regulatory interaction instruments for observed investment horizon.

## Method

All results are produced by panel regressions and event-study-style
portfolio comparisons. No new estimator is proposed. The main workhorse
estimators are:

- `panel-regression` with investment objective-by-quarter or
  industry-by-year fixed effects and two-way clustered standard errors
  (fund and quarter, or stock and year).
- `portfolio-sort` into turnover or churn quintiles for the bivariate ESG
  comparisons (Figure 1).
- `instrumental-variables` (2SLS) with two uses: (a) using Refinitiv and
  Sustainalytics ESG scores as instruments for MSCI ESG scores to address
  errors-in-variables noise in ESG ratings (Tables V-VI, following Berg et
  al. 2022); and (b) using the TreatedFunds x PostDisclosure interaction
  as an instrument for investor horizon to isolate the causal effect of
  horizon on ESG tilts (Table X).
- `difference-in-differences` around the 2004 SEC portfolio disclosure rule
  (eq. 13) and around the 2016 Morningstar sustainability globe introduction
  (eq. 16), including fund fixed effects for within-fund identification.

The churn ratio $$\text{CR}$$ is constructed at the quarterly level, using the
four-quarter moving average (Gaspar, Massa, and Matos 2005), and the
adjusted churn ratio $$\text{CR\_Adj}$$ mitigates flow bias by taking the minimum of
buy-side and sell-side churn (Yan and Zhang 2009). Both measures are used
interchangeably throughout.

## Empirical specifications

**Investor-level ESG regression (R3, R4), eq. 7, p. 615:**

$$
\text{FundESG}_{j,t+1} = \alpha_t + \beta_1 \cdot \text{Horizon}_{j,t} + \beta_2 \cdot \text{Controls}_{j,t} + \epsilon_{j,t} \tag{7}
$$

- $$\text{FundESG}_{j,t+1}$$ is the value-weighted average MSCI ESG score of fund $$j$$'s portfolio holdings at the end of the following quarter.
- $$\text{Horizon}_{j,t}$$ is the fund's annual turnover ratio, churn ratio, or adjusted churn ratio (four-quarter moving average).
- Controls include ln(Fund TNA), number of portfolio holdings, value-weighted portfolio market cap and book-to-market, past 12-month return, and fractional return rank.
- Sample: all fund-quarter observations. Fixed effects: investment objective-by-quarter. Standard errors: two-way clustered at the fund and quarter level.

**Firm-level horizon regression (R5), eq. 8, p. 619:**

$$
\text{InvestorHorizon}_{i,t+1} = \alpha_t + \beta_1 \cdot \text{ESG}_{i,t} + \beta_2 \cdot \text{Controls}_{i,t} + \epsilon_{i,t} \tag{8}
$$

- $$\text{InvestorHorizon}_{i,t+1}$$ is the weighted-average turnover ratio or churn ratio of firm $$i$$'s mutual fund or 13f shareholders.
- $$\text{ESG}_{i,t}$$ is the MSCI ESG score.
- Controls include ln(market cap), book-to-market, dividend yield, profitability, past return, return volatility, stock turnover, and fund flow volatility.
- Fixed effects: industry (two-digit SIC) and year. Standard errors: double-clustered at stock and year.

**Earnings-surprise trading regression (R6, information channel), eq. 9, p. 626:**

$$
\text{Dummy(Sell)}_{i,j,t} = \alpha_{j,t} + \beta_1 \cdot \text{EarningsShortfall}_{i,t} + \beta_2 \cdot \text{LongTermInvestor}_{j,t} \times \text{EarningsShortfall}_{i,t} + \text{Controls}_{j,t} + \epsilon_{i,j,t} \tag{9}
$$

- $$\text{Dummy(Sell)}_{i,j,t}$$ is an indicator for fund $$j$$ selling stock $$i$$ in quarter $$t$$.
- $$\text{EarningsShortfall}_{i,t}$$ is a negative earnings surprise measure for stock $$i$$.
- $$\text{LongTermInvestor}_{j,t}$$ is an indicator for funds whose four-quarter trailing churn ratio falls below the 30th percentile.
- Sample: interquarter changes for positions held in previous quarter. Fixed effects: fund-by-quarter. Standard errors: double-clustered at the fund and quarter level.

**ES-incident trading regression (R6, information channel), eq. 10, p. 628:**

$$
\text{Dummy(Sell)}_{i,j,t} = \alpha_{j,t} + \beta_1 \cdot \text{ESIncident}_{i,t-1} + \beta_2 \cdot \text{LongTermInvestor}_{j,t} \times \text{ESIncident}_{i,t-1} + \text{Controls}_{j,t} + \epsilon_{i,j,t} \tag{10}
$$

- $$\text{ESIncident}_{i,t-1}$$ is an indicator from RepRisk for severe negative environmental or social incidents in the previous quarter.
- $$\text{LongTermInvestor}_{j,t}$$ is an indicator for low-churn funds (below 30th percentile trailing churn ratio).
- Specifications include both fund-by-quarter and stock-by-quarter fixed effects (columns 4-6 of Table VIII absorb all stock-time variation, isolating differential fund responses).

**Flow-performance sensitivity and ESG (R7, limits-to-arbitrage), eqs. 11-12, p. 630:**

$$
\text{FPS}_{j,t} = \alpha_t + \beta_1 \cdot \text{Horizon}_{j,t} + \beta_2 \cdot \text{Controls}_{j,t} + \epsilon_{j,t} \tag{11}
$$

$$
\text{FundESG}_{j,t+1} = \alpha_t + \beta_1 \cdot \text{FPS}_{j,t} + \beta_2 \cdot \text{Controls}_{j,t} + \epsilon_{j,t} \tag{12}
$$

- $$\text{FPS}_{j,t}$$ is the flow-performance sensitivity of fund $$j$$, estimated from a 24-month rolling OLS of monthly net fund flows on past 12-month average monthly return.
- Fixed effects: investment objective-by-quarter. Standard errors match the investor-level specification.

**SEC 2004 DiD and 2SLS (R8, causal test), eqs. 13-15, pp. 632-633:**

$$
\text{FundESG}_{j,t+1} = \alpha_j + \gamma_t + \beta \cdot \text{TreatedFunds}_j \times \text{PostDisclosure}_t + \text{Controls} + \epsilon_{j,t} \tag{13}
$$

- Reduced-form DiD. Treated funds switched from semiannual to quarterly disclosure; control funds were already disclosing quarterly. The SEC 2004 mandatory quarterly disclosure rule is used as an exogenous shock to fund horizon following Agarwal et al. (2015).
- Sample: 2001Q1-2008Q4. Fixed effects: fund ($$\alpha_j$$) and quarter ($$\gamma_t$$). Standard errors: clustered at the fund level.

$$
\text{Horizon}_{j,t} = \alpha_j + \gamma_t + \beta \cdot \text{TreatedFunds}_j \times \text{PostDisclosure}_t + \text{Controls} + \epsilon_{j,t} \tag{14}
$$

$$
\text{FundESG}_{j,t+1} = \alpha_j + \gamma_t + \beta_1 \cdot \widehat{\text{Horizon}}_{j,t} + \beta_2 \cdot \text{Controls} + \epsilon_{j,t} \tag{15}
$$

- Eq. 14 is the first stage of 2SLS; eq. 15 is the second stage, with $$\widehat{\text{Horizon}}_{j,t}$$ the fitted value from eq. 14. Standard errors: clustered at the fund level for serial dependence.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CRSP Mutual Fund Database | Fund characteristics (TNA, turnover ratio, returns, flows, expense ratios); mutual fund sample construction (98,252 fund-years) | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| Thomson Reuters s12 (mutual fund holdings) | Quarterly equity holdings for mutual funds; portfolio ESG construction | [WRDS](/wiki/commercial/wrds/) (licensed) |
| Thomson Reuters s34 (13f institutions) | Quarterly equity holdings for 13f institutions; 166,185 institution-year observations | [WRDS](/wiki/commercial/wrds/) (licensed) |
| MSCI ESG STATs (formerly KLD) | Annual positive/negative ESG indicators for firm-years; primary ESG scoring; 26,217 firm-years | [KLD / MSCI ESG](/wiki/commercial/kld/) (licensed) |
| Refinitiv ESG (formerly ASSET4) | Alternative ESG scores for robustness (2009-2017 subsample); used as instrument in 2SLS | no page yet |
| Sustainalytics ESG | Alternative ESG scores for robustness (2009-2017); used as instrument in 2SLS | no page yet |
| Compustat annual fundamentals | Book-to-market, profitability, dividend yield; firm-level controls | [WRDS / Compustat](/wiki/commercial/wrds/) (licensed) |
| CRSP daily/monthly stock data | Market capitalization, stock turnover, past returns, return volatility | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| I/B/E/S | Analyst earnings forecasts; second measure of earnings surprise for Table VII | [I/B/E/S](/wiki/commercial/ibes/) (licensed) |
| RepRisk | Negative environmental and social (ES) incident data; used in ES-incident trading tests (Table VIII) | [RepRisk](/wiki/commercial/reprisk/) (licensed) |
| SEC EDGAR (N-CSR/N-CSRS filings) | Mutual fund shareholder reports; bag-of-words ESG mention analysis (Table III) | [EDGAR](/wiki/datasets/edgar/) |
| Bushee institutional investor classifications | Transient/dedicated/quasi-indexer classification from Bushee (1998); supplemental horizon measure for 13f institutions | no page yet |
| Climate Change News Index (CCNI) | Engle et al. (2020) index of WSJ climate reporting intensity; clientele-catering test (Table XII) | no page yet |

Sample: 2000 to 2018 for main analyses; 2001-2008 for SEC 2004 DiD;
2015-2017 for Morningstar globe DiD. Quarterly frequency for fund/institution
analyses; annual for firm-level.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.70008) if you are:
studying institutional demand heterogeneity for ESG assets; designing tests
of limits-to-arbitrage in ESG pricing; examining how regulatory shocks
(disclosure frequency) affect fund portfolio composition; or extending the
horizon-ESG nexus to non-US markets or private-asset investors. The locators
above point to the exact tables and figures.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 81(2), April 2026. This
distillation was extracted by an LLM on 2026-06-01 and is **not
human-verified or independently reproduced**. The CC BY-NC-ND 4.0 licence
permits non-commercial sharing with attribution but prohibits derivatives;
the verbatim PDF is not hosted.

> **Attribution (CC BY-NC-ND 4.0).** Starks, Laura T., Parth Venkat, and Qifei Zhu.
> "Corporate ESG Profiles and Investor Horizons."
> *The Journal of Finance* 81, no. 2 (April 2026): 603-642.
> DOI: 10.1111/jofi.70008. (c) 2026 The Author(s).
> Licensed under [CC BY-NC-ND 4.0](https://creativecommons.org/licenses/by-nc-nd/4.0/).
> This page is a distilled summary by the Institute for Automated Research;
> it is not a reproduction of or derivative from the original article.


==============================================================================
# The Dollar during the Great Recession: Stavrakeva & Tang (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jf/2026/stavrakeva-dollar-great-recession-2026/
# Distilled: U.S. forward guidance easings during the Great Recession (Dec 2008 to Sep 2012) caused the dollar to appreciate, not depreciate, against both advanced-economy and emerging-market currencies, reversing the conventional wisdom. The paper attributes this to an information channel: forward guidance signaled economic weakness, triggering a flight-to-safety effect and lower expected U.S. inflation. J. Finance 2026, paywalled. Eight core results with source locators, datasets used, the partial-equilibrium model, and the high-frequency local projection method.
# Tags: paper-summary, macro, exchange-rates, monetary-policy, flight-to-safety
==============================================================================

**What this is.** The paper's core results, the partial-equilibrium model of the
information channel, and the high-frequency local projection method: enough to
understand what was found and why, without reading all 40 pages. To replicate
or extend the results, read the full source at the
[original](https://doi.org/10.1111/jofi.70025).

## TL;DR

During the Great Recession (December 2008 to September 2012), U.S. forward
guidance (FG) easings caused the dollar to persistently appreciate against both
advanced-economy and emerging-market currencies, three to four weeks after each
FOMC announcement. This is the opposite of the conventional wisdom that
monetary easing depreciates the domestic currency. The appreciation coincided
with the release of detailed FOMC meeting minutes and was driven entirely by
FG, not QE. Outside the Great Recession, FG easings caused equity prices to fall (information effect); during the Great Recession, the pattern reversed and equity prices rose in response to FG easings, with the switch occurring about three to four weeks post-announcement. The paper builds a partial-equilibrium model showing this occurs when FG has a
dominant information effect: calendar-based FG signals economic weakness,
triggering a flight to safety, raising investor risk aversion, and lowering
expected U.S. inflation relative to foreign inflation. Cross-sectional
heterogeneity confirms the mechanism: the dollar appreciated most against
currencies that are poor hedges (those that tend to depreciate when the U.S.
economy contracts), consistent with a flight-to-safety story.

## Core results

Magnitudes and significance are as reported; `*`/`**`/`***` = 10%/5%/1%.
Locators point into the source PDF. Coefficients in Tables I-III are multiplied
by 100 for presentation.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | During the Great Recession, a positive U.S. FG surprise (unexpected easing) led to a **persistent and large dollar appreciation** against advanced-economy currencies (opposite to pre- and post-GR periods) | Table I, p. 980; Figure 1, p. 979 | Non-QE GR coefficient at horizon 50 days: -37.68\*\*\* (SE 11.89); pre-GR coefficient at same horizon: 2.26 (0.00 R2); post-GR: 8.11 |
| R2 | The sign reversal against **emerging-market currencies** mirrors the advanced-economy result, with a similarly large eventual depreciation against USD | Table II, p. 981; Figure 1, p. 979 | Non-QE GR coefficient at horizon 50 days: -34.27\*\*\* (SE 10.18); pre-GR: 3.12; post-GR: 13.99\* |
| R3 | The sign reversal is driven **entirely by FG, not QE**: using Swanson (2021) LSAP and FG factors, FG carries the appreciation and LSAP does not | Table III, p. 983; Figure 2, p. 982 | FG coefficient for advanced economies at 90 days: -0.72 (0.72 SE); LSAP coefficient: 1.07\* (0.53 SE). FG dominates at intermediate horizons (40-80 days): e.g. day 40, FG: -0.85\* (0.47), LSAP: 0.43 (0.25) |
| R4 | The dollar appreciation is **larger for currencies that are worse hedges** (higher covariance with U.S. SDF proxies), consistent with the flight-to-safety channel | Figure 3, p. 985; Figure 4, p. 986 | A one-standard-deviation worse hedge quality is associated with about 10 percentage points larger dollar appreciation at 90 days; statistically significant for SDF-covariance measures |
| R5 | FG easings during the GR caused **equity prices to rise** (S&P 500 and MSCI World peak ~30 pp for all announcements and ~100 pp for non-QE surprises at ~40-60 days post-announcement); outside GR, easing led to equity price declines (information effect) | Figure 5, p. 987; Figure 6, p. 988 | S&P 500 non-QE GR: positive response peaking near 100 pp at ~40-60 days; all-announcement GR: positive response peaking near 30 pp; conventional negative near-term response of close to 10 pp reversed after the minutes release; pre- and post-GR coefficients negative (easing → equity falls) |
| R6 | The paper interprets the response of its risk-aversion proxies (VIX and the Bekaert-Engstrom-Xu index) to FG easings during the GR as **increased risk aversion**, three to four weeks post-announcement | Figure 7, p. 1001 | Peak effect at ~60 days: estimated log VIX response is nearly -50 pp and log BEX risk aversion index about -25 pp (coefficient negative and statistically significant for non-QE surprises); the paper reads this negative estimated response, through its flight-to-safety mechanism, as increased risk aversion (p. 1001, and pp. 973-974) |
| R7 | The structural break in exchange rates is driven by **currency risk premia and inflation expectations, not the nominal interest rate component** | Figure 8, p. 1004; §III.B, p. 1002 | Interest rate component response is small and conventional in sign across all subsamples; currency premia and inflation residual carries the sign reversal during GR (Figure 8, Panel B) |
| R8 | The structural break in equity prices is driven by **equity risk premia and dividend growth expectations**, not the nominal rate component | Figure 9, p. 1005; §III.B, p. 1002 | Interest rate component of equity price response is small and pushes in the conventional direction; equity risk premia and dividend growth component carries the large positive response (equity prices rising in response to FG easings) during GR, while it is the dominant channel for the negative responses outside GR |

**Overall (paper's conclusion).** The Great Recession was a period where the
information content of U.S. forward guidance dominated the direct interest-rate
channel: investors interpreted FG promises of low rates as bad news about
growth, raising risk aversion and triggering a flight to safety. This generated
dollar appreciation in response to policy easings, contrary to standard
macroeconomic predictions. This contrasts with Rogers, Scotti, and Wright
(2018), who find conventional-sign dollar responses to unconventional policy
at longer horizons using a monthly VAR with sign restrictions that rule out
the information channel by assumption. Equity prices also rose in response to FG easings
during this period (reversing the pre- and post-GR pattern where easings led to
equity price declines due to the information effect). The finding implies that
models of exchange rates and asset prices need to incorporate the information
channel of monetary policy, especially during periods of high fundamental
uncertainty and calendar-based forward guidance.

## Theory / model

The paper proposes a partial-equilibrium daily model of the U.S. economy and
a marginal investor SDF. The model is purposefully stripped down to illustrate
how the information channel can qualitatively reconcile the empirical facts.
Building on the information-channel model of monetary policy signaling in
Tang (2015), it adds a partial-equilibrium extension with exchange rates and
stock prices.
The signaling model of monetary policy with heterogeneous agents in Melosi
(2017) co-inspires the theoretical framework, and the forward guidance with
heterogeneous beliefs in Andrade et al. (2019) informs the information-effect
model structure.

**U.S. macroeconomic block** (pp. 989-990, eqs. 3-5). Log inflation and real
output follow:

$$
\pi^{us}_t = \alpha \, y^{us}_t \tag{3}
$$

$$
y^{us}_t = -v(i^{us}_t - \pi^{us}_t) + \varepsilon^{y,us}_t \tag{4}
$$

with a Taylor rule:

$$
i^{us}_t = \phi^y y^{us}_t + \phi^\pi \pi^{us}_t + \varepsilon^{mp,us}_t
$$

where $$\varepsilon^{y,us}_t \sim N(0, \sigma^2_y)$$ is the demand shock and
$$\varepsilon^{mp,us}_t \sim N(0, \sigma^2_{mp})$$ is the monetary policy shock.
Solving the system (eq. 5, p. 990):

$$
\begin{aligned}
y^{us}_t  &= \frac{\varepsilon^{y,us}_t - v\,\varepsilon^{mp,us}_t}{1+v\kappa} \\
\pi^{us}_t &= \frac{\alpha(\varepsilon^{y,us}_t - v\,\varepsilon^{mp,us}_t)}{1+v\kappa} \\
r^{us}_t  &= \frac{\kappa\,\varepsilon^{y,us}_t + \varepsilon^{mp,us}_t}{1+v\kappa}
\end{aligned}
$$

where $$\kappa = \phi^y + \alpha(\phi^\pi - 1)$$.

**Investor SDF and pricing conditions** (pp. 990-991, eqs. 6-7). The real
log SDF is $$\text{sdf}_{t,t+1} = \ln(\beta) - \rho_t \Delta c_{t+1} - c_t \Delta \rho_t$$
with CRRA preferences. The exchange rate pricing condition is:

$$
E_t\!\left[ \text{SDF}_{t,t+1} \cdot e^{-\pi_{t+1}} \cdot \left((1+i^{us}_t) - \frac{S_{t+1}}{S_t}(1+i^i_t)\right) \right] = 0 \tag{6}
$$

and the equity pricing condition:

$$
E_t\!\left[ \text{SDF}_{t,t+1} \cdot \left(e^{r^{eq}_{t+1}} - (1+i^{us}_t)\right) \right] = 0 \tag{7}
$$

**Currency risk premium** (p. 991, eqs. 8-10). The expected excess return
from being long the U.S. bond and short currency $$i$$:

$$
\lambda_t = E_t[\Delta s_{t+1}] - \tilde{i}_t
          = \frac{\sigma^2_s}{2} + \sigma_{\pi,s} + \rho_t \, \sigma_{c,s} \tag{10}
$$

**Risk aversion is countercyclical** (eq. 12, p. 991):

$$
\rho_t = \rho_y \, y^{us}_t, \quad \rho_y < 0 \tag{12}
$$

so a negative demand shock raises risk aversion, consistent with habit
formation and intermediary asset pricing.

**Information channel mechanics** (pp. 992-995, eqs. 14-16). The central
bank receives a private signal $$\tilde{\varepsilon}^{y,us,CB}_{t+1} = \varepsilon^{y,us}_{t+1} + \hat{\varepsilon}_{t+1}$$
about future demand. Investors initially believe the signal is less precise
(market perception variance inflated by $$q^{CB}_{t+l} = \hat{q}_{CB} > 0$$ for
$$l < l^m$$, where $$l^m$$ is the date of the FOMC minutes release). At the
FG announcement, investors apply Bayes' rule:

$$
E_{t+l}[y^{us}_{t+h}] - E_t[y^{us}_{t+h}] = K_{t+l} \cdot a_{t+1}, \quad 1 \leq l < h \tag{15}
$$

where $$K_{t+l}$$ (eq. 16, p. 994) is:

$$
K_{t+l} = \frac{(\kappa+\alpha)\,\sigma^2_y/\sigma^2_{mp} - \left((\sigma^2_{\varepsilon,CB}+q^{CB}_{t+l})/\sigma^2_y + 1\right)v\eta}{(\kappa+\alpha)^2\,\sigma^2_y/\sigma^2_{mp} + \left((\sigma^2_{\varepsilon,CB}+q^{CB}_{t+l})/\sigma^2_y + 1\right)\eta^2} \tag{16}
$$

When $$\sigma^2_y/\sigma^2_{mp}$$ is large (high fundamental uncertainty relative to
policy uncertainty), $$K_{t+l} > 0$$: a negative FG surprise (lower rates
promised) causes investors to revise down expected future GDP, raising risk
aversion and triggering the flight-to-safety channel. The bias $$q^{CB} > 0$$
makes $$K < 0$$ on the FOMC day itself but turns positive once minutes are
released (at $$l = l^m$$, $$q^{CB} = 0$$), generating the delayed reversal in
estimated impulse responses.

**Exchange rate decomposition** (pp. 995-997, eqs. 17-21). The exchange
rate change is decomposed into:

$$
\Delta s_{t+l,t} = \psi^{EH}_{t+l,t} + \psi^{\lambda}_{t+l,t} + \psi^{LR}_{t+l,t}
$$

- $$\psi^{EH}_{t+l,t}$$: interest rate differential component
- $$\psi^{\lambda}_{t+l,t}$$: currency risk premium component
- $$\psi^{LR}_{t+l,t}$$: long-run inflation / real exchange rate component

The derivative of the currency risk premium component with respect to the
FG announcement (eq. 20, p. 996):

$$
\frac{d\,\psi^{\lambda}_{t+l,t}}{d\,a_{t+1}} = -\rho_y \, \sigma_{c,s} \, K_{t+l} \tag{20}
$$

Since $$\rho_y < 0$$ and $$K_{t+l} > 0$$ when the information channel dominates,
this derivative is positive (dollar appreciates) for currencies where
$$\sigma_{c,s} < 0$$ (currencies that are poor hedges: they tend to depreciate
when U.S. consumption/output falls). The interest rate component is
unconditionally positive and independent of $$K_{t+l}$$, so it pushes in the
conventional direction; the structural break requires the flight-to-safety and
inflation channels to dominate.

## Method

The estimation strategy is the lag-augmented local projection (LALP) of
Montiel Olea and Plagborg-Moller (2021), applied to high-frequency monetary
policy surprises. The high-frequency monetary policy identification using
interest-rate futures, and the "information effect" framing, follows Nakamura
and Steinsson (2018). The approach builds on `event-study` identification and
`panel-regression` for the cross-currency specification, and uses
`affine-term-structure` models disciplined by survey forecasts for the channel
decomposition.

**Monetary policy surprise** (p. 977). The surprise $$mp_\tau$$ is the change in
Eurodollar futures expiring three quarters hence (ED4) over a one-hour window
(15 minutes before to 45 minutes after) around FOMC announcements and
QE announcements made outside regular FOMC meetings. ED4 captures
unconventional policy during the ZLB period (Swanson (2021)).

**Baseline impulse response regression** (eq. 1, p. 977):

$$
\bar{s}_{\tau+n} - \bar{s}_{\tau-1} = \alpha_n + \beta^{\Delta s}_n \, mp_\tau + \gamma_n(\bar{s}_{\tau-1} - \bar{s}_{\tau-2}) + \text{error}_{\tau,n} \tag{1}
$$

- $$\bar{s}_t = \frac{1}{K}\sum_{k=1}^K s_{k,t}$$: average log exchange rate across $$K$$ currencies (units of currency $$k$$ per USD)
- $$\beta^{\Delta s}_n$$: impulse response at horizon $$n$$ ($$n = 0,\ldots,90$$ days)
- Sample: full sample with pre-GR, GR, and post-GR subsamples estimated separately
- SE: Newey-West, lag length set to the maximum number of overlapping FOMC dates within the estimation window for each horizon

**Cross-currency heterogeneity regression** (eq. 2, p. 984):

$$
s_{k,\tau+n} - s_{k,\tau-1} = \alpha_{k,n} + \beta^{\Delta s}_n \, mp_\tau + \beta^{\text{hedge}}_n \cdot \text{hedge}_k \cdot mp_\tau + \gamma_{n,h}(s_{k,\tau-1} - s_{k,\tau-2}) + \text{error}_{\tau,n} \tag{2}
$$

- $$\text{hedge}_k$$: currency $$k$$'s hedging quality (covariance of exchange rate change with the SDF proxy: log S&P 500, log intermediary capital ratio, or U.S. minus country $$k$$ average interest rate differential), standardized to unit variance
- $$\beta^{\text{hedge}}_n$$: measures how the response to tightening varies with hedge quality
- Fixed effects: currency-specific intercept $$\alpha_{k,n}$$
- Sample: GR non-QE subsample (19 dates)

**Channel decomposition** (eqs. 23-24, p. 1002). The estimated
$$\beta^{\Delta s}_n$$ is decomposed by replacing the dependent variable with each
exchange rate change component estimated from an affine term structure VAR
disciplined by Blue Chip Financial Forecasts survey data:

$$
\beta^{\Delta s}_n = \beta^{\psi^{EH}}_n + \beta^{\psi^{\lambda}+\psi^{LR}}_n \tag{23}
$$

and analogously for equity prices:

$$
\beta^{\Delta p^{eq}}_n = \beta^{\psi^{EH,eq}}_n + \beta^{\psi^{\lambda,eq}+\psi^{D,eq}}_n \tag{24}
$$

**FG vs QE separation** (pp. 977-978). Two complementary approaches: (i)
restrict to FOMC announcement dates without QE announcements (19 of 33 GR
dates); (ii) use the Swanson (2021) principal-component-identified FG and LSAP
factors with sign restrictions.

## Empirical specifications

All specifications use daily frequency data. The main sample for exchange rates
is 1990 through 2019 (173 pre-GR observations, 33 GR observations, 60 post-GR
observations). The GR subsample spans December 16, 2008 through September 13,
2012.

**Exchange rate outcome (R1, R2).** Baseline regression (eq. 1) using average
log exchange rate changes against 9 advanced-economy currencies (AUD, CAD, CHF,
EUR/DEM pre-1999, GBP, JPY, NOK, NZD, SEK) and 15 emerging-market currencies
(BRL, CLP, COP, CZK, ILS, INR, ISK, KRW, MXN, PHP, RUB, SGD, THB, TRY, ZAR).
Pegged-regime observations excluded. Newey-West SE with horizon-specific lags.
Fixed effects: date-of-announcement intercept $$\alpha_n$$.

**FG vs. LSAP factor comparison (R3).** Same dependent variable, but monetary
policy surprise replaced by the (negated) LSAP factor and the FG factor from
Swanson (2021). Sample: July 5, 1991 through June 19, 2019 with GR subsample
December 16, 2008 through September 13, 2012. 30 GR observation dates for each
factor (Table III, p. 983).

**Cross-currency heterogeneity (R4).** Panel regression (eq. 2) on GR non-QE
subsample (19 dates), with currency-specific fixed effect $$\alpha_{k,n}$$ and
interaction $$\beta^{\text{hedge}}_n$$. Three hedge quality proxies used separately
(Figure 3, p. 985; Figure 4, p. 986).

**Risk aversion outcome (R6).** Dependent variable replaced by log VIX or log
Bekaert-Engstrom-Xu (2022) risk aversion index. Otherwise same regression as
eq. 1. Full sample 1990-2019 (Figure 7, p. 1001).

**Channel decomposition (R7, R8).** Affine term structure VAR (three-factor,
discipline by Blue Chip Financial Forecasts) for advanced economies to extract
interest rate component. Residual is the currency risk premium and inflation
component (exchange rates) or the equity risk premium and dividend growth
component (equity prices). Results in Figures 8-9, pp. 1004-1005. Robustness:
measurement error in interest rate expectations proxied and found to push
against the main findings (footnote 33, p. 1003).

**Uncertainty measures (R3, §III.C).** Jurado-Ludvigson-Ng (2015) 12-month
macroeconomic uncertainty, GDP forecast dispersion from Blue Chip Financial
Forecasts (25th-75th percentile range), and Baker-Bloom-Davis (2016) monetary
policy uncertainty index, all standardized 1990-2019 (Table IV, p. 1006).
GR subsample means significantly higher for macro uncertainty and GDP
dispersion; monetary policy uncertainty declines slightly, supporting the
high-$$\sigma^2_y / \sigma^2_{mp}$$ interpretation.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Eurodollar futures (ED4), daily, 15-min window around FOMC | Monetary policy surprise measure; identification instrument | no page yet |
| Daily bilateral nominal exchange rates (24 currencies, 1990-2019) | Primary outcome variable (log changes against USD) | no page yet |
| Swanson (2021) FG and LSAP factors | Alternative surprise decomposition separating forward guidance from QE | no page yet |
| S&P 500 and MSCI World daily total return indices | Equity price outcome variable | no page yet |
| VIX (CBOE Volatility Index), daily | Risk aversion proxy; additional testable implication | [VIX](/wiki/datasets/vix/) |
| Bekaert, Engstrom, and Xu (2022) risk aversion index, daily | Risk aversion proxy (alternative to VIX) | no page yet |
| Blue Chip Financial Forecasts survey (GDP forecasts) | Discipline affine-term-structure VAR; measure GDP forecast dispersion | [no page yet] |
| Jurado, Ludvigson, and Ng (2015) macro uncertainty index | Measure macroeconomic uncertainty; explain why GR was special (Table IV) | no page yet |
| Baker, Bloom, and Davis (2016) monetary policy uncertainty index | Measure monetary policy uncertainty; complement to JLN in Table IV | no page yet |
| He, Kelly, and Manela (2017) intermediary capital ratio | Proxy for marginal investor SDF (alternative hedge quality measure) | no page yet |

Sample: 1990 through 2019 for most series; GR subsample December 16, 2008
through September 13, 2012 (33 FOMC-adjacent observations). Pegged-regime
observations excluded per Ilzetzki, Reinhart, and Rogoff (2022) classification.

## When to read the full paper

Use the [original](https://doi.org/10.1111/jofi.70025) if you are:
decomposing the channels through which FG shocks transmit to asset prices (the
Internet Appendix contains the affine term structure VAR, formal derivations,
and robustness tables); extending the model to binding-ZLB environments or
non-U.S. central banks; building on the cross-currency hedging heterogeneity
result (Figures 3-4 and the supporting regressions); or evaluating whether
the information channel was also operative outside the Dec 2008 to Sep 2012
window. The locators above point to the exact tables and figures.

## Attribution and rights

Source: peer-reviewed, *The Journal of Finance* 81(2). This distillation was
extracted by an LLM on 2026-06-01 and is **not human-verified or independently
reproduced**. The paper is paywalled; no PDF is hosted here. Extract only.

> Stavrakeva, Vania, and Jenny Tang. "The Dollar during the Great Recession:
> The Information Channel of U.S. Monetary Policy and the 'Flight to Safety'."
> *The Journal of Finance* 81, no. 2 (April 2026): 971-1010.
> DOI: 10.1111/jofi.70025. (c) 2026 the American Finance Association.
> All rights reserved. Paywalled; this page is an extract-only distillation.


==============================================================================
# Policy News and Stock Market Volatility: Baker, Bloom, Davis & Kost (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jfe/2026/baker-policy-news-stock-market-2026/
# Distilled: Baker, Bloom, Davis and Kost build newspaper-based Equity Market Volatility (EMV) trackers that track the VIX with R-squared above 0.60 in-sample and 0.55 out-of-sample through 2023; policy news accounts for 35-55% of EMV articles; category EMV trackers combined with 10-K exposures explain cross-sectional realized volatility. Journal of Financial Economics 2026, paywalled. Six core results with source locators, datasets used, the tracker construction, and empirical specifications.
# Tags: paper-summary, macro, equities, volatility, text-as-data, return-predictability
==============================================================================

**What this is.** This distillation captures the core findings, tracker construction, and empirical specifications of Baker, Bloom, Davis and Kost (2026). To replicate or extend the work, read the full source at the [original](https://doi.org/10.1016/j.jfineco.2025.104187). The EMV tracker and its extensions are updated at [www.policyuncertainty.com](https://www.policyuncertainty.com).

## TL;DR

Baker, Bloom, Davis and Kost construct an Equity Market Volatility (EMV) tracker by counting U.S. newspaper articles that discuss economic conditions, stock market movements, and volatility. Running from January 1985 to December 2023 across eleven major U.S. newspapers, the monthly EMV tracker correlates approximately 0.80 with the VIX and achieves R-squared of 0.60 in contemporaneous regressions. The methodology was finalized in 2018 and first published in a 2019 NBER working paper; data from 2019 onward are fully out-of-sample, and the tracker continues to achieve R-squared above 0.55 through year-end 2023 despite COVID-19, the Russia-Ukraine war, and multiple other episodes. The tracker is decomposed into roughly 40 category-specific EMV trackers covering macroeconomic news, monetary policy, fiscal policy, regulation, and other topics; policy-related categories collectively account for 35-55% of EMV articles, with peaks during 2001-03 (9/11 and Iraq), 2011-12 (debt-ceiling crisis), and the first Trump presidency. Combined with firm-level risk disclosures from 10-K Part 1A filings, the category EMV trackers explain cross-sectional realized volatility and co-movement in daily stock returns, even after conditioning on firm and time fixed effects.

## Core results

Magnitudes and significance as reported; `*`/`**`/`***` = 10%/5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **EMV tracker tracks monthly VIX in-sample** (1985-2023): contemporaneous OLS | Table 1, col 1, p. 6 | Slope = 0.745\*\*\* (SE 0.053), R² = 0.603, 468 monthly obs |
| R2 | **EMV tracker tracks monthly VIX out-of-sample** (2019-2023): term sets finalized 2018, data from 2019 onward used for testing only | Table 2, col 5, p. 6 | Slope = 0.714\*\*\* (SE 0.085), R² = 0.558 (vs R² = 0.606 in-sample 1985-2018) |
| R3 | **EMV lagged averages retain predictive power at multi-year VIX horizons**: even the 12-month lagged average remains significant at 10-year horizon | Table 3a, cols 4-7, p. 8 | R² = 0.691 (1-year VIX), 0.607 (3-year), 0.534 (5-year), 0.334 (10-year); Newey-West SE |
| R4 | **EMV tracker predicts future S&P 500 returns**: higher EMV foreshadows higher annualized returns at 3-month to 2-year horizons | Table 4, p. 8 | Slope = 0.0857\* at 3-month, 0.0590\*\* at 6-month, 0.0470\*\* at 1-year, 0.0298\*\* at 2-year |
| R5 | **Composite firm-level 10-K exposure explains cross-sectional realized volatility**, conditional on firm and time fixed effects | Table 5, col 1, p. 12 | Composite exposure coefficient = 2.16\*\*\* (SE 0.22); R² = 0.546; 508,447 firm-months |
| R6 | **EMV tracker explains average pairwise return correlations**: firms sharing a leading EMV category comove more strongly when that category's EMV is higher | Table 6, col 1, p. 13 | Coefficient on ln(EMV) = 4.24\*\*\* (SE 0.020); R² = 0.226; doubling ln(EMV) raises avg pairwise correlation ~4.24 pp |

**Overall (paper's conclusion).** The EMV tracker is a simple, transparent, and scalable measure of equity market volatility that correlates closely with the VIX in and out of sample. Policy news is a major and time-varying source of stock market volatility; monetary policy and tax policy are the most important policy-related sources, followed by regulation. Category-specific EMV trackers, combined with firm-level 10-K risk exposures, explain the cross-sectional structure of realized volatility and its evolution over time.

## Theory / model

The paper does not develop a formal structural model. Instead it documents empirical patterns under two competing interpretations of stock market volatility, following the framing of Shiller (1981):

1. **Efficient markets view**: equity price movements reflect genuine news about future cash flows and discount rates. Under this view, the EMV tracker provides a catalog of specific news items and economic developments that shift rational investor beliefs.

2. **Animal spirits view** (referencing Shiller 2014 and Keynes): market fluctuations are partly driven by shifts in investor mindsets unrelated to fundamentals. Under this view, the newspaper articles captured by EMV reflect and amplify these mindset shifts over time.

The paper treats both views as consistent with the data and does not attempt to resolve the debate. The core empirical claim is that EMV articles identify the proximate drivers of VIX fluctuations regardless of which interpretation is correct.

Niederhoffer (1971) was an early study linking newspaper headlines to U.S. stock market movements (from 1950 to 1966); the EMV tracker extends this approach with algorithmic term selection and a scalable multi-paper construction running to the present.

**Tested hypotheses:**
- $$H_1$$: The EMV frequency-count tracker correlates with implied and realized stock market volatility in-sample and out-of-sample across multiple horizons.
- $$H_2$$: Policy-related EMV categories (fiscal, monetary, regulation, national security) account for a major and time-varying share of overall EMV articles.
- $$H_3$$: Firm-level EMV category exposures from 10-K Part 1A text combined with category EMV trackers explain firm-level realized volatility and pairwise return correlations after conditioning on firm and time fixed effects.

## Method

**EMV tracker construction.** Following Baker et al. (2016), the tracker is built from scaled article counts in leading U.S. newspapers containing terms from three overlapping sets (pp. 3-4):

- **E** (Economic): {economic, economy, financial}
- **M** (Market): {stock market, equity, equities, S&P, "Standard and Poors" and variants}
- **V** (Volatility): {volatility, volatile, uncertain, uncertainty, risk, risky}

The best-fit permutation is selected from $$2^5 \times 2^6 = 2048$$ candidate combinations (all elements of $$\mathbf{M}' \times \mathbf{V}'$$) by maximizing the R-squared in an OLS regression of the 30-day VIX on the candidate tracker using monthly data from 1990 to 2015. For each newspaper and month, the raw count of articles containing at least one term from each of E, M, and V is divided by the total count of all articles in the same newspaper-month, standardized to unit standard deviation per newspaper, and averaged across the eleven newspapers. The series is then multiplicatively rescaled to match the mean VIX value from 1985 to 2015.

**Category-specific EMV trackers.** To decompose aggregate EMV by topic, each EMV article is classified into roughly 40 categories (approximately 20 general economic, approximately 20 policy-related) by checking whether the article contains terms from a category-specific term set $$b$$. The share of articles in category $$b$$ in month $$t$$ times the overall EMV tracker gives the category-specific tracker (p. 4):

$$\left(\frac{\#\{E \cap M \cap V \cap b\}_t}{\#\{E \cap M \cap V\}_t}\right) EMV_t$$

where $$\#\{\cdot\}_t$$ counts articles satisfying all conditions in month $$t$$. The **Monetary Policy** term set includes: monetary policy, money supply, open market operations, fed funds rate, discount window, quantitative easing, forward guidance, interest on reserves, taper tantrum, Fed chair names, central bank names, and many others (pp. 3-4, Appendix B).

**Firm-level exposure measure.** Following the approach of Davis et al. (2021), who use Part 1A of 10-K filings to explain firm-level stock price volatility in the wake of COVID-19, the paper measures each firm's exposure to EMV categories (p. 11). For firm $$i$$, fiscal year $$y$$, and EMV category $$b$$:

$$F_{iy}^b = \frac{\#\{\text{sentences pertaining to EMV category } b\}_{iy}}{\#\{\text{total sentences in Part 1A of 10K}\}_{iy}} \tag{1}$$

Firms with the largest Part 1A sentence share in a given category are treated as most exposed to that category's volatility driver. LASSO is used in one robustness specification (Table 5, col 5) to select the most informative categories from among 38 candidate exposure measures.

## Empirical specifications

**VIX tracking regression (R1, R2).** The baseline specification regresses contemporaneous implied or realized stock market volatility on the EMV tracker (Tables 1-2):

$$VIX_t = \alpha + \beta \cdot EMV_t + \varepsilon_t \tag{2}$$

with heteroskedasticity-robust standard errors. Monthly frequency, January 1985 to December 2023 (in-sample). The out-of-sample test (R2) uses data from January 2019 to December 2023 (60 monthly observations), since the methodology and term sets were finalized in 2018. Log-log specifications and daily data yield similar results (Table 1, cols 4-8).

**Long-horizon VIX regression (R3).** Time-$$t$$ implied VIX at horizons $$h$$ from 1 month to 10 years is regressed on contemporaneous EMV and lagged EMV averages (Table 3a, p. 8):

$$VIX_t^h = \alpha^h + \beta_0 EMV_t + \beta_1 \overline{EMV}_{t,3} + \beta_2 \overline{EMV}_{t,12} + \varepsilon_t^h$$

where $$\overline{EMV}_{t,k}$$ is the simple mean of $$EMV_{t-1}, \ldots, EMV_{t-k}$$. Newey-West standard errors with maximum autocorrelation lag of 2. Data: January 1996 to February 2023 (columns 1-4) and November 2002 to July 2016 (columns 5-7, restricted by availability of multi-year VIX data).

**Return predictability regression (R4).** Annualized S&P 500 returns from month $$t$$ to $$t+\tau$$ are regressed on lagged EMV (Table 4):

$$r(t \to t+\tau) = \mu + \delta \cdot EMV_{t-1} + \varepsilon_t$$

with Newey-West standard errors at lag equal to the horizon $$\tau$$ (3 months, 6 months, 1 year, 2 years). Monthly data, January 1985 to December 2023.

**Firm-level volatility panel regression (R5).** The composite firm-level exposure measure is constructed by weighting the category EMV trackers by each firm's Part 1A exposure shares (p. 11-12, specification 1):

$$\sigma_{it} = \alpha_i + \gamma_t + \beta \sum_b F_{iy}^b \cdot EMV_t^b + \varepsilon_{it} \tag{3}$$

where $$\sigma_{it}$$ is the realized volatility (standard deviation of daily equity returns) for firm $$i$$ in month $$t$$, $$\alpha_i$$ is a firm fixed effect, $$\gamma_t$$ is a time fixed effect, and $$F_{iy}^b$$ is the Part 1A exposure share for firm $$i$$ in fiscal year $$y$$ under EMV category $$b$$. Each firm-month observation is weighted by the firm's lagged log market capitalization times the square root of the number of Part 1A sentences, placing more weight on firms with more informative filings. Standard errors are clustered at the firm level. Sample: 10-K filings issued 2006 to 2019 (fiscal years 2005-2018), 508,447 firm-months. Realized volatility is winsorized at the 1% and 99% levels.

**Pairwise correlation regression (R6).** For each firm-month, the firm's "leading EMV category" $$l$$ is the category most discussed in its most recent Part 1A filing. Average pairwise daily return correlations among firms sharing leading category $$l$$ in month $$t$$ are regressed on the log of the corresponding EMV tracker (Table 6, p. 13):

$$\bar{\rho}_{lt} = \mu + \delta \ln(EMV_t^{b-l}) + \varepsilon_{lt}$$

where $$\bar{\rho}_{lt}$$ is the average pairwise correlation of daily returns in month $$t$$ among firms assigned to leading category $$l$$. All columns include firm fixed effects; some specifications also add the contemporaneous VIX and time fixed effects. The sample mean of the dependent variable is 0.21.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| 11 major U.S. newspapers (ProQuest and Newsbank databases; Boston Globe, Chicago Tribune, Dallas Morning News, Houston Chronicle, LA Times, Miami Herald, NYT, SF Chronicle, USA Today, WSJ, Washington Post) | Article-count source for EMV tracker construction; tracker publicly available at www.policyuncertainty.com | No page yet (licensed commercial newspaper archives) |
| CBOE VIX / VXO (daily 1990-2023; extended to 1985 using Berger et al. 2019) | Dependent variable in VIX tracking regressions (Tables 1, 2, 3a) | No page yet |
| S&P 500 daily and monthly returns | Realized volatility (RVol) dependent variable; future return prediction target (Tables 1, 4) | No page yet |
| SEC EDGAR 10-K filings, Part 1A (2006-2019) | Firm-level risk exposure measures $$F_{iy}^b$$ for cross-sectional volatility and correlation regressions (Tables 5, 6) | [EDGAR](/wiki/datasets/edgar/) |
| FRED (crude oil realized volatility and WTI series) | Petroleum markets EMV tracker validation (Fig. 3, Section 3.8) | [FRED](/wiki/datasets/fred/) |

Sample (EMV tracker): January 1985 to December 2023, 11 U.S. newspapers, daily and monthly frequency.
Sample (firm-level analysis): 508,447 firm-months, fiscal years 2005-2018 (10-K filings issued 2006-2019).

## When to read the full paper

Use the [original](https://doi.org/10.1016/j.jfineco.2025.104187) if you are: (a) building or extending text-based volatility trackers for equity, commodity, or country markets and need the full term-set specifications and 40-category taxonomy (Appendix B); (b) studying the sources of stock market volatility and the role of policy news vs. economic fundamentals vs. animal spirits; (c) constructing firm-level risk exposure measures from SEC filings to explain cross-sectional return variation (Tables 5-6 and Appendix D detail the firm-level data construction); or (d) comparing EMV against alternative news-based volatility measures such as the NVIX of Manela and Moreira (2017) (Section 3.6 and Appendix Figures A.4-A.6). The Internet Appendix also contains the historical EMV tracker back to 1928 using ProQuest Historical Archive, and daily EMV using the Newsbank World News database.

## Attribution and rights

Source: peer-reviewed, *Journal of Financial Economics* 175 (2026), article 104187. Paywalled: © 2025 Elsevier B.V. All rights are reserved. This distillation was extracted by an LLM on 2026-06-24 and is **not human-verified or independently reproduced**.

> Baker, Scott R., Nicholas Bloom, Steven J. Davis, and Kyle Kost.
> "Policy news and stock market volatility."
> *Journal of Financial Economics* 175 (2026) 104187.
> DOI: [10.1016/j.jfineco.2025.104187](https://doi.org/10.1016/j.jfineco.2025.104187).
> Paywalled. Extract-only; no redistribution of the verbatim PDF.


==============================================================================
# Teams and Belief Overreaction: Barahona, Cassella, Jansen & Pezone (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jfe/2026/barahona-teams-alleviate-exacerbate-overreaction-2026/
# Distilled: Preregistered lab experiments and US mutual fund data show that two-person teams reduce individual belief overreaction to past returns by 30 to 55 percent, with self-selection into team leadership accounting for roughly 70 percent of the lab effect. Journal of Financial Economics 176 (2026), paywalled. Six core results with source locators, datasets used, the measurement framework, and the estimating equations.
# Tags: paper-summary, behavioral-finance, expectations, overreaction, extrapolation
==============================================================================

**What this is.** The paper's core results, the measurement framework for belief overreaction, the decomposition of the team effect into three channels, and the key estimating equations: enough to understand what was found and how, without reading the full 17-page article. To replicate or extend it, read the original at <https://doi.org/10.1016/j.jfineco.2025.104219>.

## TL;DR

The paper addresses a fundamental question in behavioral finance: does moving from individual to team decision-making amplify or attenuate belief overreaction to recent asset returns? Using preregistered randomized experiments on the Labvanced platform with 1,512 Prolific participants, plus a within-subject field study of US equity mutual fund managers (1980-2018), the paper finds that two-person teams reduce individual overreaction by 30 to 55 percent. A quantitative decomposition, following the approach of Enke et al. (2023), partitions the lab team effect into three channels: internal reflection (the act of pre-team deliberation), self-selection (the tendency of the less-biased member to lead), and external screening (the group interaction itself). Self-selection accounts for roughly 70 percent of the reduction. LLM analysis of roughly 18,000 chat exchanges in the Group treatment corroborates this, and dynamic evidence shows that participants reduce their leadership role after making larger forecast errors. The field results, based on the approach of Bordalo et al. (2020) for identifying overreaction, are consistent: mutual fund teams attenuate extrapolative overreaction by about 55 percent relative to the individual behavior of the same managers, and this attenuation coincides with better investment performance.

## Core results

Magnitudes and significance are as reported; `\*` / `\*\*` / `\*\*\*` = 10% / 5% / 1%. Standard errors in brackets, clustered at the team level (equivalent to individual-level for the Individual treatment). Locators reference the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Teams reduce overreaction by ~30%**: Group treatment dummy lowers the individual overreaction coefficient by 0.092, significant at 1%, robust across demographics, financial sophistication, and income fixed-effects controls | Table 2, p. 6 | Group = -0.092\*\*\* [0.035] (Col 1, n = 704); stable at -0.099\*\*\* to -0.101\*\*\* in Cols 2-4; Individual mean beta = 0.311, SD = 0.424 |
| R2 | **Self-selection explains ~70% of the team effect**: the three-channel decomposition shows internal reflection contributes essentially nothing (-0.001, n.s.) and external screening contributes -0.025 to -0.031 (n.s.), while self-selection contributes -0.067 to -0.090 (significant in most specifications) | Table 3/Eq. (5), p. 8 | Most conservative spec (Col 4): IR = -0.001 (n.s.); SS = -0.068\* [0.039]; ES = -0.031 (n.s.); total team effect = -0.098 [0.035] |
| R3 | **Teams show a weaker recency effect**: fitting the exponential return-extrapolation model separately for each treatment, the recency weight (1 minus the decay parameter) is about 2.4x smaller for Group participants than Individual participants | Fig. 3/Eq. (7), p. 11 | Group lambda2 = 0.953 (recency = 0.047) vs Individual lambda2 = 0.886 (recency = 0.114); Group places more equal weight across all 40 past returns |
| R4 | **Self-selection is dynamic and driven by past errors**: participants who led the team prediction in round t-1 are significantly less likely to lead in round t after the team made a large forecast error in round t-1 | Table 4/Eq. (6), p. 10 | MostVotes\_{t-1} x \|Error\_{t-1}\| coefficient = -4.440\*\*\* [0.645] (Votes, Col 1); -0.137\*\*\* [0.018] (MostVotes dummy, Col 3); n = 7,657 round observations |
| R5 | **Teams achieve higher prediction accuracy**: Group participants generate lower mean-squared error, lower mean-absolute error, and earn a higher experimental bonus than Individual participants | Table 5, p. 10 | Group vs Individual: MSE -157.722\*\*\* [30.394]; MAE -2.015\*\*\* [0.408]; Bonus +$0.059\*\* [0.023]; full controls, n = 703 |
| R6 | **Mutual fund teams attenuate extrapolative overreaction by ~55%**: within-subject comparison of team overreaction and statistical counterfactual individual overreaction shows teams transmit only about 45% of extrapolative behavior, while contrarian (non-overreacting) behavior is fully transmitted | Table 8/Eq. (10-11), p. 15 | IV sum delta0 + delta1 = 0.45 (Col 7); null of full transmission (= 1) rejected at IV p = 0.015 (Col 7) and IV p = 0.018 (Col 8); contrarian-only delta0 not significantly below 1 (IV p = 0.568) |

**Overall (paper's conclusion).** Both in the lab and in the field, teams reduce belief overreaction relative to individuals. The dominant mechanism is self-selection: in two-person teams, the less-biased member tends to take on decision authority. This process is dynamic (driven by past forecast errors and feedback) and is confirmed by LLM-based analysis of chat exchanges. In the field, the attenuation of extrapolative trading by mutual fund teams is associated with better subsequent fund performance, while contrarian (non-overreacting) behavior is preserved.

## Theory / model

The paper has no formal structural model. It motivates belief overreaction with the representativeness-heuristic framework of Barberis (2018), which predicts that investors overextrapolate recent returns, and tests team effects on this well-documented bias.

**Cognitive task and AR(1) process.** The hypothetical stock used in the experiment follows an AR(1) process (p. 4):

$$
x_t = \rho x_{t-1} + \varepsilon_t, \qquad \varepsilon_t \sim \mathcal{N}(0, \sigma^2), \quad \rho = 0.5, \quad \sigma = 20
$$

Given that past returns have only weak predictive power for future returns, the paper sets the rational benchmark at $$\beta_i = 0$$ (the best response when $$\rho$$ is difficult to infer). Any positive $$\beta_i$$ signals overreaction: the participant over-weights the most recent return realization.

**Return-extrapolation model (recency channel).** Following Greenwood and Shleifer (2014), an exponentially-weighted extrapolation model is estimated to decompose overreaction into a level (attribute substitution) and a recency component (Eq. 7, p. 11):

$$
\hat{E}_i x_{i,t+1} = \lambda_0 + \lambda_1 \frac{\sum_{j=0}^{N} \lambda_2^j \, x_{t-j}}{\sum_{j=0}^{N} \lambda_2^j} + \varepsilon_{i,t} \tag{7}
$$

Here $$\lambda_1$$ captures the overall level of attribute substitution (sensitivity to past returns in general) and $$\lambda_2$$ governs the relative importance of more versus less recent returns: as $$\lambda_2 \to 0$$, the most recent observation receives disproportionately more weight (stronger recency effect); as $$\lambda_2 \to 1$$, all past returns receive equal weight. The quantity $$1 - \lambda_2$$ is used as the recency-effect measure.

**Team-effect decomposition.** The aggregate team effect is $$\Delta\beta_G = \bar{\beta}_G - \bar{\beta}_I$$. It is partitioned into three additive channels (Eq. 3, p. 6):

$$
\Delta\beta_G
= \underbrace{(\bar{\beta}_{IR} - \bar{\beta}_I)}_{\Delta\beta_{IR}\,(\text{internal reflection})}
+ \underbrace{(\bar{\beta}_{SS} - \bar{\beta}_{IR})}_{\Delta\beta_{SS}\,(\text{self-selection})}
+ \underbrace{(\bar{\beta}_G - \bar{\beta}_{SS})}_{\Delta\beta_{ES}\,(\text{external screening})} \tag{3}
$$

where $$\bar{\beta}_{IR}$$ is average overreaction in the Internal Reflection treatment (participants forecast individually before seeing their partner's prediction, no team interaction yet), $$\bar{\beta}_{SS}$$ is average overreaction in the Self-Selection treatment (voting mechanism picks the team forecast), and $$\bar{\beta}_G$$ is the average in the actual Group treatment. The design isolates each mechanism: IR captures the effect of deliberate pre-team reasoning; the gap between SS and IR isolates the self-selection mechanism; the residual of G vs SS measures external screening via actual discussion.

## Method

**Lab experiment.** The experiment runs on the Labvanced browser-based platform with subjects recruited via Prolific (US-based, pre-screened for 98% approval rate). Participants observe a 40-period AR(1) return series and use a vertical slider to predict the next-period return; the prediction task repeats for 20 rounds per session. Compensation uses a Brier-style scoring rule following Dwyer et al. (1993) and Afrouzi et al. (2023): $$S_t = 100 \times \max(0,\, 1 - |FE_t|/\sigma)$$, paid as a dollar bonus (mean approximately $6.09). Two key treatments:

- **Individual (I)**: each participant forecasts independently in each round.
- **Group (G)**: two randomly matched participants communicate via a live chat box and must agree on a joint forecast before advancing; both earn the group score.

Two additional preregistered treatments isolate mechanism channels:

- **Internal Reflection (IR, RCT-Id AEARCTR-0013710)**: participants forecast individually first, see their partner's forecast, then make a joint prediction; eliminates the actual chat discussion.
- **Self-Selection (SS, RCT-Id AEARCTR-0014914)**: participants each independently forecast and then allocate 100 votes across the two predictions; the prediction with the most votes becomes the team forecast.

Final sample: 1,512 participants (248 Individual, 456 Group, 405 IR, 403 SS) after quality filtering for abnormally high rates of exactly-correct predictions.

**LLM analysis of chat content.** To quantify self-selection patterns in the Group treatment, the paper uses GPT-4o-mini (07/18/2024) on the roughly 18,000 chat exchanges collected (Fig. 2, p. 9). In a "supervised" pass, the LLM records: (i) the first numeric proposal made and its author, (ii) whether the other participant accepted or counter-proposed. The number of rounds where the first proposal was accepted without counter-proposal (uncontested rounds, mean 16.37 per team) proxies for self-selection intensity. In an "unsupervised" pass, the LLM rates each team's self-selection score on a 0-10 scale (mean 6.56). The two measures correlate at 0.29 (Spearman, p < 0.01, Panel C), validating the LLM-based approach. This LLM analysis is conducted by Enke et al. (2023)-inspired methods adapted to team financial decisions.

**Within-subject field design.** The paper identifies 308 mutual fund teams in which at least one member has also managed a fund individually at some point. The statistical counterfactual $$\hat{\beta}_j^{CF}$$ is the equal-weighted average overreaction of the team's members measured when they manage individually, observing them at the same point in time as the team observation. This within-subject design isolates the team effect from compositional differences between solo- and team-managed funds. An IV strategy based on Jegadeesh et al. (2019) uses disjoint subsamples of the data to construct an instrument for $$\hat{\beta}_j^{CF}$$ that is free of measurement error.

## Empirical specifications

**Overreaction measurement (R1-R5, lab).** For each participant $$i$$, an individual overreaction coefficient $$\hat{\beta}_i$$ is estimated from the individual-level time-series regression (Eq. 1, p. 4):

$$
\hat{E}_i x_{i,t+1} = \alpha_i + \beta_i x_t + \varepsilon_{i,t} \tag{1}
$$

where $$\hat{E}_i x_{i,t+1}$$ is participant $$i$$'s stated prediction for the next-period return and $$x_t$$ is the most recent return. Rational expectations benchmark: $$\beta_i = 0$$. Larger $$\hat{\beta}_i > 0$$ signals stronger overreaction.

**Team effect regression (R1).** The participant-level overreaction coefficients $$\hat{\beta}_i$$ serve as the dependent variable in the cross-sectional regression (Eq. 2, p. 5):

$$
\hat{\beta}_i = \alpha_2 + \gamma G_i + \delta' X_i + \eta_i \tag{2}
$$

where $$G_i = 1$$ for Group treatment participants and $$X_i$$ is a vector of demographic and financial sophistication controls. Standard errors are clustered at the team level. The key estimate is $$\hat{\gamma} = -0.092$$ (Table 2, Col 1), stable from -0.099 to -0.101 with controls (Cols 2-4), corresponding to a 30 percent reduction relative to the Individual mean of 0.311.

**Mechanism decomposition regression (R2).** The sample is expanded to include all four treatments and the regression becomes (Eq. 4, p. 8):

$$
\hat{\beta}_i = \alpha_3 + \gamma_{IR} IR_i + \gamma_{SS} SS_i + \gamma_G G_i + \delta' X_i + \varepsilon_i \tag{4}
$$

where $$IR_i$$, $$SS_i$$, and $$G_i$$ are treatment dummies. The estimated treatment differences map to the three channels in Eq. (3). The most conservative four-control specification gives (Eq. 5, p. 8):

$$
\underbrace{-0.098}_{\Delta\hat{\beta}_G\;[\text{se}=0.035]}
= \underbrace{-0.001}_{\Delta\hat{\beta}_{IR}}
  \underbrace{-0.067}_{\Delta\hat{\beta}_{SS}}
  \underbrace{-0.031}_{\Delta\hat{\beta}_{ES}} \tag{5}
$$

**Dynamic self-selection (R4).** The round-level panel regression identifies how past decision-making errors affect subsequent voting behavior in the Self-Selection treatment (Eq. 6, p. 9):

$$
Y_{i,t} = \alpha + \beta \, \text{MostVotes}_{i,t-1} + \gamma |\text{Error}_{t-1}| + \delta \, \text{MostVotes}_{i,t-1} \times |\text{Error}_{t-1}| + \varepsilon_t \tag{6}
$$

where $$\text{MostVotes}_{i,t-1} = 1$$ if participant $$i$$ cast the most votes (and hence made the team decision) in round $$t-1$$, and $$|\text{Error}_{t-1}|$$ is the absolute team forecast error in round $$t-1$$, demeaned and standardized. The dependent variable $$Y_{i,t}$$ is either the number of votes cast in round $$t$$ (Cols 1-2) or an indicator for being the decision maker in round $$t$$ (Cols 3-4). The interaction term $$\hat{\delta}$$ captures the feedback loop: the decision maker in round $$t-1$$ reduces their vote count in round $$t$$ by 4.44 votes per standard-deviation increase in the team forecast error (Table 4, Col 1).

**Field overreaction measurement (R6).** For each fund $$j$$, the fund's sensitivity of trades to past returns is estimated by a panel regression (Eqs. 8-9, p. 12):

$$
\text{trade}_{s,j,t+1} = \alpha_j + \beta_j^X r_{s,t-4\to t} + \gamma_j^T C_{s,t} + \theta_{jt} + \varepsilon_{s,j,t+1} \tag{8}
$$

$$
\text{trade}_{s,j,t+1} \;\equiv\; \frac{(\text{shares}_{s,j,t+1} - \text{shares}_{s,j,t+1}^{\text{sp(t-adj)}}) P_{s,t+1}}{TNA_{j,t+1}} \tag{9}
$$

where $$r_{s,t-4\to t} = \sum_{l=0}^{3} w_l r_{s,t-l}$$ is the exponentially-weighted four-quarter past return of stock $$s$$ (weights from Greenwood and Shleifer (2014), $$\lambda = 0.56$$), $$C_{s,t}$$ is a vector of stock controls (momentum, stock characteristics), $$\theta_{jt}$$ is a fund-quarter fixed effect, and $$TNA_{j,t+1}$$ is fund net assets. A positive $$\hat{\beta}_j^X$$ characterizes extrapolators; a negative $$\hat{\beta}_j^X$$ characterizes contrarians.

**Team transmission (R6).** The team overreaction $$\hat{\beta}_j^{TM}$$ is regressed on the statistical counterfactual $$\hat{\beta}_j^{CF}$$ (Eq. 10, p. 13):

$$
\hat{\beta}_j^{TM} = \alpha + \delta_0 \hat{\beta}_j^{CF} + \delta_1 \hat{\beta}_j^{CF} \times D_j^E + \delta_2 D_j^E + \delta_3 C_j + \varepsilon_j \tag{10}
$$

where $$D_j^E = 1$$ for extrapolative teams ($$\hat{\beta}_j^{CF} > 0$$) and $$C_j$$ are fund controls. Full transmission of overreaction for extrapolative teams implies $$\delta_0 + \delta_1 = 1$$; attenuation implies $$\delta_0 + \delta_1 < 1$$. Contrarian behavior is fully transmitted if $$\delta_0 = 1$$. The IV estimate (Table 8, Col 7) gives $$\delta_0 + \delta_1 \approx 0.45$$; the null $$\delta_0 + \delta_1 = 1$$ is rejected at $$p = 0.015$$ (IV Col 7) and $$p = 0.018$$ (IV Col 8).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Lab experiment data (Labvanced / Prolific, 2024) | Primary data for Sections 2-2.5: 1,512 participants, four treatments, 20 prediction rounds each; chat transcripts analyzed by LLM | No page yet (original data; replication package on Mendeley Data) |
| CRSP monthly stock returns (via WRDS) | Quarterly stock-level returns for the fund trading regression (Eq. 8); past four-quarter return predictor | [WRDS](/wiki/commercial/wrds/) (licensed) |
| Thomson Reuters Mutual Fund Holdings (via WRDS) | Quarterly holdings of US stocks per fund; used to construct the split-adjusted trade measure (Eq. 9) | [WRDS](/wiki/commercial/wrds/) (licensed) |
| Morningstar | Fund investment objectives, fund family, expense ratios, fund age; used as controls in the field analysis | [Morningstar](/wiki/commercial/morningstar/) (licensed) |
| SEC mandatory fund filings | Fund managerial structure (team vs individual management identification); fund-level panel 1980-2018 | [EDGAR](/wiki/datasets/edgar/) |

Sample (field): 467 unique managers, 847 unique funds, 308 unique team observations, quarterly 1980-2018. Sample (lab): 1,512 participants across four treatments, run June-November 2024.

## When to read the full paper

Read the original at <https://doi.org/10.1016/j.jfineco.2025.104219> if you are studying team effects on belief formation and behavioral biases (Section 2 for the experimental design and Tables 2-5 for the core results); implementing the three-channel decomposition of team effects (Section 2.3 and Eq. 3-5 for the framework); analyzing the dynamic feedback between forecast errors and team leadership roles (Section 2.4 and Table 4); studying how organizational structure (solo vs team management) affects fund manager trading behavior and fund performance in the field (Section 3 and Tables 6-8); or looking for evidence linking overreaction to investment underperformance (Section 3.6.1 and Fig. 4). The Internet Appendix contains preregistration documents, the LLM prompts (Appendix IA3), and robustness checks.

## Attribution and rights

Source: peer-reviewed, *Journal of Financial Economics* 176 (2026), article 104219. This distillation was extracted by an LLM on 2026-06-24 and is **not human-verified or independently reproduced**. All rights reserved; this page is extract-only.

> Barahona, Ricardo, Stefano Cassella, Kristy A.E. Jansen, and Vincenzo Pezone.
> "Do teams alleviate or exacerbate overreaction in beliefs?"
> *Journal of Financial Economics* 176 (2026): 104219.
> DOI: 10.1016/j.jfineco.2025.104219. © 2025 Elsevier B.V. All rights reserved.


==============================================================================
# Social Media as a Bank Run Catalyst: Cookson et al. (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jfe/2026/cookson-social-media-bank-run-2026/
# Distilled: Banks with greater pre-run Twitter exposure lost 4.3 percentage points more stock value during the March 2023 Silicon Valley Bank run; Twitter attention at the hourly frequency predicted lower returns for high-risk banks, while Twitter sentiment did not amplify run risks. Journal of Financial Economics 176 (2026), paywalled. Eight core results with source locators, datasets used, and the estimating equations.
# Tags: paper-summary, bank-runs, social-media, fintech, banking, financial-stability
==============================================================================

**What this is.** The paper's core results, the empirical framework, and the estimating
equations, extracted for quick lookup. To replicate or extend the analysis, read the
original at [doi.org/10.1016/j.jfineco.2025.104218](https://doi.org/10.1016/j.jfineco.2025.104218).

## TL;DR

This paper quantifies social media's role in the March 2023 Silicon Valley Bank (SVB) run
and the broader US regional bank distress that followed, using Twitter data
for 277 publicly traded bank holding companies. The headline finding is that banks with
greater pre-run Twitter exposure (log number of cashtag tweets posted before SVB's failure,
January 1 to February 15, 2023) suffered 4.3 percentage points more stock market loss
during the run. The Twitter effect amplifies classical bank run risks: the interaction
between Twitter pre-exposure, the fraction of uninsured deposits, and mark-to-market asset
losses is large and statistically significant. At the hourly frequency during the run
period, Twitter attention (number of tweets in the prior four hours) predicts lower returns
for high-risk banks but not for low-risk banks. The effect is concentrated in periods when
tweets are highly retweeted, pointing to a social propagation channel. Twitter sentiment
(VADER negative score), by contrast, does not amplify run risks, distinguishing broad
attention from negativity as the operative mechanism.

## Core results

Magnitudes are as reported; `\*\*` = 5%, `\*\*\*` = 1%. Variables marked `(z)` are
standardized to mean zero, SD one.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | A 1-SD increase in Twitter pre-exposure predicts 4.3 pp more stock market loss during the SVB run | Table 4, Panel A, col. 2, p. 13 | 4.317\*\*\* pp (SE = 0.936) |
| R2 | Twitter amplifies classical run risks: triple interaction pre-exposure x %Uninsured x %LossMTM is significant | Table 4, Panel A, col. 4, p. 13 | 2.407\*\*\* (SE = 0.561); pre-exposure x %Uninsured = 2.695\*\*\* (SE = 0.701) |
| R3 | Triple interaction is robust to full controls for size, analyst coverage, deposit franchise, and market returns | Table 4, Panel B, col. 4, p. 14 | 1.660\*\*\* (SE = 0.424); pre-exposure x %Uninsured = 1.641\*\* (SE = 0.613) |
| R4 | Twitter pre-exposure predicts Q1-2023 uninsured deposit outflows through the same triple interaction | Table 5, col. 1, p. 16 | Triple interaction = 1.47\*\* (SE = 0.746); %Uninsured alone = 3.31\*\*\* (SE = 0.993) |
| R5 | At the hourly frequency, high Twitter attention predicts lower returns for high-run-risk banks, not for low-risk banks | Table 6, Panel A, col. 4-6, p. 18 | High risk: post-Mar-9 x high attention (4h lag) = -0.096\*\* to -0.108\*\* bps/hour (cols 4-6); low risk: near zero, insignificant |
| R6 | The hourly return effect is driven by broadly retweeted tweets, not low-retweet or high-follower tweets | Table 7, Panel B, col. 3 vs. col. 6, p. 20 | High-retweet: -0.124\*\*\* (SE = 0.062); low-retweet: -0.075\* (SE = 0.039, 10%) |
| R7 | Twitter sentiment (VADER negative score) does not amplify bank run risk | Table 8, col. 5-8, p. 21 | VADER Neg x High Run Risk = statistically insignificant across all specifications |
| R8 | In-run contagion tweets and run-behavior tweets predict stock losses and account for most of the pre-exposure effect | Table 4, Panel C, col. 2-5, p. 15 | Contagion tweets: 6.696\*\*\* (SE = 1.333); run tweets: 7.252\*\*\* (SE = 1.374); pre-exposure falls to near zero when both are included |

**Overall (paper's conclusion).** Social media exposure amplifies classical bank run
risks, with the channel being Twitter attention and social propagation via retweets,
not negative sentiment. The effect is consistent across stock-return and deposit-outflow
measures, and is evident at both the cross-sectional and hourly frequency. The paper
concludes that social media serves as a coordination device for depositors, linking it
to the theoretical framework of Diamond and Dybvig (1983) and Goldstein and Pauzner (2005).

## Theory / model

The paper has no formal structural model. The theoretical grounding draws on classical
bank run models. Diamond and Dybvig (1983) show that bank runs can arise as coordination
failures: even a fundamentally solvent bank can fail if a sufficient mass of depositors
withdraw simultaneously because each depositor's optimal action depends on what they
believe others will do. Goldstein and Pauzner (2005) extend this to show that
communication among depositors that reveals information about others' run intentions
can amplify the probability of a run equilibrium.

The paper's contribution is to ask whether social media functions as such a coordination
channel. Three related empirical hypotheses are tested:

1. Banks with greater pre-run social media exposure will experience more severe runs,
   conditional on traditional balance-sheet risk factors.
2. Twitter pre-exposure amplifies rather than substitutes for classical run risk
   factors (high uninsured deposits and large mark-to-market asset losses), consistent
   with amplification of the uninsured-depositor coordination problem documented
   by Jiang et al. (2024a).
3. The channel is social attention and propagation (retweets), not negative sentiment:
   broadly retweeted content raises the salience of run concerns and reaches the
   audience of likely depositors, while negativity alone does not coordinate behavior.

**Identification.** The main cross-sectional identification relies on the fact that
Twitter pre-exposure in January-February 2023 is driven primarily by bank size and
investor fanaticism (Pedersen (2022)), and is empirically near-zero correlated with
traditional run risk factors (Fig. 6b, p. 11: correlation of pre-exposure with run risk
is 0.023). The content of pre-period tweets is also unrelated to banking distress topics
(Fig. A.5 in the Appendix). Thus, cross-sectional variation in pre-exposure reflects
social media reach, not ex ante run vulnerability. For the hourly tests, a narrow
5-minute window around individual tweets (adapting the approach of Bianchi et al. (2023))
provides sharper identification by limiting the scope for confounding news and price
dynamics. For the deposit outflow tests, the quarterly FDIC Call Report data (Drechsler
et al. (2024) deposit beta measures are also used as controls) corroborates the
cross-sectional stock-return evidence.

## Method

**Twitter pre-exposure construction** (Section 2.1, p. 5). The paper uses academic
access to Twitter's API to collect 5,399,740 original tweets about 602 depository
institutions (SIC codes 602, 603, 609) from January 2020 to March 2023, filtered to
English original (non-retweeted) posts containing each bank's cashtag. A supplemental
sample of 765,224 retweets (January-March 2023) is also collected. Twitter pre-exposure
for bank $$i$$ is the log count of original cashtag tweets from January 1 to February 15,
2023, winsorized at the 95th percentile. The resulting data are available on Mendeley
Data (dataset link on p. 1 of the article).

**Sentiment scoring** (Section 2.1.1, p. 5). Each tweet is scored by the VADER
(Valence Aware Dictionary and sEntiment Reasoner) algorithm (Hutto and Gilbert (2014)),
which produces positive ($$\text{VADERPos}$$) and negative ($$\text{VADERNeg}$$) component
scores. Both components are standardized to mean zero and SD one within the estimation
sample.

**Content dictionaries** (Section 2.1.2, p. 6, Table 1). Four contextual dictionaries
classify tweet content into Balance Sheet, Run behavior, Contagion, and Tech Community
categories. Each dictionary is built iteratively using domain seed words; the top 40
most salient words by topic are selected via frequency analysis on the run-period corpus,
following the approach of Cookson et al. (2020) for investor tweets.

**Mark-to-market loss construction** (Section 2.2, p. 6, Eq. 1). Bank asset losses
are estimated from FDIC Call Report asset holdings as of January 2022 and Treasury Bond
index changes through Q1-2023:

$$
\Delta\text{Loss MTM} = \sum_m \Bigl( (\text{RMBS}_m + \text{Mortgages}_m) \times
\Delta\text{Treasury Price}_m \times \text{Multiplier} \Bigr) + \sum_m \Bigl(
\text{Treasuries}_m + \text{securities}_m + \text{loans}_m \Bigr) \times
\Delta\text{Treasury Price}_m \tag{1}
$$

where $$m$$ indexes 9 maturity-repricing breakdowns (1 month through 30 years);
$$\Delta\text{Treasury Price}_m$$ is the Q1-2022 to Q1-2023 percentage price change in
the corresponding CRSP Treasury index; and $$\text{Multiplier}$$ adjusts RMBS and
mortgages for the iShares MBS ETF return relative to the S&P U.S. Treasury Bond Index.
The resulting variable is labeled $$\%\text{ Asset Decline MTM}$$.

**Determinants of Twitter pre-exposure** (Section 2.5, p. 9, Eq. 2). To validate
identification, the cross-sectional regression

$$
\text{Twitter Pre-Exposure}_i = a + b_1 \text{Size}_i + b_2 \,\%\text{LossMTM}_i + b_3
\,\%\text{Uninsured}_i + b_4 \mathbf{X}_i + b_5 \mathbf{W}_i + b_6 \mathbf{Z}_i +
\varepsilon_i \tag{2}
$$

is estimated, where $$\text{Size}_i$$ is log market capitalization; $$\mathbf{X}_i$$
is an information-environment vector (analyst coverage from I/B/E/S, newspaper article
count from RavenPack); $$\mathbf{W}_i$$ includes Q3-Q4 2022 stock returns, the FOMC
2022 cumulative abnormal return, and the Google Trends SVI; and $$\mathbf{Z}_i$$ captures
deposit franchise characteristics (branch ZIP count, deposit beta, market-to-book,
CRE loan share, liquid assets, deposit concentration). Table 3 (p. 10) shows size
is the dominant predictor (coefficient 0.767\*\*\* in the baseline); the classical run
risk factors $$\%\text{LossMTM}$$ and $$\%\text{Uninsured}$$ are statistically
insignificant or slightly negative, supporting the identification strategy.

## Empirical specifications

**Main cross-sectional specification** (Section 3.1, p. 12, Eq. 3). Run severity is
measured as the percentage stock market loss from March 1 to March 15, 2023. The
headline estimating equation is:

$$
\text{Stock Loss}_i = \beta_1 \,\%\text{LossMTM}_i \times \%\text{Uninsured}_i \times
\text{Pre-Exposure}_i + \text{lower-order terms} + \gamma \mathbf{X}_i + \varepsilon_i
\tag{3}
$$

All variables marked $$z$$ in Table 4 are standardized to mean zero and SD one. The
vector $$\mathbf{X}_i$$ includes log market cap, analyst coverage, news article count,
2022 Q3/Q4 stock returns, FOMC CAR, Google SVI, Q4-2022 deposit flow, branch footprint,
deposit beta, market-to-book, CRE loan share, liquid assets, and deposit concentration.
The key coefficients are those on the triple interaction and the two-way interactions
involving Pre-Exposure (Table 4, Panel A, col. 4, p. 13).

**Deposit outflow specification** (Section 3.2, p. 14, Eq. 4). An analogous
cross-sectional specification uses Q1-2023 deposit outflows as the outcome:

$$
\text{Deposit Outflow}_i = \beta_1 \,\%\text{LossMTM}_i \times \%\text{Uninsured}_i
\times \text{Pre-Exposure}_i + \text{lower-order terms} + \gamma \mathbf{X}_i +
\varepsilon_i \tag{4}
$$

where $$\text{Deposit Outflow}_i = 100 \times (\text{Deposits}_{Q4\text{-}2022} -
\text{Deposits}_{Q1\text{-}2023}) / \text{Deposits}_{Q4\text{-}2022}$$, computed for
both uninsured (above the FDIC $250K threshold) and total deposits from the FDIC Call
Reports. Standard errors are robust. Sample size is 275 (Table 5, p. 16).

**Hourly panel specification** (Section 3.3, pp. 15-16, Eq. 5). For each bank-hour
in the period March 6-10, 2023 (shorter window: March 8-9), hourly stock returns
(in basis points, from FirstRate Data) are regressed on lagged Twitter attention:

$$
r_{i,t} = a + b_1 \mathbf{1}(\geq\!\text{Mar 09})_t + b_2 \mathbf{1}(\text{N Tweets High})_{i,t-1}
+ b_3 \bigl[ \mathbf{1}(\geq\!\text{Mar 09})_t \times \mathbf{1}(\text{N Tweets High})_{i,t-1} \bigr]
+ \eta \mathbf{X}_{i,t-1} + \delta_i + \gamma_t + \varepsilon_{i,t} \tag{5}
$$

where $$\mathbf{1}(\text{N Tweets High})_{i,t-1}$$ equals one if the count of cashtag
tweets about bank $$i$$ in the prior 4 hours was above the (within-sample) median;
$$\delta_i$$ and $$\gamma_t$$ are firm and day-by-hour fixed effects; and
$$\mathbf{X}_{i,t-1}$$ includes the bank's cumulative 4-hour return and lagged news
article count. Standard errors are clustered at the bank level. The specification is
estimated separately for high-run-risk banks (above-median $$\%\text{Uninsured} \times
\%\text{LossMTM}$$) and low-run-risk banks. The coefficient of interest is $$b_3$$:
whether, after the SVB run began (March 9), higher Twitter attention predicted lower
returns for high-risk banks. Table 6 (p. 18) reports Panel A (full March 6-10 window),
Panel B (shorter March 8-9 window), Panel C (excluding SVB), and Panel D
(retweet-weighted tweet counts).

**Tweet-level high-frequency specification** (Section 3.5, p. 20, Eq. 6). At the
individual tweet level (one observation per tweet), log price changes in the 5-minute
window around the tweet are regressed on VADER sentiment:

$$
\Delta p_{i,t} = a + b \times \text{VADERPos}_{i,t} + c \times \text{VADERNeg}_{i,t}
+ \eta \mathbf{X}_{i,t} + \gamma_i + \varepsilon_{i,t} \tag{6}
$$

where $$\Delta p_{i,t}$$ is the log price change from just before to 5-15 minutes after
the tweet (in basis points), following Bianchi et al. (2023, 2024); $$\gamma_i$$ is a
bank fixed effect; and $$\mathbf{X}_{i,t}$$ includes the log price change in the 10
minutes prior to the tweet, the count of RavenPack news articles in that window, and
their ESS sentiment. The interaction terms in Table 8 (p. 21) test whether VADER Neg
amplifies returns for high-run-risk banks and for specific tweet types (run tweets,
contagion tweets, tech community tweets). Standard errors are clustered at the bank-day
level.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Twitter API (original collection, Jan 2020-Mar 2023; data on Mendeley Data) | Twitter pre-exposure variable, in-run attention and sentiment, tweet content classification | No page yet (author-introduced dataset) |
| FDIC / FFIEC Call Reports (Q4-2022) | % Uninsured deposits, Q1-2023 deposit outflows, asset holdings for MTM loss | [FFIEC Call Reports](/wiki/datasets/call-reports/) |
| CRSP US Treasury and Inflation Indexes | Treasury bond price changes by maturity for % Loss MTM construction | [WRDS](/wiki/commercial/wrds/) (licensed) |
| Compustat bank fundamentals (via WRDS) | Market capitalization, market-to-book, liquid assets, CRE loan share | [WRDS](/wiki/commercial/wrds/) (licensed) |
| I/B/E/S analyst coverage (via WRDS) | Number of analysts as information-environment control | [WRDS](/wiki/commercial/wrds/) (licensed) |
| FirstRate Data (intraday) | Minute-level and 5-minute bank stock prices for hourly return construction (March 2023) | No page yet |
| RavenPack news analytics | Traditional news article count and ESS sentiment as controls in high-frequency tests | [RavenPack](/wiki/commercial/ravenpack/) (licensed) |
| Google Trends SVI | Retail investor attention to bank stocks as a market-control variable | No page yet |

Sample: 277 publicly traded bank holding companies; Twitter pre-exposure from
January 2020 to March 14, 2023; run period March 1-15, 2023; hourly panel March 6-10,
2023; tweet-level tests use approx. 36,659 tweet observations.

## When to read the full paper

Read the [original](https://doi.org/10.1016/j.jfineco.2025.104218) if you are:
studying bank run dynamics during the 2023 regional banking crisis and need the full
battery of robustness checks (Appendix Tables A.1-A.14, covering alternative pre-exposure
windows, alternative event windows, specification-curve analysis per Simonsohn et al. (2020),
and SVB-excluded subsamples); building on the Twitter data for US bank stocks (the
Mendeley dataset is linked from the article); examining the tweet-content classification
methodology in detail (Table 1 and Appendix Fig. A.5); or extending the hourly bank
panel (Table 6) or tweet-level sentiment tests (Tables 8-9) to other episodes. Table 7
(Panel B, p. 20) is particularly valuable for readers interested in the social propagation
mechanism via retweets.

## Attribution and rights

Source: peer-reviewed, *Journal of Financial Economics* 176 (2026) 104218. This
distillation was extracted by an LLM on 2026-06-24 and is **not human-verified or
independently reproduced**. The original is paywalled; extract-only redistribution applies.

> Cookson, J. Anthony, Corbin Fox, Javier Gil-Bazo, Juan F. Imbet, and Christoph Schiller.
> "Social media as a bank run catalyst."
> *Journal of Financial Economics* 176 (2026) 104218.
> DOI: [10.1016/j.jfineco.2025.104218](https://doi.org/10.1016/j.jfineco.2025.104218).
> © 2025 Elsevier B.V. All rights reserved.


==============================================================================
# Securing Technological Leadership? The Cost of Export Controls: Crosignani et al. (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jfe/2026/crosignani-securing-technological-leadership-cost-2026/
# Distilled: Crosignani, Han, Macchiavelli, and Silva (2026) document using hand-collected BIS Entity List data matched to FactSet Revere supply-chain linkages that U.S. export controls on Chinese firms cause broad-based decoupling from Chinese customers; affected U.S. suppliers suffer large stock market losses, declining revenues and employment, and tighter bank credit, while failing to form new customer relations domestically or in politically aligned countries. Journal of Financial Economics 2026, paywalled. Nine core results with source locators, datasets used, and the empirical design (stacked DiD and event study). LLM-distilled.
# Tags: paper-summary, geopolitics, trade-policy, export-controls, supply-chains, geoeconomics
==============================================================================

**What this is.** The paper's core results, the identification strategy, and the empirical
design: enough to know what it found and how, without reading all 16 pages. To replicate or
extend it, read the full source at the [original](https://doi.org/10.1016/j.jfineco.2025.104192).

## TL;DR

To safeguard its technological edge, the U.S. government has restricted domestic firms from
selling advanced technology to selected Chinese companies by adding those companies to Bureau
of Industry and Security (BIS) export control lists. Crosignani et al. (2026) use hand-collected
BIS list data matched to FactSet Revere global supply-chain linkages to document how these
restrictions affect U.S. domestic suppliers and their Chinese customers. The intended effect
materializes: affected U.S. suppliers terminate relations with their targeted Chinese customers.
However, the broader supply-chain reconfiguration that U.S. policymakers hope for (reshoring or
friendshoring) does not happen in the three years following the controls. Instead, affected U.S.
firms experience negative stock market reactions, declines in revenues, cash flow, and employment,
and tighter bank credit conditions. Chinese firms targeted by export controls, by contrast, are
more proactive: they terminate U.S. supplier relations and replace them with domestic Chinese
alternatives, consistent with the prior work of Crosignani et al. (2023) on supply-chain
propagation. Total U.S. supplier stock market losses ($158 billion across all affected suppliers) far exceed
the $18-19 billion in losses on the Chinese side; even the conservative estimate for U.S. suppliers
linked to publicly listed Chinese targets ($77 billion) is approximately four times larger (p. 15),
raising doubts about the effectiveness of export controls as a tool for preserving U.S.
technological leadership.

## Core results

Magnitudes and significance as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%.
Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Export controls lead to broad-based decoupling: affected U.S. suppliers terminate relations with targeted AND non-targeted Chinese customers | Table 4, cols 1-3 and 5-6, p. 8 | Full: coefficient 0.572\*\*\* (SE 0.209); excluding targeted Chinese customers: 0.414\* to 0.560\*\*, equivalent to +51% to +75% more terminations |
| R2 | No reshoring or friendshoring: affected U.S. suppliers form fewer new Chinese customer relations and cannot offset the loss with domestic or Asia-Pacific alternatives | Table 4, cols 7-9, p. 8; Table 5, col 1, p. 9-10; Table 6, p. 10 | Total customers: -0.145\*\* (SE 0.064); new Chinese relations: -0.473\*\* to -0.523\*\*\* (Table 4, cols 7-9); domestic, Asia, Asia-ally, and EU shares: all coefficients insignificant |
| R3 | Stock market reaction: affected U.S. suppliers lose -3.6% in cumulative abnormal return in the 30-day window around BIS list announcements | Fig. 3, p. 11-12 | -3.6% CAR over [-10, 20] (3-factor FF model); $1 billion market cap loss per firm; $158 billion total across 156 affected suppliers |
| R4 | Cash flow declines significantly: export controls reduce operating cash flow by an amount equal to 21% of average treated-firm value | Table 7, col 1 (Panel A), p. 12 | Coefficient -0.018\*\* (SE 0.007) in stacked panel regression |
| R5 | Revenue declines: export controls reduce revenues by 8.9% | Table 7, col 2 (Panel A), p. 12 | Coefficient -0.093\*\* (SE 0.032) |
| R6 | Employment falls 7.3%; capital expenditure not significantly affected | Table 7, cols 5 and 4 (Panel A), p. 12 | Employees: -0.076\*\* (SE 0.031); CapEx: 0.004 (SE 0.003, insignificant) |
| R7 | Banks tighten lending to affected U.S. suppliers: term loans fall, spreads rise, maturities shorten | Table 8, cols 2, 5, 6, p. 13 | Term loans: -0.630\*\* (SE 0.251); spread: +0.179\*\* bps (SE 0.088); maturity: -4.874\*\*\* months (SE 1.538) |
| R8 | Chinese firms targeted by export controls successfully reshore by forming new domestic supplier relations, replacing U.S. suppliers faster than U.S. firms can friendshore | Table 9, cols 3-4, p. 13 | New relations with Chinese suppliers: +0.469\*\*\* (SE 0.181) to +0.517\*\*\* (SE 0.172); U.S. supplier share of targeted Chinese firms falls significantly (Table 10, cols 5-6) |
| R9 | Targeted Chinese firms also lose market capitalization, but U.S. supplier losses are considerably larger (approximately 4x by the conservative subsample estimate) | Fig. 5, p. 14-15 | CAR [-10, 20]: -8.2% (3-factor) to -9.0% (4-factor China model); Chinese total $18-19 billion; U.S. full-sample $158 billion; U.S. subsample linked to listed Chinese targets $77 billion (approximately 4x the Chinese losses, p. 15) |

**Overall (paper's conclusion).** Export controls prompt immediate and broad-based decoupling of
U.S. suppliers from their Chinese customers, but U.S. firms struggle to find new customers in
the three years following the controls. These supply-chain rigidities translate into sizable
declines in market capitalization, revenues, cash flow, employment, and bank credit access.
Chinese firms targeted by export controls are more proactive in supply-chain reconfiguration,
increasing reliance on domestic Chinese alternatives. The total capitalization losses incurred by U.S. suppliers ($158 billion full sample;
$77 billion for the subsample linked to listed Chinese targets) are considerably larger than those
experienced by their Chinese counterparts ($18-19 billion); the conservative ratio is approximately
four times (p. 15), questioning whether export controls are an effective tool for preserving U.S.
technological leadership at acceptable cost.

## Theory / model

The paper has no formal theoretical model. It operates in the nascent theoretical literature on
geoeconomics summarized by Clayton et al. (2025b) and the literature on targeted trade
restrictions such as Liu et al. (2024), who develop a calibrated model in which comprehensive
semiconductor restrictions could raise domestic welfare by facilitating technology transfer.

The paper tests three sets of hypotheses grounded in the policy context:

1. **Decoupling hypothesis**: Export controls cause affected U.S. suppliers to terminate supply
   relations with Chinese customers named in BIS lists, and plausibly also with non-targeted
   Chinese customers ("wake-up call" or re-export concern effect).
2. **Reshoring / friendshoring hypothesis**: Following decoupling, U.S. firms form new customer
   relations with domestic or politically aligned (Asia-ally, EU) customers to offset lost
   Chinese business. The paper finds this does NOT happen within three years.
3. **Collateral damage hypothesis**: The inability to substitute customers generates measurable
   declines in firm value (stock market), operating performance (revenues, cash flow, employment),
   and credit supply (bank lending).

**Identification logic.** The BIS Entity List, Military End User (MEU) List, and Unverified List
(UVL) have been expanded continuously since the late 1990s. The staggered, firm-specific nature
of these additions constitutes a quasi-natural experiment: some U.S. suppliers become
"affected" earlier (their Chinese customer is added to a list), others later. By comparing
affected firms with not-yet-affected firms exporting to China in the same industry-size cell
around each BIS event, the paper identifies the causal effects of export controls on supply-chain
configurations and firm outcomes.

## Method

The paper uses two main estimators: a stacked difference-in-differences (DiD) panel regression
for supply-chain and balance-sheet outcomes, and a standard event study for stock-market
reactions.

**Stacked DiD.** Because BIS additions are staggered and could produce biased two-way
fixed-effects (TWFE) estimates, the paper follows the stacked regression methodology developed
by Gormley and Matsa (2011) and described in Baker, Larcker, and Wang (2022). Cohorts are
defined around each BIS inclusion event. Within each cohort, treated firms are U.S. suppliers
whose Chinese customer is included in that cohort's event; control firms are U.S. suppliers
exporting to China that are never treated or not yet treated at event time. A [-3, 3] year window
centers each cohort. The estimating equation (p. 7, Eq. 1) is:

$$
y_{ict} = \beta \, \text{Affected}_{ic} \times \text{Post}_{ict} + \mu_{ic} + \mu_{ckt} + \varepsilon_{ict} \tag{1}
$$

where $$i$$ indexes a firm, $$c$$ a cohort (round of export controls), and $$t$$ a year.
$$\text{Affected}_{ic}$$ is an indicator equal to one if export control $$c$$ is imposed on a
Chinese customer of U.S. firm $$i$$. $$\text{Post}_{ict}$$ equals one after the imposition.
$$\mu_{ic}$$ are cohort-firm fixed effects; $$\mu_{ckt}$$ are cohort-industry-size-year fixed
effects (absorbing demand shocks that hit similar firms in the same year). Standard errors are
double-clustered at the firm and year level.

For count-like outcomes (number of terminated or new relations), the paper estimates Poisson
pseudo-maximum-likelihood (PPML) regressions following Cohn, Liu, and Wardlaw (2022), with the
same fixed-effects structure. Coefficients in PPML regressions are interpreted as percentage
effects via $$\exp(\hat\beta) - 1$$.

**Event study (stock market reactions).** Cumulative abnormal returns (CARs) are estimated over a
[-10, 20] trading-day window around the BIS list announcement date (p. 11). Pre-event betas are
estimated on the [-150, -50] day window using the Fama and French (1993) 3-factor model or the
Fama and French (2015) 5-factor model. Affected suppliers are the U.S. firms that supply Chinese
entities included in the BIS lists; the sample includes 250 events involving 156 unique affected
suppliers (one firm can contribute multiple events if it exports to multiple Chinese targets added
at different times).

## Empirical specifications

**Supply-chain reconfiguration (R1, R2, Table 4-6, p. 8-10).** The outcome variables are:
(a) the total number of terminated relations with Chinese customers (including/excluding directly
targeted firms); (b) the number of new Chinese customer relations; (c) total customer count;
(d) regional customer shares (domestic, China, Asia, Asia-ally, EU). Estimated by PPML with
cohort-firm and cohort-SIC-size-year fixed effects, double-clustered SEs. Treatment requires
all control firms to export to China in the pre-treatment period; within each cohort, controls
are matched on industry (2-digit SIC) and firm-size quartile. A "Restrictive Sample" narrows
to only Entity List and MEU List events, excluding the less restrictive UVL.

**Balance sheet and real outcomes (R4-R6, Table 7, p. 12).** Outcome variables are:
cash flow (operating income before depreciation minus interest and taxes divided by lagged assets),
revenues (log total revenues), EBIT (earnings before interest and taxes divided by lagged assets),
CapEx (capital expenditures divided by lagged assets), and employees (log total employees).
Specification: OLS stacked panel regression (Eq. 1) with cohort-firm and cohort-SIC-size-year
fixed effects. Results are robust to NAICS fixed effects (Table C.1 online appendix).

**Bank lending (R7, Table 8, p. 13).** Outcome variables: committed total credit, committed term
loans, committed credit lines, utilized credit lines, the interest rate spread, and loan maturity.
Sample: 331 firms exporting to China that borrow from a total of 38 banks over 2012:Q3-2023:Q3;
71 are affected by export controls. Specification: stacked OLS panel with firm fixed effects
(absorbing time-invariant firm characteristics), industry-size-quarter fixed effects (absorbing
common demand conditions for similar firms), and bank-quarter fixed effects (capturing
bank-specific credit-supply shocks):

$$
y_{ibqt} = \beta \, \text{Affected}_{i} \times \text{Post}_{iqt} + \mu_{i} + \mu_{kqt} + \mu_{bqt} + \varepsilon_{ibqt}
$$

The coefficient $$\beta$$ is identified from within-bank-quarter comparisons of affected vs
control firms in the same industry-size cohort.

**Chinese supply-chain reconfiguration (R8, R9, Table 9-11, Fig. 5, p. 13-15).** Chinese
targeted firms' supply-chain adjustments are estimated symmetrically to the U.S. side, with
Targeted replacing Affected and the control group being Chinese firms importing from U.S. suppliers
not in the BIS lists. This documents whether Chinese firms actively reshore (R8) and how
non-U.S. third-country firms benefit (Table 11: revenues of non-U.S., non-allied suppliers to
targeted Chinese firms increase by 15.7%\*\*\* after controls). Stock market reactions for
Chinese targets (R9) use the China-specific 3-factor and 4-factor models of Liu, Stambaugh,
and Yuan (2019).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| FactSet Revere supply-chain linkages | Global firm-to-firm customer-supplier relations 2007-2023; identifies U.S. suppliers of BIS-targeted Chinese firms | [FactSet Revere](/wiki/commercial/factset-revere/) |
| BIS Entity List, MEU List, UVL (hand-collected) | Export control targets; hand-collected additions and removals of Chinese entities from federalregister.gov and ecfr.gov with dates and aliases | no page yet |
| CRSP daily stock file | U.S. equity prices and returns for event-study CAR estimation | [WRDS](/wiki/commercial/wrds/) (licensed) |
| Compustat North America fundamentals (annual) | Firm-level balance sheet characteristics (assets, revenues, employment, CapEx) | [WRDS](/wiki/commercial/wrds/) (licensed) |
| Ken French Data Library | Fama-French 3- and 5-factor daily returns for beta estimation | [Ken French library](/wiki/datasets/ken-french/) |
| Federal Reserve Y-14Q (CCAR) | Confidential quarterly loan-level data for 331 U.S. firms borrowing from 38 large banks, 2012:Q3-2023:Q3 | [FR Y-14Q](/wiki/confidential/fr-y14q/) |
| Refinitiv (Chinese stock prices) | Daily stock price data for publicly listed Chinese firms targeted by export controls | no page yet |
| S&P Capital IQ | International firm balance sheet data (EBIT, revenues) for 6,068 suppliers of targeted Chinese firms | [Capital IQ](/wiki/commercial/capital-iq/) |
| Mingshi | Chinese stock market 3-factor and 4-factor model returns for Chinese-side CAR estimation | no page yet |

Sample: U.S. supply-chain panel 2007-2023 (up to Q3); bank-lending sample 2012:Q3-2023:Q3; stock event study 250 events, 156 unique affected U.S. suppliers.

## When to read the full paper

Use the [original](https://doi.org/10.1016/j.jfineco.2025.104192) if you are:
studying the real effects of geoeconomic policy on domestic supply chains; estimating the
collateral costs of U.S. export controls on firms in the technology sector; applying stacked
DiD methods with staggered treatment to firm-level data; or examining the asymmetric
reconfiguration capacity of U.S. versus Chinese firms in a supply-chain disruption. Locators
above point to the exact tables and figures in the source PDF.

## Attribution and rights

Source: peer-reviewed, *Journal of Financial Economics* 175 (2026) 104192. This
distillation was extracted by an LLM on 2026-06-24 and is **not human-verified or
independently reproduced**. The paper is paywalled; reproduction is extract-only.

> Crosignani, Matteo, Lina Han, Marco Macchiavelli, and André F. Silva.
> "Securing technological leadership? The cost of export controls on firms."
> *Journal of Financial Economics* 175 (2026): 104192.
> DOI: 10.1016/j.jfineco.2025.104192. © 2025 Elsevier B.V. All rights reserved.
> This page is an extract-only distillation by the Institute for Automated Research.


==============================================================================
# How Costly Are Cultural Biases: D'Acunto, Ghosh & Rossi (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jfe/2026/d-acunto-costly-cultural-biases-evidence-2026/
# Distilled: Using a P2P lending platform in India paired with a robo-advising tool, D'Acunto, Ghosh, and Rossi show that unassisted lenders discriminate against out-group (Muslim) and lower-caste (Shudra) borrowers, facing 8% higher defaults and up to 7.3 pp lower returns as a result. Robo-advising reduces both biases and improves lender-level returns by 4.5 to 7.3 pp, with biased beliefs as the dominant mechanism over taste-based discrimination. Journal of Financial Economics 2026, CC BY 4.0. Eight core results with source locators, datasets used, and the estimating equations.
# Tags: paper-summary, cultural-finance, discrimination, fintech, robo-advising
==============================================================================

**What this is.** The paper's core results, the competing hypotheses it tests, and the regression specifications with their defining equations: enough to know what it found and how, without reading all 26 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1016/j.jfineco.2025.104202).

## TL;DR

Using lender-level panel data from Faircent, a peer-to-peer lending platform in India, D'Acunto, Ghosh, and Rossi compare the choices the same lenders make when unassisted and after observing suggestions from an automated robo-advising tool (Auto Invest). Unassisted Hindu lenders are 5.8 percentage points less likely to fund Muslim borrowers than Muslim lenders are, and lenders of all castes systematically under-lend to Shudra (lower-caste) borrowers relative to the platform population. These biases are costly: the disfavored borrowers default less and earn higher standardized returns on average. After adopting Auto Invest, both biases shrink substantially and lender-level total returns improve by 4.5 pp to 7.3 pp. Lenders rarely override robo-advised suggestions to previously-disfavored groups, which supports inaccurate statistical discrimination (biased beliefs) rather than taste-based discrimination, as in Becker (1957), as the dominant mechanism.

## Core results

Magnitudes and significance are as reported; `\*\*`/`\*\*\*` = 5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Hindu lenders are **5.8 pp less likely to fund Muslim borrowers** than Muslim lenders before robo-advising | Table 2, col 1, p.11 | Hindu Lender = -0.058\*\*\*, t=-3.52; baseline Muslim-borrower share is 18% for Muslim lenders and 12% for Hindu lenders |
| R2 | **Auto Invest reduces Hindu lenders' out-group bias by 4.5 pp** | Table 2, col 1, p.11 | Hindu Lender x Auto Invest = 0.045\*\*, t=2.51; after adoption, Hindu and Muslim lenders fund Muslim borrowers at the same rate as the platform population |
| R3 | **Discrimination is costly**: Muslim borrowers in Hindu lender portfolios default 2.4 pp less than Hindu borrowers | Table 3, col 1, p.17 | Muslim Borrower = -0.024\*\*, t=-2.02; in-group bias selects worse borrowers from the preferred group |
| R4 | After Auto Invest, **Hindu (in-group) borrowers' default drops 11.2 pp**; Muslim (out-group) drops 7.3 pp; overall improvement driven by eliminating low-quality in-group loans | Table 3, col 2, p.17 | Hindu Borrower x Auto Invest = -0.112\*\*\*, t=-5.21; Muslim Borrower x Auto Invest = -0.073\*\*, t=-2.49 |
| R5 | Before robo-advising, **Muslim borrowers deliver higher standardized returns** for Hindu lenders | Table 5, col 1, p.21 | Muslim Borrower = 0.282\*\*\*, t=6.25 (standardized return); confirms the financial cost of out-group bias |
| R6 | After Auto Invest, **return improvement concentrated entirely in Hindu (in-group) borrowers**: +0.222 SD; Muslim borrowers unchanged | Table 5, col 2, p.21 | Hindu Borrower x Auto Invest = 0.222\*\*\*, t=3.07; Muslim Borrower x Auto Invest = -0.012, t=-0.18 |
| R7 | **Lender-level total returns improve by 4.5 pp (in-group vs. out-group) and 7.3 pp (stereotypical discrimination)** after Auto Invest | Fig. 11, p.23 | Average lender-level return increase, value-weighted across all loans before and after Auto Invest adoption |
| R8 | **Shudra borrowers default 3.8 pp less** than other borrowers before Auto Invest, confirming stereotypical discrimination is costly | Table 3, col 4, p.17 | Shudra Borrower = -0.038\*\*\*, t=-3.03; all-lender sample; lending to Shudra borrowers increases with robo-advice adoption |

**Overall (paper's conclusion).** Cultural biases lead lenders to select borrowers from preferred social groups who systematically underperform. The evidence is most consistent with inaccurate statistical discrimination: biased ex-ante beliefs about borrower quality correlated with ethnicity and caste, not a conscious taste for discrimination. Robo-advising reduces both types of bias and generates returns improvements of approximately 6% to 12% of the average capital invested, concentrated in loans to previously-favored in-group borrowers who were over-selected relative to their quality.

## Theory / model

The paper has no formal economic model. It tests three competing hypotheses about the nature of the discrimination it documents, each with distinct predictions for lenders' behavior when assisted vs. unassisted:

**Hypothesis 1 (taste-based discrimination, Becker (1957)):** lenders are willing to pay a utility cost to avoid transacting with disfavored groups. Prediction: lenders should frequently override robo-advised matches to disfavored-group borrowers, even when their economic incentives are aligned with the tool.

**Hypothesis 2 (accurate statistical discrimination):** disfavored groups are truly riskier on average, so the favored-group lending pattern reflects correct Bayesian updating on observable signals. Prediction: in-group borrowers should outperform (lower default, higher returns) relative to out-group borrowers matched by the platform's unbiased screening.

**Hypothesis 3 (inaccurate statistical discrimination, biased beliefs):** lenders hold systematically incorrect ex-ante beliefs about out-group borrower quality, even when identical objective risk information is available. Predictions: (a) disfavored borrowers should outperform when lenders are unassisted; (b) performance should converge after robo-advising; (c) lenders should not override robo-advice, because their financial incentives do not justify doing so.

The platform design eliminates channels that could rationalize statistical discrimination: interest rates are set by the platform algorithm, borrower risk profiles are directly observable, lenders and borrowers never interact, and Faircent's screening ensures no unbanked borrowers enter the pool. Heterogeneity tests rule out monitoring advantages, social collateral, and peer effects as alternative explanations (Fisman et al. (2017), Fisman et al. (2020)). By contrast, Hjort (2014) detects strong taste-based discrimination in a setting that removes scope for inaccurate statistical discrimination; the paper argues that Faircent's FinTech context is closer in spirit to that ideal but that inaccurate beliefs nonetheless dominate. The evidence consistently supports Hypothesis 3: disfavored borrowers outperform before robo-advising, performance converges after, and lenders rarely override robo-advised matches. The paper further shows that biases are stronger for lenders in areas with higher Hindu-Muslim inter-ethnic conflict, consistent with culturally shaped priors rather than rational updating.

## Method

The primary estimator is OLS on a lender-borrower-loan triad panel with lender fixed effects and year fixed effects, clustering standard errors at the lender level throughout (Table 2 caption, p.10). The within-lender before-after variation in tool adoption identifies the de-biasing effect; the approach builds on D'Acunto and Rossi (2020) who study robo-advising effects in a savings context.

For loan returns the paper additionally estimates quantile regressions (Eq. 5, p.22) to identify which part of the return distribution drives the improvement:

$$
Q_\tau(\text{Returns}_{i,j}) = a(\tau) + \beta(\tau)\,\text{Auto Invest}_j + \mathbf{X}'_{i,j}\,\zeta(\tau) + \varepsilon_{i,j} \tag{5}
$$

where $$Q_\tau$$ is the $$\tau$$-th quantile of standardized loan returns for loan $$i$$ of lender $$j$$, and $$\mathbf{X}_{i,j}$$ are loan risk controls. Coefficient $$\hat{\beta}(\tau)$$ measures how the $$\tau$$-th quantile shifts after the lender adopts Auto Invest.

The lender-level total return is computed as a value-weighted average across loans originated before and after Auto Invest adoption (Eq. 6, p.23):

$$
\text{Lender Tot Ret}_{i,t} = 100 \times \frac{\sum_j \text{Amount Disbursed}_{i,j,t} \times \text{Loan Return}_{j,t}}{\sum_j \text{Amount Disbursed}_{i,j,t}} \tag{6}
$$

and the lender-level change is POST minus PRE (Eq. 7, p.23). A purged measure (Eq. 8, p.23-24) removes compositional effects by holding the pre-period disbursed amounts fixed and applying post-period returns, isolating the cultural-debiasing channel from other effects of Auto Invest on portfolio composition.

## Empirical specifications

**Primary specification for in-group vs. out-group discrimination (Eq. 1, p.10):**

$$
\text{Muslim Borrower}_{i,j,t} = \alpha + \beta\,\text{Auto Invest}_{j,t} + \gamma\,\text{Hindu Lender}_j + \delta\,(\text{Hindu Lender}_j \times \text{Auto Invest}_{j,t}) + \zeta\,\mathbf{x}_{i,t} + \eta_j + \eta_t + \varepsilon_{i,j,t} \tag{1}
$$

where $$\text{Muslim Borrower}_{i,j,t} = 1$$ if borrower $$i$$ funded by lender $$j$$ in year $$t$$ is Muslim; $$\text{Auto Invest}_{j,t} = 1$$ if the lender has adopted the tool by year $$t$$; $$\text{Hindu Lender}_j = 1$$ if lender $$j$$ is Hindu; $$\mathbf{x}_{i,t}$$ are loan characteristics assigned by the platform (maturity, amount, interest rate); $$\eta_j$$ are lender fixed effects; $$\eta_t$$ are year fixed effects. Coefficient $$\hat{\gamma}$$ (R1) measures the in-group bias before adoption; $$\hat{\delta}$$ (R2) measures the de-biasing effect of robo-advising.

**Stereotypical discrimination specification (Eq. 2, p.14):**

$$
\text{Shudra Borrower}_{i,j,t} = \alpha + \beta\,\text{Auto Invest}_{j,t} + \zeta\,\mathbf{x}_{i,t} + \eta_j + \eta_t + \varepsilon_{i,j,t} \tag{2}
$$

This drops the Hindu Lender interaction because all castes, including Shudra lenders, discriminate against Shudra borrowers (stereotypical discrimination is not in-group favoritism but group-wide negative stereotyping). Caste recognizability (continuous probability from the Bhagavatula et al. (2017, 2018) matrimonial-registry algorithm) is used to test whether the $$\hat{\beta}$$ coefficient grows with how easily a borrower is identifiable as Shudra (Fig. 5-6, pp.14-15).

**Performance specifications (Eqs. 3-4, p.16 and p.19):**

$$
\text{Delinquent Loan}_{i,j,t} = \alpha + \gamma\,\text{Muslim Borrower}_j + \delta\,(\text{Muslim Borrower}_j \times \text{Auto Invest}_{j,t}) + \theta\,(\text{Hindu Borrower}_{i,j} \times \text{Auto Invest}_{j,t}) + \zeta\,\mathbf{x}_{i,t} + \eta_j + \eta_t + \varepsilon_{i,j,t} \tag{3}
$$

$$
\text{Loan Return}_{i,j,t} = \alpha + \gamma\,\text{Muslim Borrower}_j + \delta\,(\text{Muslim Borrower}_j \times \text{Auto Invest}_{j,t}) + \theta\,(\text{Hindu Borrower}_{i,j} \times \text{Auto Invest}_{j,t}) + \zeta\,\mathbf{x}_{i,t} + \eta_j + \eta_t + \varepsilon_{i,j,t} \tag{4}
$$

where $$\text{Delinquent Loan}_{i,j,t} = 1$$ if the loan is closed delinquent (more than 90 days past due at closure) and $$\text{Loan Return}_{i,j,t}$$ is the standardized return. Coefficient $$\hat{\gamma}$$ tests whether disfavored borrowers outperformed before the tool was adopted (the cost of bias, R3 and R5); $$\hat{\delta}$$ and $$\hat{\theta}$$ capture the differential change in performance across borrower groups after adoption (R4 and R6). The falsification column (col 3 and col 6 in Table 3, p.17) adds loan risk controls to show that default changes are not driven by compositional shifts in borrower riskiness.

**Heterogeneity tests** use cross-sectional proxies for the salience of cultural stereotypes: state-level Hindu-Muslim riots (Ticku (2015)), BJP vote shares (Bhavnani (2014)), and birth-cohort exposure to the rise of Hindu-Muslim conflict (Fig. 4, p.13). The bias is about twice as large for lenders residing in high-riot states (6.4 pp vs. small and insignificant elsewhere), and the de-biasing effect is correspondingly stronger in these states.

**Sample:** lender-borrower-loan triads from the Faircent platform, January 2018 to March 2020. Main sample: 113,283 triads involving 2,818 unique Hindu and Muslim lenders and borrowers. Caste sub-sample: 62,831 triads for which Hindu varna (caste category) of the borrower can be inferred from the matrimonial registry. Loan maturity averages 22 months; median maturity 24 months; average loan amount is approximately Rs.130,000 (~$1,770); average annual interest rate is 24%. Standard errors are clustered at the lender level.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Faircent P2P lending platform | Main panel: lender-borrower-loan triads; loan amounts, interest rates, maturity, delinquency; Auto Invest adoption and fund-allocation share; lender and borrower demographic characteristics (name, state, date of birth, occupation) | No page yet |
| Marriage registry (Bhagavatula et al. 2017, 2018) | Religion and caste inference for lenders and borrowers; 2.5 million individuals from online matrimonial agencies; used to assign religion and varna probabilities from surname, date of birth, and location | No page yet |
| Ticku (2015) Hindu-Muslim riots data | State-level count of large-scale riots between Hindus and Muslims (1980-2000); proxy for inter-ethnic conflict salience in heterogeneity tests | No page yet |
| Bhavnani (2014) BJP election data | Average BJP candidate vote shares across national and state elections (1977-2015) per Indian state; proxy for ideological salience of Hindu nationalism | No page yet |
| National Crime Records Bureau (NCRB 2019) | Crimes against Scheduled Castes per 100,000 inhabitants per Indian state (2018); proxy for salience of caste discrimination in stereotypical bias heterogeneity tests | No page yet |

Sample: January 2018 to March 2020. Roughly 60% of loans issued in 2019 and 19% in the first three months of 2020. Median lender disburses funds to borrowers across 13 different Indian states; 90% of lenders serve borrowers in at least 5 different states.

## When to read the full paper

Use the [original](https://doi.org/10.1016/j.jfineco.2025.104202) if you are: studying the mechanism of discrimination (taste-based vs. inaccurate statistical) in financial markets; examining how robo-advising corrects culturally biased lending choices; studying in-group vs. out-group or stereotypical discrimination in a high-stakes economic setting; or extending the Faircent platform setting to other FinTech platforms or demographic groups. Table 2 (p.10) gives the main lending-bias results; Table 3 (p.17) the default-performance results; Tables 5-6 (pp.21-22) the return results; and Figs. 4, 6, and 8 the heterogeneity and mechanism tests.

## Attribution and rights

Source: peer-reviewed, *Journal of Financial Economics* 175 (2026), article 104202. This distillation was extracted by an LLM on 2026-06-24 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** D'Acunto, Francesco, Pulak Ghosh, and Alberto G. Rossi.
> "How costly are cultural biases? Evidence from FinTech."
> *Journal of Financial Economics* 175 (2026) 104202.
> DOI: 10.1016/j.jfineco.2025.104202. © 2025 The Authors. Published by Elsevier B.V.
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Institutions' Return Expectations: Dahlquist & Ibert (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jfe/2026/dahlquist-institutions-return-expectations-assets-2026/
# Distilled: Institutional investors' subjective risk premia across equity, cash, and credit track objective (model-based) risk premia one-to-one and are countercyclical, but cross-sectional disagreement across institutions exceeds time-series variation and is driven mainly by heterogeneous views about long-term price-earnings ratio mean reversion. J. Fin. Econ. 2026, CC BY 4.0. Six core results with source locators, datasets used, the regression specifications, and the building-block decomposition of return expectations.
# Tags: paper-summary, asset-pricing, expectations, beliefs, macro, pensions
==============================================================================

**What this is.** The paper's core results, the empirical framework (panel regression of subjective on objective risk premia), and the building-block decomposition of institutional equity return expectations: enough to understand what was found and how, without reading all 22 pages. To replicate or extend, read the original at [doi.org/10.1016/j.jfineco.2025.104188](https://doi.org/10.1016/j.jfineco.2025.104188).

## TL;DR

This paper documents the subjective risk premia of institutional investors (asset managers, investment consultants, wealth advisors, public pension funds) and professional forecasters (Survey of Professional Forecasters, Livingston survey) across five asset classes: US equities, developed markets ex-US equities, emerging markets equities, US cash, and US high-yield corporate bonds. Extending Dahlquist and Ibert (2024), who cover asset managers' US equity and term premia from 1997 to 2021, this paper spans multiple institution types through 2024.

The main finding is that these subjective risk premia vary one-to-one with objective, model-based risk premia that are available in real time and are countercyclical. Despite this high time-series co-movement, several subjective equity risk premia vary more in the cross-section of institutions than in the time series. This cross-sectional disagreement persists across asset classes and is primarily driven by heterogeneous expectations about long-term price-earnings (P/E) ratio mean reversion. Institutions that expect the P/E ratio to mean-revert report low equity premiums; those that treat it as a near-random walk report higher premiums.

## Core results

Magnitudes as reported; locators point into the published article.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Subjective risk premia track objective risk premia one-to-one across institution types and asset classes | Table 3, p. 9; Table 4, p. 10; Fig. 1, p. 7 | b ranges 0.538-1.615 across 23 cells; AUM-weighted b = 1.055 (US equity), 1.034 (DM equity), 0.856 (EM equity), 0.656 (US cash), 1.608 (US credit); cannot reject b = 1 in 22/23 cells at 5% |
| R2 | Cross-sectional disagreement across institutions exceeds time-series variation for US and DM equities | Table 5, p. 15 | Asset managers US equity: CS SD = 2.93%, TS SD = 1.04%; wealth advisors: CS SD = 1.32%, TS SD = 0.90%; pension funds: CS SD = 1.26%, TS SD = 1.17% |
| R3 | Institution fixed effects explain more variance than time fixed effects for most equity cases | Table 5, p. 15 | Asset managers US equity: institution FE = 80.6%, time FE = 12.6% (p = 0.000); wealth advisors US equity: institution FE = 68.6%, time FE = 19.4% (p = 0.002) |
| R4 | Repricing expectations (views on long-term P/E ratio) drive 70% of cross-sectional disagreement in US equity return expectations | Table 8, p. 16 | Repricing = 69.5%, income+growth = 25.3%, inflation = 5.2% (cross-sectionally demeaned decomposition, Spec. II); repricing share significantly exceeds income+growth share (p = 0.018) |
| R5 | Optimism about one asset class spills over to all others: all pairwise cross-asset correlations of institution fixed effects are positive | Table 10, p. 21 | US equity vs. DM equity = 0.806 (p = 0.000); range across 10 pairs: 0.221-0.806; all significant at 5% |
| R6 | Disagreement among professional forecasters is U-shaped in the CAPE: largest when valuations are very high or very low | Table 9, p. 18 | SPF IQR regression: Log(CAPE) = -2.056 (SE = 0.486), D(CAPE > Mean) = -20.300 (SE = 4.031), interaction = 6.100 (SE = 1.153) |

**Overall (paper's conclusion).** Institutional investors' and professional forecasters' subjective risk premia are countercyclical and track objective risk premia one-to-one, consistent with rational expectations asset pricing models. At any given point in time, however, institutions disagree substantially about future returns. This disagreement is primarily driven by heterogeneous beliefs about long-term P/E ratio mean reversion and persists across all five asset classes, suggesting that institutions form coherent sets of return expectations anchored to common macroeconomic primitives (p. 20-21).

## Theory / model

The paper has no formal asset-pricing model. It tests two related empirical hypotheses about institutional return expectations.

**Hypothesis 1 (time-series, countercyclical expectations).** Standard rational expectations asset pricing models imply that expected returns are countercyclical: high in recessions, low in expansions. The paper tests this for institutional investors by regressing subjective risk premia on objective model-based benchmarks. If $$b = 1$$, institutions' expectations move one-to-one with the objective benchmark; $$b < 1$$ indicates underreaction; $$b > 1$$ overreaction relative to the benchmark. In contrast, the behavioral finance literature of Greenwood and Shleifer (2014) documents procyclical expectations for retail investors ($$b < 0$$). The paper's central regression is (p. 2):

$$
\text{F}_{i,t}\!\bigl[r^e_{t \to t+h}\bigr] = a_i + b\,\text{E}_t\!\bigl[r^e_{t \to t+h}\bigr] + \varepsilon_{i,t} \tag{1}
$$

where $$\text{F}_{i,t}[r^e_{t \to t+h}]$$ is the subjective risk-premium forecast of institution $$i$$ on day $$t$$ over horizon $$[t, t+h]$$, $$a_i$$ is a forecaster fixed effect, $$\text{E}_t[r^e_{t \to t+h}]$$ is the corresponding objective (model-based) risk premium, and $$\varepsilon_{i,t}$$ is an error term.

**Hypothesis 2 (cross-sectional heterogeneity).** Institutions may persistently disagree about future returns. To decompose variation in subjective risk premia into persistent institution-level optimism/pessimism versus aggregate time-series co-movement, the paper uses the panel specification (p. 9-10):

$$
\text{F}_{i,t}\!\bigl[r^e_{t \to t+h}\bigr] = a_i + c_m + \eta_{i,t} \tag{3}
$$

where $$c_m$$ denotes time (year-month) fixed effects. The implied variance decomposition identity is (p. 10):

$$
1 = \frac{\text{Cov}(\text{F}_{i,t},\, a_i)}{\text{Var}(\text{F}_{i,t})} + \frac{\text{Cov}(\text{F}_{i,t},\, c_m)}{\text{Var}(\text{F}_{i,t})} + \frac{\text{Cov}(\text{F}_{i,t},\, \eta_{i,t})}{\text{Var}(\text{F}_{i,t})} \tag{4}
$$

where $$\text{F}_{i,t}$$ abbreviates the full forecast notation. The first term is the institution-fixed-effect share and the second is the time-fixed-effect share of total variance.

**Identification.** The regression in Eq. (1) identifies $$b$$ from time-series variation in the countercyclical objective benchmarks. Cross-sectional identification of disagreement comes from differences in institution fixed effects $$a_i$$ in Eq. (3). No causal claim is made; the design is descriptive.

## Method

**Objective risk premia.** For each equity market, the paper constructs the objective risk premium using a present-value model based on the regional CAPE (p. 4-5):

$$
\text{E}_t\!\bigl[r_{t \to t+10}\bigr] = \ln\!\Bigl(1 + 1/\text{CAPE}_t\Bigr) - r^f_{t \to t+10} \tag{2}
$$

where $$\text{CAPE}_t$$ is the cyclically adjusted price-earnings ratio for the relevant equity market and $$r^f_{t \to t+10}$$ is the 10-year real yield (Federal Reserve, FRED code REAINTRATREART10Y). For US cash, the objective risk premium is the negative of the Kim and Wright (2005) term premium. For US credit, it is the Gilchrist and Zakrajsek (2012) excess bond premium, available since 2011 (p. 5).

**Building-block decomposition.** The paper uses a standard log-return decomposition to connect institution-level forecasts to their components. Log nominal equity returns decompose as (p. 13):

$$
r_{t+1} = \underbrace{dp_{t+1}}_{\text{income}} + \underbrace{\Delta e_{t+1}}_{\text{real earnings growth}} + \underbrace{\pi_{t+1}}_{\text{inflation}} + \underbrace{\Delta pe_{t+1}}_{\text{repricing}} \tag{5}
$$

where $$dp_{t+1} = \log(D_{t+1}/P_{t+1})$$ is the log dividend-price ratio, $$\Delta e_{t+1} = \log(E_{t+1}/E_t)$$ is log real earnings growth, $$\pi_{t+1}$$ is the inflation rate, and $$\Delta pe_{t+1} = \log(P_{t+1}/E_{t+1}) - \log(P_t/E_t)$$ is the log change in the P/E ratio (the "repricing" component). Taking conditional expectations at time $$t$$, the decomposition holds ex ante as well (p. 13):

$$
\text{E}_t(r_{t+1}) = \text{E}_t(dp_{t+1}) + \text{E}_t(\Delta e_{t+1}) + \text{E}_t(\pi_{t+1}) + \text{E}_t(\Delta pe_{t+1}) \tag{7}
$$

43 out of 64 institutions in the sample explicitly reference this building-block approach in their white papers and capital market assumption documents (Table 6, p. 16).

**Estimation.** Eq. (1) is estimated as a panel with forecaster fixed effects identified from time-series variation. Twenty-three separate regressions are run: five asset classes times up to six institution types. Standard errors use a wild cluster bootstrap (Roodman et al., 2019), bootstrapping by forecaster and double-clustering the variance-covariance matrix by year-month and forecaster (p. 7-8). For robustness, observations are also weighted by discretionary AUM from Form ADV (Table 4, p. 10).

## Empirical specifications

**Time-series regression (R1).** For each of 23 institution-type x asset-class cells, Eq. (1) is estimated as a panel regression with forecaster fixed effects. The LHS is the annualized subjective risk premium (computed as subjective nominal expected return minus the horizon-matched Treasury yield); the RHS is the objective risk premium for the corresponding asset class and horizon. Both the null $$b = 0$$ (acyclicality) and $$b = 1$$ (one-to-one tracking) are tested. Sample sizes range from 88 (wealth advisors, US credit) to 1,425 (Livingston survey, US equity) panel observations (Table 3, p. 9).

**Variance decomposition (R2, R3).** Eq. (3) is estimated with year-month fixed effects for each cell, and the forecaster and time shares in Eq. (4) are computed via their covariance contributions to total variance (p. 10). For pension funds, year-level rather than year-month-level time fixed effects are used. Wild cluster bootstraps provide p-values for whether the institution share significantly differs from the time share (Table 5, p. 15).

**Building-block variance decomposition (R4).** For the subsample of asset managers, investment consultants, and wealth advisors who provide individual building-block forecasts (26-64 institutions depending on the component), the paper decomposes variance in US equity return expectations into the income+growth, inflation, and repricing components of Eq. (5). Table 8 (p. 16) reports three specifications: pooled (Spec. I), cross-sectionally demeaned (Spec. II), and time-series demeaned (Spec. III). In Spec. II, which directly addresses cross-sectional disagreement, repricing explains 69.5% of the variation.

**Cross-asset correlations (R5).** Institution fixed effects from Eq. (3) are estimated separately for each of the five asset classes. All 10 pairwise correlations between the institution fixed effects across asset classes are positive and statistically significant at the 5% level, ranging from 0.221 (EM equity vs. US cash) to 0.806 (US equity vs. DM equity) (Table 10, p. 21). Andonov and Rauh (2022) study a related mechanism through which pension funds' expectations affect their portfolio allocation. The correlations hold within institution type (controlling for institution-type fixed effects in a second specification).

**CAPE-level disagreement (R6).** For SPF forecasters (annual, 33 time-series obs.) and Livingston forecasters (semi-annual, 67 obs.), the paper regresses cross-sectional disagreement measures (interquartile range and standard deviation of expectations at each point in time) on the log CAPE, a dummy $$D(\text{CAPE} > \text{Mean})$$, and their interaction. The significant positive interaction coefficient confirms a U-shaped pattern: disagreement is largest when valuations are either very low or very high, consistent with heterogeneous priors about the long-term mean of the P/E ratio. Nagel and Xu (2023) document related evidence for a broader set of forecasters; Couts, Gonçalves and Loudis (2024b) study cross-forecaster disagreement in the risk-return trade-off across 19 asset classes.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Hand-collected institutional return expectations (asset managers, investment consultants, wealth advisors; 1997-2024) | Primary outcome: subjective risk premia across five asset classes; 64+ institutions | Introduced by this paper; data hosted on Mendeley Data |
| Public pension fund CAFR / GASB statements (207 funds, 2014-2023) | Subjective risk premia for public pension funds (US equity, DM equity, EM equity, US cash, US credit) | No page yet |
| Survey of Professional Forecasters (SPF, Philadelphia Fed, 1991-2024) | Professional forecasters' US equity and T-bill return expectations | No page yet (`data:spf`) |
| Livingston survey (Philadelphia Fed, 1990-2023) | Professional forecasters' one-year S&P 500 price targets, converted to return expectations | No page yet (`data:livingston`) |
| Robert Shiller CAPE data (US equities, 1881 to present) | Objective US equity risk premium via Eq. (2) | [shiller-data](/wiki/datasets/shiller-data/) |
| Research Affiliates CAPE (DM and EM equities) | Objective equity risk premium for developed and emerging markets via Eq. (2) | No page yet (`data:research-affiliates`) |
| Kim and Wright (2005) term structure model (FRED: REAINTRATREART10Y) | Objective cash risk premium (negative of the 10-year term premium) | [FRED](/wiki/datasets/fred/) |
| Gilchrist and Zakrajsek (2012) excess bond premium | Objective credit risk premium for US high-yield corporate bonds | No page yet |
| Gurkaynak, Sack and Wright (2007) Treasury yield curve | Horizon-matched Treasury yields used to compute risk premia from nominal return forecasts | No page yet |
| Public Plans Data / Form ADV | AUM of pension funds and investment advisors for AUM-weighted regressions (Table 4) | No page yet |

Sample: five asset classes; institution expectations collected from earliest available dates (1990 for Livingston, 1991 for SPF, 1997 for most asset managers) through 2024. Long-run expectations (approximately ten-year horizon) for most institutions; Livingston survey is one-year.

## When to read the full paper

Use the [original](https://doi.org/10.1016/j.jfineco.2025.104188) if you are:
studying how institutional investors form and report return expectations across asset classes; assessing whether institutional or professional forecasters have rational or extrapolative return expectations; building a model of heterogeneous beliefs about long-run valuation levels; using institutional capital market assumptions as data for asset allocation research; or extending the analysis to additional asset classes or newer institution types. Tables 3 and 4 contain the core time-series evidence; Table 5 the variance decompositions; Table 8 the building-block breakdown; Table 10 the cross-asset correlations.

## Attribution and rights

Source: peer-reviewed, *Journal of Financial Economics* 175 (2026) 104188. This distillation was extracted by an LLM on 2026-06-24 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Dahlquist, Magnus, and Markus Ibert.
> "Institutions' return expectations across assets and time."
> *Journal of Financial Economics* 175 (2026) 104188.
> DOI: 10.1016/j.jfineco.2025.104188. Copyright 2025 The Author(s).
> Published by Elsevier B.V. Licensed under
> [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Bank Consolidation and Uniform Pricing: Granja & Paixão (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jfe/2026/granja-bank-consolidation-uniform-pricing-2026/
# Distilled: After bank mergers, deposit and loan rates at acquired branches converge toward the acquirer's network-wide rate because banks price uniformly across their branch networks; pre-merger rate differences between acquirer and acquired predict post-merger rate changes far better than local HHI changes; and forced branch divestitures reduce consumer welfare by about 7% in markets where the acquirer offered better deposit rates. J. Fin. Econ. 2026, paywalled. Seven core results with source locators, datasets used, the structural demand-and-supply model, and the empirical specifications.
# Tags: paper-summary, banking, bank-mergers, deposit-markets, uniform-pricing
==============================================================================

**What this is.** The paper's core results, the structural model of deposit demand and bank pricing, and the identifying empirical specifications: enough to know what it found and how, without reading all 27 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1016/j.jfineco.2025.104204).

## TL;DR

Granja and Paixão document that U.S. banks set deposit and loan rates with a high degree of uniformity across their branch networks, and show that this practice fundamentally shapes the evolution of interest rates at acquired branches following a merger. Building on the earlier documentation of Radecki (1998) and Park and Pennacchi (2009) that U.S. banks price uniformly, they use a broader dataset and more products to confirm the pattern and trace its consequences for mergers and antitrust policy. Using branch-level RateWatch data on roughly 100,000 branches and 2,006 acquired banks over 2006-2019, they find that the absolute rate difference between acquired and acquirer branches falls by 40-70% within twelve months of a deal, regardless of changes in local market concentration. Pre-merger rate differences between the acquirer and acquired branches explain far more of the cross-sectional variation in post-merger rate changes than HHI-based concentration indicators. A structural model of monopolistic competition in banking that features uniform pricing fits the observed rate changes significantly better than a local-pricing model, even for acquirer branches in markets where the two banks never competed. Counterfactual welfare analysis shows that forced branch divestitures, which antitrust authorities commonly require, reduce consumer welfare by about 7% on average in markets where the acquirer offered better deposit rates than the acquired bank, raising questions about the adequacy of concentration-only merger review.

## Core results

Magnitudes are as reported; \*\*\*/\*\* = 1%/5%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | After a bank merger, the absolute difference in 12-month CD rates between acquired and acquirer branches falls by ~11 bps on average | Table 3, col. 1, p. 11 | Post-Acquisition coeff = -0.108\*\*\* (SE=0.009); pre-merger average absolute difference 25.7 bps |
| R2 | Rate convergence holds across all four products: savings deposits, HELOCs, and personal loans show declines of 3.7 bps, 43 bps, and 139 bps in absolute rate differences | Table 3, cols. 2-4, p. 11 | Post-Acquisition coefficients: SAV100K -0.037\*\*\* (SE=0.003), HELOC -0.429\*\*\* (SE=0.047), Personal -1.385\*\*\* (SE=0.142) |
| R3 | Rate adjustments are driven primarily by the acquired branches, not the acquirer; both higher-rate and lower-rate acquired branches converge toward the acquirer's median | Table 5, p. 14 | CD: higher-rate acquired branches fall 12.2 bps\*\*\*; acquirer median rises 3.3 bps\*\*\*. Lower-rate acquired branches rise 9.5 bps\*\*\*; acquirer median falls 2.9 bps\*\*\* |
| R4 | Acquired branches lose 7-11% of their deposit volume after a merger; deposit losses are larger when the acquired branch had higher rates than the acquirer | Table 6, cols. 1-2, p. 15 | Post-Acquisition on ln(total deposits): 12MCD10K -0.109\*\*\* (SE=0.012), SAV100K -0.069\*\*\* (SE=0.014); 1-SD higher pre-merger rate diff adds -2.5% (CD) and -3.8% (SAV) |
| R5 | Pre-merger rate differences between acquired and acquirer branches explain 27.6% of residual variation in post-merger deposit rates; HHI changes explain near-zero additional variation | Table 7, Panel A, p. 16 | CD rate top quintile (highest pre-diff) falls 28.5 bps\*\*\* (SE=0.026); bottom quintile rises 24.2 bps\*\*\* (SE=0.031); adj. R2 rises from 0.783 to 0.843 vs 0.784 for HHI bins |
| R6 | The structural uniform pricing model fits observed post-merger rate changes significantly better than the local pricing model, including for acquirer branches that never competed with the acquired bank | Fig. 11, Panel A-C, p. 25 | Uniform pricing slope beta=0.0818 (SE=0.0296, sig.); local pricing beta=0.0144 (SE=0.0365, insig.) for all branches; beta=0.0766 (sig.) vs 0.000 (insig.) for non-overlapping acquirer branches |
| R7 | Forced branch divestitures required by antitrust authorities reduce consumer welfare by 7.2% on average in markets where the acquirer offered lower deposit rates (and where uniform pricing would have benefited depositors) | Table 9, p. 26 | DeltaW^{Divestitures} = -7.209 pp in markets with Pre-Merger Rate Dif<0 (39 markets); no-divestiture welfare gain approx. 0.77%; in 5 markets with higher acquirer rates, divestitures add +2.167 pp |

**Overall (paper's conclusion).** Standard merger review based on local market concentration (HHI) fails to capture the most consequential channel through which bank mergers affect deposit rates: the convergence of rates at acquired branches toward the acquirer's uniform network rate. This uniform pricing practice means that pre-merger rate differences between banks are far stronger predictors of post-merger deposit rate changes than local concentration indices. Antitrust regulators who require branch divestitures without accounting for uniform pricing risk imposing welfare losses on depositors in markets where the merger would have been beneficial.

## Theory / model

The paper models demand and supply in the retail deposit market. On the demand side, consumers in each local banking market choose among available bank branches to maximize utility. On the supply side, the bank chooses deposit rates under one of two pricing regimes.

**Consumer preferences.** Each consumer in zip code $$z$$, banking market $$m$$, derives total indirect utility from depositing with branch $$j$$ of bank $$b$$ (p. 21, Eq. 3):

$$
V_{j,z,m,b,t} = a_m r_{j,t} + \beta_0 X_{j,t} + \beta_1 H_{b,m,t} + \beta_2 W_{b,t} + \xi_b + \gamma_z \tag{3}
$$

where $$a_m$$ is the market-specific deposit-rate semi-elasticity (the rate coefficient allowed to vary by market to capture heterogeneous depositor clienteles), $$r_{j,t}$$ is the branch deposit rate, $$X_{j,t}$$ is the branch age (capturing local relationship capital), $$H_{b,m,t}$$ is a vector of bank characteristics in market $$m$$ (branch network density, years of experience), $$W_{b,t}$$ captures time-varying bank-wide attributes (size, ROA, NPL ratio, Tier 1 capital), $$\xi_b$$ is a bank fixed effect (capturing unobserved service quality and brand), and $$\gamma_z$$ is a zip code fixed effect.

**Market shares.** Following the standard discrete-choice procedure of Berry et al. (1995), with Type-I extreme-value utility shocks, the aggregate share of branch $$j$$ in market $$m$$ is (p. 21, Eq. 4):

$$
s_{j,z,m,b,t} = \frac{\exp(V_{j,z,m,b,t})}{\sum_{k \in \Gamma^m} \exp(V_{k,z,m,b,t}) + \exp(V_{O,m,t})} \tag{4}
$$

where $$\Gamma^m$$ is the set of branches available in market $$m$$ and $$V_{O,m,t}$$ is the outside option value (normalized so $$\xi_O = 0$$).

**Bank profit and pricing regimes.** A multi-market bank $$b$$ owns branches $$j$$ across markets $$m \in \Omega_b$$, paying marginal cost $$c_b$$ per unit of deposits and fixed cost $$C_{jbm}$$ per branch. Let $$R_{bm} = \tilde{R}_{bm} - c_b$$ denote bank-market returns net of marginal costs. Bank profits are (p. 22, Eq. 6):

$$
\Pi_b = \sum_{m \in \Omega_b} \sum_{j \in m} \left[ (R_{bm} - r_{jbm}) \, s_{jbm} \, D_m - C_{jbm} \right] \tag{6}
$$

Under **local pricing**, each branch rate $$r_{jbm}$$ is chosen independently. The first-order condition sets the deposit spread equal to the inverse of the local demand semi-elasticity (p. 23, Eq. 7):

$$
R_{bm} - r_{bm} = \frac{1}{a_m \bigl(1 - s_{bm}(r_{bm},\, \mathbf{r}_{b'/m})\bigr)} \tag{7}
$$

Under **uniform pricing**, the bank sets a single rate $$r_b$$ across all branches and markets. The optimal uniform rate satisfies (p. 23, Eq. 8):

$$
R_b - r_b = \frac{1}{\displaystyle\sum_{m \in \Omega_b} a_m \bigl(1 - s_{bm}(r_b,\, \mathbf{r}_{b'/m})\bigr)\, \xi_{b,m}} \tag{8}
$$

where $$\xi_{b,m} = s_{bm} D_m \big/ \sum_{m' \in \Omega_b} s_{bm'} D_{m'}$$ is the share of bank $$b$$'s total deposits held in market $$m$$. The uniform rate thus equals the inverse of the deposit-weighted average of local demand semi-elasticities. This has the key implication that any merger changes $$r_b$$ for all markets where bank $$b$$ operates, including markets that have no overlap with the acquired bank. In contrast, the local pricing model predicts no adjustment in non-overlapping markets.

**Consumer welfare.** Following Small and Rosen (1981), the change in consumer welfare in market $$m$$ induced by a policy intervention is (p. 24):

$$
\Delta W_m = \ln\!\left(\sum_{j \in \Gamma_m^{post}} \exp V_j^{post}\right) - \ln\!\left(\sum_{j \in \Gamma_m^{pre}} \exp V_j^{pre}\right)
$$

where $$V_j^{pre}$$ and $$V_j^{post}$$ are pre- and post-policy indirect utilities. This formula is applied to evaluate two counterfactuals: (i) merger approved without divestitures, and (ii) merger with the divestitures antitrust authorities actually required.

## Method

**Demand estimation.** The market-specific rate semi-elasticities $$a_m$$ in Eq. (3) are estimated by taking logarithms of market shares relative to the outside option and writing the logit share equation in a linear specification (p. 22, Eq. 5):

$$
\ln s_{j,z,m,b,t} = \sum_m (a_m \, r_{j,t} \times I_m) + \beta_0 X_{j,t} + \beta_1 H_{b,m,t} + \xi_b + \gamma_z + \chi_{m,t} \tag{5}
$$

where $$I_m$$ are market indicator variables and $$\chi_{m,t}$$ are year-market fixed effects that absorb the unobservable outside-option characteristics $$r_{O,t}$$ and $$H_{O,m,t}$$. The specification restricts the sample to banking markets with at least 200 observations to obtain reliable market-level elasticity estimates.

The key endogeneity concern is that deposit rates $$r_{j,t}$$ are chosen by banks in response to local demand conditions, creating a simultaneity bias. Following the Hausman (1996) instrument approach used by Dellavigna and Gentzkow (2019) and Egan et al. (2017), the paper instruments the deposit rate offered at a branch with the average rate offered by other branches of the same bank in other markets. This instrument is strong because branches of the same bank share an equilibrium uniform rate, but it is excludable because the rate differences across markets are driven by the bank-wide uniform rate, not local demand.

Estimated market-specific semi-elasticities average 0.134 (standard deviation 0.111) and are Winsorized and empirical-Bayes-shrunk to reduce sampling noise in thin markets (Fig. 10, p. 23). This average semi-elasticity is lower than estimates from Abrams (2019) (approximately 0.3) and Egan et al. (2017) (0.16-0.60), partly reflecting different sample selection and the use of a finer market-level identification.

**Counterfactual simulation.** The structural model is used to simulate all mergers in the sample and compute post-merger equilibrium deposit rates under both pricing regimes. The procedure: (i) recover bank-market net returns $$R_{bm}$$ from the pre-merger first-order conditions; (ii) reassign acquired branches to the acquirer; (iii) solve the fixed-point system of first-order conditions (Eq. 7 or 8) for all banks under the new ownership structure. In both counterfactuals, acquired branches adopt the non-price characteristics of the acquirer, so quality convergence is held constant and the equilibrium rate differences reflect only pricing conduct.

## Empirical specifications

**Rate convergence around mergers (Table 3, p. 11, R1-R2).** The main reduced-form specification is OLS on the panel of acquired branches (p. 9, Eq. 1):

$$
Y_{i,t,s} = \gamma_{s,t} + \theta_i + \beta \cdot \text{Post-Acquisition}_{i,t,s} + \varepsilon_{i,t,s} \tag{1}
$$

where $$Y_{i,t,s}$$ is the absolute value of the difference between the acquired branch rate and the acquirer's median rate, $$\gamma_{s,t}$$ are state-by-month fixed effects, $$\theta_i$$ are branch fixed effects, $$s \in \{0, 1, \ldots, 12\}$$ post-merger and $$s \in \{-12, \ldots, -1\}$$ pre-merger. The coefficient $$\beta$$ measures the average impact on the absolute rate difference in the twelve months after acquisition. Standard errors are clustered at the merger level. Results shown for four products (12MCD10K, SAV100K, HELOC, Personal) and for both levels and percent differences.

**Deposit volume around mergers (Table 6, p. 15, R4).** To assess how uniform pricing affects branch deposits over a five-year window, Eq. (2) interacts the post-merger indicator with the pre-merger percent rate difference (p. 14, Eq. 2):

$$
Y_{i,t,s} = \gamma_{s,t} + \theta_i + \beta_0 \cdot \text{Post-Acq}_s + \beta_1 \cdot \text{Post-Acq}_s \times \left(\frac{\text{Acquired Branch Rate} - \text{Acquirer Rate}}{\text{Acquirer Rate}}\right)^{\!\text{Pre}} + \varepsilon_{i,t,s} \tag{2}
$$

where $$Y_{i,t,s}$$ is the natural logarithm of total deposits at acquired branch $$i$$ in year $$t$$, $$s$$ years from the merger. The coefficient $$\beta_1$$ measures the average percent change in deposits per one-unit increase in the pre-merger rate difference. Branch- and state-by-year fixed effects are included throughout.

**Rate convergence decomposition by HHI and pre-merger rate quintile (Tables 7-8, p. 16, R5).** To compare the predictive power of HHI changes versus pre-merger rate differences, the following flexible specification is estimated:

$$
Y_{i,t,s} = \gamma_{s,t} + \theta_i + \sum_{k=1}^{5} \beta_k \cdot \text{Post-Acq}_s \times \mathbf{1}[\text{Pre-Diff Rate}_i \in Q_k] + \varepsilon_{i,t,s}
$$

where $$\mathbf{1}[\text{Pre-Diff Rate}_i \in Q_k]$$ are dummy variables for each quintile of pre-merger rate differences $$(\text{Branch Rate} - \text{Acq Median Rate}) / \text{Acq Median Rate}$$. A parallel specification interacts $$\text{Post-Acq}$$ with indicators for HHI change bins (zero, 0-200, 200+ basis points). The incremental adjusted $$R^2$$ from adding quintile interactions is 0.060 (0.783 to 0.843), explaining 27.6% of the unexplained variation from the base specification; the HHI bins add only 0.001. Following Liebersohn (2020), the same pattern holds for the subsample of mergers with predicted post-merger HHI between 1,300 and 1,800 (Table 8), which are below the antitrust trigger threshold and thus free from selection by regulatory intervention.

Robustness includes: large vs. small acquirer partitions (Table 4 Panel B), market overlap vs. non-overlap partitions (Panel C), matched-control difference-in-differences using branches not involved in mergers (Internet Appendix E), and extending the analysis to two and three years post-merger (Internet Appendix F).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| RateWatch (S&P Global Market Intelligence) | Branch-level weekly posted rates for CDs, savings, HELOC, personal loans (4 products); primary source for all interest-rate analysis | [RateWatch](/wiki/commercial/ratewatch/) |
| FDIC Summary of Deposits (SOD) | Annual branch-level deposit balances; used to compute market shares, deposit volumes, and bank market shares for the structural model | [FDIC Summary of Deposits](/wiki/datasets/fdic-summary-of-deposits/) |
| FFIEC Call Reports | Quarterly bank balance-sheet data (total assets, deposits, ROA, NPL, Tier 1 capital); used as bank-level characteristics in demand estimation | [Call Reports](/wiki/datasets/call-reports/) |
| Federal Reserve NIC (public structure data) | Panel of all bank M&A events with dates and FDIC branch identifiers; used to identify ownership transfers and the sample of mergers; only the public BHC ownership and merger-history layer was used, not confidential CAMELS ratings | No page yet (public NIC structure data) |

Sample: January 2006 to December 2019 (merger analysis); 2,177 M&A deals for 12MCD10K, 9,370 acquired branches across 49 states. Uniform pricing documentation (Section 3) uses 2004-2019.

## When to read the full paper

Use the [original](https://doi.org/10.1016/j.jfineco.2025.104204) if you are: designing antitrust review procedures that account for uniform pricing (Section 7.5, Table 9); calibrating a structural model of retail deposit demand and estimating market-specific semi-elasticities (Section 7.1-7.2); assessing the empirical prevalence of uniform pricing in banking (Section 3 and Fig. 2-4); or extending the event-study convergence analysis to other banking products or jurisdictions. The Internet Appendices contain robustness tests on alternative product definitions, sample restrictions, alternative estimation approaches (Appendix G), and the role of HHI at longer horizons (Appendix F).

## Attribution and rights

Source: peer-reviewed, *Journal of Financial Economics* 176 (2026) 104204. DOI: 10.1016/j.jfineco.2025.104204. This distillation was extracted by an LLM on 2026-06-24 and is **not human-verified or independently reproduced**. The article is paywalled (Elsevier B.V., 2025); this page reproduces only short extracts for academic commentary under fair use. No verbatim PDF is hosted here.

> Granja, João, and Nuno Paixão. "Bank consolidation and uniform pricing."
> *Journal of Financial Economics* 176 (2026) 104204.
> DOI: 10.1016/j.jfineco.2025.104204.
> © 2025 Published by Elsevier B.V. All rights reserved.
> Extract-only. This page is an LLM-distilled summary by the Institute for Automated Research.


==============================================================================
# Prospect Theory in the Field: Han, Sui & Yang (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jfe/2026/han-prospect-theory-field-revealed-2026/
# Distilled: Funds whose past returns generate higher prospect theory value attract larger future flows, confirmed by panel regressions and account-level trading data from January 1981 to June 2022. A revealed preference analysis recovers loss aversion of 1.824 and curvature of 0.745, aligned with lab-based studies. Journal of Financial Economics 2026, CC BY 4.0. Seven core results with source locators, datasets used, the prospect theory valuation framework, and the empirical specifications.
# Tags: paper-summary, behavioral-finance, prospect-theory, mutual-funds, fund-flows, revealed-preference
==============================================================================

**What this is.** The paper's core results, the prospect theory valuation framework it applies, and the empirical specifications behind the results: enough to know what it found and how, without reading all 20 pages. To replicate or extend it, read the original at the [DOI](https://doi.org/10.1016/j.jfineco.2025.104221).

## TL;DR

The paper tests whether prospect theory governs mutual fund investors' choices by constructing a prospect theory value measure (termed TK after Tversky and Kahneman 1992) for each fund from its past 60-month return distribution, then linking it to future flows. Analyzing roughly 2,698 active US equity mutual funds per month from January 1981 to June 2022, the paper finds that funds with higher TK values attract significantly larger subsequent flows, with a high-minus-low TK decile spread of about 1.4 percentage points per month. All four components of prospect theory (loss aversion, concavity/convexity, probability weighting, and reference dependence) independently and jointly predict flows. A revealed preference analysis using a discrete choice model on quarterly fund subscription data estimates a field loss-aversion coefficient of 1.824, between the lab-based values from Tversky and Kahneman (1992) and more recent meta-analytic estimates. Account-level evidence confirms that individual investors hold more and net-buy more of high-TK funds. TK-driven flows are followed by negative subsequent fund performance, pointing to a "dumb money" pattern consistent with non-fully rational demand. The paper extends the scope of Barberis et al. (2021), who document prospect theory in stock market anomalies, to the mutual fund investor demand setting, and provides a broader analysis than concurrent work by Gu and Yoo (2021).

## Core results

Magnitudes and significance are as reported; `\*\*`/`\*\*\*` = 5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | High-minus-Low TK decile fund flow spread is positive, monotone, and significant | Table 2, p. 6 | EW: H-L = 1.3% per month (t=8.50); TNA-weighted: H-L = 1.4% per month (t=10.47); monotone from Low (-0.4%) to High (+0.9%) EW |
| R2 | TK positively and significantly predicts future fund flows in panel regressions | Table 3 col (4), p. 8 | Full-controls (fund+date FE): coeff=0.383\*\*\* (t=11.27); univariate: 0.604\*\*\* (t=26.89); one-SD increase in TK raises monthly flow by ~$11.16 million for median-TNA fund |
| R3 | Each feature of prospect theory independently predicts flows; all remain significant when combined | Table 5, p. 9 | LA alone: 0.839\*\*\* (t=14.17); CC alone: 0.974\*\*\* (t=15.12); PW alone: 0.623\*\*\* (t=9.59); joint regression: LA=0.415\*\*\* (t=5.77), CC=0.598\*\*\* (t=7.78), PW=0.233\*\*\* (t=3.55), TK=0.383\*\*\* (t=11.27) |
| R4 | Revealed preference analysis from field data recovers prospect theory parameters aligned with lab estimates | Table 8 Panel A, p. 13 | lambda (loss aversion)=1.824 (99% CI [1.529, 2.119]); alpha (curvature)=0.745; gamma (PW gain)=0.110; delta (PW loss)=0.228 |
| R5 | Account-level evidence confirms individual investors hold more and net-buy more of high-TK funds | Table 7, p. 11 | Holdings/balance%: coeff=45.357\*\*\* (t=4.73); NetBuy/balance%: 6.260\*\*\* (t=3.95) |
| R6 | TK predicts new fund subscriptions (purchase decisions) but not redemptions (sell decisions) | Table 6, p. 10 | New subscriptions: 0.214\*\*\* (t=2.31); Redemptions: -0.047 (t=-0.43, insignificant) |
| R7 | TK-driven flows predict negative future fund performance; non-TK-driven flows predict positive future performance | Table 17 Panels B and C, p. 18 | TK-driven flow: -0.001\*\* on 1-month four-factor alpha (t=-2.07); non-TK-driven flow: +0.006\*\*\* (t=5.62) |

**Overall (paper's conclusion).** Prospect theory offers a new framework for understanding mutual fund investor demand. Fund-level and account-level evidence confirm that investors' allocation choices are significantly shaped by each feature of prospect theory. The field-based parameter estimates align with lab-based findings, supporting the external validity of prospect theory as a description of investor preferences. Investors acting on prospect theory allocate more capital to funds with high prospect theory values, but those funds do not subsequently outperform, pointing to a "dumb money" pattern. The predictive power of TK is strongest for retail and broker-sold funds (less sophisticated investors) and weakens during recessions.

## Theory / model

The paper has no equilibrium model. It applies the cumulative prospect theory framework of Tversky and Kahneman (1992), as operationalized for the mutual fund context by Barberis et al. (2016). Kahneman and Tversky (1979) established the foundational properties this framework builds on: reference dependence, loss aversion, and probability weighting. The paper tests the hypothesis that investors evaluate mutual funds using this framework when making capital allocation choices.

**Representation.** For each mutual fund with at least 60 months of past returns, the 60 monthly excess returns (returns minus the risk-free rate) are sorted in ascending order and assigned equal probability 1/60. The resulting mental representation of the fund's future prospects is (p. 4, Eq. 1):

$$\left(r_{-m},\,\tfrac{1}{60};\;\ldots;\;r_{-1},\,\tfrac{1}{60};\;r_1,\,\tfrac{1}{60};\;\ldots;\;r_{n-1},\,\tfrac{1}{60};\;r_n,\,\tfrac{1}{60}\right) \tag{1}$$

where $$r_{-m}$$ to $$r_{-1}$$ are loss outcomes (negative excess returns) and $$r_1$$ to $$r_n$$ are gain outcomes. The reference point is the risk-free rate.

**Valuation: value function.** Each outcome is evaluated by the prospect theory value function (p. 4, Eq. 5):

$$v(x) = \begin{cases} x^{\alpha} & x \geq 0 \\ -\lambda(-x)^{\alpha} & x < 0 \end{cases} \tag{5}$$

where $$\alpha \in (0,1)$$ governs the curvature (concave over gains, convex over losses, giving an S-shaped function) and $$\lambda > 1$$ is the loss aversion coefficient (losses loom larger than gains of equal magnitude).

**Valuation: probability weighting.** Objective probabilities are replaced by transformed decision weights using the weighting functions (p. 4, Eq. 6):

$$w^+(P) = \frac{P^{\gamma}}{(P^{\gamma} + (1-P)^{\gamma})^{1/\gamma}}, \qquad w^-(P) = \frac{P^{\delta}}{(P^{\delta} + (1-P)^{\delta})^{1/\delta}} \tag{6}$$

where $$\gamma, \delta \in (0,1)$$, with smaller values indicating stronger overweighting of tail events. The $$w^+$$ function applies to gains and $$w^-$$ to losses.

**TK measure.** Combining value function and probability weights gives the prospect theory value of a fund (p. 4, Eq. 7):

$$\text{TK} = \sum_{i=-m}^{-1} v(r_i)\!\left[w^-\!\!\left(\tfrac{i+m+1}{60}\right) - w^-\!\!\left(\tfrac{i+m}{60}\right)\right] + \sum_{i=1}^{n} v(r_i)\!\left[w^+\!\!\left(\tfrac{n-i+1}{60}\right) - w^+\!\!\left(\tfrac{n-i}{60}\right)\right] \tag{7}$$

The standard parameter values from Tversky and Kahneman (1992) are used as defaults (p. 4, Eq. 8):

$$\alpha = 0.88, \quad \lambda = 2.25, \quad \gamma = 0.61, \quad \delta = 0.69 \tag{8}$$

**Testable hypothesis.** If prospect theory describes how investors evaluate mutual funds, then funds with higher TK will attract more capital flows, holding other known determinants constant.

## Method

**TK construction.** For each fund-month, the preceding 60 monthly excess returns are sorted ascending, the risk-free rate is used as the reference point, and TK is computed by applying Eq. (7). The paper is robust to alternative reference points (market returns, style benchmarks, zero, and expectation-based benchmarks per Koszegi and Rabin 2006 and 2009) and alternative windows (55, 57, 59, 61, 63, 65 months, and quarterly or semi-annual compounding), as documented in Internet Appendix Tables B1-B9.

**Mutual fund flows.** Fund flow is the percentage growth of net new assets following the standard definition (p. 5, Eq. 9):

$$\text{Flow}_{i,t} = \frac{TNA_{i,t} - TNA_{i,t-1}(1 + r_{i,t})}{TNA_{i,t-1}} \tag{9}$$

where $$TNA_{i,t}$$ is total net assets of fund $$i$$ at the end of month $$t$$ and $$r_{i,t}$$ is the net return. Flows are winsorized at the 5th and 95th percentiles. Factor loadings and alphas use a rolling 60-month window.

**Revealed preference analysis (Section 4.3).** To estimate the prospect theory parameters $$\theta = [\alpha, \lambda, \gamma, \delta]$$ from field data rather than laboratory experiments, the paper specifies a discrete choice model in which each investor in a given quarter chooses among $$J$$ actively managed equity funds plus a Vanguard index fund baseline (fund 0, utility normalized to zero). The utility of fund $$i$$ is (p. 11, Eq. 13):

$$\delta_i = \beta TK_i(\theta, R_i) + c_x \sum_k x_k^i + e_i \tag{13}$$

where $$TK_i(\theta, R_i)$$ is the TK value computed from the fund's historical returns $$R_i$$ under parameter vector $$\theta$$, $$x_k^i$$ are fund characteristics (size, turnover ratio, expense ratio), and $$e_i$$ is an iid type-I extreme-value error. The multinomial logit probability is $$\text{Prob}_j = e^{\delta_j} / \sum_{j=0}^J e^{\delta_j}$$. Quarterly market shares $$s_j = f_j / \sum_{j=0}^J f_j$$ (new subscription shares, constructed from CRSP subscription data) are used to estimate $$\theta$$ by maximum likelihood (p. 11, Eq. 14):

$$\ln L = \ln \prod_{j=0}^{J} \text{Prob}_j^{f_j} \tag{14}$$

MLE is run separately for each quarter from 2013 to 2022, yielding a time series of estimated parameter vectors $$\hat{\theta}_t = [\hat{\alpha}_t, \hat{\lambda}_t, \hat{\gamma}_t, \hat{\delta}_t]$$. Averages and standard errors across quarters are reported in Table 8 Panel A. The approach draws on the `revealed-preference` framework applied to consumer discrete choices.

## Empirical specifications

**Main panel regression (R2, baseline for all robustness).** The central specification regresses monthly fund flow on lagged TK with fund and date fixed effects (p. 6, Eq. 10):

$$\text{Flow}_{i,t} = \beta TK_{i,t-1} + \gamma X_{i,t-1} + \phi_i + \eta_t + \varepsilon_{i,t} \tag{10}$$

where $$\phi_i$$ are fund fixed effects, $$\eta_t$$ are date (year-month) fixed effects, and $$X_{i,t-1}$$ is a vector of controls: cumulative 60-month returns, CAPM alpha, FF4 alpha, factor loadings on market, SMB, HML, and MOM, FF4 R-squared, return volatility, log age, log TNA, expense ratio, and turnover ratio. Standard errors are two-way clustered at the fund and date levels (Table 3). Sample: January 1981 to June 2022, averaging ~860,000 fund-month observations.

**Portfolio sorts (R1).** Each month, funds with at least 60 months of history are sorted into deciles by $$TK_{t-1}$$ and average flows in the subsequent month are computed per decile with both equal weights and TNA weights (Table 2). Time-series means are compared; t-statistics use Newey-West standard errors with 12 lags. The sample spans January 1986 to June 2022.

**Prospect theory feature decomposition (R3).** Three alternative prospect theory values are constructed by fixing the parameters that turn off each feature: LA (loss aversion only: $$\alpha=1, \gamma=1, \delta=1$$); CC (concavity/convexity only: $$\lambda=1, \gamma=1, \delta=1$$); PW (probability weighting only: $$\alpha=1, \lambda=1$$). Each component is substituted for TK in Eq. (10). Table 5 reports results for each component separately and jointly (Table 5 col 5 repeats the baseline with full TK as a reference).

**Alternative specifications for persistent regressors (Table 4).** Three approaches address Stambaugh (1999) bias from the persistent nature of TK: (1) first-difference estimator; (2) recursive demeaning from Hjalmarsson (2010) and Pastor, Stambaugh, and Taylor (2015), using the backward-demeaned TK as an instrument; (3) Amihud and Hurvich (2004) bias correction. TK coefficients range from 0.450 to 0.643 (all significant), confirming the baseline.

**Account-level regressions (R5).** Using retail brokerage data from Barber and Odean (2000) at the investor-fund-month level with account and date fixed effects (p. 10, Eqs. 11-12):

$$\text{AmtHeld}_{i,j,t} = \beta TK_{j,t-1} + \gamma X_{j,t} + \lambda_i + \eta_t + \varepsilon_{i,j,t} \tag{11}$$

$$\text{NetBuy}_{i,j,t} = \beta TK_{j,t-1} + \gamma X_{j,t} + a_i + \eta_t + \varepsilon_{i,j,t} \tag{12}$$

where $$\text{AmtHeld}_{i,j,t}$$ is dollars held in fund $$j$$ by investor $$i$$ at date $$t$$ scaled by account balance (%) or fund size (bps), and $$\text{NetBuy}_{i,j,t}$$ is the net transaction amount similarly scaled. Controls are the same as in Eq. (10) but at the share-class level. Standard errors cluster at account and date levels. Sample: 1991-1996, ~1.3 million investor-fund-month observations.

**Subsequent fund performance (R7).** Fund flows are decomposed into a TK-driven component and a non-TK-driven component by regressing fund-level flows on lagged TK. TK-driven flows are the fitted values $$\hat{\delta}TK_{i,t-1}$$; non-TK-driven flows are the residuals $$\hat{u}_{it}$$. Fama-MacBeth regressions with future Fama-French four-factor alpha as the dependent variable test predictability at horizons of 1, 3, and 12 months (Table 17). Non-TK-driven flows consistently predict positive future alpha; TK-driven flows predict negative future alpha at short horizons.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CRSP Survivor-Bias-Free US Mutual Fund Database | Primary sample: monthly fund returns, TNA, expenses, share classes, new subscriptions, and redemptions | [CRSP Mutual Funds](/wiki/commercial/crsp-mutual-funds/) |
| Thomson Reuters Mutual Fund Holdings (via MFLINKS) | Holdings-based TK measure in Section 5.2; linked to CRSP via MFLINKS | [CRSP Mutual Funds](/wiki/commercial/crsp-mutual-funds/) |
| Morningstar fund ratings (5-year star rating) | Control variable; Morningstar Risk Adjusted Return (MRAR) used in Table 12 | [Morningstar](/wiki/commercial/morningstar/) |
| Kenneth French Data Library | Four-factor returns (market, SMB, HML, MOM) for computing alphas and factor loadings; value-weighted market return | [Ken French library](/wiki/datasets/ken-french/) |
| Barber and Odean (2000) retail brokerage account data | Account-level holdings and transaction data for Section 4.2 account-level evidence | No page yet |

Sample: January 1981 to June 2022 (41 years, monthly). Equity mutual funds only (excluding ETFs, ETNs, variable annuities, index funds). Funds require at least 60 months of history for TK construction. Account-level subsample: 1991 to 1996 from the retail brokerage.

## When to read the full paper

Use the [original](https://doi.org/10.1016/j.jfineco.2025.104221) if you are: testing prospect theory in other financial markets or with other investor populations; replicating the revealed preference parameter estimation (Internet Appendix B and Tables B1-B10 document construction details and robustness); studying investor heterogeneity across fund distribution channels or sophistication levels (Sections 6.1-6.2 and Tables 14-16); or investigating the supply-side response of fund managers to prospect-theory-driven investor demand (Section 5.2 and Table 11). The locators above point to the exact tables.

## Attribution and rights

Source: peer-reviewed, *Journal of Financial Economics*, volume 176 (2026), article 104221. This distillation was extracted by an LLM on 2026-06-24 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Han, Bing, Pengfei Sui, and Wenhao Yang.
> "Prospect theory in the field: Revealed preferences from mutual fund flows."
> *Journal of Financial Economics* 176 (2026): 104221.
> DOI: 10.1016/j.jfineco.2025.104221. © 2025 The Authors. Published by Elsevier B.V.
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Demand Disagreement: Heyerdahl-Larsen & Illeditsch (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jfe/2026/heyerdahl-larsen-demand-disagreement-2026/
# Distilled: An OLG model where investors disagree about future demand for savings (driven by heterogeneous time preferences and a false consensus bias) generates stochastic yield volatility, time-varying bond risk premia, and an upward-sloping yield curve, resolving both the correlation puzzle and the disagreement correlation puzzle without relying on disagreement about macroeconomic fundamentals. Journal of Financial Economics 2026, CC BY 4.0. Six core results with source locators, datasets used, the model (equilibrium SDF and consumption share dynamics), and the method (SPF-based demand disagreement proxy + UKF estimation).
# Tags: paper-summary, asset-pricing, bond-risk-premia, yield-curve, heterogeneous-beliefs
==============================================================================

**What this is.** The paper's core results, the OLG model it proposes (equilibrium SDF, consumption share dynamics, and bond pricing), and the empirical method (SPF-based demand disagreement proxy and UKF estimation): enough to know what it found and how, without reading all 16 pages. To replicate or extend, read the original at [doi.org/10.1016/j.jfineco.2025.104191](https://doi.org/10.1016/j.jfineco.2025.104191).

## TL;DR

Standard heterogeneous-beliefs models predict that disagreement about asset prices should be tightly linked to disagreement about macroeconomic fundamentals. The data contradict this: a large fraction of yield disagreement remains after conditioning on every macro variable the Survey of Professional Forecasters (SPF) tracks, as documented by Giacoletti, Laursen and Singleton (2021). Albuquerque, Eichenbaum, Luo and Rebelo (2016) highlight a related puzzle for equity returns. Heyerdahl-Larsen and Illeditsch call this the "disagreement correlation puzzle" and resolve it with a model where investors disagree not about fundamentals but about future demand for savings, driven by differing time preferences and a false consensus bias. In equilibrium, this demand disagreement introduces a priced demand shock, generating stochastic yield volatility, time-varying bond risk premia, and an unconditionally upward-sloping yield curve. Using the component of SPF yield disagreement orthogonal to macro disagreement as their proxy, the paper confirms that demand disagreement is positively related to yield volatility and predicts future excess bond returns, consistent with the model.

## Core results

Magnitudes are as reported; locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Macro-fundamental disagreement explains only a modest portion of yield disagreement, especially in the changes specification | Table 2, p. 4 | Changes R² ranges from 0.22 (3Q ahead, lowest) to 0.32 (1Q ahead, highest); Level R² = 0.72-0.77; substantial residual unexplained by fundamentals |
| R2 | Demand disagreement is strongly positively associated with nominal yield volatility across all maturities (1-5Y), both in model and data | Table 3, p. 12 | Standardized coefficients in data: 0.66-0.73 (t-stats 11.5-14.1); model values: 0.99 (near one-for-one) |
| R3 | Demand disagreement positively predicts excess bond returns across maturities (2-5Y), both in model and data | Table 4, p. 12 | Standardized coefficients in data: 0.38-0.44 (t-stats 4.9-5.8); model coefficients: 0.15 |
| R4 | Model-implied real yields closely track observed real yields; slope coefficient of one not rejected | Table 5, p. 13 | Regression coeff 0.52-0.62, R² = 0.63-0.69 across maturities 2-10Y |
| R5 | Model-implied nominal bond risk premia significantly predict realized premia | Table 5, p. 13 | Regression coeff 1.28-1.54, R² = 0.08-0.18 across maturities 2-5Y |
| R6 | Baseline calibration generates a positive equity premium and excess bond return that would be essentially zero without demand disagreement | Fig. 3, p. 8 | Equity risk premium = 3.6%, stock market volatility = 22.8% at sigma\_d = 0.8; without disagreement (sigma\_d = 0) equity premium = 0.11% |

**Overall (paper's conclusion).** Demand disagreement, driven by investors' incomplete knowledge of the evolving mix of patient and impatient agents in the economy, accounts for a large share of observed yield disagreement and explains stylized facts in bond and equity markets that macro-fundamentals-based models leave unresolved: positive and time-varying risk premia, yield volatility, and an upward-sloping term structure. The single-shock OLG structure is parsimonious but delivers good fits to yields, yield volatilities, and bond risk premia when estimated on SPF and TIPS data.

## Theory / model

The economy is a continuous-time OLG model in the tradition of Blanchard (1985) and Garleanu and Panageas (2015). The paper builds on Ehling, Gallmeyer, Heyerdahl-Larsen and Illeditsch (2018a), who study inflation disagreement and the yield curve, shifting the focus to time-preference (demand) disagreement. A continuum of agents is born at rate $$\nu > 0$$ and dies at the same rate. There are two investor types: patient (type $$a$$, low discount rate $$\rho^a$$) and impatient (type $$b$$, high discount rate $$\rho^b > \rho^a$$). The fraction of patient newborns is $$\alpha_t$$, where $$\alpha_t = 1/(1+\exp(-l_t))$$ and $$l_t$$ follows a mean-reverting Ornstein-Uhlenbeck process (Section 3.1, p. 5):

$$
dl_t = \kappa(\bar{l} - l_t)\,dt + \sigma_l\,dZ_{a,t} \tag{4}
$$

This demand shock $$Z_{a,t}$$ is independent of the output supply shock $$Z_{Y,t}$$. Aggregate output $$Y_t$$ follows geometric Brownian motion with drift $$\mu_Y$$ and volatility $$\sigma_Y$$. Each agent has log utility over consumption, common endowment, and trades in four securities: the stock, a risk-free bond, and a consol bond (zero-net-supply) driven by $$Z_{Y,t}$$ and $$Z_{a,t}$$ respectively.

**False consensus bias and belief dynamics.** Both types observe $$\alpha_t$$ and $$l_t$$, but disagree about the long-run mean $$\bar{l}$$: patient investors are optimistic ($$\bar{l}^a \geq \bar{l}$$), impatient investors pessimistic ($$\bar{l}^b \leq \bar{l}$$) (Section 3.2, p. 5). The likelihood ratio capturing belief disagreement is (Eq. 8, p. 6):

$$
\eta_t \equiv \frac{\eta_t^b}{\eta_t^a} = \exp\!\left(-\tfrac{1}{2}\sigma_d^2\,t - \sigma_d\,Z_{a,t}\right), \quad \sigma_d = \frac{2\kappa}{\sigma_l}(\bar{l} - \bar{l}^b) \tag{8}
$$

The parameter $$\sigma_d \geq 0$$ measures the degree of demand disagreement between types.

**Equilibrium SDF.** Proposition 1 (p. 6) establishes the equilibrium SDF:

$$
\xi_t = \frac{X_t}{Y_t}, \qquad X_t = \int_{-\infty}^t \nu e^{-\nu(t-s)}\!\left(\alpha_s\beta_s^a e^{-\rho^a(t-s)}\frac{\eta_t^a}{\eta_s^a} + (1-\alpha_s)\beta_s^b e^{-\rho^b(t-s)}\frac{\eta_t^b}{\eta_s^b}\right)X_s\,ds
$$

The SDF is inversely proportional to aggregate output $$Y_t$$ and depends on the process $$X_t$$, which captures heterogeneity in time discount rates and beliefs. The key state variable that fully describes all asset prices is the consumption share of patient investors, $$f_t$$.

**Risk-free rate and market price of demand risk.** Proposition 3 (p. 6) gives the equilibrium short rate and market prices of risk (Eqs. 12-13):

$$
r_t = \mathcal{E}_f[\rho] + \mu_Y - \sigma_Y^2 + \nu\bigl(1 - \alpha_t\beta_t^a - (1-\alpha_t)\beta_t^b\bigr) \tag{12}
$$

$$
\theta_{a,t} = \sigma_d\!\left(\tfrac{1}{2} - f_t\right) \tag{13}
$$

where $$\mathcal{E}_f[\rho] = f_t\rho^a + (1-f_t)\rho^b$$ is the consumption-share-weighted discount rate and $$\beta_s^i = (\rho^i + \nu)\phi_t$$ is the initial consumption-wealth ratio. The market price of demand shock risk $$\theta_{a,t}$$ is strictly decreasing in $$f_t$$: when patient investors dominate (high $$f_t$$), the consol appears overpriced and $$\theta_a < 0$$; when impatient investors dominate (low $$f_t$$), the market price of demand risk is positive.

**Consumption share dynamics.** Proposition 5 (p. 9) gives the stochastic differential equation for $$f_t$$ (Eq. 15):

$$
df_t = \mu_{f,t}\,dt + \sigma_{f,t}\,dZ_{a,t} \tag{15}
$$

$$
\mu_{f,t} = \nu\bigl(\alpha_t\beta_t^a(1-f_t) - (1-\alpha_t)\beta_t^b f_t\bigr) + (\rho^b - \rho^a)f_t(1-f_t) + \sigma_d^2\!\left(\tfrac{1}{2} - f_t\right)f_t(1-f_t)
$$

$$
\sigma_{f,t} = f_t(1-f_t)\sigma_d
$$

The diffusion $$\sigma_{f,t}$$ is maximized at $$f_t = 0.5$$ and vanishes at the extinction boundaries, so demand shocks have the largest impact when the economy is balanced between types.

**Consol price and stock return.** Corollary 2 (p. 7) establishes that the consol price equals the wealth-consumption ratio $$\phi_t$$, and its return is exposed only to the demand shock $$Z_{a,t}$$. The stock price $$P_{s,t} = Y_t\phi_t$$ and its return dynamics are (Eq. 14):

$$
d\log P_{s,t} = d\log Y_{s,t} + d\log B_t^C \tag{14}
$$

separating cash-flow exposure from exposure to the consol. Both the stock and the consol inherit stochastic risk premia through their exposure to the demand shock price $$\theta_{a,t}$$.

## Method

**Demand disagreement proxy (Eq. 2-3).** Because demand disagreement is not directly observable, the paper constructs a proxy from SPF data (Section 2.3, p. 4). Every forecaster $$i$$ in the SPF reports a forecast for the three-month T-bill rate $$\hat{y}^i_{t,t+\Delta t}$$ and for macro fundamentals $$\hat{X}^{i,LF}_{t,t+\Delta t}$$ for horizons $$\Delta t \in \{1Q,2Q,3Q,4Q\}$$. The paper first regresses the yield forecast on macro forecasts in a pooled cross-section:

$$
\hat{y}^i_{t,t+\Delta t} = \beta_0 + \beta'_X \hat{X}^{i,LF}_{t,t+\Delta t} + \beta_y \hat{y}_{t,t+\Delta t} + \varepsilon^i_{t,t+\Delta t} \tag{2}
$$

The residual $$\varepsilon^i_{t,t+\Delta t}$$ captures yield disagreement unrelated to macro fundamentals. The demand disagreement proxy for period $$t$$ is the cross-sectional standard deviation of this residual:

$$
DD_{t,t+\Delta t} = \text{SD}_t(\varepsilon^i_{t,t+\Delta t}) \tag{3}
$$

This proxy accounts for 65% of total yield disagreement at the one-quarter horizon and 76% at the four-quarter horizon (Fig. 1, p. 4).

**UKF state estimation.** To estimate the two latent state variables $$(f_t, \alpha_t)$$, the paper uses the Unscented Kalman Filter (UKF) (Section 5.7, p. 12; Fig. 6, p. 13). The two observables used for filtering are the demand disagreement proxy $$DD_t$$ and the two-year TIPS yield, covering Q1 1999 to Q2 2024. The observation equations are:

$$
y_t = h(l_t, f_t) + e_{y,t}, \qquad DD_t = g(l_t, f_t) + e_{DD,t}
$$

with $$\alpha_t = 1/(1+\exp(-l_t))$$ and diagonal noise covariances. The UKF is preferred over the EKF because it avoids linearization: instead of approximating the nonlinear observation equations, the UKF propagates the state distribution through them directly using a set of deterministic sigma points.

**Empirical tests.** Yield volatility regressions (Table 3) use an AR(1)-GARCH(1,1) model to estimate nominal yield volatility, then run panel regressions of volatility on the demand disagreement proxy with Newey-West standard errors (4 lags). Bond risk premium regressions (Table 4) define the excess return of a $$T$$-maturity bond (T in quarters) as $$rx^{(T)}_{t,t+4} = Ty^{(T)}_t - (T-4)y^{(T)}_{t+4} - y^{(1)}_t$$ and regress on $$DD_t$$.

## Empirical specifications

**Yield volatility regression (R2, Table 3).** For each maturity $$T \in \{1,2,3,4,5\}$$ years, the paper estimates:

$$
\sigma^{(T)}_{y,t} = \gamma_0 + \gamma_1\,DD_t + \epsilon_t \tag{Table 3}
$$

where $$\sigma^{(T)}_{y,t}$$ is the conditional standard deviation of the $$T$$-year yield from an AR(1)-GARCH(1,1) fitted to actual quarterly yields. Coefficients are standardized (both variables normalized) and standard errors are Newey-West with 4 lags. Standardized coefficients range from 0.66 to 0.73 in the data (t-stats 11.5-14.1), and the model produces standardized coefficients of 0.99, consistent with the empirical magnitudes.

**Bond risk premium regression (R3, Table 4).** For each maturity $$T \in \{2,3,4,5\}$$ years (equivalently $$T \in \{8,12,16,20\}$$ quarters), the one-year holding period excess bond return is:

$$
rx^{(T)}_{t,t+4} = \gamma_0 + \gamma_1\,DD_t + \epsilon_t \tag{Table 4}
$$

where $$rx^{(T)}_{t,t+4} = Ty^{(T)}_t - (T-4)y^{(T)}_{t+4} - y^{(1)}_t$$ (T in quarters; positive when bonds appreciate). Standardized coefficients range from 0.38 to 0.44 in the data (t-stats 4.9-5.8). The model predicts smaller but same-sign coefficients (0.15). The predictive relationship is robust to controlling for yield levels and macroeconomic disagreement (Tables 12-15, Internet Appendix).

**Goodness-of-fit (R4-R5, Table 5).** Using filtered state variables $$(f_t, \alpha_t)$$, the paper generates model-implied time series for real yields (maturities 2, 3, 5, 7, 10Y and slope), real yield volatilities, and nominal bond risk premia, then regresses observed values on model-implied counterparts. The regression $$y^{\text{Data}} = a + b\,y^{\text{Model}} + u$$ should have $$b = 1$$ and high $$R^2$$ if the model captures the data. Real yields: $$b = 0.52-0.62$$, $$R^2 = 0.63-0.69$$; nominal bond risk premia: $$b = 1.28-1.54$$, $$R^2 = 0.08-0.18$$. The slope coefficient of one is not rejected for real yields or nominal bond risk premia.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Survey of Professional Forecasters (SPF), Philadelphia Fed | Yield and macro forecasts; source of the demand disagreement proxy DD (Q3 1981-Q2 2024, quarterly) | no page yet |
| TIPS yields (2-year, Federal Reserve/FRED) | Observable for UKF state estimation (Q1 1999-Q2 2024) | [FRED](/wiki/datasets/fred/) |
| Shiller long-run U.S. stock market and macro data | Correlation puzzle motivation (Table 1): annual stock returns, dividends, consumption, one-year yield (1891-2009) | [Shiller data](/wiki/datasets/shiller-data/) |

Sample: SPF quarterly data from Q3 1981 to Q2 2024 (172 quarters) for the main empirical results; Shiller annual data from 1891 to 2009 for the motivating correlation puzzle in Table 1. Replication data and code are deposited at Mendeley Data and Zenodo (links on the article landing page).

## When to read the full paper

Use the [original](https://doi.org/10.1016/j.jfineco.2025.104191) if you are: building or calibrating a heterogeneous-beliefs OLG model for bond markets; constructing a survey-based measure of disagreement orthogonal to macro fundamentals; extending the framework to recursive preferences, production, or learning from experience (Section 6 outlines these; the Internet Appendix contains the full derivations); or evaluating the UKF filtering approach for latent state estimation in a nonlinear OLG model. Table 3 (yield volatility) and Table 4 (bond return predictability) give the cleanest empirical entry points; Fig. 3 (unconditional moments vs. sigma\_d) is the key calibration diagnostic.

## Attribution and rights

Source: peer-reviewed, *Journal of Financial Economics* 175 (2026) 104191. This distillation was extracted by an LLM on 2026-06-24 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Heyerdahl-Larsen, Christian, and Philipp Illeditsch.
> "Demand disagreement."
> *Journal of Financial Economics* 175 (2026) 104191.
> DOI: 10.1016/j.jfineco.2025.104191. © 2025 The Authors. Published by Elsevier B.V.
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Implicit Extrapolation and the Beliefs Channel: Liu & Palmer (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jfe/2026/liu-implicit-extrapolation-beliefs-channel-2026/
# Distilled: Households extrapolate past home-price returns into investment allocations beyond what their stated expectations reveal, roughly tripling the estimated effect of past returns on investment relative to a beliefs-only channel. J. Fin. Econ. 2026, paywalled. Seven core results with source locators, datasets used, the Merton portfolio framework, and the main regression specifications.
# Tags: paper-summary, household-finance, beliefs, extrapolation, housing
==============================================================================

**What this is.** The paper's core results on implicit extrapolation in real-estate investment decisions, the theoretical framework (Merton portfolio choice), and the regression specifications that identify the gap between stated beliefs and decision-relevant beliefs: enough to know what it found and how, without reading all 17 pages. To replicate or extend, read the original at [https://doi.org/10.1016/j.jfineco.2025.104172](https://doi.org/10.1016/j.jfineco.2025.104172).

## TL;DR

Liu and Palmer document that households extrapolate from perceived past home-price returns when making real-estate investment decisions even after conditioning on their stated expected future returns and stated risk aversion. They call this gap "implicit extrapolation." Using the Survey of Consumer Expectations (SCE) housing module (2015-2021), they find that a 5 percentage-point increase in perceived past HPA raises housing investment by roughly 4.6 pp when allowing for the direct channel, versus only 1.56 pp if past returns only matter through stated expected returns. The confidence mechanism is key: investors who are more confident about their perceived past returns than about their return forecasts rely more heavily on past returns at the investment stage. The findings are consistent with reinforcement learning (Barberis and Jin (2023)) and with ambiguity aversion, and are inconsistent with measurement error in stated beliefs.

## Core results

Magnitudes and significance are as reported; `*`/`**`/`***` = 10%/5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Perceived past HPA **strongly predicts stated one-year-ahead HPA forecasts** even conditional on individual controls and forecasted fundamentals | Table 2, col 4, p. 7 | coeff = 0.24\*\*\* (SE 0.014); N = 6,993; R² = 0.222 |
| R2 | Perceived past HPA **remains a significant, independent predictor of housing investment** conditional on stated forecasted returns | Table 3, col 3, p. 7 | Past returns coeff = 0.71\*\*\* (0.11); Forecasted returns coeff = 0.88\*\*\* (0.15); N = 2,963 |
| R3 | **Confidence in past returns** is an independent predictor of housing investment beyond past returns and forecasted returns | Table 3, col 3, p. 7 | Confidence in Past Returns = 5.48\*\*\* (1.29) pp per unit; above-median risk aversion = -9.63\*\*\* (1.29) |
| R4 | Allowing for implicit extrapolation **multiplies the estimated investment response to past HPA by roughly three** | Text, p. 8 (computed from Tables 2-3) | Via beliefs only: 1.56 pp per 5 pp HPA; combined channels: 4.6 pp per 5 pp HPA |
| R5 | Investors **shift weight from past returns to forecasted returns** as their relative confidence in forecasts rises | Table 5, col 1, p. 9 | (Conf Forecast - Conf Past) x Past Returns = -0.56\*\*\* (0.17); N = 925 |
| R6 | **Forward-looking investors rely on forecasts; backward-looking investors rely on past returns**, each ignoring the other signal | Table 6, cols 1-2, p. 10 | Forward-looking: forecasted returns 1.41\*\*\* (0.27), past returns 0.19 (insig); Backward-looking: past returns 1.16\*\*\* (0.25), forecasted returns 0.42 (insig) |
| R7 | **Rent and inflation forecasts predict stated HPA expectations but are downweighted to zero at the investment stage**; only past returns matter for both | Table 7, cols 1-3, p. 10 | Forecasted rent growth 0.14\*\*\* (0.02) on HPA forecast but -0.09 (0.11) on investment; inflation 0.12\*\*\* (0.03) on HPA forecast but -0.17 (0.15) on investment |

**Overall (paper's conclusion).** Consumers extrapolate from perceived past returns even beyond what their stated expectations reveal, a pattern the paper calls implicit extrapolation. The implied magnitude of the beliefs channel of investment demand is larger than previously estimated. The pattern is driven by differential confidence across signals: investors are systematically more confident about their recalled past returns than about their forecasted returns, leading them to rely on past returns as a conservative or more reliable guide at the investment stage. Andries et al. (2022) document the information-to-beliefs and information-to-decisions channels separately; this paper shows the former does not capture the latter even when the full stated distribution is recorded.

## Theory / model

The classical portfolio choice benchmark is the Merton (1969) model. For an investor with constant absolute risk aversion allocating between a risky asset with return $$r_{t+1}$$ and a risk-free rate $$r_f$$, the optimal risky-asset share is (Eq. 1, p. 2):

$$
\phi_i = \frac{E_i[r_{t+1}] - r_f}{a \sigma_i^2} \tag{1}
$$

where $$E_i[r_{t+1}]$$ is the investor's conditional expectation of the risky asset's return, $$\sigma_i^2$$ is the conditional variance, and $$a$$ is the coefficient of constant absolute risk aversion. Under this model, if stated beliefs are a sufficient statistic for decision-relevant beliefs, the prior period's realized return $$r_t$$ enters $$\phi_i$$ only through its effect on $$E_i[r_{t+1}]$$: there is no direct channel.

The paper's central empirical question is whether stated $$E_i[r_{t+1}]$$ and $$\sigma_i^2$$ are indeed sufficient statistics. If investors hold latent decision-relevant beliefs that differ from stated ones, then $$r_t$$ can affect $$\phi_i$$ even conditional on stated $$E_i[r_{t+1}]$$. The paper documents this as implicit extrapolation: extrapolation that goes beyond what is revealed by expectations surveys.

The paper has no formal model of its own; the formal microfoundation is adapted from Barberis and Jin (2023), which proposes model-free reinforcement learning as the channel. Related evidence in housing comes from Glaeser and Nathanson (2017), who show that extrapolative belief formation can arise endogenously in housing markets. The intuition for the confidence mechanism is that investors engage in both model-based learning (forming explicit forecasts) and model-free learning (using past performance directly to guide decisions), with the relative weight depending on investors' confidence in each signal. Giglio et al. (2021a) show that investors react more to their return forecasts when they are more confident in those forecasts; the present paper extends this by showing that confidence asymmetry across past and future returns drives differential weighting at the investment stage.

## Method

The paper is empirical, using the Survey of Consumer Expectations (SCE) housing module. No structural estimation is performed; estimation is cross-sectional OLS on survey data. The primary estimating equation (Eq. 2, p. 7) is:

$$
Y_{i,t} = \beta_0 + \beta_1 \hat{r}_{i,t} + \beta_2 \hat{E}_i[r_{i,t+1}] + X'_{i,t}\psi + \varepsilon_{i,t} \tag{2}
$$

where $$\hat{r}_{i,t}$$ is respondent $$i$$'s perceived past home-price appreciation (HPA) in their zip code over the prior 12 months, $$\hat{E}_i[r_{i,t+1}]$$ is their stated expected HPA over the next 12 months, $$Y_{i,t}$$ is the investment outcome (primarily housing fund share), and $$X_{i,t}$$ is a vector of demographic controls. The null hypothesis of rational beliefs being a sufficient statistic is $$\beta_1 = 0$$.

The paper builds on `panel-regression` and `randomized-survey-experiment` primitives. Standard errors are heteroskedasticity-robust (Huber-White) throughout. The investment experiment was originally designed by Armona et al. (2018) for the 2015 SCE wave; the current paper reuses it and extends the design to 2020 and 2021 waves. Each wave is fielded to the rotating panel of approximately 1,200 respondents (with a larger cross-section when stacking waves).

Robustness approaches include:
- IV for survey noise in stated beliefs (Appendix F): perceived past returns instrumented with actual CoreLogic zip-code HPA to strip measurement error
- Non-parametric controls for the full distribution of expected future returns (bin fixed effects)
- Interactions with risk tolerance, wealth, and housing equity to rule out risk-aversion confounds
- ACS-SCE reweighted sample for population representativeness (Appendix Table A6)

## Empirical specifications

**Belief formation (Table 2, p. 7):** Regresses stated one-year-ahead HPA forecast on perceived past HPA, progressively adding individual controls and forecasted fundamentals (rent growth, inflation). The estimating equation is the same Eq. 2 above with the HPA forecast as the outcome. This identifies explicit extrapolation: the degree to which stated forecasts reflect past returns.

**Main investment result (Table 3, p. 7):** Runs Eq. 2 with housing fund share as the outcome. Key columns:
- Col 1: Forecasted returns alone (coefficient = 1.30***)
- Col 2: Perceived past returns alone (coefficient = 1.01***)
- Col 3: Both jointly (forecasted returns = 0.88***, past returns = 0.71***), plus confidence dummies; N = 2,963; R² = 0.047

Columns 4-6 add full individual controls. The coefficient on perceived past returns in column 6 is 0.54*** (SE 0.11), and on forecasted returns is 0.93*** (0.14).

**Role of risk aversion (Table 4, p. 8):** The risk-tolerance score (1-10) enters positively and significantly (3.70*** in bivariate), confirming risk aversion matters, but past returns remain significant after nonparametric controls for the risk-tolerance distribution and the full distribution of expected returns.

**Confidence mechanism (Table 5, p. 9):** Adds the confidence-gap variable (Confidence in Forecast Returns minus Confidence in Past Returns, scaled 1-5) and its interactions with both return signals. Key interaction on past returns is -0.56*** (0.17), showing higher relative confidence in forecasts reduces reliance on past returns. The 2020-2021 subsample of 925 respondents received the confidence elicitation module.

**Forward vs. backward-looking investors (Table 6, p. 10):** Splits the 2020-2021 sample by self-reported reliance on past vs. expected returns. Forward-looking respondents (N = 772) show significance only on forecasted returns (1.41***); backward-looking respondents (N = 613) show significance only on past returns (1.16***). Column 3 of Table 6 pools both and includes a Forward-Looking indicator and its interactions.

**Factor reweighting (Table 7, p. 10):** Column 1 estimates the belief-formation equation for HPA forecasts (with forecasted rent growth 0.14*** and inflation 0.12***); columns 2-3 run the investment equation, finding both rent and inflation are insignificant predictors of investment allocation. This rules out pure measurement-error explanations: noise would not cause only certain belief factors (rent, inflation) to lose relevance while past returns retain it.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Survey of Consumer Expectations (SCE), FRBNY, housing module | Primary data: investment allocations, perceived past HPA, stated forecasted HPA, confidence, demographics; 2015, 2020, 2021 waves | No page yet |
| CoreLogic repeat-sales price index | Zip-code-level actual HPA for constructing Perception Gap; used as IV in robustness | No page yet |

Sample: SCE respondents 2015-2021, N up to ~7,000 per wave; main investment-experiment sample is N = 2,963-3,015 (2015 wave); confidence subsample is N = 925 (2020-2021).

## When to read the full paper

Read the [original](https://doi.org/10.1016/j.jfineco.2025.104172) if you are studying: how stated beliefs and decision-relevant beliefs diverge (Appendix C formalizes the confidence mechanism); heterogeneity in implicit extrapolation across demographic groups (Tables 8-9); robustness to IV, bin-fixed-effects, and wealth-channel alternatives (Appendix Tables A2-A14); the open-ended survey evidence on why investors rely on past returns (Section 4.1, Figure 3); or other real-estate investment outcomes beyond the fund-share experiment (Table A10). The locators above point to the exact tables.

## Attribution and rights

Source: peer-reviewed, *Journal of Financial Economics* 175 (2026), article 104172. This distillation was extracted by an LLM on 2026-06-24 and is **not human-verified or independently reproduced**. The paper is paywalled; extract-only applies. For text-mining inquiries see the Elsevier TDM licence at https://www.elsevier.com/tdm/userlicense/1.0/.

> Liu, Haoyang, and Christopher Palmer. "Implicit extrapolation and the beliefs channel of investment demand." *Journal of Financial Economics* 175 (2026): 104172. DOI: 10.1016/j.jfineco.2025.104172. © 2025 Elsevier B.V. All rights reserved.


==============================================================================
# Discount Factors and Monetary Policy: Vandeweyer, Yang & Yannelis (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jfe/2026/vandeweyer-discount-factors-monetary-policy-2026/
# Distilled: Exploiting dual-listed stocks in Mainland China and Hong Kong to isolate the discount factor channel of monetary policy, the paper finds that US FOMC surprises cause significant revisions in investors' risk-adjusted discount factors: a 100 bp surprise shifts the A/H share-price ratio by about 30 bp within five trading days, driven exclusively by cycle-amplifying surprises. J. Fin. Econ. 2026, paywalled. Six core results with source locators, datasets used, the conceptual model, and the estimating equations.
# Tags: paper-summary, monetary-policy, asset-pricing, equities, macro, international-finance
==============================================================================

**What this is.** The paper's core results, the conceptual model that motivates the A/H
ratio design, and the estimating equations: enough to understand what the paper found and
how the discount factor channel is identified, without reading all 18 pages. To replicate
or extend, read the original at [https://doi.org/10.1016/j.jfineco.2025.104190](https://doi.org/10.1016/j.jfineco.2025.104190).

## TL;DR

The paper uses dual-listed stocks (A-shares in Mainland China, H-shares in Hong Kong) as a
laboratory to isolate the discount factor channel of monetary policy transmission to stock
prices. Since both share types represent claims on the same firm cash flows, the ratio of
A-share to H-share prices (the A/H ratio) cancels out cash flow news and reflects only
differences in investors' discount rates across the two segmented markets. US Federal
Open Market Committee (FOMC) monetary policy surprises, as measured by Kuttner (2001)
fed-funds futures, significantly shift the A/H ratio, with a 100 basis-point surprise
causing roughly a 30 basis-point change within five trading days. The effect is concentrated
in cycle-amplifying surprises (surprise rate cuts during easing cycles and surprise rate
hikes during tightening cycles), while contradictory surprises are insignificant. Cross-
sectionally, value stocks, small stocks, and high-beta stocks respond more strongly,
consistent with standard asset pricing theory in which discount rate revisions have
disproportionately large effects on risky, short-duration cash flows.

## Core results

Magnitudes and significance as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%.
Locators refer to the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **A 100 bp Fed surprise shifts the A/H ratio by ~30 bp within five trading days**, isolating the discount factor channel | Table 3, cols 1-5, p. 8 | Surprise x Post = 0.293\*\*\* (SE 0.0297) with company FE and post-announcement indicator (col 3); stable across specifications: range 0.276 to 0.306, all significant at 1% |
| R2 | **The effect is asymmetric: only cycle-amplifying surprises matter**. Surprise rate cuts during easing cycles and surprise rate hikes during tightening cycles move the A/H ratio significantly; contradictory surprises do not | Table 4, Panels A-B, p. 8 | Amplifying-hike: 0.390\*\*\* (0.112); amplifying-cut: 0.355\*\*\* (0.0400); contradictory-hike: -0.0964 (0.107); contradictory-cut: 0.0467 (0.179) |
| R3 | **US monetary policy passes through ~1-for-1 to Hong Kong interbank rates (HIBOR) but leaves Mainland China interbank rates (SHIBOR) unaffected**, validating the market segmentation assumption | Figure 3, p. 6 | 1 pp Fed funds surprise generates ~1 pp increase in 3-month HIBOR by day 1; LIBOR (USD) also rises; SHIBOR coefficient near zero and statistically insignificant |
| R4 | **Value firms (low PE ratio) show roughly twice the discount-factor sensitivity of growth firms** | Table 6, Panel A, p. 11 | Below-median PE: Surprise x Post = 0.346\*\*\* (0.0651); above-median PE: 0.162\*\*\* (0.0485); triple interaction Surprise x Post x Charact = -0.161\* (0.0908) |
| R5 | **Small firms (low market capitalization) are more sensitive to discount-factor revisions than large firms** | Table 6, Panel B, p. 11 | Below-median MC: Surprise x Post = 0.367\*\*\* (0.0575); above-median MC: 0.237\*\*\* (0.0346); triple interaction = -0.0889\*\*\* (0.0210) |
| R6 | **High-CAPM-beta stocks react more strongly than low-beta stocks**, consistent with discount rate revisions hitting riskier cash flows disproportionately | Table 6, Panel C, p. 11 | Below-median beta: Surprise x Post = 0.220\*\*\* (0.0638); above-median beta: 0.381\*\*\* (0.0730); triple interaction = 0.362\*\* (0.146) |

**Overall (paper's conclusion).** Monetary policy announcements cause investors to revise
their discount factors and impact stock prices. The discount factor channel is substantial
and survives controls for cash-flow news: professional analysts' EPS forecasts for the
same firms do not diverge across the two regions following FOMC announcements (ruling out
the information channel of Nakamura and Steinsson (2018)), and non-exporting firms (with
no US revenue channel) react identically to exporters. The
asymmetry toward cycle-amplifying surprises suggests that higher-frequency event-study
strategies that do not control for business-cycle context may understate the effects of
monetary policy on asset prices.

## Theory / model

The conceptual framework follows Section 2.2 (p. 4). Consider an economy with two
segmented regions, A (Mainland China) and H (Hong Kong), populated by different investors
but trading integrated firms whose shares represent claims on the same future cash flows.
The monetary policy stance in each region is captured by $$m_A$$ and $$m_H$$, interpreted as
the central bank's reaction function across all policy instruments.

Following Cochrane (2005), the price of stock $$i$$ in region $$J \in \{A, H\}$$ equals the
expected discounted cash flow under the region's risk-adjusted discount factor, and following
Cieslak and Pang (2021) the discount factor decomposes into a risk-free rate and a risk
premium component (Eq. 1, p. 4):

$$
P_J^i = \frac{1}{R_J^i(m_J)} E\!\left[x^i(m_A, m_H)\right], \tag{1}
$$

where $$R_J^i(m_J)$$ is the risk-adjusted discount factor applied by region-$$J$$ investors to
stock $$i$$, and $$E[x^i(m_A, m_H)]$$ is the common cash-flow expectation (identical across
regions under integrated trade). The discount factor decomposes into a risk-free component
and a firm-specific risk premium:

$$
R_J^i = r^f(m_J) + r^{p,i}(m_J), \qquad J \in \{A, H\}.
$$

Taking the ratio of A-share price to H-share price (Eq. 2, p. 4):

$$
\frac{P_A^i}{P_H^i} = \frac{R_H^i(m_H)}{R_A^i(m_A)}. \tag{2}
$$

The common cash-flow term $$E[x^i]$$ cancels exactly. The A/H ratio therefore depends only
on the ratio of discount factors across the two regions, not on firms' cash flows. A change
in the A/H ratio around an FOMC announcement reflects changes in relative discount
factors attributable to monetary policy, not information about earnings or dividends.

The identification logic rests on two institutional facts. First, the Hong Kong dollar is
pegged to the USD through the Linked Exchange Rate System (LERS), so the Hong Kong
overnight interbank rate (HIBOR) closely tracks the US federal funds rate (Section 2.1,
p. 3). Second, Mainland China maintains strict capital controls, so investors there are
largely insulated from US monetary policy and have an independent monetary policy stance
via the People's Bank of China. A US monetary policy surprise therefore affects
$$m_H$$ but not $$m_A$$, generating an exogenous shift in $$R_H^i$$ that moves the A/H ratio
without contaminating the Mainland discount factor $$R_A^i$$. Figure 3 (p. 6) confirms this
empirically: a 1 pp Fed funds surprise shifts 3-month HIBOR by approximately 1 pp but
leaves SHIBOR (the Mainland equivalent) near zero.

## Method

The paper applies two estimating frameworks from the macro-finance event-study tradition,
building on `difference-in-differences` and `event-study` designs, with `panel-regression`
for inference.

**Main DiD specification (Eq. 3, p. 4).** The primary specification regresses the A/H ratio
on the interaction of the Kuttner (2001) monetary policy surprise with a post-announcement
indicator. The approach extends Bernanke and Kuttner (2005), who found a roughly 1%
stock-index response to a 100 bp surprise, by isolating the discount factor channel via
the A/H design:

$$
(P_A/P_H)_{ist} = \alpha_i + \eta_s + \lambda_t + \beta\,\text{Surprise}_s \times \text{Post}_t + \epsilon_{ist}, \tag{3}
$$

where $$(P_A/P_H)_{ist}$$ is the A/H share-price ratio for stock $$i$$ around announcement $$s$$
at event time $$t$$; $$\alpha_i$$ are company fixed effects absorbing time-invariant
characteristics (size, location, sector); $$\eta_s$$ are announcement fixed effects absorbing
macroeconomic conditions on each FOMC date; $$\lambda_t$$ are event-time fixed effects;
$$\text{Surprise}_s$$ is the surprise component of the change in the target rate computed
from federal-funds futures following Kuttner (2001); and $$\text{Post}_t$$ equals one for
event times $$t \geq 0$$. The coefficient $$\beta$$ measures the change in the A/H ratio per
percentage-point surprise. Standard errors are clustered at the company level (Section 2.3,
p. 4). The sample covers FOMC announcements from June 2000 to September 2024, excluding
ZLB periods (December 2008 to December 2015) and COVID (March 2020 to March 2022).

**HIBOR passthrough specification (Eq. 4, p. 5).** To validate the transmission assumption,
the paper estimates event-by-event passthrough of the Fed surprise to Hong Kong interbank
rates:

$$
\text{HIBOR}_{st} = \eta_s + \lambda_t + \sum_{\tau=-4}^{5} \beta_\tau^{HB}\,\text{Surprise}_s \times \mathbf{1}(\text{Time Since Announcement}_{st} = \tau) + \epsilon_{st}. \tag{4}
$$

**Dynamic A/H event study (Eq. 5, p. 5).** For graphical assessment of pre-trends and
post-announcement dynamics, the main specification is extended to a full coefficient path:

$$
(P_A/P_H)_{ist} = \alpha_i + \eta_s + \sum_{\tau=-4}^{5} \beta_\tau\,\text{Surprise}_s \times \mathbf{1}(\text{Time Since Announcement}_{st} = \tau) + \epsilon_{ist}. \tag{5}
$$

Figure 4 (p. 8) shows no pre-trend ($$\beta_\tau \approx 0$$ for $$\tau < 0$$), with the effect
peaking around day 3 and showing some mean-reversion by day 5.

## Empirical specifications

**Baseline results (R1, Table 3, p. 8).** Equation (3) is estimated on 93,414
company-announcement observations (143 stocks, 129 announcements). Column (3) adds company
FE and a post-announcement indicator for event-time trend; Column (4) adds company FE and
full event-time FE; Column (5) adds separately trending company-specific time controls.
Across all five specifications the Surprise x Post coefficient ranges from 0.276 to 0.306,
all significant at the 1% level.

**Asymmetry test (R2, Table 4, p. 8).** Announcements are split by interest rate cycle phase
(hiking vs. cutting vs. flat, defined by Figure 2) and by surprise direction (increase vs.
decrease vs. near-zero). This yields a 3 x 3 grid. The key finding is that Panels A and B
show significant coefficients only for amplifying configurations: surprise increases during
hike periods (0.390\*\*\*) and surprise decreases during cut periods (0.355\*\*\*). All other
cells are statistically insignificant, including surprise hikes during cutting cycles and
surprise cuts during hiking cycles (contradictory surprises). Column (3) in each panel
(near-zero surprises) is also insignificant.

**Cross-sectional triple-difference (R4-R6, Eq. 6, p. 10).** To examine heterogeneity in
discount-factor sensitivity, the paper estimates a triple-difference specification adding a
stock characteristic $$\text{Charact}_{ist}$$ (lagged one year) to equation (3). Following
Fama and French (1992), the characteristics tested include price-to-earnings ratio and
market capitalization, which predict cross-sectional stock returns and are interpreted as
proxies for risk-factor exposure:

$$
\begin{aligned}
(P_A/P_H)_{ist} &= \alpha_i + \eta_s + \lambda_t
  + \beta_1 \text{Surprise}_s \times \text{Post}_t
  + \beta_2 \text{Charact}_{ist} \\
  &\quad + \beta_3 \text{Surprise}_s \times \text{Charact}_{ist}
  + \beta_4 \text{Post}_t \times \text{Charact}_{ist} \\
  &\quad + \beta_5 \text{Surprise}_s \times \text{Post}_t \times \text{Charact}_{ist}
  + \epsilon_{ist}.
\end{aligned} \tag{6}
$$

The coefficient $$\beta_5$$ captures differential discount-factor sensitivity by firm
characteristic. Characteristics tested: price-to-earnings ratio (PE, lagged 1 year);
log market capitalization (MC); CAPM beta from 5-year rolling monthly regressions on the
Shanghai Composite Index; and the lagged A/H ratio. Table 6 (p. 11) reports results for
splits below and above the cross-sectional median of each characteristic.

**Robustness (Table 7, p. 11).** Results are stable when including ZLB announcements
(Column 1, coefficient 0.130\*\*\*), excluding near-zero surprises (Column 2, 0.291\*\*\*),
excluding holiday event times (Column 3, 0.286\*\*\*), and combining both holiday and
near-zero exclusions (Column 4, 0.295\*\*\*). Appendix D applies the stacked DiD approach
of Baker et al. (2022), finding larger point estimates (Table D.1, all specifications
approximately 0.445-0.447\*\*\*) consistent with the main results.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Wind Information (China) | Daily closing prices for A-shares and H-shares of 143 dual-listed firms; primary source for the A/H ratio | [no page yet](/wiki/datasets/) |
| HKMA (Hong Kong Monetary Authority) | Historical 3-month HIBOR rates on each trading day; core channel-validation data | [no page yet](/wiki/datasets/) |
| FRED (Federal Reserve Economic Data) | Fed target rate before December 2008; target-rate series for computing Kuttner surprises | [FRED](/wiki/datasets/fred/) |
| Datastream (Refinitiv) | 1-month fed-funds futures; 3-month HIBOR futures; LIBOR historical series (USD) | [no page yet](/wiki/datasets/) |
| Bloomberg | Historical USD-denominated LIBOR series used for placebo test (Figure 3, bottom-left panel) | [no page yet](/wiki/datasets/) |
| SHIBOR (Shanghai Interbank Offered Rate) | Mainland China interbank rate; placebo test confirming capital controls prevent Fed passthrough | [no page yet](/wiki/datasets/) |
| I/B/E/S via WRDS | EPS forecasts for dual-listed firms from Hong Kong and Mainland China brokers; used in robustness (Appendix B) | [WRDS](/wiki/commercial/wrds/) |
| FactSet GeoRev | Firm-level US export revenue share; used to rule out the cash-flow channel in robustness (Appendix A) | [no page yet](/wiki/datasets/) |
| China Foreign Exchange Trade System (CFETS) | HKD-CNY and USD-CNY 6-month forward exchange rates; used to rule out exchange-rate channel (Appendix C) | [no page yet](/wiki/datasets/) |

Sample: June 2000 to September 2024 (daily), 143 dual-listed firms, 129 FOMC announcements
(excluding ZLB and COVID periods), window of 5 trading days before and after each
announcement. Stock prices converted to CNY using daily HKD-CNY exchange rates from Wind.

## When to read the full paper

Use the [original](https://doi.org/10.1016/j.jfineco.2025.104190) if you are:
studying the transmission mechanism of monetary policy to stock prices; working on
macro-finance models that need to distinguish the cash-flow and discount-rate channels;
examining the role of the "Fed Put" and belief updating about the Fed's reaction function;
or designing event studies around FOMC announcements. The appendices (pp. 13-18) contain
robustness tests on US export share (Appendix A), EPS forecast divergence (Appendix B),
exchange rates (Appendix C), stacked DiD (Appendix D), and alternative clustering
(Appendix E). The replication package (pseudo-data) is available via the Mendeley Data
link on the article page.

## Attribution and rights

Source: peer-reviewed, *Journal of Financial Economics* 175 (2026) 104190.
This distillation was extracted by an LLM on 2026-06-24 and is **not human-verified
or independently reproduced**. The article is paywalled (copyright 2025 Elsevier B.V.);
no CC licence was found in the Crossref metadata. Reproduction is extract-only.

> Vandeweyer, Quentin, Minghao Yang, and Constantine Yannelis.
> "Discount factors and monetary policy: Evidence from dual-listed stocks."
> *Journal of Financial Economics* 175 (2026): 104190.
> DOI: 10.1016/j.jfineco.2025.104190. Copyright 2025 Elsevier B.V.


==============================================================================
# Policy Uncertainty Reduces Green Innovation: Wang, Wurgler & Zhang (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jfe/2026/wang-policy-uncertainty-reduces-green-2026/
# Distilled: Exogenous weather-driven variability in Chinese environmental subsidy allocations reduces firms' green R&D investment and green R&D employment, with stronger effects on green-tech and subsidy-reliant firms. Journal of Financial Economics 2026, paywalled. Six core results with source locators, datasets used, a mean-variance model of investment under subsidy uncertainty, and a two-stage IV specification using weather volatility as an instrument for policy uncertainty.
# Tags: paper-summary, climate-finance, green-innovation, policy-uncertainty
==============================================================================

**What this is.** The paper's core results, the simple mean-variance model of investment under uncertain subsidies, and the two-stage IV design that instruments policy uncertainty with weather volatility: enough to understand what was found and how, without reading all 14 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1016/j.jfineco.2025.104189).

## TL;DR

The paper documents that policy uncertainty about environmental subsidies suppresses the green R&D that those subsidies are designed to stimulate. The setting is China, where provincial and local governments allocate environmental subsidies partly on the basis of Air Quality Index (AQI) readings, and weather (wind and rain) drives AQI in predictable but noisy ways. Weather variability over a rolling six-year window therefore generates exogenous variation in how predictable a city's subsidy allocation will be, forming an instrument for policy uncertainty. A one-standard-deviation increase in weather variability reduces city-level green R&D by 0.132 standard deviations (IV estimate, Table 4 Model 5, p. 9). Effects are stronger for green-tech firms, subsidy-reliant firms, and firms under financial pressure. Placebo tests on non-green R&D and capital expenditure show no significant effect, consistent with the channel being specific to subsidized green activities. The paper extends Gulen and Ion (2016) by identifying a behavioral source of policy uncertainty: policymakers respond to salient recent pollution readings, which are themselves influenced by transient weather, creating a recency-bias channel that Bernanke (1983) and Dixit and Pindyck (1994) frameworks predict will deter investment.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Policy uncertainty (IV) reduces city-level green R&D | Table 4 Model 5, p. 8 | PU coefficient = -2.23\*\*\* (t = -2.64); one-SD weather variability reduces green R&D by 0.132 SD; OLS comparison: -0.29\*\* (Table 4 Model 1) |
| R2 | Higher average subsidies stimulate green R&D | Table 4 Model 5, p. 8 | Avg Green Subsidy = 0.34\*\*\* (t = 2.76); one-SD subsidy increase raises green R&D by 0.158 SD |
| R3 | Policy uncertainty reduces green R&D employment | Table 8 Model 1, p. 12 | PU coefficient = -1.07\*\*\* (t = -3.41); one-SD weather variability reduces employment by 0.153 SD |
| R4 | Green-tech firms are more sensitive to policy uncertainty | Table 6 Model 2, pp. 10-11 | PU × Green Tech = -0.58\*\*\* (t = -7.47), adding to average-firm effect of -0.40\*\*\* (Table 6 Model 1, t = -2.72) |
| R5 | Subsidy-reliant firms are more sensitive to policy uncertainty | Table 6 Model 5, p. 11 | PU × Subsidy Reliant = -0.62\*\*\* (t = -7.77) |
| R6 | Policy uncertainty does not affect non-green R&D or capital expenditure (placebo) | Table 5 Models 6-8, p. 10 | Non-green R&D: PU = 0.83 (t = 1.48, n.s.); capex: PU = -0.00 (t = -0.44, n.s.); real estate capex: PU = 0.03 (t = 0.32, n.s.) |

**Overall (paper's conclusion).** Policy uncertainty about environmental subsidies has real effects on green innovation: it reduces both the R&D investment and the technical employment that subsidies are intended to promote. Effects concentrate on the firms theory predicts should be most sensitive (green-tech, subsidy-reliant, financially pressured), and placebos confirm the result is not a general investment or uncertainty shock but is specific to the green component tied to the subsidized activities.

## Theory / model

The paper develops a stylized mean-variance model to derive qualitative predictions about how subsidy uncertainty affects green investment relative to traditional investment (Section 2.2, p. 4). A firm raises capital to invest in two technologies at the start of a period. The first, $$G$$, is an environmental technology receiving an uncertain government subsidy $$S \sim N(\bar{S}, \sigma_S^2)$$. The total per-dollar payoff from $$G$$ is $$\bar{R}_G + \bar{S}$$ with variance $$\sigma_G^2 + \sigma_S^2$$. The second, $$X$$, is a traditional technology with per-dollar payoff $$\bar{R}_X \sim N(\bar{R}_X, \sigma_X^2)$$. Both investments are subject to a gross financing cost $$C$$. The firm maximizes mean-variance utility (equation 1, p. 4):

$$
\max_{G,X} \; G(\bar{R}_G + \bar{S} - C) - \frac{\gamma}{2}G^2(\sigma_G^2 + \sigma_S^2) + X(\bar{R}_X - C) - \frac{\gamma}{2}X^2\sigma_X^2 \tag{1}
$$

where $$\gamma$$ is risk aversion, which stands in for irreversibility, adjustment costs, or managerial risk aversion in the spirit of Bernanke (1983), Dixit and Pindyck (1994), and Julio and Yook (2012). The first-order conditions yield optimal investments (equation 2, p. 4):

$$
G^* = \frac{\bar{R}_G + \bar{S} - C}{\gamma(\sigma_G^2 + \sigma_S^2)}, \qquad X^* = \frac{\bar{R}_X - C}{\gamma\sigma_X^2} \tag{2}
$$

The ratio of green to traditional investment is (equation 3, p. 4):

$$
\frac{G^*}{X^*} = \frac{\bar{R}_G + \bar{S} - C}{\bar{R}_X - C} \times \frac{\sigma_X^2}{\sigma_G^2 + \sigma_S^2} \tag{3}
$$

The model yields three comparative statics that the paper tests empirically:
1. Higher subsidy policy uncertainty $$\sigma_S^2$$ reduces both the level $$G^*$$ and the share $$G^*/X^*$$.
2. Higher average subsidies $$\bar{S}$$ increase the level and share of green R&D.
3. Higher financing cost $$C$$ (a proxy for financial pressure or over-indebtedness) increases the sensitivity of $$G^*/X^*$$ to $$\sigma_S^2$$, predicting that financially constrained firms are more affected by policy uncertainty.

This is not a structural model; it is a stylized device for generating falsifiable comparative statics. The paper also discusses irreversibility and adjustment-cost channels and a recency-bias channel, in which policymakers over-weight recent salient pollution conditions and under-account for transient weather, creating avoidable policy noise.

**Identification logic.** The empirical chain is: weather volatility drives AQI volatility, which drives subsidy volatility, which creates policy uncertainty that depresses green R&D. The exclusion restriction is that weather volatility affects green R&D only through this subsidy-uncertainty channel, not through other routes such as direct productivity effects of air quality on R&D effort. Placebos against non-green R&D and capex (R6) support this restriction.

## Method

**Step 1: Subsidies depend on AQI.** To show that green subsidy allocations respond to observed pollution, the paper estimates a Fama and MacBeth (1973) cross-sectional regression of city-level green subsidy on lagged AQI (equation 4, p. 6):

$$
\text{Green Subsidy}_{jt} = a_0 + a_1 \times AQI_{j,t-1} + A \times X_{j,t-1} + \varepsilon_{jt} \tag{4}
$$

where $$\text{Green Subsidy}_{jt}$$ is the log RMB of total environmental subsidies in city $$j$$ in year $$t$$, $$AQI_{j,t-1}$$ is the lagged average AQI, and $$X_{j,t-1}$$ contains city and firm-level controls. Coefficients are averaged across cross-sections with standard errors robust to cross-sectional correlation. A one-SD higher AQI is associated with a 0.17 SD higher subsidy the following year (Table 2 Model 2, coefficient = 0.05\*\*\*, t = 4.88, p. 6).

**Step 2: AQI depends on weather.** The paper documents the physical link between weather and AQI using a city-season-level specification (equation 5, p. 7):

$$
AQI_{jt,s} = b_0 + b_1 \times \text{Wind Speed}_{jt,s} + b_2 \times \text{Rain Volume}_{jt,s} + B \times X_{jt} + \delta_j + \delta_t + \varepsilon_{jts} \tag{5}
$$

where $$AQI_{jt,s}$$ is the average daily AQI in city $$j$$ in season $$s$$ of year $$t$$; city and year fixed effects are included; standard errors are clustered by city and year. Wind speed and rain volume are both negative and significant (Table 3, p. 7), confirming that weather disperses particulates and lowers AQI.

**Step 3: Instrument construction.** Policy uncertainty $$PU_{j,t-5:t}$$ is defined as the standard deviation of the characteristics-adjusted residuals from equation (4) over a six-year rolling window ending at $$t$$. The first stage regresses $$PU$$ on six-year rolling standard deviations of windy days and rainy days (equation 6, p. 8):

$$
PU_{j,t-5:t} = c_0 + c_1 \times \text{SD Windy Days}_{j,t-6:t-1} + c_2 \times \text{SD Rainy Days}_{j,t-6:t-1} + C \times X_{jt} + \delta_j + \delta_t + \varepsilon_{jt} \tag{6}
$$

The fitted weather component alone forms the instrument (equation 7, p. 9):

$$
\hat{PU}_{j,t-5:t} = \hat{c}_1 \times \text{SD Windy Days}_{j,t-6:t-1} + \hat{c}_2 \times \text{SD Rainy Days}_{j,t-6:t-1} \tag{7}
$$

First-stage estimates: $$\hat{c}_1 = 0.01^{***}$$ (t = 3.27), $$\hat{c}_2 = 0.16^{***}$$ (t = 3.71); first-stage F = 14.5 (Table 4 Model 2, p. 8). One-SD weather variability raises measured policy uncertainty by 0.228 SD. The estimation builds on `instrumental-variables` and `panel-regression` primitives.

## Empirical specifications

**City-level green R&D (R1, R2).** The second stage regresses next-year city-level green R&D on instrumented policy uncertainty and average subsidy (equation 8, p. 9):

$$
\text{Green R\&D}_{j,t+1} = d_0 + d_1 \times \hat{PU}_{j,t-5:t} + d_2 \times \text{Avg Subsidy}_{j,t-5:t} + D \times X_{jt} + \delta_j + \delta_t + \eta_{j,t+1} \tag{8}
$$

where $$\text{Green R\&D}_{j,t+1}$$ is log aggregate green R&D of city $$j$$'s listed firms, $$\hat{PU}$$ is the weather-instrumented policy uncertainty (equation 7), $$\text{Avg Subsidy}_{j,t-5:t}$$ is the six-year moving average of the characteristics-adjusted subsidy, and $$X_{jt}$$ includes AQI and city and firm-level controls. City and year fixed effects; standard errors clustered by city and year. Headline estimates: $$d_1 = -2.23^{***}$$ (t = -2.64) and $$d_2 = 0.34^{***}$$ (t = 2.76), Table 4 Model 5.

The one-standard-deviation translation: $$(0.01 \times 25.3 + 0.16 \times 1.95) \times (-2.23) / 9.54 = -0.132$$ SD in green R&D per SD in weather variability (footnote 18, p. 9).

**Firm-level heterogeneity (R4, R5).** The city-level instrument enters a firm-level regression interacted with firm-type indicators (equation 9, p. 10):

$$
\text{Green R\&D}_{ic,j,t+1} = e_0 + e_1 \times \hat{PU}_{j,t-5:t} + e_2 \times \text{Firm Type}_{ic,jt} \times \hat{PU}_{j,t-1} + e_3 \times \text{Avg Subsidy}_{j,t-5:t} + E \times X_{ic,jt} + \delta_i + \delta_t + \eta_{ic,j,t+1} \tag{9}
$$

with firm and year fixed effects; standard errors clustered by city and year. Firm type indicators include green tech (CSRC industry N77), metropolitan area, large size, state ownership, subsidy reliance (above-median ratio of environmental subsidy to total assets), cross-city registration, and industry affiliation (manufacturing, chemical, metal, public facility). The green-tech and subsidy-reliant interactions are the most significant (Table 6, pp. 10-11).

**Green R&D employment (R3).** Equation (8) replaces the dependent variable with $$\log(1 + \text{total R\&D and technical employees in city } j)$$. Result: PU = -1.07\*\*\* (t = -3.41), Table 8 Model 1.

**Placebos (R6).** Equation (8) with three alternative dependent variables: total R&D minus green R&D, city capital expenditure, and city capital expenditure of real estate firms (Table 5 Models 6-8). None shows a significant effect of $$\hat{PU}$$, supporting the exclusion restriction.

**Robustness.** Table 5 reports eight robustness checks including: excluding city-years with zero green R&D, two-year average AQI, weather-adjusted AQI, subsidy starting from 2007, and a subsample from 2013 onward. The main result is stable across all variations (PU coefficients ranging from -2.22 to -5.14, all significant at 5% or better except in Model 6 which is a placebo). Baker, Bloom and Davis (2016) style policy uncertainty indices serve as a conceptual foil; the paper distinguishes its weather-based channel from those political and fiscal sources.

The estimation sample is 2009-2019, pre-Covid. City-year panel: 1,340 observations (up to 352 cities). Firm-year panel: 26,189 observations (3,168 firms).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CSMAR (China Stock Market & Accounting Research Database) | Firm-level green R&D (from financial statement footnotes), environmental subsidy receipts, stock return | No page yet |
| CNRDS (China Research Data Services) | City-level GDP, GDP growth, consumption, population growth; firm characteristics digitized from China Statistical Yearbook | No page yet |
| WIND Financial Terminal | Firm-level turnover, ROA, leverage, size, cash holdings, capex, profit margin | No page yet |
| Ministry of Ecology and Environment (MEPC) | Annual average AQI for major cities from the national monitoring network | No page yet |
| China Meteorological Administration (CMA) | Daily city-level wind speed (m/s) and rain volume (mm) used to construct Windy Days, Rainy Days, and their rolling standard deviations | No page yet |
| EPSnet | City-level industrial Waste Gas Treatment expenditure; used to confirm the AQI-subsidy relationship independently (Table 2 Models 5-6) | No page yet |

Sample: weather and subsidy data back to 2003 (needed for six-year rolling windows); main regressions: 2009-2019 (pre-Covid). 352 Chinese cities; 3,168 listed firms; 1,340 city-year observations in primary city-level regressions; 26,189 firm-year observations in firm-level regressions.

## When to read the full paper

Read the [original](https://doi.org/10.1016/j.jfineco.2025.104189) if you are:
- studying how uncertainty about government policies, not just their level, can undermine their intended effects on investment;
- building a weather-based IV design for policy uncertainty in settings where subsidies are tied to observable environmental indicators;
- interested in the institutional mechanics of Chinese environmental policy and AQI-linked subsidy allocation;
- examining heterogeneous investment responses to policy uncertainty across firm types (green-tech, subsidy reliance, financial constraints);
- researching the real-side effects of climate policy uncertainty for the growing green finance literature.

The online appendix (available at the DOI) contains additional institutional background, variable definitions, and further robustness checks.

## Attribution and rights

Source: peer-reviewed, *Journal of Financial Economics* 175 (2026) 104189. This distillation was extracted by an LLM on 2026-06-24 and is **not human-verified or independently reproduced**. The paper is paywalled; only extract-only redistribution is permitted.

> Wang, Mengyu, Jeffrey Wurgler, and Hong Zhang. "Policy uncertainty reduces green innovation." *Journal of Financial Economics* 175 (2026) 104189. DOI: [10.1016/j.jfineco.2025.104189](https://doi.org/10.1016/j.jfineco.2025.104189). © 2025 Published by Elsevier B.V.


==============================================================================
# Intraday Proprietary Traders and Short-Term Mispricing: Anshuman et al. (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jfm/2026/anshuman-intraday-proprietary-traders-mispricing-2026/
# Distilled: Using trader-level BSE transaction data and hand-collected Indian TV analyst recommendations, the paper shows only intraday proprietary traders trade contrarian against short-term recommendation-induced mispricing, earning informed-trading profits while bearing liquidity costs; overnight proprietary traders provide liquidity but do not exploit the mispricing. Journal of Financial Markets 2026, paywalled. Six core results with source locators, datasets used, and the empirical specifications.
# Tags: paper-summary, market-microstructure, equities, price-discovery, emerging-markets, panel-regression
==============================================================================

**What this is.** The paper's core results, the empirical design, and the main regression specifications: enough to know what it found and how, without reading all 28 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1016/j.finmar.2025.101028).

## TL;DR

Using a hand-collected dataset of Indian TV analyst recommendations (CNBC Awaaz Stock 20/20, July 2009 to March 2016) matched with BSE trader-level intraday order and trade data, the paper documents that TV analyst recommendations create temporary mispricing that fully reverts within about two weeks, extending the US evidence of Engelberg et al. (2012) to an emerging-market setting with trader-level data. Among the investor categories studied (individuals, institutions, intraday proprietary traders, overnight proprietary traders), only intraday proprietary traders trade contrarian in the first half hour, selling buy-recommended stocks and buying sell-recommended stocks. Their trades earn positive returns from informed trading and negative returns from liquidity provision (they pay a small liquidity cost), consistent with the interpretation that they are informed arbitrageurs rather than liquidity providers. Overnight proprietary traders do the opposite: they provide liquidity and earn positive returns from it but earn no significant returns from informed trading. Intraday proprietary traders account for 12 to 22 percent of the contemporaneous price correction in the first half hour, and they continue their contrarian activity over the following five trading days.

## Core results

Magnitudes and significance are as reported; `\*\*` / `\*\*\*` = 5% / 1%; ns = not significant. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | TV recommendations create temporary mispricing; prices fully revert to pre-recommendation levels within ~10 days | Table 2, Fig. 1, p. 7 | Buy recs: DGTW-adj Day 0 return vs matched control = +0.80 bps\*\*\* (t=27.77); sell recs: -0.41 bps\*\*\* (t=-10.06); complete reversal by day ~10 |
| R2 | Only intraday (not overnight) prop traders trade contrarian in the first half hour; institutions are not contrarian | Fig. 5, pp. 12-13; Fig. 6, p. 14 | Intraday NTB significantly negative (selling) for buy recs, positive (buying) for sell recs in first half hour; overnight NTB near zero; result robust to matched-set, trader, and stock-by-date fixed effects |
| R3 | Intraday prop traders earn significant informed-trading profits even after bearing a negative liquidity cost | Table 6, p. 17 | Buy recs: informed trading returns = +23.1 bps\*\*\* (t=3.64), liquidity provision = -7.22 bps\*\*\* (t=-12.1); sell recs: informed = +24.5 bps\*\* (t=2.55), liquidity = -11.9 bps\*\*\* (t=-10.1) |
| R4 | Overnight prop traders earn positive returns from liquidity provision and no significant informed-trading returns | Table 6, p. 17 | Buy recs: liquidity = +3.81 bps\*\*\* (t=3.82), informed = -8.4 bps (t=-0.88, ns); sell recs: liquidity = +9.38 bps\*\*\* (t=3.65), informed = -19.0 bps (t=-1.04, ns) |
| R5 | Intraday prop traders account for 12-22% of first-half-hour price correction | Table 8, p. 22 | Buy recs: contribution = -1.94 bps\*\*\* (t=-7.45), 12-13% of first-half-hour return; sell recs: +3.37 bps\*\*\* (t=7.69), 17-22% of first-half-hour return |
| R6 | Intraday prop traders continue contrarian activity on Days 1-5 after recommendation; overnight traders remain inactive | Fig. 9, pp. 24-25 | Intraday NTB remains significantly contrarian in first half hour on Days 1-5; first-half-hour returns continue reversing in tandem; overnight NTB insignificant throughout |

**Overall (paper's conclusion).** Intraday proprietary traders are the only category that actively exploits short-term mispricing caused by TV analyst recommendations. They act as informed arbitrageurs, earning positive informed-trading returns despite paying a small liquidity cost. Overnight proprietary traders provide liquidity and earn compensation for it but do not exploit the mispricing, consistent with the evidence in Biais et al. (2016) that some proprietary traders specialize in liquidity supply. Policies designed to curtail short-term trading must account for the beneficial role of intraday proprietary traders in price efficiency.

## Theory / model

The paper tests no formal model of its own. It draws on theoretical models of short-term speculation positing that short-term investors specialize in information about the behavior of other market participants rather than fundamental value (Tirole (1982), De Long et al. (1990), Froot et al. (1992)). The central empirical question is which investor category corrects temporary mispricing and how they profit from doing so.

**Identification.** TV analyst recommendations on CNBC Awaaz Stock 20/20 are treated as an exogenous shock to stock prices. The program targets individual retail investors whose trading creates price pressure in the first half hour. Selection of stocks into the program is addressed using propensity score matching via a conditional logit model (Table A.1, Internet Appendix), matching each recommended stock to a control stock with similar lagged return, volume, market capitalization, book-to-market ratio, analyst coverage, index membership, and individual investor ownership fraction (p. 6). Results are robust to three increasingly stringent fixed-effect structures that absorb unobservable time-invariant and time-varying confounders at the matched-set, trader, and stock levels.

**Hypotheses tested.**

1. TV recommendations create temporary mispricing: prices deviate on Day 0 and revert over the following days.
2. Among sophisticated investors, only intraday proprietary traders trade contrarian in the first half hour against the recommendation direction.
3. Intraday prop traders earn returns from informed trading (private information about other participants' behavior) rather than from liquidity provision.
4. Overnight prop traders earn returns from liquidity provision rather than from informed trading.

## Method

The paper applies panel regression, propensity-score `matching` via `conditional-logit`, and an `event-study` design. It builds on `panel-regression` for the NTB and return-decomposition specifications.

**Trader classification (eq. 2, p. 8).** Proprietary traders (BSE client code "OWN") are split into intraday vs overnight groups using a trader-specific inventory measure: for trader $$k$$, stock $$i$$, day $$d$$,

$$
\text{Inventory}_{k,i,d} = \frac{\bigl|\text{No.\,shares bought}_{k,i,d} - \text{No.\,shares sold}_{k,i,d}\bigr|}{\text{Total no.\,shares traded}_{k,i,d}} \tag{2}
$$

The trader-specific inventory is the median of $$\text{Inventory}_{k,i,d}$$ across all stock-days on which trader $$k$$ was active. Traders below (at or above) the cross-trader median are intraday (overnight) proprietary traders.

**Net total buying (NTB) regression (eq. 1, p. 10).** For each investor category, NTB in half hour $$h$$ is the rupee value of buyer-initiated minus seller-initiated trades as a percentage of total volume. The baseline specification:

$$
\text{NTB}_{i,d,h} = \sum_{h=1}^{13} \beta_h I_h + \sum_{h=1}^{13} \gamma_h I_h \cdot \text{Treated}_{i,d} + \varepsilon_{i,d,h} \tag{1}
$$

where $$I_h$$ is a half-hour indicator and $$\text{Treated}_{i,d} = 1$$ if stock $$i$$ received a recommendation on day $$d$$. The $$\gamma_h$$ coefficients capture the difference in NTB between recommended and control stocks in half hour $$h$$. Three fixed-effect extensions absorb unobservables:

$$
\text{NTB}_{k,i,j,d,h} = \sum_{h=1}^{13} \beta_h I_h + \sum_{h=1}^{13} \gamma_h I_h \cdot \text{Treated}_{i,d} + \delta_{j,d,h} + \varepsilon_{k,i,j,d,h} \tag{3, D-MS-HH}
$$

$$
\text{NTB}_{k,i,j,d,h} = \sum_{h=1}^{13} \beta_h I_h + \sum_{h=1}^{13} \gamma_h I_h \cdot \text{Treated}_{i,d} + \delta_{k,d,h} + \varepsilon_{k,i,j,d,h} \tag{4, D-Tr-HH}
$$

where $$\delta_{j,d,h}$$ = matched-set-by-date-by-half-hour FE and $$\delta_{k,d,h}$$ = trader-by-date-by-half-hour FE.

**Return decomposition (eq. 8, p. 16).** Following Kaniel et al. (2012), the return from the end of the first half hour to the day's close ($$\text{Ret\_2Cl}$$) is regressed on the net total buying of both proprietary trader groups, using only control stocks:

$$
\text{Ret\_2Cl}_{c,d} = \alpha + \beta_1 \text{SNTB}_{c,d,1,\text{ID}} + \beta_2 \text{SNTB}_{c,d,1,\text{ON}} + \theta \text{Ret}_{c,d,1} + \varepsilon_{c,d} \tag{8, Model B}
$$

where $$\text{SNTB}$$ is net total buying scaled by mean volume, subscripts ID and ON denote intraday and overnight prop traders, and $$\text{Ret}_{c,d,1}$$ controls for first-half-hour return (reversal). Fitted values give the estimated return to normal liquidity provision; the residual (actual return minus fitted) is the informed-trading return (eqs. 9-10, p. 17).

**Price impact (eq. 11, p. 18).** Per-unit price impact for each trader category is estimated by regressing half-hour returns on net aggressive buying (RNAB, market orders):

$$
\text{Ret}_{i,d,h} = \alpha + \sum_{\text{cat}} \beta_{\text{cat},h} \cdot \text{RNAB}_{i,d,\text{cat},h} + \theta_h \text{Ret\_ClOp}_{i,d} + \varepsilon_{i,d,h} \tag{11, Model O}
$$

estimated separately for recommended and control stocks. Trader category contributions to price correction are the product of estimated $$\hat{\beta}_{\text{cat},h}$$ and actual $$\text{RNAB}_{i,d,\text{cat},h}$$ (eq. 12, p. 20), then regressed on a Treated indicator with matched-set-by-date FE (eq. 13) to get the treatment-minus-control difference.

## Empirical specifications

All headline results use a propensity-score-matched sample from July 2009 to March 2016 (excluding the program hiatus July 2010 to September 2011). Standard errors are clustered at the date level throughout.

**R1 (temporary mispricing, Table 2).** Calendar-time portfolio approach: long in recommended stocks, short in matched control stocks, held for $$n$$ days. The DGTW characteristic-adjusted return (Daniel et al. (1997)) is the outcome measure. Day $$-1$$ close to Day 0 open captures the announcement effect; Day 0 open to close captures the intraday reversal; the combination gives the full Day 0 return. Long-run return to 252 trading days is reported as further evidence of no permanent price effect.

**R2 (contrarian trading, Figs. 4-6).** Equation (1) without FE followed by equations (3) and (4) and a stock-by-date FE variant (D-Stk). Results are multiplied by the average number of traders per group to obtain total category-level NTB.

**R3-R4 (return decomposition, Table 6).** Stocks are sorted daily into terciles of abnormal net total buying (ANTB) in the first half hour. Tercile 3 (most intensive buying/selling) minus Tercile 1 difference is reported separately for each return component. Returns are winsorized at 1%. Decomposition from Models B, R, and RI (eq. 8) is estimated on control stocks and applied to treatment minus control differences.

**R5 (price correction, Table 8).** Price impact estimated from equation (11) separately for buy and sell recommendations. Contribution computed via equation (12) and then equation (13) with matched-set-by-date FE. Reported as basis points and as a fraction of the total first-half-hour treatment-minus-control return.

**R6 (multi-day contrarian activity, Fig. 9).** Equation (1) applied to the first half hour on Days 1-5 post-recommendation. First-half-hour returns (difference treatment vs control) are plotted alongside intraday NTB to document joint behavior.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CNBC Awaaz Stock 20/20 TV recommendations (hand-collected) | Exogenous mispricing shock; buy/sell treatment assignment for 1,230 unique stocks over 968 days | No page yet |
| BSE intraday order and trade data | Trader-level NTB, inventory classification, price impact estimation; all BSE orders and trades July 2009 to March 2016 | No page yet |
| CMIE Prowess database | Daily stock prices, financial statement data (market cap, book-to-market, volume) for propensity score matching | No page yet |

Sample: 26,827 recommendations (20,345 buy, 6,482 sell) on 1,230 unique BSE-listed stocks, July 2009 to March 2016. After propensity score matching: 26,341 matched treatment stock-days (19,927 buy, 6,414 sell) paired with 24,657 control stock-days (18,776 buy, 5,881 sell). Intraday data at the half-hour level; daily stock characteristics from Prowess. The authors explicitly state the data cannot be shared (Data availability, p. 27).

## When to read the full paper

Read the [original](https://doi.org/10.1016/j.finmar.2025.101028) if you are: studying who corrects short-term mispricing in equity markets; working on intraday trading behavior, proprietary traders, or market microstructure in emerging markets; applying the Kaniel et al. (2012) return decomposition to a new setting; examining the policy trade-off between curbing short-term trading and maintaining price efficiency; or studying the role of media (TV/social-media) recommendations in creating temporary price deviations.

## Attribution and rights

Source: peer-reviewed, *Journal of Financial Markets* 78 (2026) 101028. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. The paper is paywalled (copyright 2025 Elsevier B.V., all rights reserved); no Creative Commons license was found. Extract-only.

> Anshuman, V. Ravi, Prachi Deuskar, Krishnamurthy V. Subramanian, and Ramabhadran S. Thirumalai.
> "Intraday Proprietary Traders and Short-Term Mispricing."
> *Journal of Financial Markets* 78 (2026) 101028.
> DOI: 10.1016/j.finmar.2025.101028. copyright 2025 Elsevier B.V.


==============================================================================
# Tick Size and Market Quality: Barardehi, Dixon, Liu & Lohr (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jfm/2026/barardehi-tick-size-market-quality-2026/
# Distilled: The U.S. Tick Size Pilot (TSP) harmed liquidity for stocks with quoted spreads below 10 cents but improved it for stocks with spreads above 15 cents, explaining mixed results across prior studies that pool stocks with very different prevailing spreads. Journal of Financial Markets 2026, CC BY 4.0. Seven core results with source locators, datasets used, the DiD specification, and empirical specifications.
# Tags: paper-summary, equities, market-microstructure, tick-size, liquidity
==============================================================================

**What this is.** The paper's core results, the economic framework, and the DiD specification used: enough to know what was found and how, without reading all 17 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1016/j.finmar.2025.101024).

## TL;DR

Using the U.S. Tick Size Pilot (TSP), which raised the minimum tick from 1 cent to 5 cents for roughly 1,200 small-cap pilot stocks in 2016, the paper shows that tick size effects on market quality are heterogeneous and depend on a stock's prevailing quoted spread. For stocks with average quoted spreads below approximately 10 cents, the larger tick harmed liquidity: quoted and effective spreads widened, consistent with pricing fidelity deteriorating. For stocks with spreads above 15 cents, the larger tick improved liquidity: spreads narrowed, consistent with undercutting being reduced. Stocks in between experienced indeterminate or conflicting effects. These opposing effects explain why prior TSP studies that pool narrow- and wide-spread stocks together in a single non-tick-constrained group find muddled or conflicting results. The paper recommends that researchers using the TSP for causal inference bifurcate samples at the 10 cent prevailing spread threshold.

## Core results

Magnitudes are as reported; `\*\*`/`\*\*\*` = 5%/1%. QR = quantile (median) regression; OLS results are similar unless noted. All columns correspond to spread bins by prevailing May-June 2016 quoted spread: (1) Bin 1 = tick-constrained (<5 cents), (2) Bin 2 = near-tick-constrained (5-10 cents), (3) Bin 3 = intermediate-spread (10-15 cents), (4) Bin 4 = wide-spread (>15 cents). Pilot x Event is the DiD coefficient from Eq. (1).

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | TSP widened quoted spreads for tick-constrained and near-tick-constrained stocks; narrowed them for wide-spread stocks; effect indeterminate for intermediate stocks | Table 3, Panel A, p. 9 | QR Pilot x Event: Bin 1 +0.032\*\*\* (t=19.34); Bin 2 +0.033\*\*\* (t=12.81); Bin 3 +0.0079\* (t=1.69); Bin 4 -0.048\*\*\* (t=-3.63) |
| R2 | TSP widened effective spreads for stocks with spreads below 10 cents and narrowed them for wide-spread stocks | Table 3, Panel A, p. 9 | QR Pilot x Event: Bin 1 +0.033\*\*\* (t=26.19); Bin 2 +0.025\*\*\* (t=16.46); Bin 3 +0.0075\*\* (t=2.84); Bin 4 -0.019\*\*\* (t=-3.13) |
| R3 | TSP reduced undercutting (QID ratio) across all four spread bins; QID rebounded upon TSP conclusion, consistent with the undercutting channel | Table 2, Panels A-B, p. 8 | Imposition: Bin 1 -0.27\*\*\* (t=-17.90); Bin 2 -0.23\*\*\* (t=-15.01); Bin 3 -0.16\*\*\* (t=-10.83); Bin 4 -0.11\*\*\* (t=-10.81); all four bins significantly positive at TSP conclusion |
| R4 | TSP increased NBBO quoted depth for all spread bins; largest log-increases for tick-constrained and near-tick-constrained stocks | Table 3, Panel A, pp. 9-10; Fig. 2, p. 10 | Ln(NBBO Depth) QR: Bin 1 +1.06\*\*\* (t=20.89); Bin 2 +0.91\*\*\* (t=20.34); Bin 3 +0.49\*\*\* (t=9.88); Bin 4 +0.21\*\*\* (t=5.99) |
| R5 | TSP increased round-trip transaction costs (CRT) for tick-constrained stocks and reduced them for wide-spread stocks; at very large trade sizes the TSP reduced CRT near-uniformly (Bin 1 remains positive at 2500 shares; paper: "almost uniformly" with that exception) | Table 4, Panel A, p. 12; Fig. 5, p. 15 | 10-round-lot CRT QR: Bin 1 +0.023\*\*\* (t=13.98); Bin 4 -0.075\*\*\* (t=-4.10); 2500-share (y=25) Bin 1 +0.008\*\* (Table 4); Bins 2-4 negative at 2500 shares |
| R6 | Rolling-bin threshold analysis shows ~10 cents as the key breakpoint where the TSP switches from harming to benefiting effective spreads; the threshold for small-trade CRT is also approximately 10 cents | Fig. 3, p. 13; §4.4, pp. 11-13 | TSP effect on effective spread positive and significant for pre-shock spreads below ~10 cents; switches to negative above ~15 cents; 10-15 cent zone indeterminate in both imposition and conclusion windows |
| R7 | Applying the Werner et al. (2023) median-bifurcation methodology to TSP imposition data yields a null result for the non-tick-constrained group; this reflects the cancellation of near-tick-constrained (+) and wide-spread (-) effects when pooled | Table 5, p. 16 | TSC (tick-constrained) QR +0.046\*\*\* (t=23.57); TSU (non-tick-constrained) -0.0046 (t=-1.17, insignificant) |

**Overall (paper's conclusion).** The TSP provides direct evidence of the tick size tradeoff between pricing fidelity and undercutting. Narrow-spread stocks were harmed because the larger tick distorted the bid-ask spread relative to the latent competitive spread, worsening pricing fidelity. Wide-spread stocks benefited because the larger tick made undercutting less attractive, encouraging displayed liquidity provision. The opposing effects within the broad non-tick-constrained category explain why Hu et al. (2018), Griffith and Roseman (2019), Rindi and Werner (2019), Chung et al. (2020), and Werner et al. (2023) reach different conclusions: they employ different sample constructions that weight the two opposing groups differently.

## Theory / model

The paper has no formal theoretical model of its own. It tests the "tick size tradeoff" framework of Werner et al. (2023), which characterizes tick size effects as a tradeoff between two competing channels.

**Pricing fidelity channel.** A smaller tick enables a finer price grid, allowing the equilibrium bid-ask spread to narrow toward the latent competitive spread. When a tick is too large relative to the competitive spread, the tick forces the quoted spread above its competitive level, inflating transaction costs and reducing allocative efficiency. This channel dominates for stocks whose competitive spreads are narrow (tick-constrained or near-tick-constrained): a tick increase makes pricing fidelity worse.

**Undercutting channel.** A smaller tick lowers the cost of front-running a resting limit order with a trivially better price: posting a quote that is just one tick better is cheaper when the tick is smaller. Greater undercutting risk discourages displayed liquidity provision (adverse selection for slower traders and higher monitoring costs), reduces fill rates on limit orders, and lowers overall depth. A larger tick reduces this activity and can thereby improve market quality. This channel dominates for stocks with wide spreads (many ticks intra-spread): a tick increase reduces undercutting and improves market quality.

**Testable prediction.** The two channels predict opposite effects for stocks with different prevailing spreads. The paper identifies two empirical thresholds: at ~10 cents, the net effect on effective spreads and small-trade CRT switches from negative (pricing fidelity channel dominates) to ambiguous; at ~15 cents, the net effect turns positive (undercutting channel dominates).

**Identification.** The TSP is a natural experiment: the SEC randomly assigned approximately 2,400 stocks with market capitalization below $3 billion to pilot (tick = 5 cents) and control (tick = 1 cent) groups. This design satisfies the parallel trends assumption: within each spread bin, treated and control stocks are comparable in market capitalization, dollar volume, and return volatility before the TSP (Table 1, p. 7).

## Method

The paper uses difference-in-differences (DiD) quantile regressions (Callaway and Li, 2019) to estimate the causal effect of the tick size change. Quantile (median) regression is used as the primary estimator because microstructure outcomes such as quoted spreads are positively skewed: OLS means are sensitive to large outliers and can diverge from medians in ways that depend on how researchers handle extreme observations. The paper also reports OLS estimates after winsorizing at the 5th and 95th percentiles by tick-constraint category; results are qualitatively similar.

The DiD specification (equation 1, p. 6) is:

$$
Y_{j,t} = \alpha_0 + \alpha_p \text{Pilot}_j + \alpha_e \text{Event}_{j,t} + \beta \left( \text{Pilot}_j \times \text{Event}_{j,t} \right) + u_t + \varepsilon_{j,t} \tag{1}
$$

where $$\text{Pilot}_j$$ equals 1 for treated stocks (G1 or G2) and 0 for control stocks; $$\text{Event}_{j,t}$$ equals 0 before the tick size change and 1 after it (with enforcement-date-based assignment for the staggered TSP rollout across October 2016); $$u_t$$ is a date fixed effect; and $$\varepsilon_{j,t}$$ is the error. Standard errors are double-clustered at the stock and date level.

The coefficient of interest is $$\beta$$, the DiD estimate of the treatment effect. Because G1 and G2 differed only in whether the 5-cent increment applied to trading in addition to quoting (the authors find no reliable differences between G1 and G2 and so pool them), while G3 included a trade-at rule, the primary analysis uses G1 and G2 pilot stocks versus control stocks.

The paper estimates equation (1) separately for each of four spread bins based on prevailing pre-shock quoted spreads: bin 1 (tick-constrained, below 5 cents), bin 2 (near-tick-constrained, 5-10 cents), bin 3 (intermediate, 10-15 cents), and bin 4 (wide-spread, above 15 cents). For the threshold analysis in Section 4.4, the paper uses overlapping 6-cent bins shifted by 1 cent ({(0, 6c), (1c, 7c), ...}) and plots the DiD coefficient versus the bin's median pre-shock quoted spread. The resulting figure traces the spread-level transition from harmful to beneficial TSP effects.

## Empirical specifications

**Spread and depth outcomes (R1, R2, R4).** The dependent variable $$Y_{j,t}$$ is, in turn: time-weighted average quoted dollar spread; size-weighted average effective dollar spread; and natural log of time-weighted average quoted depth at the NBBO (using WRDS Intraday Indicators). Equation (1) is estimated via quantile regression (median) with date fixed effects and double-clustered SEs. Sample for TSP imposition: 08/11/2016-12/15/2016 (pre-period before 10/03/2016, treatment period from 10/24/2016 onwards). Sample for TSP conclusion: 08/07/2018-11/20/2018 (pre-period before 10/01/2018). Results in Table 3 (p. 9).

**Undercutting and cancel-to-trade (R3).** The primary undercutting measure QID (quote-improvement-to-deterioration ratio) is from Barardehi et al. (2025): the frequency of NBBO quote improvements divided by the sum of improvement and trade-driven NBBO deterioration frequencies, computed daily from TAQ data. A higher QID indicates more undercutting. The cancel-to-trade ratio (from MIDAS data) is total canceled orders divided by total executed trades. Equation (1) estimated the same way. Results in Table 2 (p. 8).

**Round-trip transaction costs (R5).** The dependent variable $$\text{CRT}_{y,j,t}$$ is the average per-share cost of a round-trip trade of $$y$$ round lots (where $$y \in \{1, 2.5, 5, 10, 25, 50, 100\}$$), computed from MIDAS order book snapshots taken every 15 minutes. The CRT is the absolute difference between the average per-share purchase price and the midpoint, averaged across buy and sell sides and across the trading day. Equation (1) estimated separately for each $$y$$. Results in Table 4 (p. 12) and Figs. 4-5.

**Threshold analysis (R6).** Rolling overlapping bins of width 6 cents shifted by 1 cent across the pre-shock quoted spread distribution. For each bin, equation (1) yields a DiD coefficient on effective spread and on 500-share CRT. A bin is classified as having a determinable positive (negative) effect when both the imposition and conclusion estimates have the same sign and at least one is significant; otherwise the effect is labelled indeterminate. Results in Fig. 3 (p. 13).

**Bifurcation reconciliation (R7).** The paper replicates the Werner et al. (2023) sample-bifurcation approach: stocks split by median quoted spread, with stocks below the median classified as tick-constrained (TSC) and above as non-tick-constrained (TSU). Equation (1) is estimated for the TSP imposition on quoted and effective spreads. Results in Table 5 (p. 16).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| TAQ (Trade and Quote) daily data via WRDS | Daily time-weighted quoted spreads; size-weighted effective spreads; daily QID undercutting measure; computed from tick-level trades and quotes during regular trading hours | [TAQ](/wiki/commercial/taq/) |
| WRDS Intraday Indicators | Daily time-weighted NBBO depth measures for ~2,400 TSP pilot and control stocks | [WRDS](/wiki/commercial/wrds/) |
| MIDAS (SEC Market Information Data Analytics System) | 15-minute order book snapshots used to compute cumulative depth at price levels, round-trip transaction costs (CRT), and cancel-to-trade ratios; covers all exchanges and price levels | no page yet |

Sample: TSP imposition window 08/11/2016-12/15/2016 (with enforcement dates 10/03/2016-10/23/2016 as staggered treatment onset); TSP conclusion window 08/07/2018-11/20/2018 (simultaneous return to 1-cent tick 10/01/2018). Approximately 2,400 securities: ~1,200 pilot stocks (G1+G2 combined) and ~1,200 control stocks.

## When to read the full paper

Use the [original](https://doi.org/10.1016/j.finmar.2025.101024) if you are: (i) designing or evaluating tick size reform policies and need the specific threshold estimates (10 cents, 15 cents) and the spread-bin CRT gradients; (ii) using the TSP as a causal instrument for microstructure or price efficiency questions and need guidance on sample construction; (iii) reconciling conflicting TSP findings in the literature (Section 5 provides a systematic comparison to Hu et al. (2018), Griffith and Roseman (2019), Rindi and Werner (2019), Chung et al. (2020), and Werner et al. (2023)); or (iv) studying undercutting dynamics and the cancel-to-trade channel using MIDAS data.

## Attribution and rights

Source: peer-reviewed, *Journal of Financial Markets* 78 (2026) 101024. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Barardehi, Yashar H., Peter Dixon, Qiyu Liu, and Ariel Lohr.
> "When does the tick size help or harm market quality? Evidence from the Tick Size Pilot."
> *Journal of Financial Markets* 78 (2026) 101024.
> DOI: 10.1016/j.finmar.2025.101024. © 2025 The Authors.
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Stock Market Indexing and Option Market Conditions: Chang, Ge, Lin & Ma (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jfm/2026/chang-indexing-option-market-conditions-2026/
# Distilled: Stocks at the top of the Russell 2000 Index have smaller put-call parity deviations, higher options trading volume, and narrower bid-ask spreads than similar-sized stocks at the bottom of the Russell 1000 Index, documented via the annual Russell 1000/2000 reconstitution as a regression discontinuity design (local linear regressions, 1998-2006). Journal of Financial Markets 2026, CC BY-NC-ND 4.0. Six core results with source locators, datasets used, the measure construction, and the identification approach.
# Tags: paper-summary, options, asset-pricing, market-microstructure, stock-indexing
==============================================================================

**What this is.** The paper's core results, the measure it constructs (put-call parity deviation as a proxy for option market conditions), and the regression discontinuity design it applies (local linear regressions around the Russell 1000/2000 threshold): enough to know what it found and how, without reading the full paper. To replicate or extend it, read the full source at the [original](https://doi.org/10.1016/j.finmar.2025.101026).

## TL;DR

Using the annual Russell 1000/2000 Index reconstitution as a regression discontinuity design (following Chang et al. (2015)), the paper finds that stocks at the top of the Russell 2000 Index exhibit better option market conditions than similar-sized stocks at the bottom of the Russell 1000 Index, over 1998-2006. Specifically, indexed stocks have smaller put-call parity deviations, higher options trading volume, and narrower options bid-ask spreads. Boone and White (2015) document that these same threshold stocks have higher stock liquidity and lower information asymmetry. The paper argues the channel is a supply-side liquidity spillover: improved stock liquidity reduces the hedging costs of options market makers, who hedge by trading the underlying stock, making them more willing to provide liquidity in options. The lending-fee channel is ruled out because call options bid-ask spreads are more affected than put options bid-ask spreads, opposite to the prediction under a short-selling-cost mechanism. The results are consistent with Kamara and Miller (1995), who show that higher options liquidity reduces put-call parity violations.

## Core results

Magnitudes as reported; `\*\*` = 5%, `\*\*\*` = 1%. All from local linear regressions around the Russell 1000/2000 threshold, bandwidth ±50 stocks, year and industry fixed effects, 1998-2006. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Stocks at top of Russell 2000 have **smaller put-call parity deviations** (Absoptivspread) than similar-sized stocks at bottom of Russell 1000 | Table 2, col. 3, p. 7 | Dum2000 = -0.004\*\* (t = -2.42); N = 199 |
| R2 | Indexed stocks have **higher options trading volume** (log total contracts) | Table 3, col. 3, p. 8 | Dum2000 = 0.561\*\*\* (t = 5.28); N = 368 |
| R3 | Indexed stocks have **narrower options bid-ask spreads** (open-interest-weighted, %) | Table 4, col. 3, p. 9 | Dum2000 = -2.781\*\*\* (t = -4.59); N = 368 |
| R4 | **Call bid-ask spreads are more affected than put bid-ask spreads**, inconsistent with the lending-fee channel prediction | Tables 5 and 6, pp. 10-11 | Volume: put Dum2000 = 0.566\*\*\* (t = 4.53), call Dum2000 = 0.524\*\*\* (t = 3.53); Spreads: put Dum2000 = -1.731\*\*\* (t = -2.80), call Dum2000 = -3.565\*\*\* (t = -6.25) |
| R5 | Alternative liquidity measure: indexed stocks have **fewer zero trading volume days** (OptNZVD) | Table 8, col. 3, p. 13 | Dum2000 = -32.774\*\*\* (t = -5.68); N = 321 |
| R6 | Alternative liquidity measure: indexed stocks have **lower options ILLIQ** (Optilliq) | Table 9, col. 3, p. 14 | Dum2000 = -0.050\*\*\* (t = -5.32); N = 303 |

**Overall (paper's conclusion).** The stock market indexing effect, identified via the Russell 1000/2000 reconstitution, improves option market conditions through liquidity spillovers from equity markets to options markets. The supply-side channel, where options market makers face lower hedging costs when underlying stock liquidity improves, dominates demand-side and lending-fee alternatives. Results are robust to alternative bandwidths (±25 and ±75 stocks, Table 7) and alternative liquidity measures (OptNZVD, Optilliq, Tables 8-9).

## Theory / model

The paper has no formal economic model. It tests two empirical hypotheses derived from the stock indexing mechanism.

**H1 (option market conditions).** If stock market indexing improves stock liquidity and reduces information asymmetry for stocks just above the Russell 1000/2000 threshold (documented by Boone and White (2015)), then put-call parity deviations for options on those stocks should be smaller. The put-call parity deviation reflects demand pressure for call options relative to put options (Cremers and Weinbaum, 2010), and better-informed or more active market making reduces such imbalances (Rösch et al. (2017)).

**H2 (supply-side liquidity spillover).** Options market makers hedge their positions by trading the underlying stock. When stock liquidity improves, their hedging costs fall, making them more willing to provide options liquidity (wider coverage, narrower spreads). This supply-side shift predicts higher options volume and narrower bid-ask spreads simultaneously.

**Mechanism distinction.** A demand-side story, where informed arbitrageurs are more attracted to better-liquid stocks and increase options activity, would predict higher volume but wider bid-ask spreads (adverse selection). A lending-fee channel, where lower short-selling costs reduce the replication cost of put options, would predict stronger effects on put options relative to call options. The paper tests these alternatives using separate call and put regressions (Tables 5 and 6): call bid-ask spreads are more affected than put spreads, ruling out the lending-fee channel and pointing toward the supply-side market-making mechanism.

**Identification.** Each year, Russell constructs the Russell 1000 and Russell 2000 indexes based on market capitalization at end of May; portfolio weights are released in June. Stocks just above the cutoff enter the Russell 2000 with high portfolio weights (the 2000 index has a smaller aggregate market cap denominator), while stocks just below enter the Russell 1000 with low portfolio weights. Market capitalizations around the threshold are continuous, but portfolio weights jump discontinuously, providing near-random assignment in a narrow bandwidth. Chang et al. (2015) establish that this generates significant stock price effects. The paper follows the same setting, focusing on stocks within ±50 ranks of the threshold (1998-2006, ending before Russell's 2007 banding policy change).

## Method

The paper follows Lin et al. (2019) and adopts local linear nonparametric regression as the estimator. This avoids the boundary bias that kernel regression (Nadaraya-Watson) produces near the support boundary of the running variable (p. 5). The option market conditions measure follows Rösch et al. (2017): the absolute deviation from put-call parity for stock $$i$$ on day $$t$$ is the open-interest-weighted average absolute difference in implied volatilities across matched call-put pairs (equation (1), p. 4):

$$
\text{Absoptivspread}_{i,t} = \left| IV_{i,t}^{\text{calls}} - IV_{i,t}^{\text{puts}} \right| = \left| \sum_{j=1}^{N_{i,t}} w_{j,t}^{i} \left( IV_{j,t}^{i,\text{call}} - IV_{j,t}^{i,\text{put}} \right) \right| \tag{1}
$$

where $$N_{i,t}$$ is the total number of valid call-put pairs (same strike, same maturity) for stock $$i$$ on day $$t$$; $$w_{j,t}^{i}$$ is the open-interest weight; and $$IV_{j,t}^{i,\text{call}}$$, $$IV_{j,t}^{i,\text{put}}$$ are the implied volatilities. Only short-term options with time to maturity of 10 to 60 days are used. Daily $$\text{Absoptivspread}_{i,t}$$ is then averaged from July to the next May (11 months post-reconstitution) to form the annual dependent variable. A smaller value indicates better option market conditions (put-call parity more closely obeyed).

The alternative liquidity measure OptNZVD follows Liu (2006) and is the standardized turnover-adjusted number of zero trading volume days over $$x$$ months (equation (5), p. 12):

$$
\text{OptNZVD}_{x} = \left( N_{0,x} + \frac{1/(x\text{-month turnover})}{\text{Deflator}} \right) \times \frac{21x}{\text{NoTD}} \tag{5}
$$

where $$N_{0,x}$$ is the number of zero-volume trading days, $$x\text{-month turnover}$$ is the sum of daily options turnover (volume in shares divided by shares outstanding), NoTD is the number of trading days, and the Deflator ensures the fractional term lies in $$(0,1)$$. The paper uses $$x = 11$$ months (July through next May).

## Empirical specifications

**Main regression (option market conditions), equation (2), p. 5:**

$$
\text{Absoptivspread}_{i,t} = \tau \cdot \text{Dum2000}_{i,t} + \delta X_{i,t-1} + \text{FixedEffects} + \xi_{it} \tag{2}
$$

where $$\text{Dum2000}_{i,t}$$ equals one when stock $$i$$ is at the top of the Russell 2000 in year $$t$$ (zero when at the bottom of the Russell 1000). $$X_{i,t-1}$$ is a vector of controls following Roll et al. (2010) and Lin and Lu (2015): log market capitalization in May (LNMAYSIZE), book-to-market ratio (B/M), cumulative daily stock return (LAGSTOCKRET), skewness of daily stock returns (LAGSTOCK\_SKEW), log number of analysts (ANALYSTS), standard deviation of analyst earnings forecasts (DISPERSION), average stock bid-ask spread (STKSPREAD), average stock trading volume (STKVOL), mean open-interest-weighted implied volatility (IMPLIEDVOL), cumulative daily S&P 500 return (SP500), and average daily VIX (VIXYEAR). All controls are constructed over the same 11-month window. Fixed effects include year and industry. The coefficient $$\tau$$ captures the indexing effect on option market conditions.

**Liquidity spillover regressions, equations (3) and (4), pp. 7-8:**

$$
\text{Optvol}_{i,t} = \tau \cdot \text{Dum2000}_{i,t} + \delta X_{i,t-1} + \text{FixedEffects} + \xi_{it} \tag{3}
$$

$$
\text{Optspread}_{i,t} = \tau \cdot \text{Dum2000}_{i,t} + \delta X_{i,t-1} + \text{FixedEffects} + \xi_{it} \tag{4}
$$

where $$\text{Optvol}_{i,t}$$ is the log of total options contracts and $$\text{Optspread}_{i,t}$$ is the open-interest-weighted daily average bid-ask spread in percent. Equations (3) and (4) are also estimated separately for call and put options (Tables 5 and 6, pp. 9-10) to test the lending-fee channel.

**Alternative measures.** The same specifications replace the dependent variable with OptNZVD (equation (5), Table 8) and Optilliq, the average daily change in options prices divided by dollar trading volume, adjusted for mechanical price changes due to the underlying (Table 9). Both follow from Amihud (2002) and Liu (2006) adapted for options.

**Robustness.** Table 7 (p. 12) repeats the main regressions with bandwidths of ±25 and ±75 stocks. In both cases the Dum2000 coefficient is negative and statistically significant for put-call parity deviation, positive and significant for options volume, and negative and significant for bid-ask spreads, consistent with the main results at ±50.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| OptionMetrics | Daily options data: strike price, trading volume, price, open interest, maturity, implied volatility, bid-ask spread, delta (for all exchange-traded options on sample stocks) | [OptionMetrics](/wiki/commercial/optionmetrics/) (licensed) |
| CRSP | Daily stock price and trading volume for underlying stocks; used for control variables (STKVOL, LAGSTOCKRET, LAGSTOCK\_SKEW) | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| Compustat | Annual accounting data for control variables (book-to-market ratio B/M) | [WRDS / Compustat](/wiki/commercial/wrds/) (licensed) |
| Russell 1000/2000 Index membership lists | Annual index membership and ranking data (from Russell); used to identify Dum2000 and the threshold cutoff | No page yet |

Sample: 1998-2006 (9 annual reconstitutions), 11-month estimation window per reconstitution (July through next May). Bandwidth: ±50 stocks around the Russell 1000/2000 cutoff. Final samples range from 199 (put-call parity regressions, Table 2) to 559 (options volume, Table 3) stock-year observations depending on options data coverage.

## When to read the full paper

Use the [original](https://doi.org/10.1016/j.finmar.2025.101026) if you are: studying how equity market structure changes spill over into derivatives markets; extending the analysis to post-2007 reconstitutions or other index settings; distinguishing supply-side (market-making cost), demand-side (informed trading), and lending-fee channels in options markets; or building on the put-call parity deviation measure of Rösch et al. (2017) in an RDD setting. Tables 2-4 (pp. 7-9) contain the primary results; Tables 5-6 (pp. 9-10) contain the mechanism tests; Tables 7-9 (pp. 12-14) contain robustness.

## Attribution and rights

Source: peer-reviewed, *Journal of Financial Markets* 78 (2026) 101026. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. CC BY-NC-ND 4.0 license permits redistribution of verbatim copies; derivative works are not permitted. This page is an extract-only distillation; the PDF is not hosted.

> **Attribution (CC BY-NC-ND 4.0).** Chang, Eric C., Li Ge, Tse-Chun Lin, and Xiaorong Ma.
> "The effect of stock market indexing on option market conditions."
> *Journal of Financial Markets* 78 (2026): 101026.
> DOI: 10.1016/j.finmar.2025.101026. © 2025 The Authors.
> Published by Elsevier B.V. Licensed under [CC BY-NC-ND 4.0](https://creativecommons.org/licenses/by-nc-nd/4.0/).
> This page is an extract-only distillation by the Institute for Automated Research.


==============================================================================
# Domestic Funds and Price Informativeness: Chen, Wu, Yang & Zhong (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jfm/2026/chen-domestic-funds-price-informativeness-2026/
# Distilled: Using Chinese listed companies (2005-2019), domestic fund ownership alone has no significant effect on stock price informativeness, but incentive-weighted domestic fund ownership significantly improves it through two channels: information processing and information provision. J. Financial Markets 2026, paywalled. Seven core results with source locators, datasets, the empirical design, and the firm-level price informativeness decomposition.
# Tags: paper-summary, asset-pricing, equities, price-informativeness, fund-behavior
==============================================================================

**What this is.** The paper's core results, the empirical design, and the two proposed mechanisms (information processing and information provision) with their estimating equations: enough to know what it found and why, without reading the full article. To replicate or extend it, read the original at [https://doi.org/10.1016/j.finmar.2025.101027](https://doi.org/10.1016/j.finmar.2025.101027).

## TL;DR

Using a panel of Chinese A-share listed companies from 2005 to 2019 (21,242 firm-level observations), the paper shows that plain domestic institutional fund ownership has no significant effect on stock price informativeness, consistent with prior literature (Kacperczyk, Sundaresan and Wang (2021)). However, once incentives are accounted for, incentive-weighted domestic fund ownership significantly improves price informativeness: a greater incentive-weighted shareholding ratio is associated with higher future earnings forecastability of stock prices. Price informativeness is measured following Carpenter, Lu and Whitelaw (2021) as the sensitivity of future earnings to current stock prices. Two mechanisms drive this: (1) incentivized funds process information better (their shareholding changes predict future stock returns more accurately), and (2) they provide more and higher-quality information to firms (more frequent and more positively toned site visits). A novel MO-OLS-based firm-level decomposition traces approximately 43-44% of the improvement attributable to the fund manager, consistent with Fang, Kempf and Trapp (2014) who find fund managers play a predominant role in funds' price discovery; year fixed effects and fund-stock matching account for the remainder.

## Core results

Magnitudes and significance as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Domestic fund ownership (Dom) has **no significant effect** on price informativeness at either horizon | Table 2, cols 2 and 7, p. 6 | Dom x log(M/A) = -0.014 (0.034) [h=1]; -0.013 (0.025) [h=3]; neither significant |
| R2 | Incentive-weighted domestic ownership (Dom\_Inc) significantly **raises** price informativeness | Table 2, cols 4 and 9, p. 6 | Dom\_Inc x log(M/A) = 0.038\*\* (0.015) [h=1]; 0.032\*\* (0.016) [h=3] |
| R3 | Synchronicity robustness: Dom\_Inc **reduces price synchronicity** (more idiosyncratic, more informative) | Table 3, Panel A, col 2, p. 8 | Dom\_Inc on SYNCH = -0.149\*\*\* (0.032); Dom insig at -0.042 (0.029) |
| R4 | PIN robustness: Dom\_Inc **raises probability of informed trading** | Table 3, Panel B, col 8, p. 8 | Dom\_Inc on PIN = 0.053\*\*\* (0.004) |
| R5 | Information processing channel: changes in Dom\_Inc **predict future returns** (funds buy/sell before price moves) | Table 5, cols 2 and 4, p. 11 | delta Dom\_Inc on Ret_{t+1} = 0.023\*\* (0.010); on Ret_{t+4} = 0.144\*\* (0.062); delta Dom is negative (-0.088\*\*\*, col 1) |
| R6 | Information provision: Dom\_Inc **raises site visit quantity and quality** | Table 6, cols 2 and 8, p. 12 | VisitNum coeff = 4.532\*\*\* (0.696); VisitTone coeff = 0.320\*\*\* (0.122) |
| R7 | Decomposition: **fund manager** explains ~43% of the incentive-driven price informativeness improvement | Fig. 1, p. 14 | Fund manager: 43.74% (h=1), 43.49% (h=3); year FE: ~30-35%; matching: ~22-25% |

**Overall (paper's conclusion).** Incentives reconcile the puzzle that domestic funds appear passive despite holding local informational advantages. Once direct and flow incentives are folded into a weighted ownership measure, domestic funds demonstrably improve price efficiency through both superior information processing and active information provision to firms.

## Theory / model

The paper has no formal theoretical model. It proposes two empirical hypotheses motivated by prior theory (Grossman and Stiglitz (1980), Kyle (1985), Holmstrom and Tirole (1993)):

**Hypothesis 1: Information processing channel.** Incentive-aligned fund managers process publicly available information more efficiently. Under this channel, an increase in incentive-weighted domestic fund ownership should align with subsequent stock-price movements, i.e., the change in Dom\_Inc should positively predict future stock returns (p. 9). Formally: if $$\beta_1 > 0$$ in the return-prediction regression (equation 11, p. 9), funds are proficient at anticipating future price movements.

**Hypothesis 2: Information provision channel.** Incentive-aligned funds actively seek and transmit private information to firm managers through corporate site visits (Chen, Goldstein and Jiang (2007); Bond, Edmans and Goldstein (2012)). Under this channel, higher Dom\_Inc should increase (a) the number of site visits (VisitNum) and (b) the quality of those visits (VisitTone), which in turn should improve the earnings-forecasting content of stock prices. A positive coefficient of Dom\_Inc in the visit regression and a significant interaction of VisitNum with log(M/A) in the earnings-forecastability regression (Table 6, cols 3-6) would support this channel (p. 11-12).

**Identification caveat.** The paper uses panel OLS with industry and period fixed effects plus firm-level controls; no instrument or quasi-natural experiment is employed. The identification rests on selection-on-observables. Results are robust to alternative price informativeness measures (synchronicity, PIN, RPE) and to lagged ownership, but a causal reading requires this conditional-ignorability assumption.

## Method

**Price informativeness measure (MO-OLS).** Following Keane and Neal (2020), the paper estimates a mean-observation OLS (MO-OLS) framework, building on `mo-ols`, that allows the coefficient linking stock prices to future earnings to vary over both firm and time dimensions. The procedure nests three regression levels:

Pooled regression to obtain $$\tilde{b}^{\text{Dom\_Inc},t+h}$$ (equation 16, p. 13):

$$
\frac{E_{i,t+h}}{A_{it}} = a^t + b^{\text{Dom\_Inc},t+h} \log\!\left(\frac{M}{A}\right)_{it} \times Dom\_Inc_{it} + \gamma' X_{it} + \varepsilon_{it}
\tag{16}
$$

Time-specific cross-sectional regression to obtain $$\hat{b}_t^{\text{Dom\_Inc},t+h}$$ (equation 17, p. 13):

$$
\frac{E_{i,t+h}}{A_{it}} = a^t + b_t^{\text{Dom\_Inc},t+h} \log\!\left(\frac{M}{A}\right)_{it} \times Dom\_Inc_{it} + \gamma_t' X_{it} + v_{it}
\tag{17}
$$

Unit-specific time-series regression to obtain $$\hat{b}_i^{\text{Dom\_Inc},t+h}$$ (equation 18, p. 13):

$$
\frac{E_{i,t+h}}{A_{it}} = a^t + b_i^{\text{Dom\_Inc},t+h} \log\!\left(\frac{M}{A}\right)_{it} \times Dom\_Inc_{it} + \gamma_i' X_{it} + u_{it}
\tag{18}
$$

A preliminary estimate (equation 19, p. 13) averages these three, and an iterative correction (equation 20) yields the consistent coefficient $$\hat{\beta}_{it}^{\text{Dom\_Inc},t+h}$$. The firm-level FPE measure is then $$\text{FPE}_{it}^{\text{Dom\_Inc},t+h} = \hat{\beta}_{it}^{\text{Dom\_Inc},t+h} \times \sigma_t(\log(M/A))$$ (equations 13-15, pp. 12-13), where $$\sigma_t(\log(M/A))$$ is the cross-sectional standard deviation of log price-to-asset ratios in year $$t$$.

**Decomposition.** Following Abdulkadiroglu, Pathak and Schellenberg (2020), the fund-level contribution of fund $$j$$ to firm $$i$$'s FPE is decomposed via a cross-sectional regression of fund-level contributions on fund-year intercepts and firm characteristics (equation 23, p. 15, using the `fund-pi-decomposition` approach):

$$
\text{FPE}_{ijt}^{t+h} = \alpha_{jt} + X_{it} \beta_{jt} + \varepsilon_{ijt}
\tag{23}
$$

The decomposition of equation 22 (p. 15) into three additive components then yields:

$$
\text{FPE}_{ijt}^{t+h} = \underbrace{\bar{\alpha}_t}_{\text{Year}_t} + \underbrace{(\alpha_{jt} - \bar{\alpha}_t)}_{\text{Manager}_{jt}} + \underbrace{X_{it}\beta_{jt} + \varepsilon_{ijt}}_{\text{Match}_{it}}
\tag{22}
$$

where $$\bar{\alpha}_t = \frac{1}{J}\sum_{j=1}^J \alpha_{jt}$$ is the year-average fund intercept.

## Empirical specifications

**Baseline (Table 2, equations 6, pp. 5-6).** The price informativeness regression follows Kacperczyk, Sundaresan and Wang (2021) and Bai, Philippon and Savoy (2016):

$$
\left(\frac{E}{A}\right)_{i,t+h} = \alpha + \beta_1 \log\!\left(\frac{M}{A}\right)_{it} + \beta_2 \log\!\left(\frac{M}{A}\right)_{it} \times \text{Fund\_ownership}_{it} + \beta_3 \text{Fund\_ownership}_{it} + \gamma' X_{it} + \lambda_{is} + \tau_t + \varepsilon_{it}
\tag{6}
$$

where $$(E/A)_{i,t+h}$$ is firm $$i$$'s earnings/assets in period $$t+h$$ ($$h=1,3$$); $$\log(M/A)_{it}$$ is the log price-to-asset ratio; Fund\_ownership is Dom, For, Dom\_Inc, or For\_Inc depending on the column; $$X_{it}$$ includes current earnings, insider ownership, leverage, tangibility, listed years, cash, and ROA; $$\lambda_{is}$$ are industry fixed effects; $$\tau_t$$ are period fixed effects. Standard errors are clustered at industry-period level. The coefficient of interest is $$\beta_2$$: the average price informativeness conditional on fund ownership type.

**Synchronicity robustness (Table 3, Panel A, equation 7-8, pp. 7-8).** Price synchronicity is estimated as the logistic transformation of $$R^2$$ from regressing firm $$i$$'s A-share return on market and industry factors:

$$
\text{SYNCH}_{it} = \log\!\left(\frac{R^2_{it}}{1 - R^2_{it}}\right)
\tag{8}
$$

A lower SYNCH implies more firm-specific information in prices. Fund ownership enters as a level regressor; a significantly negative coefficient on Dom\_Inc confirms the baseline result (Table 3 Panel A, col 2: -0.149\*\*\*).

**PIN robustness (Table 3, Panel B, equation 9, p. 7).** The probability of informed trading (VPIN) follows Easley, Kiefer, O'Hara and Paperman (1996):

$$
\text{VPIN} = \frac{1}{nV} \sum_{\tau=1}^{n} \left| V^{\tau}_{\text{buy}} - V^{\tau}_{\text{sell}} \right|
\tag{9}
$$

where $$V^{\tau}_{\text{buy}}$$ and $$V^{\tau}_{\text{sell}}$$ are buy and sell volumes in volume bucket $$\tau$$, $$n$$ is the number of buckets, and $$V$$ is the uniform bucket volume. The coefficient of Dom\_Inc on PIN (0.053\*\*\*, Table 3 Panel B, col 8) confirms the baseline.

**Information processing channel (Table 5, equation 11, pp. 9-10).** A return-prediction regression assesses whether fund ownership changes predict future stock returns:

$$
\text{Return}_{i,t+h} = \alpha + \beta_1 \Delta\text{Fund\_ownership}_{it} + \gamma' X_{i,t} + \lambda_{is} + \tau_t + \varepsilon_{it}
\tag{11}
$$

where $$h = 1$$ or $$4$$ periods; $$\Delta\text{Fund\_ownership}_{it}$$ is $$\Delta\text{Dom}_{it}$$, $$\Delta\text{For}_{it}$$, or $$\Delta\text{Dom\_Inc}_{it}$$; $$X_{it}$$ includes log size, book-to-market, past 12-month volatility, and momentum. A positive $$\beta_1$$ on $$\Delta\text{Dom\_Inc}$$ (0.023\*\*, col 2; 0.144\*\*, col 4) indicates forward-looking information content.

**Information provision channel (Table 6, equation 12, p. 11).** Firm site visits are regressed on fund ownership:

$$
\text{Visit}_{it} = \alpha + \beta_1 \text{Fund\_ownership}_{it} + \gamma' X_{i,t} + \lambda_{is} + \tau_t + \varepsilon_{it}
\tag{12}
$$

where Visit is proxied by VisitNum (annual count of site visits from all domestic funds, 2011-2019 from SZSE) or VisitTone (tone of the Investor Relations Activity Log, constructed using a Chinese Financial Sentiment Dictionary following Gordon, Loeb and Shu (2013) and Brockman, Cicon, Li, Price and Shu (2017)). A positive coefficient on Dom\_Inc in both (4.532\*\*\* and 0.320\*\*\*) supports the information provision channel.

**Fund incentive construction.** Fund $$j$$'s incentive to promote firm $$i$$'s value is the total management fee gain from a 1% increase in firm $$i$$'s value (equation 1, p. 3):

$$
\text{Incentives}_{ijt} = \text{Direct incentives}_{ijt} + \text{Flow incentives}_{ijt}
\tag{1}
$$

$$
\text{Direct incentives}_{ijt} = p \times \text{AUM}_{jt} \times w_{ijt}
\tag{2}
$$

$$
\text{Flow incentives}_{ijt} = p \times \text{AUM}_{jt} \times \beta^j \times (w_{ijt} - v_{(-j)it})
\tag{3}
$$

where $$p = 0.828\%$$ is the average management fee rate, $$w_{ijt}$$ is the value weight of stock $$i$$ in fund $$j$$'s portfolio, $$\beta^j$$ is fund $$j$$'s estimated inflow-to-performance sensitivity, and $$v_{(-j)it}$$ is the period-$$t$$ average value weight of stock $$i$$ across peer funds. The incentive-weighted domestic ownership variable is:

$$
\text{Dom\_Inc}_{it} = \sum_{j=1}^{J} \text{Incentives}_{ijt} \times \text{Dom}_{ijt}
\tag{5}
$$

Net fund inflows are estimated via (equation 4, p. 4):

$$
\text{Net Inflow}_{jt} = \frac{\text{AUM}_{jt} - \text{AUM}_{j,t-1}(1 + R_{jt})}{\text{AUM}_{j,t-1}}
\tag{4}
$$

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CSMAR (China Stock Market Accounting Research Database) | Firm financial characteristics, stock returns, market capitalization; Chinese A-listed non-financial corporates, 2005-2019 | no page yet |
| Resset (Renmin University CSMAR Economic Research Data System) | Semi-annual open-end fund holdings (3,400+ equity, hybrid, and index funds), 2005-2019 | no page yet |
| SZSE Investor Relations Activity Log | Firm site visit records (quantity and tone), 2011-2019; Shenzhen Stock Exchange only | no page yet |

Sample: 21,242 firm-period observations (Table 1), semi-annual frequency, 2005-2019 for baseline; site visit subsample 2011-2019. Observations with missing data, delisted firms, and firms with fewer than 90 trading days in the prior six months are excluded.

## When to read the full paper

Use the [original](https://doi.org/10.1016/j.finmar.2025.101027) if you are:
studying why domestic institutional investors in emerging markets fail to improve price efficiency despite having local informational advantages; extending the fund-incentive and price-discovery literature to non-US markets; replicating or extending the MO-OLS firm-level FPE measure or the three-component decomposition; or benchmarking the relative magnitudes of foreign vs. domestic institutional investors' impact on price efficiency in China.

## Attribution and rights

Source: peer-reviewed, *Journal of Financial Markets* 78 (2026) 101027. All rights reserved (© 2025 Elsevier B.V.). This distillation was extracted by an LLM (claude-sonnet-4-6) on 2026-06-25 and is **not human-verified or independently reproduced**. Extract-only: no verbatim PDF is hosted here.

> Chen, Shaoling, Xi Wu, Haisheng Yang, and Jiaying Zhong.
> "Incentives matter: Domestic funds and price informativeness improvement."
> *Journal of Financial Markets* 78 (2026) 101027.
> DOI: [10.1016/j.finmar.2025.101027](https://doi.org/10.1016/j.finmar.2025.101027).
> © 2025 Elsevier B.V. All rights reserved.


==============================================================================
# Geography and Hedge Fund Activism: Faleye (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jfm/2026/faleye-geography-hedge-fund-activism-2026/
# Distilled: Activist hedge funds disproportionately target firms located closer to their headquarters, yet activism returns are lower for nearer targets by 1.2 percentage points per one-standard-deviation decrease in distance. Economic explanations (activism costs, target selection, employee wealth transfers) are ruled out; results are consistent with familiarity bias in targeting decisions. Journal of Financial Markets 2026, CC BY-NC-ND 4.0. Eight core results with source locators, datasets used, the tested hypotheses, and the estimation specifications.
# Tags: paper-summary, equities, institutional-investors, hedge-fund-activism
==============================================================================

**What this is.** The paper's core results, the competing hypotheses it tests (information vs. familiarity), and the regression specifications with their key findings: enough to understand what the paper found and how, without reading all 23 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1016/j.finmar.2025.101005).

## TL;DR

The paper asks whether geographical proximity between a hedge fund and a potential target affects activism targeting and returns. Using 2,106 activism events involving 693 U.S. hedge funds from 2000 to 2014, Faleye finds that activist funds disproportionately target firms located closer to their headquarters: targets are 43.1% more likely than matched controls to be within 100 km of the fund. Despite this proximity preference, activism returns are lower for nearer targets: mean CAR[-10,+10] is 3.84% for below-median-distance targets versus 6.89% for those farther away. After ruling out lower activism costs, target selection effects, and reduced employee wealth transfers as explanations, the results align with familiarity bias in hedge fund targeting decisions. Additional tests show the proximity effect is stronger among small targets, openly hostile campaigns, and significant passive investments (Schedule 13G filings), all of which reinforce the behavioral explanation.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Activist hedge funds prefer geographically closer targets: distance is negative and significant in probit predicting targeting likelihood | Table 1, col (1), p. 5 | Distance coeff = -0.0225`\*\*\*` (SE 0.009); 1-SD decrease in distance raises targeting probability by 7.4% relative to unconditional mean |
| R2 | Firms within 100 km are significantly more likely to be activism targets | Table 1, col (3), p. 5 | Local firm (<=100 km) coeff = 0.1977`\*\*\*` (SE 0.035); targets 19.8% more likely to be local to the hedge fund |
| R3 | Activism returns are lower for closer targets in univariate comparisons | Table 2, Panel B, p. 6 | Mean CAR[-10,+10]: 3.84% (below-median distance) vs 6.89% (above-median); within 100 km: 4.36% vs beyond 100 km: 5.53% |
| R4 | Lower returns for closer targets hold in multivariate OLS controlling for target characteristics | Table 3, Panel A, cols (3)-(4), p. 7 | Nearer firm = -0.0307`\*\*\*` (SE 0.009); local firm = -0.0291`\*\*` (SE 0.014); continuous distance: 1-SD decrease = 1.2 pp decline in CAR (21.9% relative to full-sample mean) |
| R5 | Employment reductions post-activism are similar for proximate and distant targets: wealth-transfer explanation ruled out | Table 6, Panels A-B, p. 12 | Target x post-activism = -0.9816`\*\*\*` for full sample; Chow test nearer vs farther: p=0.951; no statistically significant difference by proximity |
| R6 | Proximity preference is significantly stronger among small targets than large ones, consistent with the mere-exposure channel being stronger for locally covered small firms | Table 9, Panel C, p. 16 | Small firms: local firm = 0.2517`\*\*\*`, 28.9% more likely to be local target; large: 0.1240`\*\*`, only 7.7%; Chow test p=0.046 |
| R7 | Proximity increases the likelihood of an openly hostile campaign more than a non-hostile one | Table 11, Panel A, p. 18 | Distance coeff for hostile = -0.0402`\*\*` (SE 0.019) vs non-hostile = -0.0172`\*` (SE 0.009) |
| R8 | Passive 13G investments also show a proximity preference and lower announcement returns for local investees | Tables 13-14, pp. 19-21 | Local firm (<=100 km) = 0.0945`\*\*` (SE 0.039) in 13G likelihood probit; local investee CAR coefficient = -0.0334`\*\*\*` (SE 0.012), 3.3 pp lower return |

**Overall (paper's conclusion).** Economic factors including activism costs, target selection effects, and reduced employee wealth transfers at nearby firms do not explain why hedge funds earn lower returns on geographically closer targets. Results are consistent with familiarity bias: hedge funds disproportionately target closer firms while realizing suboptimally lower returns from such activism, suggesting that even sophisticated investors are not immune to distortions created by a bias toward the familiar. This extends the hedge fund geography literature, where Teo (2009) documents that funds with regional investment focus outperform funds without a physical presence in the region, but at the portfolio level rather than the activist-targeting level.

## Theory / model

The paper has no formal economic model. It evaluates two competing hypotheses for why geographical proximity matters in hedge fund activism:

**Information hypothesis.** Proximity reduces the cost of obtaining localized "soft" information about potential targets, including local media coverage, interactions with employees and managers, relationships with suppliers and clients, and the ability to directly observe target facilities. If proximity provides an information advantage, activism returns should be higher (not lower) for closer targets. Brav et al. (2008) establish that most activism events are ones where the hedge fund believes the target is undervalued and that better monitoring, advising, and communication can unlock value; these actions are easier and less costly for closer entities, which is the kernel of the information hypothesis.

**Familiarity hypothesis.** Proximity reflects the familiarity bias, a cognitive bias arising from the "mere exposure" effect first identified by Zajonc (1968). Repeated exposure to a stimulus creates affective preference for it even without substantive informational content. Huberman (2001) shows this bias is pervasive among investors, including the tendency of individuals to hold shares in their regional Bell Operating Company. A related bias is the "illusion of control" (Langer, 1975): hedge fund managers may overestimate their ability to influence outcomes at nearby firms because they perceive them as more familiar and hence easier to control. This predicts a proximity preference in targeting but lower (suboptimal) returns for closer targets because target selection is, by definition, biased. To distinguish the two hypotheses from rational wealth-transfer explanations, the paper tests whether wealth transfers from employees explain the return gap. Agrawal and Lim (2022) document significant post-activism pension underfunding; the paper tests and rules out this as the driver of lower proximity returns, finding that such wealth transfers are equally present for proximate and distant targets.

The two hypotheses have opposite predictions for activism returns: the information hypothesis predicts higher returns for closer targets; the familiarity hypothesis predicts lower returns. The empirical analysis is organized around distinguishing these predictions and ruling out rational economic alternatives.

## Method

The paper applies standard empirical methods from the corporate governance and behavioral finance literatures; it proposes no new estimator. The headline empirical exercises are:

**Targeting propensity model.** For each activism event, five control firms are selected by propensity score matching on firm size (log market cap), Tobin's q, three-year average sales growth, institutional ownership, and industry, within the same two-digit SIC code industry and within 0.2 standard deviations of the target's score. Matching is with replacement. This builds on `matching` to construct a counterfactual set of potential targets.

**Probit regressions.** Targeting likelihood is estimated via probit (`probit-regression`) over the 1,794-target matched sample with 3,679 unique control firms. The model includes year and two-digit SIC code industry fixed effects; standard errors are clustered at the firm level.

**Activism return OLS.** For the 2,060-event return subsample, CAR[-10,+10] is regressed on proximity measures and controls via OLS (`panel-regression`) with year and industry or hedge-fund fixed effects; standard errors are clustered at the firm level.

**Difference-in-differences (employment/pensions).** For the wealth-transfer tests in Section 5.3, a DiD estimator (`difference-in-differences`) is applied. Each target is matched (propensity score, no replacement) to a control firm in the same two-digit SIC industry. Treated and control firms are compared over the [-5, +5] year window around the activism event.

## Empirical specifications

**Targeting probit (Table 1).** The dependent variable is an indicator equal to 1 for the target in the targeting fiscal year and 0 for matched controls in all years. The proximity variable in column (1) is distance in thousands of kilometers; columns (2)-(3) use indicator variables for below-median and within-100-km distance, respectively. Equation estimated:

$$
\text{Pr}(\text{Target}_{it} = 1) = \Phi\!\left(\alpha + \beta_1 \text{Proximity}_{it} + \beta_2' X_{it} + \gamma_t + \delta_j\right)
$$

where $$\Phi$$ is the standard normal CDF, $$\text{Proximity}_{it}$$ is one of three distance measures, $$X_{it}$$ is a vector of firm controls (size, Tobin's q, excess return, ROA, capital expenditures, cash holdings, institutional ownership, R&D, total debt, sales growth, SG&A, dividends, dividend yield), $$\gamma_t$$ are year fixed effects, and $$\delta_j$$ are industry fixed effects. Sample: 87,134 observations (1,794 targets + 3,679 control firms x multiple years).

**Activism return OLS (Table 3).** CAR[-10,+10] is the dependent variable (p. 7). The key coefficients are on the proximity measures:

$$
\text{CAR}_{i}[-10,+10] = \alpha + \beta_1 \text{Proximity}_{i} + \beta_2' X_i + \gamma_t + \delta_j + \varepsilon_i
$$

Controls $$X_i$$ include market cap, Tobin's q, cash holdings, total debt, dividend yield, ROA, dissident ownership, board representation, and a wolf-pack indicator. Standard errors clustered at the firm level. Panel A uses year and industry FE (N=2,056); Panel B uses hedge-fund FE (N=652, serial activists only). The continuous distance coefficient in column (1) is 0.0082`\*\*\*` (SE 0.003); the nearer-firm indicator in column (3) is -0.0307`\*\*\*`.

**DiD employment regression (equation 1, p. 10).** The paper's only numbered equation. The dependent variable $$Y_{it}$$ is employment in thousands. The estimation window is [-5, +5] years relative to the activism fiscal year:

$$
Y_{it} = \beta_0 + \beta_1 \text{Post}_{it} + \beta_2 \left(\text{Target}_i \times \text{Post}_{it}\right) + \gamma \text{Controls}_{it} + \varphi \text{Year}_t + \delta \text{Firm}_i + \varepsilon_{it} \tag{1}
$$

where $$\text{Post}_{it}$$ equals 1 for post-activism fiscal years, $$\text{Target}_i$$ equals 1 for target firms (0 for matched controls), and $$\delta \text{Firm}_i$$ are firm fixed effects. The coefficient of interest is $$\beta_2$$, capturing the differential change in employment for target firms relative to controls after activism. $$\beta_2 = -0.9816$$`\*\*\*` (SE 0.189) in the full sample (Table 6 Panel A), implying a 982-employee reduction (14.6% of mean). Panels B and C split the sample by proximity; Chow tests find no statistically significant difference (p-values 0.951 and 0.916).

**Hostile and non-hostile campaign probits (Table 11).** The probit specification mirrors Table 1, estimated separately for 402 hostile and 1,704 non-hostile campaigns. Coefficients on the continuous distance variable: hostile = -0.0402`\*\*` vs non-hostile = -0.0172`\*`.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Boyson and Pichler (2019) activism sample | Base sample of 2,206 hedge fund activism events (2000-2014), from which 2,106 events in the 48 contiguous US states are retained | No page yet |
| SEC EDGAR (Schedule 13D/13G filings) | Identify hedge fund and target ZIP codes; define passive 13G investments; locate fund headquarters over time | [SEC EDGAR](/wiki/datasets/edgar/) |
| CRSP (via WRDS) | Market model for CAR/BHAR estimation; stock returns; value-weighted market portfolio (estimation window [-301, -46]) | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| Compustat (via WRDS) | Firm characteristics: market cap, Tobin's q, ROA, sales growth, capital expenditures, cash holdings, R&D, total debt, SG&A, dividends; DB pension data (Compustat Pension files) | [WRDS / Compustat](/wiki/commercial/wrds/) (licensed) |
| DOL Private Pension Plan Research Files (Form 5500) | Employer contributions to defined contribution (DC) pension plans, matched to Compustat via EIN; 1,001 events for 769 unique targets | No page yet |

Sample scope: 2,106 activism events, 693 U.S. hedge funds, 2000-2014; return tests use 2,060 events; 13G tests use 5,568 filings.

## When to read the full paper

Use the [original](https://doi.org/10.1016/j.finmar.2025.101005) if you are: studying behavioral biases in institutional investors with a focus on hedge fund targeting; evaluating whether geographical proximity is an economic or behavioral factor in activism; replicating the DiD employment or pension tests (the paper matches each event to one control firm without replacement, so the exact match set matters for robustness); or extending the analysis to other dimensions of familiarity (language, industry familiarity, social networks). Locators above point directly to the tables.

## Attribution and rights

Source: peer-reviewed, *Journal of Financial Markets* 77 (2026) 101005. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. The CC BY-NC-ND 4.0 licence permits reuse for non-commercial purposes with attribution and no derivatives; the verbatim PDF is not hosted in this batch.

> Faleye, Olubunmi. "Does Familiarity Breed Activism? Geography and Hedge Fund Activism."
> *Journal of Financial Markets* 77 (2026): 101005.
> DOI: [10.1016/j.finmar.2025.101005](https://doi.org/10.1016/j.finmar.2025.101005).
> © 2025 The Author. Published by Elsevier B.V. under [CC BY-NC-ND 4.0](https://creativecommons.org/licenses/by-nc-nd/4.0/).
> This page is a distillation by the Institute for Automated Research: core results summarized and re-expressed; **not a substitute for the original**.


==============================================================================
# Dealer Competition in OTC Markets: Singer (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jfm/2026/singer-dealer-competition-otc-markets-2026/
# Distilled: A model of OTC dealer competition as a first-price sealed-bid common-value auction shows that information heterogeneity arises endogenously and generates core-periphery market structures in which better-informed core dealers quote tighter bid-ask spreads, earn higher margins, and trade more frequently. Journal of Financial Markets 2026, CC BY 4.0. Six core results with source locators and the formal model equations.
# Tags: paper-summary, otc-markets, market-microstructure, dealer-competition
==============================================================================

**What this is.** The paper's core results, the game-theoretic model of OTC dealer competition, and the main propositions with their formal equations: enough to know what it finds and how, without reading all 23 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1016/j.finmar.2025.101004).

## TL;DR

The paper models dealer competition in over-the-counter (OTC) markets as a first-price sealed-bid common-value auction under endogenous uncertainty. An investor simultaneously asks $n$ dealers to quote bid and ask prices for one unit of a risky asset. Neither the asset's true value nor rivals' private signals are observed by any dealer, creating a winner's curse problem: the dealer whose quote is accepted is most likely to have overestimated (underestimated) the value. Dealers mitigate this by investing in costly information acquisition, raising signal accuracy. However, when dealer 1 (the most-informed) tightens its bid-ask spread, this intensifies price competition for others, deterring them from matching dealer 1's accuracy. Equilibrium information heterogeneity emerges endogenously: for intermediate information-acquisition costs, a unique core-periphery equilibrium exists in which one well-informed core dealer coexists with less-informed peripheral dealers. The core dealer quotes the tightest bid-ask spread to individual investors yet earns the highest trading margins and loss rates that are zero for a wide range of cost parameters. This is consistent with centrality premia documented in U.S. municipal bond markets (Li and Schürhoff (2019)) and U.S. corporate bond markets.

## Core results

Locators point into the source PDF (23 pages). All results are theoretical propositions and theorems; no empirical data was used.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Better-informed dealers quote tighter bid-ask spreads (Proposition 2, part i) | Prop. 2, §2.2, p. 7; Prop. 1, p. 6 | Bid-ask spread equals $$-2b_i$$; normalized bid $$b_i$$ is strictly decreasing in uncertainty level $$\varepsilon_i$$ (Proposition 1): better-informed dealers set lower $$b_i$$, narrowing their spread |
| R2 | Better-informed dealers earn higher expected trading margins (Proposition 2, part ii) | Prop. 2, §2.2, p. 7; Fig. 4, p. 10 | Expected trading margin $$\Delta_i = R_i / P_i$$ is strictly decreasing in $$\varepsilon_i$$; in the $$n=3$$ core-periphery equilibrium, core dealer margin exceeds peripheral margin by up to a factor of 2 (Fig. 4, left panel) |
| R3 | Better-informed dealers trade more frequently (Proposition 2, part iii) | Prop. 2, §2.2, p. 7; Fig. 4, p. 10 | Trading probability $$P_i(\varepsilon_i, \varepsilon_{-i})$$ strictly decreasing in $$\varepsilon_i$$; core dealer trading probability exceeds peripheral dealer's (Fig. 4, right panel) |
| R4 | Better-informed dealers incur fewer trading losses (Proposition 2, part iv) | Prop. 2, §2.2, p. 7; Fig. 5, p. 11 | Loss probability $$P^L_i(\varepsilon_i, \varepsilon_{-i})$$ strictly increasing in $$\varepsilon_i$$; in $$n=3$$ core-periphery equilibrium, core dealer loss probability is zero for a wide range of the cost parameter $$a$$ (Fig. 5) |
| R5 | Core-periphery dealer structures emerge endogenously for intermediate information-acquisition costs (Result 2) | Result 2, §3.2, p. 9; Table 1, p. 9; Fig. 2, p. 9 | For $$n=3$$: unique core-periphery equilibrium for $$a\varepsilon_I \in (a_l, a_h)$$ with $$a_l = 0.031$$, $$a_h = 0.974$$; unique symmetric equilibrium for $$a\varepsilon_I \geq a_s = 0.733$$; no equilibrium for $$a\varepsilon_I < a_l$$ |
| R6 | Investor transaction costs decrease monotonically in the number of competing dealers (Section 5, Fig. 7) | §5, Fig. 7, p. 12-13; Table A.1, p. 22 | At $$a\varepsilon_I = 0.2$$: $$C_I(2) / C_I(10) \approx 4.05$$; ratio is above 1 for all $$n \in \{2, \ldots, 10\}$$ (Fig. 7) |

**Overall (paper's conclusion).** The friction of opaque market prices in OTC markets, modeled as simultaneous first-price sealed-bid common-value auctions, endogenously generates both (i) the core-periphery dealer structures observed in real markets and (ii) the cross-dealer differences in bid-ask spreads, trading margins, trading frequencies, and loss rates documented in U.S. fixed-income markets. Core dealers earn centrality premia despite quoting tighter spreads, because their superior information leads to fewer mispriced trades offset by higher-margin trades.

## Theory / model

The model has $n \geq 2$ dealers indexed by $\mathcal{N} = \{1, \ldots, n\}$, ordered so dealer 1 has the lowest uncertainty level ($\varepsilon_1 \leq \varepsilon_j$ for all $j$), and one investor (p. 4). The asset has a common value $\theta$ for dealers and a private value $\theta_I$ for the investor. The investor's private value is drawn from $\theta_I \sim U[\theta - \varepsilon_I, \theta + \varepsilon_I]$. Dealer $i$ observes a private signal $\theta_i \sim U[\theta - \varepsilon_i, \theta + \varepsilon_i]$, where $\varepsilon_i \in (0, \varepsilon_I/2]$ is dealer $i$'s uncertainty level (inverse signal accuracy). Both $\theta$ and $\theta_I$ are unknown to the dealers; this is the common-value auction environment.

The investor asks all $n$ dealers simultaneously to quote a binding bid price $B_i(\theta_i)$ (the price at which dealer $i$ will buy) and ask price $A_i(\theta_i)$ (the price at which dealer $i$ will sell). The investor accepts the best bid $\bar{B} \equiv \max\{B_i(\theta_i) : i \in \mathcal{N}\}$ if $\bar{B} \geq \theta_I$, or the best ask $\underline{A} \equiv \min\{A_i(\theta_i) : i \in \mathcal{N}\}$ if $\underline{A} \leq \theta_I$. Dealer $i$ earns trading margin $\theta - B_i(\theta_i)$ when buying, or $A_i(\theta_i) - \theta$ when selling (p. 4).

The bid and ask problems are mathematically symmetric; the paper solves the bid side. Lemma 1 (p. 5) establishes that all equilibrium bid strategies are linear with slope $c_i = 1$ in the dealer's signal, so $B_i(\theta_i) = b_i + \theta_i$ and it suffices to find the normalized bid prices $b_i = B_i(\theta_i) - \theta_i$ (each a scalar independent of $\theta_i$). Letting the normalized common value be $\bar{\theta}^i \equiv \theta - \theta_i$, dealer $i$'s expected bid-price profit simplifies to (eq. 4, p. 5):

$$
\pi_i^B(b_i, b_{-i}) \equiv \frac{1}{2\varepsilon_i} \int_{-\varepsilon_i}^{\varepsilon_i} (\bar{\theta}^i - b_i) \, f_i(b_i, b_{-i} \mid \bar{\theta}^i) \, d\bar{\theta}^i, \tag{4}
$$

where the conditional buying-probability factor is $f_i(b_i, b_{-i} \mid \bar{\theta}^i) \equiv g_i(b_i - \bar{\theta}^i) \prod_{k \in \mathcal{N} \setminus \{i\}} g_k(b_i - b_k - \bar{\theta}^i)$, and the function $g_j$ (eq. 5, p. 5) is:

$$
g_j(x) \equiv \begin{cases} 1, & x > \varepsilon_j, \\ \dfrac{\varepsilon_j + x}{2\varepsilon_j}, & -\varepsilon_j < x \leq \varepsilon_j, \\ 0, & x \leq -\varepsilon_j. \end{cases} \tag{5}
$$

The four dealer statistics derived from the equilibrium (pp. 6-7) are:

- **Bid-ask spread**: $A_i(\theta_i) - B_i(\theta_i) = -2b_i > 0$ (always positive by Lemma A.1 in Appendix).
- **Trading margin** (eq. 8): $\Delta_i(\varepsilon_i, \varepsilon_{-i}) \equiv R_i(\varepsilon_i, \varepsilon_{-i}) / P_i(\varepsilon_i, \varepsilon_{-i})$, where $R_i = 2\pi_i^B$ is expected revenue.
- **Trading probability** (eq. 9): $P_i(\varepsilon_i, \varepsilon_{-i}) \equiv 2 \int_{-\varepsilon_i}^{\varepsilon_i} f_i(b_i, b_{-i} \mid \bar{\theta}^i) \, d\bar{\theta}^i / (2\varepsilon_i)$.
- **Loss probability** (eq. 10): $P_i^L(\varepsilon_i, \varepsilon_{-i})$, the probability that dealer $i$ buys (sells) the asset and later resells (buys back) at a loss.

The framing of dealers as first-price sealed-bid common-value auctioneers builds on the analysis of Milgrom and Weber (1982b) and on the information-acquisition model of Persico (2000), and is complementary to the search-and-bargaining OTC model of Duffie, Garleanu, and Pedersen (2005) and to the endogenous-structure model of Farboodi, Jarosch, and Shimer (2022).

## Method

The model is solved via backward induction through a two-stage game (p. 7). Stage 2 (price competition): dealers simultaneously quote bid and ask prices given fixed uncertainty levels $(\varepsilon_1, \ldots, \varepsilon_n)$. Stage 1 (information acquisition): dealers choose $\varepsilon_i \in (0, \varepsilon_I/2]$ anticipating the Stage-2 equilibrium.

**Stage 2 equilibrium.** The equilibrium first-order condition for dealer $i$'s normalized bid price (eq. 6, p. 5-6) requires marginal expected bid-price profit to be zero:

$$
\frac{d\pi_i^B(b_i, b_{-i})}{db_i} = \frac{1}{2\varepsilon_i} \Big[ (-\varepsilon_i - b_i) f_i(b_i, b_{-i} \mid -\varepsilon_i) - (\varepsilon_i - b_i) f_i(b_i, b_{-i} \mid \varepsilon_i) \Big] = 0. \tag{6}
$$

Theorem 1 (p. 6) establishes a unique equilibrium: for all dealers $i \in \mathcal{N} \setminus \{1\}$ the equilibrium is simply $b_i = -\varepsilon_i$. For dealer 1 (most-informed):
- if $\varepsilon_1 \geq \varepsilon_I/2$, then $b_1 = -\varepsilon_1$;
- otherwise $b_1 \in [-\min(\varepsilon_2, \varepsilon_I/2), -\varepsilon_1]$ satisfies eq. (6) implicitly.

Theorem 2 (p. 6): the equilibrium ask-price strategy is $A_i(\theta_i) = -b_i + \theta_i$ for all dealers.

Proposition 1 (p. 6) establishes comparative statics: $db_1/d\varepsilon_1 < 0$ (dealer 1 quotes tighter spreads when more informed) and $db_i/d\varepsilon_i = -1$ for $i \neq 1$ (peripheral dealers' spreads move one-for-one with uncertainty). Proposition 3 (p. 7) shows dealer 1 has market power: $db_1/d\varepsilon_k \leq 0$ for $k \in \mathcal{N} \setminus \{1\}$, meaning dealer 1 narrows its spread as rivals become better informed, deterring rivals from information acquisition.

**Stage 1 equilibrium.** The information-acquisition cost for dealer $i$ is (p. 7):

$$
C(\varepsilon_i) = \frac{a}{2} \left( \frac{\varepsilon_I}{2} - \varepsilon_i \right)^2,
$$

with cost parameter $a > 0$. Dealer $i$'s overall expected profit is (eq. 11, p. 7):

$$
\Pi_i(\varepsilon_i, \varepsilon_{-i}) \equiv R_i(\varepsilon_i, \varepsilon_{-i}) - C(\varepsilon_i). \tag{11}
$$

Theorem 3 (p. 8) establishes that symmetric equilibria in uncertainty levels (all dealers equally informed) do not exist if the cost parameter $a$ is sufficiently low: dealer 1's market power prevents ex ante identical dealers from choosing the same information level. When a symmetric equilibrium exists (Result 1, eq. 15, p. 8), the equilibrium uncertainty ratio satisfies:

$$
\frac{\varepsilon_1}{\varepsilon_I} = \frac{(n(n+1)a\varepsilon_I - 4)(n+2)}{2(n+1)(n+2)a\varepsilon_I - 8}, \tag{15}
$$

valid for $a\varepsilon_I \geq a_s$ where $a_s$ is a threshold that equals 0.507 at $n=4$ and decreases as $n$ grows (Table 1, p. 9). For $a\varepsilon_I \in (a_l, a_h)$, a unique core-periphery equilibrium exists with $\varepsilon_1 < \varepsilon_j$ for all $j \neq 1$. The phase boundaries $a_l$, $a_s$, $a_h$ are derived numerically via the fixed-point procedure described in Appendices A.8-A.9 (pp. 19-21).

## Empirical specifications

This is a pure-theory paper with no empirical estimation. No data was used (Data availability statement, p. 22). The paper conducts a numerical equilibrium analysis for $n \in \{2, \ldots, 15\}$ dealers (Table 1, p. 9; Appendices A.8-A.9) that:

1. Determines phase boundaries $(a_l, a_s, a_h)$ for each $n$ by solving for symmetric and core-periphery equilibria via a four-step best-response search over $a$ and $\varepsilon_I$ (Appendix A.8, p. 19).
2. Verifies Result 2 for the $n = 3$ case: a symmetric equilibrium exists for $a\varepsilon_I \geq 11/15$; a core-periphery equilibrium for $a\varepsilon_I \in (0.031, 0.974)$; the thresholds are $a_s = 11/15 \approx 0.733$, $a_l = 0.031$, $a_h = 0.974$ (Appendix A.9, p. 20-21).
3. Shows transaction costs $C_I(n)$ decrease monotonically in dealer count for $a\varepsilon_I \in \{0.2, 0.8\}$ (Table A.1, p. 22; Fig. 7, p. 12).

Model predictions are compared qualitatively with cross-dealer statistics in real OTC markets: Li and Schürhoff (2019) for U.S. municipal bond markets, Di Maggio et al. (2017) for U.S. corporate bond markets, and Hasbrouck and Levich (2021) for foreign exchange markets.

## Datasets used

The paper is entirely theoretical. No empirical datasets are analyzed.

| Dataset | Role in paper | Wiki page |
|---|---|---|
| None | Pure theory model; stylized facts from Li and Schürhoff (2019), Di Maggio et al. (2017), Hasbrouck and Levich (2021) are cited as motivation in the introduction but no data is analyzed here | n/a |

## When to read the full paper

Read the [original](https://doi.org/10.1016/j.finmar.2025.101004) if you are: (1) modeling OTC dealer competition with simultaneous price competition rather than sequential search; (2) building on the first-price common-value auction framework for market microstructure; (3) studying endogenous information acquisition in dealer markets; or (4) extending the welfare and transaction-cost analysis of Sections 4-5 to policy settings (transaction taxes, subsidies). Formal proofs of all lemmas, propositions, and theorems are in the Appendix (pp. 15-22).

## Attribution and rights

Source: peer-reviewed, *Journal of Financial Markets* 77 (2026) 101004. This distillation was extracted by an LLM on 2026-06-25 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Singer, Alexander. "Dealer Competition in Over-the-Counter Markets." *Journal of Financial Markets* 77 (2026) 101004. DOI: 10.1016/j.finmar.2025.101004. Copyright 2025 The Author. Published by Elsevier B.V. Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). This page is an **adaptation** by the Institute for Automated Research: core results extracted and re-expressed; **changes were made**.


==============================================================================
# Ambulance Taxis: Eliason, League, Leder-Luis, McDevitt & Roberts (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jpe/2024/eliason-et-al-ambulance-taxis-impact-regulation-2024/
# Distilled: Prior authorization for Medicare ambulance rides to dialysis facilities reduced nonemergency rides by 68% and payments by 67.7%, far outperforming criminal and civil pay-and-chase litigation. Journal of Political Economy 2025 (May 2025), paywalled. Eight core results with source locators, the stylized fraud-deterrence model, and the staggered difference-in-differences specifications.
# Tags: paper-summary, health-economics, public-finance, regulation, fraud, medicare
==============================================================================

**What this is.** This is a distilled skeleton of Eliason, League, Leder-Luis, McDevitt, and Roberts (2025). Read the [original paper](https://doi.org/10.1086/734134) or the [NIH Public Access version](https://pmc.ncbi.nlm.nih.gov/articles/PMC12331087/) to replicate or extend.

## TL;DR

Between 2003 and 2017, Medicare spent $7.7 billion on 37.5 million nonemergency ambulance rides transporting dialysis patients to and from treatment facilities, most of which did not satisfy Medicare's medical necessity criteria. Using the staggered rollout of a prior authorization requirement across US states and the differential timing of 69 criminal and civil DOJ lawsuits across 26 federal judicial districts, the paper identifies the causal effects of two anti-fraud approaches. Prior authorization, which requires ambulance companies to obtain physician sign-off before providing a ride and receiving payment, caused an immediate and persistent 68% drop in nonemergency rides. Criminal litigation reduced rides by roughly 24%, while civil litigation had no statistically significant effect. No evidence is found that prior authorization harmed patient health. The paper estimates the federal government would have saved $4.8 billion had it imposed prior authorization in 2003 at an administrative cost of only $28.6 million per year.

## Core results

| # | Result | Locator | Magnitude as reported |
|---|---|---|---|
| R1 | Prior authorization reduces total Medicare payments for nonemergency dialysis rides | Table 2, col. 1, p. 1676 | beta = -1.129 log points (SE = 0.350)\*\*, dep. mean = 9.934; -67.7% in levels |
| R2 | Prior authorization reduces total nonemergency ride count | Table 2, col. 3, p. 1676 | beta = -0.913 log points (SE = 0.176)\*\*\*, dep. mean = 5.357 |
| R3 | Criminal enforcement reduces total ride payments (modest, delayed) | Table 3, col. 3, p. 1678 | beta = -0.211 log points (SE = 0.106)+; Fig. 5 shows gradual decline over 24 months |
| R4 | Criminal enforcement reduces ride count | Table 3, col. 4, p. 1678 | beta = -0.280 log points (SE = 0.099)\*\* |
| R5 | Civil enforcement has no significant effect on payments | Table 3, col. 1, p. 1678 | beta = -0.042 (SE = 0.110), n.s. |
| R6 | Civil enforcement has no significant effect on ride count | Table 3, col. 2, p. 1678 | beta = 0.026 (SE = 0.066), n.s. |
| R7 | Prior authorization reduces number of active ambulance firms by 24.9% | Table 6, col. 1, p. 1681 | beta = -0.286 log points (SE = 0.066)\*\*\*, dep. mean = 2.152 |
| R8 | No evidence of adverse patient health effects from prior authorization | Table 4, p. 1680; Table 5, p. 1680 | mortality: 0.000372 (SE 0.000580) n.s.; dialysis sessions: -0.026 (SE 0.019) n.s.; hospitalizations: -0.00132 (SE 0.00136) n.s. |

**Overall (paper's conclusion).** Prior authorization was far more effective than pay-and-chase litigation for reducing Medicare ambulance fraud because it prevents fraudulent payments from being made in the first place, bypassing the twin obstacles of limited defendant liability and low prosecution probability that make realized enforcement ineffective against a large population of small fraudulent firms.

## Theory / model

The paper develops a stylized model in Section VI (pp. 1690-1696) to frame why prior authorization dominated litigation. A firm commits fraud if the gain exceeds expected penalties:

$$G(\text{Reg}) > P_{\text{Crim}} F_{\text{Crim}} + P_{\text{Civ}} F_{\text{Civ}} \tag{5}$$

where $$G(\text{Reg}) = R(\text{Reg}) - C(\text{Reg})$$ is the net gain from fraud (fraudulent Medicare revenue minus operating costs) under regulatory regime Reg, $$P_{\text{Crim}}$$ and $$P_{\text{Civ}}$$ are the perceived probabilities of facing criminal or civil enforcement, and the penalties (p. 1690) are:

$$F_{\text{Civ}} = \min(3R(\text{Reg}),\, \text{Assets}) \quad \text{and} \quad F_{\text{Crim}} = \min(3R(\text{Reg}),\, \text{Assets}) + J$$

reflecting the False Claims Act's treble-damages rule bounded by the firm's assets, plus a jail-disutility term $$J$$ available only under criminal enforcement. Prior authorization sets $$\text{Reg} = 1$$, sharply reducing $$R(1)$$ through claim denials (denial rate jumped from 5.7% to 22.7% in January 2015, Table 10) while modestly raising $$C(1)$$ through paperwork. Litigation operates through $$P_{\text{Crim}}$$ and $$P_{\text{Civ}}$$, but the model identifies two constraints that keep these low in practice. This framework builds on Becker (1968), who established that optimal enforcement balances penalties and detection probability, and extends it to the setting of financial fraud by many small actors against the government.

**Limited liability** (Section VI.A, p. 1691): Shavell (1984) and Polinsky and Shavell (2000) showed that litigation may fail to curtail illicit behavior when severe penalties cannot be enforced. Here, firms can draw down assets before prosecution, so the effectively enforceable penalty $$\min(3R, \text{Assets})$$ is far smaller than the statutory treble damages. Of 27 observed criminal prosecution cases, the government recovered on average less than $1.2 million, or 51% of penalties owed (p. 1691).

**Low probability of detection** (Section VI.B, p. 1692): From 2007 to 2014, only 28 firms faced criminal suits and 44 civil suits while roughly 1,150 firms may have provided fraudulent rides, implying only a 2.4% detection probability for criminal enforcement and 3.8% for civil enforcement. The model implies the expected monetary cost of fraud detection was approximately $72,000 (p. 1694), making fraud profitable for any firm with revenue margins above 1.4%.

Callaway and Sant'Anna (2021) and Behrer et al. (2021) provide the econometric and theoretical benchmarks the paper tests against. The paper also connects to Glaeser and Shleifer (2003), who find that administrative rules optimally complement litigation when courts can be subverted or when specialized enforcement has informational advantages over generalist prosecutors.

## Method

The identification strategy (Section IV, pp. 1673-1675) exploits two sources of staggered quasi-random variation:

1. **Prior authorization rollout**: Medicare implemented prior authorization in December 2014 in New Jersey, South Carolina, and Pennsylvania, then expanded in January 2016 to Delaware, DC, Maryland, North Carolina, Virginia, and West Virginia. This staggered rollout creates treatment and control districts within a DiD framework.

2. **Litigation timing**: The DOJ filed 43 criminal and 26 civil lawsuits against ambulance companies across 26 federal judicial districts at different dates. The paper compares districts in the 48-month window around each action, using not-yet-treated districts as controls.

The primary approach uses TWFE. For district-month outcomes (equations 1 and 2, p. 1673), the event-study specification is:

$$Y_{dt} = \sum_{e=-K}^{-2} \beta_e T_{dt}(e) + \sum_{e=0}^{L} \beta_e T_{dt}(e) + \alpha_d + \alpha_t + \Gamma X_{dt} + \epsilon_{dt} \tag{1}$$

and the scalar post-treatment estimator is:

$$Y_{dt} = \sum_{e=-K}^{-2} \beta_e T_{dt}(e) + \beta \sum_{e=0}^{L} T_{dt}(e) + \alpha_d + \alpha_t + \Gamma X_{dt} + \epsilon_{dt} \tag{2}$$

where $$T_{dt}(e)$$ is an indicator for district $$d$$ being $$e$$ months from its treatment date, $$\alpha_d$$ and $$\alpha_t$$ are district and year-month fixed effects, and $$X_{dt}$$ is a matrix of indicators for prior exposure to a different enforcement type. The estimator sets $$K = 24$$ and $$L = 23$$ (a 48-month symmetric window) to avoid compositional issues flagged by Callaway and Sant'Anna (2021). For untreated districts, $$T_{dt}(e) = 0$$ for all $$e$$. The $$\beta$$ in equation (2) captures the average treatment effect over the first $$L$$ months post-treatment relative to the period immediately before treatment ($$e = -1$$), rather than the entire pre-period.

For patient-month outcomes (equations 3 and 4, p. 1673-1674), individual characteristics $$X_{idt}$$ replace district-level controls, facility fixed effects are added, $$K = 12$$ and $$L = 11$$:

$$Y_{idt} = \sum_{e=-K}^{-2} \beta_e T_{dt}(e) + \sum_{e=0}^{L} \beta_e T_{dt}(e) + \alpha_d + \alpha_t + \Gamma X_{idt} + \epsilon_{idt} \tag{3}$$

$$Y_{idt} = \sum_{e=-K}^{-2} \beta_e T_{dt}(e) + \beta \sum_{e=0}^{L} T_{dt}(e) + \alpha_d + \alpha_t + \Gamma X_{idt} + \epsilon_{idt} \tag{4}$$

Robustness checks in Appendix B use the Borusyak, Jaravel, and Spiess (2017); Cengiz et al. (2019); and Callaway and Sant'Anna (2021) estimators that address staggered-treatment compositional issues, finding similar results. Standard errors are clustered at the district level throughout.

## Empirical specifications

**R1, R2: Prior authorization on payments and rides (Table 2, p. 1676)**

Outcome: log(1 + total payments) or log(1 + total rides) at the district-month level. Treatment: a binary indicator for districts in states subject to prior authorization in the given month. Estimating equation (2) above with $$K = 24, L = 23$$. District and year-month fixed effects; standard errors clustered at the district level. 7,272 district-month observations (2011-2017). Dependent variable means: payments log 9.934 (levels $415,286), rides log 5.357 (levels 2,005 rides/district-month).

The -1.129 log point coefficient (Table 2, col. 1) implies a reduction of 67.7% in payments. The dynamic event-study coefficient path (Figure 3, p. 1677) shows no pre-trends in the 24 months before adoption and an immediate, persistent drop after adoption, with the effect growing to approximately -1.5 log points by month 24.

**R3, R4: Criminal enforcement on payments and rides (Table 3, p. 1678)**

Same specification as R1/R2 but the treatment date is the earliest criminal action filed in the district. Sample: 14,436 district-months (2003-2017). Criminal enforcement reduces ride payments by 0.211 log points (p < 0.10) and ride count by 0.280 log points (p < 0.01). The dynamic event study (Figure 4B, p. 1678) shows gradual rather than immediate effects, consistent with deterrence operating through learning about enforcement.

**R5, R6: Civil enforcement on payments and rides (Table 3, p. 1678)**

Same specification with the earliest civil action as the treatment date. 14,160 district-month observations. Civil enforcement has no statistically significant effect on payments (-0.042, SE = 0.110) or rides (+0.026, SE = 0.066). The dynamic event study (Figure 4A, p. 1678) shows no post-treatment decline.

**R7: Prior authorization on active firms (Table 6, p. 1681)**

Outcome: log(1 + number of firms providing nonemergency dialysis rides) at the district-month level. Estimating equation (2) with year-month and district fixed effects. Sample: 6,336 district-months (2012-2017 for firm identifiers). Prior authorization reduced the number of active ambulance firms by 0.286 log points (24.9%). Firms with a higher pre-period nonemergency dialysis share were most likely to exit; firms providing only nonemergency rides increased in number (93 to 120), indicating market specialization.

**R8: Patient health outcomes (Table 4, p. 1680)**

Outcome: dialysis sessions per month, mortality indicator, all-cause hospitalization indicator, fluid hospitalization indicator. Estimated at the patient-month level with equation (4), $$K = 12, L = 11$$. Controls include patient characteristics (age, sex, race, comorbidities), tenure on dialysis, facility fixed effects and facility characteristics. Sample: 15,077,158 patient-months (2011-2017). All four health outcome coefficients on prior authorization are small and statistically insignificant, ruling out a 0.6% decrease in monthly dialysis sessions at the 95% confidence level. Results hold for frequent riders (Table 5, p. 1680, restricted to patients with at least 100 prior rides).

**General deterrence test (Table 9, p. 1689)**

To test whether enforcement capacity alone (without realized actions) deters fraud via general deterrence, the paper regresses rides and payments on log personnel hours devoted to civil and criminal court cases in each district-year using DOJ National Caseload Data. No statistically significant elasticities are found, ruling out an elasticity of payments with respect to civil enforcement capacity of -0.20 and with respect to criminal capacity of -0.32 at the 95% level.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| USRDS (United States Renal Data System), 100% Medicare ESRD sample | Primary outcome data: 37.5 million nonemergency ambulance rides billed to Medicare for dialysis patients, 2003-2017; patient demographics, comorbidities, dialysis treatment histories, facility identifiers | [no page yet](/wiki/datasets/) |
| DOJ litigation data (hand-collected) | Novel dataset of 69 lawsuits (43 criminal, 26 civil) against ambulance companies for dialysis fraud, 2003-2017; court records from PACER, press releases from DOJ; includes filing dates, jurisdictions, defendant names | no page yet |
| Medicare 20% sample, all beneficiaries | Firm-level outcomes (active firm count, incapacitation effects); 2007-2019; supplements USRDS which began recording firm identifiers only in 2012 | no page yet |
| DOJ National Caseload Data | Enforcement capacity: log personnel hours of US Attorneys' Offices devoted to civil and criminal cases by district-year, 2001-2021; used for general deterrence test (Table 9) | no page yet |
| FOIA responses on financial recoveries | Actual financial recoveries from 27 ambulance fraud prosecutions; obtained via Freedom of Information Act requests to US Attorneys' Offices; average recovery $1.2 million (51% of penalties owed) | no page yet |

Sample scope: US, 2003-2017 (main), 2007-2019 (firm-level). Unit of observation: district-month (main), patient-month (health outcomes), firm-month (incapacitation). Monthly frequency. The dialysis industry context draws on Eliason et al. (2020), who document how acquisitions by large chains affect dialysis facility behavior.

## When to read the full paper

Read this paper if you are studying: (1) the empirical effectiveness of administrative regulation versus ex post litigation for deterring financial fraud, using a setting with clean staggered quasi-random variation in both; (2) Medicare fraud in the dialysis sector, including Table 1 (p. 1671) for patient summary statistics and Figure 1 (p. 1669) for the time series of rides; (3) the identification and robustness literature on staggered DiD, including comparisons with Callaway and Sant'Anna (2021) estimators (Appendix B); or (4) the economic theory of why limited liability and low detection probability undermine pay-and-chase enforcement (Section VI, pp. 1690-1696). The counterfactual savings calculation ($4.8 billion at $28.6 million/year administrative cost) is in Appendix K (p. 1698).

## Attribution and rights

This page is a LLM-distilled summary, not human-verified, and not a reproduction of the research.

> Eliason, P., League, R., Leder-Luis, J., McDevitt, R. C., and Roberts, J. W. (2025). "Ambulance Taxis: The Impact of Regulation and Litigation on Health-Care Fraud." *Journal of Political Economy* 133(5): 1661-1702. [https://doi.org/10.1086/734134](https://doi.org/10.1086/734134)

Paywalled (c) 2025 The University of Chicago. All rights reserved. Published by the University of Chicago Press. An NIH Public Access accepted manuscript is available at [https://pmc.ncbi.nlm.nih.gov/articles/PMC12331087/](https://pmc.ncbi.nlm.nih.gov/articles/PMC12331087/). Replication data and code: Harvard Dataverse, [https://doi.org/10.7910/DVN/QAGBDM](https://doi.org/10.7910/DVN/QAGBDM). Extract-only: do not reproduce tables or figures without permission from the University of Chicago Press.


==============================================================================
# Market Structure, Investment, and Technical Efficiencies in Mobile Telecommunications: Elliott et al. (2024)
# https://instituteforautomatedresearch.org/wiki/papers/jpe/2024/elliott-et-al-market-structure-investment-technical-2024/
# Distilled: A structural model of mobile telecommunications quantifies the trade-off between market power and scale efficiency from consolidation. Applied to the French market, consumer surplus is maximized at eight firms while total surplus peaks at four; all bilateral mergers among France's four operators decrease consumer surplus. Marginal social value of spectrum is approximately five times a firm's auction willingness to pay. Journal of Political Economy 2024, paywalled. Five core results with source locators, the full model, estimation method, and datasets used.
# Tags: paper-summary, industrial-organization, telecommunications, antitrust, market-structure
==============================================================================

**What this is.** The paper's core results, the structural model, and the estimation method are described here with key equations and source locators. To replicate or extend the analysis, read the full source at [doi.org/10.1086/734132](https://doi.org/10.1086/734132).

## TL;DR

Elliott, Houngbonon, Ivaldi, and Scott develop a structural model of competition among mobile network operators (MNOs) in which firms simultaneously choose prices and infrastructure investment. The novel supply-side feature is an engineering-based model of data transmission: download speeds depend on spectrum allocation (via Hata path loss and Shannon-Hartley channel capacity) and network congestion (via M/M/1 queuing). Fewer firms mean each firm serves a denser user base, reducing path loss, and pools more spectrum and customers in one queue, reducing congestion waste -- two scale efficiencies that offset the market-power cost. This tension is what Williamson (1968) posed theoretically for antitrust; the paper quantifies it in a calibrated equilibrium.

Estimating the demand system on French mobile data for October 2015 and simulating symmetric-firm counterfactuals, the paper finds that consumer surplus is maximized at eight firms while total surplus peaks at four. Monopoly delivers faster download speeds than a six-firm market, a reversal of intuition from models without scale economies that Spence (1975) and standard monopoly quality theory do not predict. All bilateral mergers among France's four MNOs decrease consumer surplus even accounting for spectrum-pooling efficiencies, because the market-power effect dominates when infrastructure is fixed.

## Core results

| # | Result | Locator | Magnitude as reported |
|---|---|---|---|
| R1 | Download speeds are monotonically decreasing in the number of firms; economies of pooling and density generate higher quality under concentration | Figure 7, p. 40 | At 1 firm delivered speeds reach ~90 Mbps; at 6 firms speeds collapse near 0 Mbps; channel capacity changes little across firm counts at the commune level |
| R2 | Consumer surplus is maximized at 8 firms; total surplus is maximized at 4 firms; the two peaks diverge because quality declines and producer surplus also falls with more firms | Figure 9, p. 43 | CS peaks around 29.5 EUR/person/month above monopoly; TS peaks around 14.1 EUR/person/month above monopoly at 4 firms; producer surplus declines monotonically in firm count |
| R3 | All 6 bilateral MNO mergers decrease consumer surplus in the short run when infrastructure is held fixed | Figure 12, p. 49 | Reductions range from -0.22 EUR/person/month (Orange-Bouygues) to -1.24 EUR/person/month (Orange-SFR); merger-specific spectrum pooling is insufficient to offset market-power effects |
| R4 | The marginal social value of spectrum is approximately five times a firm's auction willingness to pay at the four-firm equilibrium | Figure 11, p. 46 | Marginal consumer surplus from bandwidth = 1.24 EUR/person/MHz; firm marginal WTP = 0.25 EUR/person/MHz; the 2015 French 700 MHz auction price was ~0.70 EUR/person/MHz, between the two |
| R5 | Asymmetric spectrum allocations are inefficient: symmetric allocation raises both consumer and total surplus relative to asymmetric, though asymmetric raises producer surplus | Table 6, p. 44 | 3-firm symmetric CS = -0.668 EUR/person; 3-firm asymmetric CS = -0.701 EUR/person (both relative to 4-firm symmetric benchmark); asymmetric TS = -0.076, symmetric TS = -0.054 EUR/person |

**Overall finding.** Consolidation in mobile telecommunications presents a genuine trade-off between market power and scale efficiency. The welfare-maximizing market structure depends critically on the welfare criterion: antitrust policy targeting consumer surplus implies a more competitive market than policy targeting total surplus. All bilateral mergers among France's 2015 MNOs decrease consumer surplus, and the social return to spectrum is substantially above what firms reveal through auction bids.

## Theory and model

The paper builds a two-stage model: (i) a discrete-continuous consumer demand system and (ii) an engineering-based industry model of data transmission and firm competition.

**Demand.** Consumer $i$'s indirect utility for phone plan $j$ in municipality $m$, at chosen data consumption $x$, is (equation 1, p. 15):

$$u_{jm}(x;\, \vartheta_i,\varepsilon_{ij},\theta_{pi}) = w_j(x,Q_{f(j),m},\vartheta_i) + \theta_v v_j - \theta_{pi} P_j + \xi_{jm} + \varepsilon_{ij} \tag{1}$$

where $w_j(\cdot)$ is the utility of data consumption, $\theta_v$ is the voice-quality valuation, $\theta_{pi}$ is a price sensitivity parameter that decreases in household income, $\xi_{jm}$ is an unobserved product-market demand shock, and $\varepsilon_{ij}$ follows a nested logit structure (two nests: outside good vs. all mobile plans). The data consumption utility is (equation 2, p. 16):

$$w_j(x,Q,\vartheta_i) = \vartheta_i \log(1+x) - c_j(x,Q) \tag{2}$$

where $\vartheta_i \sim \mathrm{Exp}(\theta_{di})$ is the consumer's data valuation shock and $c_j(x,Q)$ is the time cost of downloading $x$ GB at speed $Q$. Below the data limit $\bar{d}_j$ the cost is linear in $x/Q$; above $\bar{d}_j$ the throttled speed $Q^L = 128$ Kbps applies (equation 3, p. 17). Optimal data consumption is pinned down by the first-order condition, which yields a closed-form for $x^*(P_j,Q,\vartheta_i,v_j)$.

**Download speed from engineering.** Channel capacity (maximum average speed absent congestion) is derived from hexagonal cell geometry and the Hata path loss model (equation 14, p. 26):

$$\bar{Q}_{fm}(R_{fm},B_{fm}) = \frac{B_{fm}\, A(R_{fm})}{\displaystyle\int_{\ell\in\mathcal{L}(R_{fm})} q_{m\ell}^{-1}\,d\ell} \tag{14}$$

where $R_{fm}$ is the cell radius (the firm's infrastructure choice), $B_{fm}$ is spectrum bandwidth, $A(R_{fm}) = \frac{3\sqrt{3}}{2}R_{fm}^2$ is the hexagonal cell area, and $q_{m\ell} = \gamma_m \log_2(1+SINR_\ell)$ is the location-specific data rate in bits per second per Hz via the Shannon-Hartley theorem. The signal-to-noise-and-interference ratio $SINR_\ell$ follows the Hata model with estimated path loss exponent 3.522.

Actual delivered speeds also depend on congestion. Following an M/M/1 queuing model (equation 17, p. 28):

$$Q_{fm} = \bar{Q}_{fm} - Q_{fm}^D \tag{17}$$

where $Q_{fm}^D$ is the Poisson arrival rate of download requests from consumers on operator $f$'s network in municipality $m$. This yields two sources of scale efficiency from consolidation: economies of density (fewer firms means each serves a denser population per cell, shortening the average distance and reducing path loss) and economies of pooling (combined spectrum and customer base in a single queue cuts queuing waste).

**Firm competition.** Each operator unilaterally and simultaneously sets a national price vector $\mathbf{P}_f$ and a vector of cell radii $\mathbf{R}_f$ (one per municipality) to maximize profits (equation 24, p. 34):

$$\Pi_f(\mathbf{P},\mathbf{R},\mathbf{B}) = \sum_m \Pi_{mf}(\mathbf{P},\mathbf{R}_m,\mathbf{B}_m) - \sum_m C_{fm}(R_{fm},B_{fm}) \tag{24}$$

with infrastructure cost proportional to both cell area covered and bandwidth operated per base station (equation 22, p. 33):

$$C_{fm}(R_{fm},B_{fm}) = c_{fm}^s\,\frac{A_m}{A(R_{fm})}\,B_{fm} \tag{22}$$

The Nash equilibrium prices $\mathbf{P}^*$ and cell radii $\mathbf{R}^*$ satisfy first-order conditions for both price and infrastructure simultaneously; the paper solves these jointly using numerical methods across 589 municipalities.

Counterfactual equilibria are computed on a single representative commune (population-weighted mean density 2,792 people/km$^2$) with $n$ symmetric firms each holding $B_0/n$ MHz of spectrum (Section 6.1, pp. 40-43). Merger counterfactuals add a second and third representative commune (to capture all population density categories) and combine the merging MNOs' bandwidth allocations while keeping infrastructure fixed in the short run (Section 6.3, pp. 47-49).

## Method

**Demand estimation.** The paper follows Berry, Levinsohn and Pakes (1995) extended to a discrete-continuous demand system. Market shares observed at two levels of aggregation (plan-level for Orange, firm-level for other MNOs) require a modified contraction mapping in which the inner loop solves for $\boldsymbol{\xi}(\theta)$ that simultaneously matches both data sources (Appendix B, equations B.1-B.4). The outer loop searches over demand parameters $\theta = (\theta_{pi},\theta_{di},\theta_c,\theta_v,\sigma)$ minimizing a two-stage efficient GMM objective.

Price variation in the October 2015 cross-section is insufficient to separately identify demand elasticities (prices were stable 2013-2015). The calibration approach from Bourreau, Sun and Verboven (2021) is used: the own-price elasticity of Orange (-2.36) and the diversion ratio to the outside good (0.036) are imposed as additional moment conditions (equations 11-12, p. 23), anchoring the demand system. The income heterogeneity parameter $\theta_{dz}$ is identified from the covariance of municipality median income with average data consumption. Download speed is instrumented by log population density to correct for measurement error and congestion endogeneity.

**Cost estimation.** Variable costs per user $c_j^u$ are recovered from the pricing first-order condition (equation 26, p. 35) evaluated at observed equilibrium prices and estimated demand. Infrastructure costs per base station per MHz $c_{fm}^s$ are recovered from the infrastructure first-order condition (equation 27, p. 35), obtained by numerically differentiating the marginal operating income with respect to cell radius $R_{fm}$ and equating it to the marginal cost implied by equation 22.

**Identification summary.** Demand parameters: identified by price variation across plans, consumption-income covariance (income heterogeneity), download speed variation across municipalities (speed coefficient), and calibrated elasticity and diversion moments. Supply cost parameters: identified from firms' optimal pricing and investment decisions, exploiting cross-municipality variation in population density and land area.

## Empirical specifications

**Demand GMM moments.** Nine moment conditions (p. 23): the price elasticity calibration (equation 11), the diversion ratio calibration (equation 12), mean data consumption across markets, the covariance of Orange demand shocks with median income, the covariance of Orange demand shocks with log population density, the demand-shock covariance with data limit indicators, the demand-shock covariance with voice plan indicator, and two cross-plan mean-consumption moments.

**Sample.** Cross-section of 589 French communes with population above 10,000 (covering 43.5% of France's total population). Data period: October 2015 for customer shares and consumption; Q2 2016 for Ookla speed measurements. The commune is the unit of observation for the demand estimation; the representative commune is the unit for counterfactual simulations.

**Counterfactual design.** Symmetric counterfactuals: $n \in \{1, 2, 3, 4, 5, 6, 7, 8\}$ firms each with $B_0/n$ MHz. Asymmetric counterfactual (3 firms): replicates 2011 French spectrum allocation (ANFR frequencies). Merger counterfactuals: 6 pairs from France's 4 MNOs (Orange, SFR, Bouygues, Free); three representative communes weighted by France's population distribution within each density class.

## Datasets used

| Dataset | Role | Access | Wiki page |
|---|---|---|---|
| Orange Mobile customer database (proprietary) | Plan-level market shares and average data consumption per municipality, October 2015 | proprietary-confidential | No page yet |
| Ookla Speedtest data (proprietary) | Over 1 million measured download speeds per operator per municipality, France Q2 2016 | proprietary-confidential | No page yet |
| ANFR base station database | Locations of all mobile antennas and frequencies per operator per municipality | public | No page yet |
| GSMA Intelligence | National-level market shares for SFR, Bouygues Telecom, Free Mobile | licensed-commercial | No page yet |
| INSEE 2011 population census | Income deciles per municipality; municipality land area | public | No page yet |
| OSIRIS (Orange internal traffic data) | Total data traffic per network cell, used to calibrate Poisson demand arrival rates for the queuing model | proprietary-confidential | No page yet |

The most restrictive source (Orange subscriber database, OSIRIS, Ookla) makes this a `proprietary-confidential` paper that cannot be replicated without access to those proprietary datasets. The estimation code at https://github.com/jonathantelliott/mobile-telecommunications is fully available.

## When to read the full paper

Use the [original](https://doi.org/10.1086/734132) if you are: building a structural model of network infrastructure competition; quantifying welfare trade-offs in mobile market consolidation or spectrum auction design; studying how engineering-based scale economies interact with product market competition; or applying a discrete-continuous demand system to a market with heterogeneous quality. The online Internet Appendix contains technical derivations of the Hata path loss formula, the modified BLP contraction mapping, the network-sharing deviation from equilibrium, and robustness checks for alternative cost specifications.

## Attribution and rights

Source: peer-reviewed, *Journal of Political Economy* 133(5), May 2025, pp. 1401-1459. Paywalled; no open-access license found in Crossref metadata. This distillation was extracted by an LLM on 2026-06-26 and is **not human-verified or independently reproduced**. Only the content of this page may be used; the underlying article requires a library subscription or individual purchase.

> Elliott, Jonathan T., Georges V. Houngbonon, Marc Ivaldi, and Paul T. Scott. "Market Structure, Investment, and Technical Efficiencies in Mobile Telecommunications." *Journal of Political Economy* 133, no. 5 (May 2025): 1401-1459. https://doi.org/10.1086/734132.


==============================================================================
# Theory of Fiscal Responsibility and Irresponsibility: Halac & Yared (2024)
# https://instituteforautomatedresearch.org/wiki/papers/jpe/2024/halac-yared-theory-fiscal-responsibility-irresponsibility-2024/
# Distilled: A political economy model in which successive deficit-biased governments facing private i.i.d. fiscal shocks endogenously cycle between a fiscally responsible regime (maximally enforced deficit limit) and a fiscally irresponsible regime (maximally enforced surplus limit), with transitions triggered by extreme shocks and only when governments' bias is large enough. Journal of Political Economy 133(5), May 2025, paywalled. Six core results with source locators, the full model, equilibrium programs, and the factorization algorithm.
# Tags: paper-summary, macro, fiscal-policy, political-economy, public-finance
==============================================================================

**What this is.** The paper's core propositions, the equilibrium model, and the factorization method with the defining equations: enough to know what it found and how, without reading all 47 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1086/734131).

## TL;DR

The paper presents a political economy model of fiscal policy in which successive deficit-biased governments, each privately observing an i.i.d. shock to the social value of deficit-financed spending, interact dynamically. The best equilibrium for society is characterized by exactly two regimes: a fiscally responsible regime (a maximally enforced deficit limit that restrains overborrowing) and a fiscally irresponsible regime (a maximally enforced surplus limit that pushes governments to overborrow even more as punishment). Transitions between regimes are triggered by high enough fiscal shocks, and fiscal policy is therefore history-dependent even though shocks are i.i.d. Fiscal regimes arise only if governments' deficit bias is sufficiently large, providing a theoretical foundation for the empirical pattern that fiscal consolidations cluster at crisis periods. Related dynamic models by Battaglini and Coate (2008) and Yared (2010) feature observable shocks and Markov outcomes; private information and limited commitment together are what generate regime history-dependence here.

## Core results

| # | Result | Locator | Magnitude as reported |
|---|---|---|---|
| R1 | **Two-Regime Bang-Bang**: in the best equilibrium with deficit bias and private information, continuation values jump to either the highest or lowest feasible value, inducing exactly two fiscal regimes | Proposition 1, p. 1588 | $$V_{t+1}(h^{t-1}, b_t) \in \{\underline{V}(b_t),\bar{V}(b_t)\}$$ at every on-path history; no interior continuation values in the best equilibrium |
| R2 | **Fiscally Responsible Regime = Deficit Limit**: the welfare-maximizing regime is a maximally enforced deficit limit; low-shock governments borrow at their flexible level, mid-shock governments are constrained at the limit, high-shock governments break the limit and trigger a transition to irresponsibility | Proposition 2, p. 1590; Figure 2 (left panel), p. 1591 | Three-threshold structure parameterized by $$\theta^*$$ and $$\theta^{**}$$: borrowing is at the limit $$b'(\omega,\theta^*)$$ for $$\theta \in [\theta^*,\theta^{**}]$$ and breaks the limit for $$\theta > \theta^{**}$$ |
| R3 | **Fiscally Irresponsible Regime = Surplus Limit**: the welfare-minimizing punishment regime is a maximally enforced surplus limit that induces all types to overborrow relative to their own preferred level; high-shock governments borrow flexibly, mid-shock governments are constrained, low-shock governments break the limit and return to responsibility; regime is always temporary | Proposition 3, p. 1592; Figure 2 (right panel), p. 1591 | Surplus limit threshold satisfies $$\theta_n^{**} < \bar{\theta}$$, ensuring that government types $$\theta \in [\theta_n^{**}, \bar{\theta}]$$ respect the surplus limit with positive probability, so the fiscally irresponsible regime is never absorbing |
| R4 | **Regime Transitions at Crises**: transitions from fiscal responsibility to irresponsibility occur when a shock exceeds $$\theta^{**}$$; transitions from irresponsibility to responsibility occur when a shock is at least $$\theta_n^{**}$$; fiscal policy depends on history, not just current conditions | Section IV.E, pp. 1594-1595 | Starting in the responsible regime: $$\theta_t \leq \theta^{**}$$ keeps the economy responsible at $$t+1$$; $$\theta_t > \theta^{**}$$ triggers irresponsibility. Starting in the irresponsible regime: $$\theta_t \geq \theta_n^{**}$$ triggers return to responsibility |
| R5 | **Deficit Bias Threshold for Fiscal Regimes** (analytical example, log utility): fiscal regimes arise if and only if governments' deficit bias $$\alpha$$ exceeds a threshold $$\bar{\alpha}$$ and the discount factor $$\delta$$ exceeds a threshold $$\tilde{\delta}$$; the Markov equilibrium is unique when bias is small | Proposition 4, p. 1597; Corollary 1, p. 1598 | Threshold $$\bar{\alpha} \in (1,\infty)$$ is decreasing in $$\delta$$; fiscal regimes exist iff $$T'(0) > 1$$, which holds iff $$\alpha > \bar{\alpha}$$; transitions occur on path iff $$\alpha \in (\bar{\alpha}, \tilde{\alpha})$$ for some $$\tilde{\alpha} > \bar{\alpha}$$ |
| R6 | **Numerical Simulation** (calibrated to US data, 1970-2020): simulated fiscal path shows two extended shaded periods of fiscal irresponsibility; best-equilibrium spending is constrained below the flexible Markov level during responsible periods and exceeds the first-best during irresponsible periods | Figure 4, p. 1602 | Parameters: $$\alpha = 1.151$$, $$\delta = 0.943$$, lognormal shocks (mean 0, $$\sigma = 0.175$$), $$R = 1.05$$. Deficit limit threshold = 0.0736 (flexible spending rate of $$\theta^*$$); surplus limit threshold = 0.0867 (flexible spending rate of $$\theta_n^*$$) |

**Overall (paper's conclusion).** The same deficit bias that leads governments to overaccumulate debt is also the key factor behind the emergence of fiscal regimes: when bias is large enough and the discount factor is high enough, the threat of fiscal irresponsibility in the future makes fiscal responsibility in the present sustainable. Regime transitions are punctuated by crises, and fiscal policy depends on the history of past decisions, consistent with econometric evidence of two-regime fiscal dynamics documented for the United States and the European Union.

## Theory / model

The model is an infinite-horizon small open economy with periods $$t = \{0, 1, \ldots\}$$ in which a new government takes power each period (§II.A, p. 1581). At the beginning of period t, an i.i.d. shock to the social value of spending is drawn from:

$$\theta_t \in \Theta \equiv [\underline{\theta}, \bar{\theta}], \quad f(\theta) > 0, \quad F(\cdot) \text{ its CDF}$$

The shock is privately observed by the government in power at date t (its "type"). The **government budget constraint** (eq. 1, p. 1581) is:

$$g_t = \tau - Rb_{t-1} + b_t \tag{1}$$

where $$\tau > 0$$ is exogenous tax revenue, $$R > 1$$ is the exogenous gross interest rate on government bonds, $$b_t \geq 0$$ is new borrowing, and $$g_t \geq 0$$ is government spending. Letting $$\omega(h^{t-1}) \equiv \tau - Rb_{t-1}(h^{t-1})$$ denote available resources given inherited debt, spending satisfies $$g_t = \omega + b_t$$.

**Social welfare** at date t (p. 1581):

$$V_t = \mathbb{E}_t\!\left[\sum_{k=0}^{\infty} \delta^k \theta_{t+k} U(g_{t+k})\right] = \mathbb{E}_t\!\left[\theta_t U(g_t) + \delta V_{t+1}\right]$$

where $$\delta \in (0,1)$$ is the social discount factor and $$U(\cdot)$$ is strictly increasing and strictly concave. A large shock $$\theta_t$$ represents a high social value of deficit-financed spending, as in an economic crisis.

**Government welfare** at date t (eq. 2, p. 1582):

$$W_t = \alpha \theta_t U(g_t) + \delta V_{t+1} \tag{2}$$

where $$\alpha > 1$$ is the **deficit bias**: the government overvalues current utility of spending relative to society by a factor $$\alpha$$, capturing private benefits from directing resources to preferred spending categories or constituencies (Aguiar and Amador 2011). Crucially, the government at date t shares society's preferences from date $$t+1$$ onward.

There are three frictions: (i) deficit bias ($$\alpha > 1$$), (ii) private information ($$\theta_t$$ observed only by the period-t government), and (iii) limited commitment (no external enforcement; only future equilibrium behavior can reward or punish). Lemma 1 (p. 1586) shows that either deficit bias or private information alone is insufficient to generate history dependence: both frictions together are necessary.

**Equilibrium** (§II.B, p. 1582-1583). A strategy for the period-t government is $$\sigma_t(h^{t-1}, \theta_t)$$: a feasible debt level for each public history $$h^{t-1} = \{b_{-1}, b_0, \ldots, b_{t-1}\}$$ and type $$\theta_t$$. Two incentive constraints must hold for all histories:

*Private information constraint* (eq. 3, p. 1583): for all $$\theta, \theta' \in \Theta$$,

$$\alpha\theta U(\omega + b(\theta)) + \delta V(b(\theta)) \geq \alpha\theta U(\omega + b(\theta')) + \delta V(b(\theta')) \tag{3}$$

*Limited commitment constraint* (eq. 4, p. 1583): for all $$\theta \in \Theta$$ and all $$b'_t$$ not prescribed for any type,

$$\alpha\theta U(\omega + b(\theta)) + \delta V(b(\theta)) \geq \alpha\theta U(\omega + b'_t) + \delta V(b'_t) \tag{4}$$

Given debt is bounded and shocks i.i.d., there exist highest and lowest continuation values $$\bar{V}(b)$$ and $$\underline{V}(b)$$ achievable in equilibrium from debt level b. The government's **flexible borrowing** conditional on the worst punishment is (eq. 5, p. 1584):

$$b^p(\omega,\theta) \in \arg\max_{b \in [\underline{b},\bar{b}]} \{\alpha\theta U(\omega + b) + \delta \underline{V}(b)\}$$

A necessary condition for the limited commitment constraint (4) to hold is (eq. 6, p. 1584):

$$\alpha\theta U(\omega + b(\theta)) + \delta V(b(\theta)) \geq \alpha\theta U(\omega + b^p(\omega,\theta)) + \delta \underline{V}(b^p(\omega,\theta)) \tag{6}$$

Constraints (3) and (6) are necessary and sufficient for a debt sequence to be supported by equilibrium strategies.

**Best equilibrium recursive program** ($$\mathcal{P}_{\max}$$, §II.C, p. 1584-1585):

$$\bar{V}(b_{-1}) = \max_{b(\theta),\, V(b(\theta))} \mathbb{E}[\theta U(\omega + b(\theta)) + \delta V(b(\theta))] \tag{$\mathcal{P}_{\max}$}$$

subject to private information (7), limited commitment (8), and feasibility (9): $$b(\theta) \in [\underline{b}(b_{-1}), \bar{b}(b_{-1})]$$ and $$V(b(\theta)) \in [\underline{V}(b(\theta)), \bar{V}(b(\theta))]$$ for all $$\theta \in \Theta$$. The **worst-case program** $$\mathcal{P}_{\min}$$ minimizes welfare subject to the same constraints, yielding $$\underline{V}(b_{-1})$$. Assumption 1 (p. 1585) ensures these value functions are continuously differentiable and concave with $$\bar{V}(b) > \underline{V}(b)$$; this holds for CARA and CRRA preferences.

**Virtual welfare rewrite** (§IV.A, p. 1587). Following Amador, Werning, and Angeletos (2006), substituting the envelope condition of the private-information constraint into the objective of $$\mathcal{P}_{\max}$$ yields the virtual welfare representation (eq. 11, p. 1587):

$$\alpha\underline{\theta} U(\omega + b(\underline{\theta})) + \delta V(b(\underline{\theta})) + \alpha \int_{\underline{\theta}}^{\bar{\theta}} U(\omega + b(\theta)) Q(\theta)\, d\theta \tag{11}$$

where the **virtual welfare weight** is:

$$Q(\theta) \equiv 1 - F(\theta) - \theta f(\theta)\!\left(1 - \tfrac{1}{\alpha}\right)$$

The first term $$1 - F(\theta)$$ is the standard virtual surplus from mechanism design (Myerson 1981); the second term $$-\theta f(\theta)(1 - 1/\alpha)$$ captures society's cost of prescribing higher borrowing when governments are biased. Assumption 2 (p. 1589) requires $$Q'(\theta) < 0$$ for $$\theta < \hat{\theta}$$ and $$Q'(\theta) > 0$$ for $$\theta > \hat{\theta}$$, which holds for uniform, exponential, lognormal, gamma, and beta distributions over a range of parameters.

## Method

**Recursive representation** (§II.C, p. 1584). Because shocks are i.i.d. and debt is bounded, the best equilibrium can be characterized via a static per-period optimization ($$\mathcal{P}_{\max}$$ above) by assigning each type $$\theta$$ a debt level $$b(\theta)$$ and a continuation value $$V(b(\theta))$$ drawn from the feasible set $$[\underline{V}(b),\bar{V}(b)]$$. The per-period problem embeds the infinite-horizon structure entirely through the continuation value set. This recursive representation uses the `promised-utility-recursion` technique generalized to the Markovian adverse-selection setting.

**Bang-bang structure** (Proposition 1, p. 1588). Perturbation arguments on $$\mathcal{P}_{\max}$$ show that interior continuation values are suboptimal: for any solution with interior $$V(b(\theta))$$, there is a perturbation that either compresses borrowing (when $$Q'(\theta) < 0$$) or steepens it (when $$Q'(\theta) > 0$$), strictly increasing social welfare. Hence the optimal continuation values take only the two extreme values $$\{\underline{V}(b),\bar{V}(b)\}$$. The proof uses the `mechanism-design` and `principal-agent` perturbation technique from Halac and Yared (2022), extended to the dynamic limited-commitment setting.

**Maximally enforced limits** (Propositions 2-3, pp. 1590, 1592). Given the bang-bang property, the best equilibrium prescribes a threshold structure. For the fiscally responsible regime ($$\mathcal{P}_{\max}$$), the solution satisfies (eq. 12, p. 1590):

$$\bigl(b(\theta), V(b(\theta))\bigr) = \begin{cases} \bigl(b'(\omega,\theta),\, \bar{V}(b'(\omega,\theta))\bigr) & \text{if } \theta < \theta^*, \\ \bigl(b'(\omega,\theta^*),\, \bar{V}(b'(\omega,\theta^*))\bigr) & \text{if } \theta \in [\theta^*,\theta^{**}], \\ \bigl(b^p(\omega,\theta),\, \underline{V}(b^p(\omega,\theta))\bigr) & \text{if } \theta > \theta^{**}, \end{cases} \tag{12}$$

where $$b'(\omega,\theta) \equiv \arg\max_{b} \{\alpha\theta U(\omega+b) + \delta\bar{V}(b)\}$$ is the flexible borrowing conditional on the highest continuation value. The limited commitment constraint binds with equality at the threshold type $$\theta^{**}$$ (eq. 13, p. 1590):

$$\alpha\theta^{**} U(\omega + b'(\omega,\theta^*)) + \delta\bar{V}(b'(\omega,\theta^*)) = \alpha\theta^{**} U(\omega + b^p(\omega,\theta^{**})) + \delta\underline{V}(b^p(\omega,\theta^{**})) \tag{13}$$

The fiscally irresponsible regime ($$\mathcal{P}_{\min}$$, Proposition 3) takes the mirror form: a maximally enforced surplus limit at $$b'(\omega,\theta_n^*)$$, with high types $$\theta > \theta_n^*$$ borrowing at their flexible level, mid-types constrained at the surplus limit, and low types $$\theta < \theta_n^{**}$$ breaking the limit and returning to the responsible regime.

**Factorization algorithm** (§V.C, p. 1599-1600). In the analytical example with $$U(\cdot) = \log(\cdot)$$, the gap between highest and lowest value functions is constant in debt: $$\bar{V}(b) - \underline{V}(b) = P^*$$ for some $$P^* \geq 0$$ (eq. 17, p. 1596). The equilibrium with fiscal regimes corresponds to the largest fixed point of the operator $$T(P)$$, which computes the largest self-enforceable punishment today given a punishment of size P available tomorrow (eq. 21, p. 1599):

$$T(P) = \max_{\theta^*, \theta^{**}, \theta_n^*, \theta_n^{**}} \left\{ \delta P + \alpha \!\left[ \int_{\theta^*}^{\theta^{**}} \!\bigl(U(g^f(\theta^*)) - U(g^f(\theta))\bigr) Q(\theta)\, d\theta - \int_{\theta_n^{**}}^{\theta_n^*} \!\bigl(U(g^f(\theta_n^*)) - U(g^f(\theta))\bigr) Q(\theta)\, d\theta \right] \right\} \tag{21}$$

subject to limited commitment binding at the deficit and surplus thresholds (eqs. 22-23, p. 1599):

$$\delta P \geq \alpha \int_{\theta^*}^{\theta^{**}} \bigl[U(g^f(\theta)) - U(g^f(\theta^*))\bigr]\, d\theta \tag{22}$$

$$\delta P \geq \alpha \int_{\theta_n^{**}}^{\theta_n^*} \bigl[U(g^f(\theta_n^*)) - U(g^f(\theta))\bigr]\, d\theta \tag{23}$$

where $$g^f(\theta) \equiv b'(\omega,\theta) + \omega$$ is the flexible spending level. The operator $$T$$ is increasing and concave with $$T(0) = 0$$ and $$\lim_{P\to\infty} T'(P) < 1$$. A positive fixed point $$P^* > 0$$ with $$T(P^*) = P^*$$ exists if and only if $$T'(0) > 1$$, which is the condition for fiscal regimes. This algorithm is analogous to Abreu, Pearce, and Stacchetti (1990) but applies from below (starting at the Markov outcome $$P = 0$$) rather than from above; this difference is key for finding a condition for the fixed point to exceed the Markov outcome (p. 1600-1601).

## Empirical specifications

The paper's main results are theoretical (no empirical estimation). Section V.D (p. 1601-1603) presents a **numerical simulation** to illustrate regime dynamics.

**Parameter calibration.** Parameters are chosen so that the mean and variance of the flexible spending rate match the mean and variance of US government spending over 1970-2020, using data on federal debt, receipts, and outlays from the Federal Reserve Bank of St. Louis (FRED, cited p. 1601):
- Lognormal shock distribution: mean 0, variance $$\sigma = 0.175$$, truncated to support $$[\underline{\theta}, \bar{\theta}] = [0.01, 100.01]$$
- Social discount factor: $$\delta = 0.943$$
- Gross interest rate: $$R = 1.05$$
- Deficit bias: $$\alpha = 1.151$$, calibrated so that $$1/\alpha \approx 0.87$$ corresponds to a reelection probability implying average government duration of 7.6 years, matching the average time the same party held the US presidency from 1944 to 2020 (footnote 26, p. 1601)

**Simulation results.** The factorization algorithm, applied to these parameters, yields a unique best equilibrium with fiscal regimes. The implied thresholds are:
- Deficit limit threshold: flexible spending rate of type $$\theta^* = 0.0736$$
- Surplus limit threshold: flexible spending rate of type $$\theta_n^* = 0.0867$$

Figure 4 (p. 1602) plots the simulated best-equilibrium spending rate alongside the first-best and flexible (Markov) rates. Two extended shaded periods (fiscal irresponsibility) are visible. During fiscally responsible periods, the best-equilibrium rate coincides with the flexible rate when the latter is below the deficit limit, and is constrained at the threshold when the flexible rate slightly exceeds it; for high enough shocks the government breaks the limit and the shaded irresponsible period begins. During irresponsible periods, the best-equilibrium rate coincides with the flexible rate when above the surplus limit, and is constrained at the threshold otherwise; sufficiently low shocks break the surplus limit, ending the irresponsible episode.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| FRED (Federal Reserve Bank of St. Louis): US federal debt, receipts, outlays | Calibration targets for flexible spending rate mean and variance (1970-2020, annual) | [FRED](/wiki/datasets/fred/) |

Sample: United States, 1970-2020, annual frequency, for calibration only. The paper's main propositions are theoretical and hold for any economy satisfying Assumptions 1 and 2.

## When to read the full paper

Use the [original](https://doi.org/10.1086/734131) if you are:
working on dynamic mechanism design with adverse selection and limited commitment (Appendices A-C contain the formal perturbation proofs for Propositions 1-3);
applying or extending the factorization algorithm to other games with one-dimensional state and adverse-selection incentive constraints (§V.C);
studying the theoretical foundations for observed fiscal consolidation patterns and the role of political biases in driving debt accumulation;
or connecting the model to quantitative fiscal analysis (the paper's conclusion §VI identifies key extensions: persistent shocks, richer fiscal instruments, and multiple-period government bias).

## Attribution and rights

Source: peer-reviewed, *Journal of Political Economy* 133(5), May 2025, pp. 1574-1620. Copyright 2025 The University of Chicago. All rights reserved. This distillation was extracted by an LLM on 2026-06-26 and is **not human-verified or independently reproduced**. The paper is paywalled; the PDF is not hosted here. Extract only.

> Halac, Marina, and Pierre Yared. "A Theory of Fiscal Responsibility and Irresponsibility."
> *Journal of Political Economy* 133, no. 5 (May 2025): 1574-1620.
> DOI: 10.1086/734131.


==============================================================================
# Laws and Norms: Bénabou & Tirole (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jpe/2025/benabou-tirole-laws-norms-2025/
# Distilled: A unified theory of how intrinsic motivation, material incentives, and social norms jointly shape compliance and optimal public policy. Derives modified Pigou-Ramsey taxation correcting for reputational rents, and characterizes when the expressive content of law makes incentives softer or tougher than the symmetric-information optimum. Journal of Political Economy 2025, paywalled. Eight core results with proposition locators, the model equations, and the signaling-equilibrium analysis.
# Tags: paper-summary, social-norms, public-goods, optimal-taxation, expressive-law, signaling, behavioral-economics, peer-reviewed, unreplicated
==============================================================================

**What this is.** The paper's core propositions, the image-concern model with its equations, and the signaling-game analysis of expressive law: enough to understand what it proved and how, without reading all 42 pages. To extend or replicate the formal results, read the original at [doi.org/10.1086/738343](https://doi.org/10.1086/738343).

## TL;DR

Bénabou and Tirole (2006a) introduced the image-concern model; this paper completes that program by deriving optimal taxation and the expressive role of law. A continuum of agents with heterogeneous intrinsic motivation v choose a prosocial action, earn reputational returns from peers who infer their type from their choice (honor if they act, stigma if they do not), and respond to material incentives y set by a principal. Under symmetric information, the optimal first-best incentive departs from the standard Pigouvian subsidy by subtracting the reputational rent that the marginal contributor extracts, and is hump-shaped in society's overall prosociality and in the compliance cost: norms substitute for incentives at both extremes of compliance. Under asymmetric information, where the principal knows the social environment θ but agents do not, the law becomes expressive. The principal uses her choice of y to signal θ, leading to softer law when she wants to signal norm strength, and tougher law when she signals the magnitude of the social externality. Extensions cover societies' resistance to economists' prescriptions ("commodification"), zero-tolerance policies, broken-windows theory, norm-based interventions, and the avoidance of cruel punishments.

## Core results

All locators refer to the version with DOI 10.1086/738343 (HAL preprint hal-05577272, 35 pp.).

| # | Result | Locator | Content as stated |
|---|---|---|---|
| R1 | First-best optimal incentive subtracts the reputational rent from the standard Pigouvian subsidy | Prop. 1(i), eq. (14), p. 14 | $$y^{FB}_\theta = \epsilon_\theta - \Delta_\theta\!\left(\tfrac{c_\theta - \epsilon_\theta}{e_\theta}\right)$$; the second term is the reputational rent extracted by the marginal contributor |
| R2 | Second-best optimal incentive is always strictly below the first-best and decreases with the shadow cost of funds | Prop. 1(ii), eq. (15), p. 14 | $$y^{SI}_\theta < y^{FB}_\theta$$ for all λ > 0; prosocial behavior is always underprovided; social multiplier amplifies but does not fully replace y |
| R3 | Optimal incentive is hump-shaped (bell-shaped) in the overall prosociality of society θ and in the compliance cost c | Prop. 2, Fig. 2, pp. 14-15 | When f_θ is strictly unimodal, y^{FB}_θ is single-peaked at θ₀ = (c − ε)/e; high prosociality (respectable act) and low prosociality (admirable act) both reduce the optimal incentive relative to the modal case |
| R4 | Soft law results when the principal's private information (M⁺, P⁻) or (M⁻, P⁺); tough law when (M⁺, P⁺) or (M⁻, P⁻) | Prop. 5, Fig. 3, pp. 20-21 | y^{AI}_θ < y^{SI}_θ on the off-diagonal of Table 1; y^{AI}_θ > y^{SI}_θ on the diagonal; always underprovision of prosocial behavior |
| R5 | A separating equilibrium (expressive law) exists when θ shifts societal values with a norm, operates a right truncation, affects the externality ε, or affects compliance cost under an anti-norm | Prop. 7, p. 23 | SOC₁ and SOC₂ satisfied strictly; existence proved in Online Appendix for cases (a)-(d) |
| R6 | Full pooling is the equilibrium outcome when θ indexes social monitoring intensity or performs a left-truncation; no separating equilibrium when θ is a distributional-shift parameter under an anti-norm or a cost parameter under a norm | Prop. 8-9, p. 23 | Prop. 8: A = B = 0 when θ = μ (social monitoring) or θ is a left-truncation parameter; full pooling preferred. Prop. 9: A < 0 when θ is a distributional-shift parameter with anti-norm (Δ' > 0) or a cost parameter with norm (Δ' < 0); no separating equilibrium exists |
| R7 | A principal with private information about θ selectively discloses: reveals good news, withholds bad; more disclosure when the probability of obtaining information is higher | Prop. 10, p. 24 | Disclosure iff θ ≥ θ̃ (under M⁺) or θ ≤ θ̃ (under M⁻); threshold θ̃ is decreasing in the probability q of observing θ |
| R8 | Commodification spillovers make soft law optimal when strong incentives on a formally-controlled activity would signal low prosociality and erode the norm in a non-incentivized activity | Prop. 11, p. 27 | For λ small enough, the high-type principal sets y^{AI}_{θ_H} < y^{SI}_{θ_H}; the least-cost separating equilibrium is D1-robust |

**Overall.** Optimal policy corrects the Pigouvian subsidy in two directions: subtract the reputational rent (norms already motivate, so over-incentivizing is wasteful and crowds out esteem), and use the signal sent by incentive choice itself to harness agents' intrinsic motivation and image concerns. The expressive content of law is soft when signaling social norms and tough when signaling social costs.

## Theory / model

The core model (Section II, pp. 6-9) has a continuum of agents of mass 1, each choosing a ∈ {0,1}. Choosing a = 1 costs c_θ, creates externality ε_θ, and earns a material incentive y from the principal. Agent types v are private information, distributed with continuous density f_θ(v) > 0 on V_θ = [v^min_θ, v^max_θ]. Intrinsic motivation for compliance is e_θ ≡ γε_θ + 1 − γ, capturing both consequentialist concern (weight γ on externalities) and warm glow. The utility function (eq. 1, p. 6) is:

$$
U = (ve_\theta - c_\theta + y)\,a + \epsilon_\theta\,\bar{a}_\theta + \mu_\theta\!\left(E_\theta[\tilde{v} \mid a, y] - \bar{v}_\theta\right). \tag{1}
$$

Here ā_θ is aggregate participation, μ_θ is the weight on reputational concerns, and $$E_\theta[\tilde{v} \mid a, y]$$ is others' posterior mean belief about the agent's type. The third term captures image concerns: the agent values being perceived as high-v by peers (or, via self-signaling, by himself). Reputation is a positional good.

The two conditional moments that generate honor and stigma (eqs. 2-3, pp. 7-8) are:

$$
E^+_\theta(v) = E_\theta[\tilde{v} \mid \tilde{v} \geq v], \quad E^-_\theta(v) = E_\theta[\tilde{v} \mid \tilde{v} < v], \tag{2}
$$

$$
\Delta_\theta(v^*) \equiv \mu_\theta\!\left[E^+_\theta(v^*) - E^-_\theta(v^*)\right]. \tag{3}
$$

$$E^+_\theta(v^*)$$ is the honor conferred on those who contribute when v* is the cutoff, and $$E^-_\theta(v^*)$$ is the stigma borne by abstainers. An agent chooses a = 1 iff $$ve_\theta \geq c_\theta - y - \Delta_\theta(v^*)$$, so the equilibrium cutoff $$v^*_\theta(y)$$ satisfies (eq. 4, p. 8):

$$
v^*_\theta(y)\,e_\theta - c_\theta + y + \Delta_\theta(v^*_\theta(y)) = 0. \tag{4}
$$

Lemma 1 (attributed to Jewitt; Harbaugh and Rasmusen; Adriani and Sonderegger; p. 9) characterizes Δ_θ: when f_θ is unimodal, Δ_θ is strictly quasi-convex. The equilibrium then exhibits a norm (strategic complements, Δ'_θ < 0) for respectable behaviors and an anti-norm (strategic substitutes, Δ'_θ > 0) for admirable, rare behaviors. Multiple equilibria can arise when complementarity is strong; uniqueness is ensured by $$e_\theta + \Delta'_\theta(v) > 0$$ for all v. The social multiplier (eq. 6, p. 8),

$$
s_\theta(y) \equiv -\frac{\partial v^*_\theta}{\partial y} = \frac{1}{e_\theta + \Delta'_\theta(v^*_\theta(y))}, \tag{6}
$$

amplifies the compliance response to a unit increase in y through the reputational feedback. It exceeds $$1/e_\theta$$ for respectable (norm-driven) behaviors and falls below $$1/e_\theta$$ for admirable ones.

Under asymmetric information (Sections IV-VI), θ ∈ [θ₁, θ₂] is privately known to the principal. Agents infer θ from y and form long-run reputations. The informational multiplier (Section IV.B, pp. 18-19) captures the additional channel: the policy signals θ, shifting agents' beliefs about ε, c, or the distribution of values, and hence their intrinsic motivation and reputational incentives. The principal's objective under asymmetric information is (eq. 18, p. 19):

$$
W^{AI}_\theta(y) = \int_{v^*_{\hat\theta(y)}(y)}^{+\infty} \!\left[ve_\theta + \epsilon_\theta - c_\theta - \lambda y\right] f_\theta(v)\,dv, \tag{18}
$$

where $$\hat\theta(y)$$ is agents' belief about θ on the equilibrium path.

Building on Bénabou and Tirole (2006a) and Bénabou and Tirole (2003), the paper also extends the framework in Section V to direct communication: a principal who can selectively disclose hard information about θ will reveal good news and conceal bad news (Proposition 10), with credibility limited by the sender's incentives. Besley and Ghatak (2005) provide related background on motivated-agent settings that informs the optimal-policy analysis.

## Method

The paper derives optimal policy in two settings and establishes the existence of separating equilibria for expressive law. It builds on `signaling-game-pbe`, `principal-agent`, and `mechanism-design`.

**Symmetric information: modified Pigou-Ramsey (Section III).** The principal maximizes social welfare $$W^{SI}_\theta(y)$$ subject to eq. (4). The first-order condition (eq. 12, p. 13) equates the net social marginal benefit to the deadweight loss from paying all inframarginal agents:

$$
\frac{\epsilon_\theta + v^*_\theta(y)\,e_\theta - c_\theta - \lambda y}{e_\theta + \Delta'_\theta(v^*_\theta(y))} = \frac{\lambda}{h_\theta(v^*_\theta(y))}, \tag{12}
$$

where $$h_\theta(v) = f_\theta(v)/[1-F_\theta(v)]$$ is the monotone hazard rate. The first-best formula (eq. 14, p. 14) subtracts from the standard Pigouvian subsidy ε_θ the reputational rent $$\Delta_\theta((c_\theta - \epsilon_\theta)/e_\theta)$$ that the marginal contributor extracts. The second-best (eq. 15) further discounts for fiscal cost.

**Asymmetric information: the expressive-law signaling problem (Section IV).** In a separating equilibrium, the principal of type θ chooses $$y^{AI}_\theta$$ and agents invert y to learn θ exactly. The first-order condition (eq. 19, p. 19) adds an informational multiplier to eq. (12):

$$
\left(\frac{\epsilon_\theta + v^*_\theta(y)\,e_\theta - c_\theta - \lambda y}{e_\theta + \Delta'_{\hat\theta(y)}(v^*_{\hat\theta(y)}(y))}\right)\!\!\left(1 + \left(v^*_{\hat\theta}\gamma\frac{\partial\epsilon_\theta}{\partial\theta} - \frac{\partial c_\theta}{\partial\theta} + \frac{\partial\Delta_\theta}{\partial\theta}(v^*_{\hat\theta(y)})\right)\hat\theta'(y)\right) = \frac{\lambda}{h_\theta(v^*_{\hat\theta(y)}(y))}. \tag{19}
$$

The second bracket is the informational multiplier: $$\hat\theta'(y) = 1/(y^{AI}_\theta)'$$ is the inverse slope of the separating schedule; the term in parentheses captures how a belief shift about θ changes motivation (via M⁺ or M⁻) and reputational pressure (via P⁺ or P⁻). When these signs align (diagonal of Table 1), the multiplier exceeds 1 and calls for tougher law; when they oppose (off-diagonal), it falls below 1 and calls for softer law.

The paper extends Mailath (1987)'s classic analysis to non-monotone payoffs. The key second-order condition for a separating equilibrium (Proposition 6, SOC₁, p. 22) is:

$$
\mathcal{A}(\theta, \hat\theta) \equiv y'(\hat\theta)\,b(\theta,\hat\theta,y(\hat\theta))\,\frac{\partial\!\left[b(\theta,\hat\theta,y(\hat\theta))\,h_\theta(v^*_{\hat\theta}(y(\hat\theta)))\right]}{\partial\theta} \geq 0, \tag{SOC_1}
$$

where $$b(\theta,\hat\theta,y)$$ is the social benefit of a marginal contribution. A ≥ 0 ensures no principal type wants to mimic another. Proposition 7 (p. 23) then identifies the four cases where SOC₁ and a complementary SOC₂ hold strictly, establishing existence of a separating equilibrium. Propositions 8-9 (p. 23) characterize knife-edge (full pooling) and impossible (no separating equilibrium) cases.

## Empirical specifications

This paper contains no empirical analysis of its own. All results are propositions with formal proofs in the Online Appendix. Section II.E (pp. 11-12) surveys empirical applications by other researchers that test the model's comparative-statics predictions: Besley, Jensen and Persson (2023) use eq. (4) to study tax evasion in local British and Welsh councils 1980-2009 and document persistence of social-multiplier effects; Jia and Persson (2021) exploit Chinese affirmative-action policy changes to test predictions on ethnic-identity choice; Chen (2016) studies deterrence effects of WWI executions on Irish vs British soldiers to test the social-multiplier comparative static.

## Datasets used

This paper uses no datasets. All results are mathematical propositions.

| Dataset | Role in paper | Wiki page |
|---|---|---|
| (none) | Theory paper | none |

## When to read the full paper

Read the original at [doi.org/10.1086/738343](https://doi.org/10.1086/738343) if you are: (1) working through the formal proofs, all of which are in the Online Appendix; (2) studying the extension sections (Section VI on spillovers, zero-tolerance policies, broken-windows theory, and cruel punishments; the Online Appendix on reciprocity, conformity, and status); (3) applying the framework to specific policies, where Table 1 (p. 16) and Figure 2 (p. 14) summarize the comparative statics; or (4) building on the expressive-law signaling equilibrium, where Propositions 6-9 and their appendix proofs are the required inputs.

## Attribution and rights

Source: peer-reviewed, *Journal of Political Economy* 134(2), pp. 731-772. This distillation was extracted by an LLM on 2026-06-26 and is **not human-verified or independently reproduced**. Paywalled; extract-only. An accepted-manuscript version is available on HAL at [hal.science/hal-05577272v1](https://hal.science/hal-05577272v1).

> Bénabou, Roland, and Jean Tirole. "Laws and Norms." *Journal of Political Economy* 134, no. 2 (2025): 731-772. DOI: 10.1086/738343.


==============================================================================
# Parenting with Patience: Del Boca, Flinn, Verriest & Wiswall (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jpe/2025/boca-et-al-parenting-patience-parental-incentives-2025/
# Distilled: A Markov Perfect Equilibrium model of joint parent-child cognitive skill investment estimates that Conditional Cash Transfers reduce child patience by 13-17% and that intrinsic-motivation crowding-out is the primary reason parents limit their use. Journal of Political Economy 134(1), 2026, paywalled. Seven core results with source locators, the parent-child dynamic game (utility, skill production, CCT design, discount factor transition), the Method of Simulated Moments estimator, and three datasets (PSID-CDS, Steinberg et al. 2009, Osaka PPS).
# Tags: paper-summary, child-development, parenting, household-economics, non-cognitive-skills, time-preferences, structural, paywalled, peer-reviewed, unreplicated
==============================================================================

**What this is.** The paper's core results, the dynamic model of parent-child interaction (utility functions, cognitive skill production, CCT design, discount factor dynamics), and the Method of Simulated Moments estimation: enough to understand what the paper found and how, without reading all 76 pages. To replicate or extend the model, read the full source at [10.1086/738481](https://doi.org/10.1086/738481) and use the replication package at [Harvard Dataverse](https://doi.org/10.7910/DVN/F7QVQ5).

## TL;DR

The paper builds a Markov Perfect Equilibrium model in which parents and children jointly determine cognitive skill formation over childhood (ages 3-17). Parents choose how to allocate their own time, expenditure, and whether to use a Conditional Cash Transfer (CCT) that links child consumption to study time; the child simultaneously chooses self-investment time given parental decisions. The novel feature is that the child's discount factor (patience) is endogenous: it evolves stochastically with age but is stochastically reduced by CCT use, capturing the intrinsic-motivation crowding-out effect documented by Deci, Koestner, and Ryan (1999). Estimated by the Method of Simulated Moments on PSID-CDS household data, cross-national discount factor data from Steinberg et al. (2009), and adult patience data from the Osaka Preference Parameter Survey, the model finds: CCTs reduce child patience by 13-17%; the primary deterrent to CCT use is this crowding-out cost rather than the direct disutility; maternal time inputs are most productive in early childhood while child self-investment time dominates by adolescence; and SES gaps in child outcomes are primarily explained by differences in parental time productivity and initial discount factor distributions, not by income or wage differences.

## Core results

Magnitudes are as reported in the paper. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | CCT use stochastically reduces child patience: expected discount factor falls 13% at age 11 | Sect. 5.3, p. 60 | From 0.493 (no CCT) to 0.427 (CCT use) |
| R2 | CCT use reduces child patience by 17% at age 17 | Sect. 5.3, p. 60 | From 0.503 (no CCT) to 0.418 (CCT use) |
| R3 | Removing CCT access raises final patience to 0.88 but reduces final cognitive skills by 37% of a SD | Table 13, col. 1, p. 69 | Cognitive skills fall 0.17 log-units (37% of SD); patience rises from 0.81 to 0.88 at age 17 |
| R4 | Removing the patience crowding-out channel causes CCT use to jump to 90% and cognitive skills to rise 85% of a SD | Table 13, col. 3, p. 70 | Cognitive skills rise 0.38 log-units (85% of SD); patience rises to 0.88 |
| R5 | Maternal time is most productive in early childhood | Sect. 5.2, p. 58 | 1 SD more maternal time at age 6 raises Letter Word score 9-11% of a SD |
| R6 | Child self-investment time surpasses parental time in productivity by adolescence | Sect. 5.2, p. 58 | 1 SD more study time at age 15 raises Letter Word score more than 5% of a SD |
| R7 | SES gaps in cognitive skills are driven mainly by parental time productivity and discount factor heterogeneity, not wages or income | Table 14, col. 4, p. 73 | Homogenizing productivity and discount factors closes 84% of simulated high/low-SES cognitive-skills gap and 79% of patience gap |

**Overall (paper's conclusion).** Parents rationally limit CCT use because CCTs stochastically reduce child patience, not primarily because of the direct utility cost. The model unifies cognitive and non-cognitive skill formation: study incentives boost cognitive outcomes but erode patience, creating a tradeoff that explains the declining use of CCTs with child age and the lower CCT use among college-educated parents whose children are already more patient. SES gaps in child outcomes are primarily rooted in heterogeneous parental time productivity and patience distributions, suggesting that income or wage redistribution alone would close little of the gap.

## Theory / model

The model covers ages $$ t = t_0, t_0+1, \ldots, 17 $$ with the terminal period at $$ M+1 $$. State variables at each age are the current cognitive skill stock $$ k_t $$, the child's current discount factor $$ \beta_{c,t} $$, and a set of parental characteristics (wages, education, non-labor income) collected in $$ \Gamma_t $$.

**Preferences.** The child's instantaneous utility over leisure $$ l_{c,t} $$, private consumption $$ x_t $$, and cognitive skill $$ k_t $$ is (p. 11):

$$
u_{c,t} = \lambda_1 \ln l_{c,t} + \lambda_2 \ln x_t + \lambda_3 \ln k_t
$$

Parents are altruistic toward the child. Their composite utility (Eq. 1, p. 12), combining own consumption and altruistic terms over child leisure, child consumption, and skill, is:

$$
\tilde{u}_{p,t} = \tilde{\alpha}_1 \ln l_{1,t} + \tilde{\alpha}_2 \ln l_{2,t} + \tilde{\alpha}_3 \ln c_t + \tilde{\alpha}_4 \ln k_t + \tilde{\alpha}_5 \ln l_{c,t} + \tilde{\alpha}_6 \ln x_t \tag{1}
$$

where $$ l_{1,t}, l_{2,t} $$ are mother's and father's leisure, $$ c_t $$ is parental consumption, and the $$ \tilde{\alpha} $$ parameters embed a convex combination of parental own preferences and child preferences weighted by the altruism parameter $$ \varphi $$. When the parent uses a CCT, a random fixed utility cost $$ \zeta $$ drawn from an exponential distribution is deducted (Eq. 6, p. 19):

$$
\tilde{u}_{p,t}(\mathbf{a}_{p,t}) = \tilde{\alpha}_1 \ln l_{1,t} + \tilde{\alpha}_2 \ln l_{2,t} + \tilde{\alpha}_3 \ln c_t + \tilde{\alpha}_4 \ln k_t + \tilde{\alpha}_5 \ln l_{c,t} + \tilde{\alpha}_6 \ln x_t - \zeta \cdot \mathbf{1}[CCT_t = 1] \tag{6}
$$

**Terminal values.** At period $$ M+1 $$ the model closes with an infinite-horizon continuation. The child's terminal value (Eq. 2, p. 13) is:

$$
V_{c,M+1}(k_{M+1}, \beta_{c,M+1}) = \frac{\lambda_3 \ln k_{M+1}}{1 - \beta_{c,M+1}} \tag{2}
$$

reflecting that with patience $$ \beta_{c,M+1} $$ the child will continue to value the skill stock at rate $$ \lambda_3 / (1 - \beta_{c,M+1}) $$ into the adult perpetuity. The parent's terminal value (Eq. 3, p. 14) combines the parent's own long-run discount factor with the imputed value of the child's terminal stock, weighting both $$ (1-\varphi)\alpha_4 / (1 - \beta_p) $$ (parent's own valuation of $$ k_{M+1} $$) and the altruistic share $$ \varphi \lambda_3 / (1 - \beta_{c,M+1}) $$.

**Cognitive skill production.** Skill evolves via a Cobb-Douglas log-linear production function following Cunha, Heckman, and Schennach (2010) (Eq. 4, p. 14):

$$
\ln k_{t+1} = \ln R_t + \delta_{1,t} \ln \tau_{1,t} + \delta_{2,t} \ln \tau_{2,t} + \delta_{3,t} \ln \tau_{c,t} + \delta_{4,t} \ln e_t + \delta_{5,t} \ln k_t \tag{4}
$$

where $$ \tau_{1,t}, \tau_{2,t} $$ are mother's and father's time with the child, $$ \tau_{c,t} $$ is child self-investment time, $$ e_t $$ is monetary expenditure on the child, $$ k_t $$ is the lagged skill stock, $$ R_t $$ is age-specific total factor productivity, and the $$ \delta $$ parameters are age-varying input elasticities. The persistence parameter $$ \delta_{5,t} $$ ("skills beget skills") is estimated at 0.79 in early childhood and rises to 0.84 by age 16.

**CCT incentive contract.** When a parent uses a CCT, child consumption is linked to study time (Eq. 5, p. 17):

$$
\ln x_t(\tau_{c,t};\, r_t, b_t) = b_t + r_t \ln \tau_{c,t} \tag{5}
$$

with floor parameter $$ b_t $$ (base consumption) and slope $$ r_t > 0 $$ (study-time elasticity of consumption). A higher $$ r_t $$ provides stronger incentives to study. The parent jointly chooses $$ CCT_t \in \{0,1\} $$ and, when $$ CCT_t = 1 $$, the contract parameters $$ (r_t, b_t) $$.

**Endogenous discount factor.** The child's discount factor evolves via a Markov chain (Eq. 7, p. 22) with $$ Z $$ discrete support points:

$$
\Pr(\beta_{c,t+1,h} = \beta_c^{j'} \mid \beta_{c,t,h} = \beta_c^j,\, t,\, CCT_{t,h}) \quad \forall\, (j,j') = 1,\ldots,Z \tag{7}
$$

The transition probabilities depend on current patience, age, and whether a CCT was used. The key restriction, identified from the Steinberg et al. and Osaka PPS data, is that CCT use stochastically shifts probability mass toward lower patience states: the estimated CCT-on transition matrix puts more weight on low-$$ \beta_c $$ values than the CCT-off matrix.

**Equilibrium.** The game is a Markov Perfect Equilibrium in the tradition of Del Boca, Flinn, and Wiswall (2014) and Doepke and Zilibotti (2017). Each period the parent announces actions $$ \mathbf{a}_{p,t} = (\tau_{1,t}, \tau_{2,t}, e_t, CCT_t, r_t, b_t) $$ before the child chooses $$ \tau_{c,t} $$. The child's Bellman equation is (p. 24):

$$
V_{c,t}(\Gamma_t \mid \mathbf{a}_{p,t}) = \max_{\tau_{c,t}} \Bigl[ u_c(l_{c,t}, x_t, k_t) + \beta_{c,t}\, \mathbb{E}_t V_{c,t+1}(\Gamma_{t+1} \mid \tau_{c,t}, \mathbf{a}_{p,t}, \Gamma_t) \Bigr]
$$

The parent's Bellman equation is (p. 25):

$$
V_{p,t}(\Gamma_t) = \max_{\mathbf{a}_{p,t}} \Bigl[ \tilde{u}_p(\mathbf{a}_{p,t}) + \beta_p\, \mathbb{E}_t V_{p,t+1}(\Gamma_{t+1} \mid \mathbf{a}_{p,t}, \Gamma_t) \Bigr]
$$

Under the Cobb-Douglas structure the child's optimal study time has a closed-form linear reaction function: without a CCT it is proportional to available non-parental time $$ (\tilde{T}_t - \tau_{p,t}^0) $$ at a state-dependent rate $$ \gamma_t^0(\Gamma_t) $$; with a CCT the proportionality constant $$ \gamma_t^1(r_t, \Gamma_t) $$ also depends on the CCT elasticity $$ r_t $$ (Eqs. 8 and 10, pp. 26-27). Parental CCT choice then follows from comparing $$ V_{p,t}(\Gamma_t \mid \mathbf{a}_{p,t}^0) $$ and $$ V_{p,t}(\Gamma_t \mid \mathbf{a}_{p,t}^1) $$ at each state.

## Method

**Solution.** The model is solved by backward induction (value function iteration starting at $$ t = M+1 $$ and rolling back to $$ t = t_0 $$). At each age-state grid node, the child's reaction function and the parent's optimality conditions are solved in closed form given the known next-period value function, then rolled back. The key analytic result (Proposition in Section 2.4, pp. 28-29) is that in equilibrium, child study time is a fixed fraction of remaining non-parental time, which makes the parent's optimization tractable. The `value-function-iteration` approach uses a discretized state space for $$ k_t $$ and $$ \beta_{c,t} $$ (integrating out the Markov transition at each step).

**Estimation.** Parameters are estimated by the Method of Simulated Moments. Let $$ \boldsymbol{\theta} $$ be the vector of structural parameters. Given a draw of household heterogeneity, the model simulates histories of all endogenous variables; simulated moments $$ \hat{m}(\boldsymbol{\theta}) $$ are matched to empirical counterparts $$ m_{\text{data}} $$ by minimizing:

$$
\hat{\boldsymbol{\theta}} = \arg\min_{\boldsymbol{\theta}} \bigl(\hat{m}(\boldsymbol{\theta}) - m_{\text{data}}\bigr)' \hat{W} \bigl(\hat{m}(\boldsymbol{\theta}) - m_{\text{data}}\bigr)
$$

where $$ \hat{W} $$ is a diagonal weighting matrix. Standard errors are computed from the Jacobian of simulated moments; 1,000 simulated households are used per parameter evaluation. The `method-of-simulated-moments` estimator builds on the `principal-agent` parent-child game framework and the `life-cycle-model` specification of parental wages and non-labor income.

## Empirical specifications

**Cognitive skill production function.** The age-specific input elasticities $$ \boldsymbol{\delta}_t $$ and persistence $$ \phi_t $$ are first estimated from a reduced-form log-linear specification (Eq. 12, p. 49):

$$
\ln k_{h,t+1} = \ln R_t + \mathbf{Z}_{h,t} \boldsymbol{\delta}_t + \phi_t \ln k_{h,t} + \varepsilon_{h,t} \tag{12}
$$

where $$ \mathbf{Z}_{h,t} $$ is the vector of log time and expenditure inputs for household $$ h $$ at age $$ t $$. These first-stage estimates are used as calibrated starting values and production-function moments in the structural estimation.

**Structural moment conditions.** The MSM matches 211 moments from three datasets:

- PSID-CDS moments: age-conditional means and correlations of parental time inputs, child study time, CCT use (conditional-allowance indicator), test scores (Woodcock-Johnson Letter Word and Applied Problems), wages, household income, and CCT use by parental education group.
- Steinberg et al. (2009) moments: age-conditional means and variances of elicited child discount factors (ages 10-30), and correlation of discount factors with IQ scores across sites.
- Osaka PPS moments: mean and variance of adult discount factors by age group (ages 25-65), and education-conditional means.

The CCT-patience crowding-out parameters (the off-diagonal elements of the CCT-on relative to the CCT-off discount factor transition matrix) are identified from the conditional covariance between CCT use and changes in the simulated patience distribution, cross-validated against the Steinberg et al. and Osaka PPS moments.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| PSID-CDS (Panel Study of Income Dynamics, Child Development Supplement) | Primary structural estimation: parental time inputs (Childhood Activity Study modules), child study time, CCT use (conditional-allowance questions), Woodcock-Johnson Letter Word and Applied Problems test scores, household income, wages, demographics | No page yet |
| Steinberg et al. (2009) experimental data | Discount factor moments: age profile of patience from 935 individuals ages 10-30 across 11 study sites; pins down $$ \beta_{c,t} $$ age dynamics and CCT crowding-out parameters | No page yet |
| Osaka Preference Parameter Survey (Osaka PPS) | Adult discount factor moments: mean and variance for 4,625 adults ages 25-65; anchors the terminal patience distribution used in the structural model | No page yet |

Sample: 247 PSID-CDS households, three waves (1997, 2002, 2007), children ages 3-16. Replication code and processed data: Del Boca, Flinn, Verriest, and Wiswall (2025), [Harvard Dataverse (10.7910/DVN/F7QVQ5)](https://doi.org/10.7910/DVN/F7QVQ5).

## When to read the full paper

Read the source at [doi.org/10.1086/738481](https://doi.org/10.1086/738481) if you are:

- Building or extending structural models of child development that treat children as active players with endogenous time preferences; Section 2.4 derives the closed-form equilibrium reaction functions.
- Studying parenting-style economics (CCT vs. unconditional transfers) and need the MPE solution method and identification argument.
- Calibrating age-varying skill production elasticities from PSID-CDS; Table 12 (p. 57) lists all input-elasticity estimates by age group.
- Running SES-heterogeneity counterfactuals: Table 14 (pp. 72-73) decomposes the simulated high/low-SES gap into wage, productivity, time-preference, and initial-condition channels.
- Extending the Doepke and Zilibotti (2017) or Cunha, Heckman, and Schennach (2010) frameworks to a game-theoretic setting with endogenous patience.

The comparative statics of CCT cost and crowding-out parameters (Table 13, pp. 69-70) and the SES decomposition (Table 14) are the headline policy-relevant outputs; the formal game solution and MSM algorithm are in Appendices A-C.

## Attribution and rights

Source: peer-reviewed, *Journal of Political Economy* 134(1), January 2026. This distillation was extracted by an LLM on 2026-06-26 and is **not human-verified or independently reproduced**. The VoR is paywalled (University of Chicago Press); this page is extract-only.

> Del Boca, Daniela, Christopher Flinn, Ewout Verriest, and Matthew Wiswall. "Parenting with Patience: Parental Incentives and Child Development." *Journal of Political Economy* 134, no. 1 (January 2026): 210-284. DOI: 10.1086/738481.
> Replication data: Del Boca, Flinn, Verriest and Wiswall (2025), Harvard Dataverse, https://doi.org/10.7910/DVN/F7QVQ5.


==============================================================================
# Selecting Penalty Parameters: Chetverikov & Sørensen (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jpe/2025/chetverikov-rensen-selecting-penalty-parameters-high-2025/
# Distilled: Chetverikov and Sørensen (2025) propose bootstrapping after cross-validation (BCV), a method for selecting the penalty parameter of l1-penalized M-estimators in high dimensions that yields valid l1 and l2 error bounds; post-BCV is the only method in simulations whose studentized estimates converge to N(0,1), and an empirical illustration confirms Fryer Jr (2019) findings on racial differences in police use of force are robust to model choice and expanded controls. J. Polit. Econ. 2025, paywalled. Seven core results with source locators, the M-estimation framework, and the BCV algorithm with its defining equations.
# Tags: paper-summary, machine-learning, lasso, penalized-regression, inference, peer-reviewed, unreplicated, data:ppcs
==============================================================================

**What this is.** The paper's core results, the M-estimation framework it builds on, and the bootstrap-after-cross-validation (BCV) algorithm with its defining equations: enough to understand what is proposed and what is shown, without reading all 41 pages. To replicate or extend, read the original at [doi.org/10.1086/736770](https://doi.org/10.1086/736770).

## TL;DR

Chetverikov and Sørensen (2025) develop bootstrapping after cross-validation (BCV): a practical, broadly applicable method for choosing the penalty parameter λ of ℓ₁-penalized M-estimators when p >> n. Unlike cross-validation (CV), BCV yields both ℓ₁ and ℓ₂ estimation error bounds at minimax-optimal rates, works for any convex loss including the probit (non-Lipschitz), and in simulations is the only approach whose three-step debiased estimator consistently approximates N(0,1). The key idea: run K-fold CV to obtain out-of-fold residual estimates, then apply a Gaussian multiplier bootstrap on those residuals to estimate the quantile of the score process and set λ accordingly. Post-BCV (refitting non-zero coefficients without penalty after BCV) performs best for inference. An empirical illustration revisits Fryer Jr (2019) on racial differences in police use of force, confirming his logit findings under probit loss and with 327 interaction controls (versus 30 in the original); the debiased t-statistics for the Black dummy exceed 18 in all four specifications.

## Core results

Magnitudes as reported; simulation results average over 2,000 draws; η_n = √(ln(pn)/n).

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **BCV convergence rates match the minimax-optimal LASSO rate** in the high-dimensional M-estimation setting | Theorem 4.1, eq. (4.14), p. 18 | ‖θ̂ − θ₀‖₂ ≲_P √(s_q η_n^{2-q}), ‖θ̂ − θ₀‖₁ ≲_P s_q η_n^{1-q}; exact sparsity (q = 0): ℓ₂ rate √(s₀ ln(pn)/n), ℓ₁ rate s₀ √(ln(pn)/n) |
| R2 | **Post-BCV debiased estimator is asymptotically normal** for any scalar component β₀ under stated regularity conditions | Theorem 5.1, eq. (5.6), p. 24 | √n(β̂ − β₀)/σ₀ →d N(0,1); asymptotic variance σ₀² is consistently estimable via a plug-in formula |
| R3 | **Post-CV refitting fails to converge in ~15% of cases; post-BCV fails in ~0.01%** across all simulation designs | §6.3.1, p. 29 | Post-CV non-existence rate ~15% (up to 47% in some DGPs); post-BCV non-existence rate ~0.01% (74 of 540,000 draws); post-CV estimators are therefore dropped from further comparison |
| R4 | **Post-BCV studentized densities converge to N(0,1) as n grows; BCV and CV densities do not** across all sparsity patterns | Figure 6.4, pp. 33–34 | Post-BCV densities collapse toward N(0,1) at n = p = 200, 400 for exactly sparse, intermediate, and approximately sparse patterns; BCV and CV show persistent leftward shrinkage bias that does not vanish as n increases |
| R5 | **Basic controls: Fryer Jr (2019) confirmed; post-BCV t-statistics similar to unpenalized ML** | Table 1, p. 36 | Post-BCV logit t = 10.5, probit t = 9.6; unpenalized ML logit t = 8.8, probit t = 8.7 (n = 9,930; p = 30) |
| R6 | **Full control set (327 regressors): post-BCV t-stats 20.7 (logit) and 18.9 (probit); unpenalized ML does not exist** | Table 1, p. 36 | Post-BCV logit t = 20.7, probit t = 18.9 (n = 9,930; p = 327); unpenalized ML infeasible due to complete separation in the data |
| R7 | **Average partial effect of Black race roughly doubles when interactions are added** | Table 2, p. 37 | Post-BCV logit APE = 3.2 pp, probit APE = 2.8 pp (with interactions); basic controls only: 1.1–1.4 pp (both unpenalized ML and post-BCV); unconditional force rate for white civilians: 0.7 percent |

**Overall (paper's conclusion).** BCV yields minimax-optimal ℓ₁ and ℓ₂ estimation rates for high-dimensional M-estimators and supports valid debiased inference via Neyman orthogonality. Post-BCV dominates plain BCV and CV in inference quality (substantially better size control) across all simulation designs, at near-zero additional computational cost versus plain cross-validation. The empirical illustration shows that Fryer Jr (2019)'s conclusion about racial differences in police use of force is robust to switching from logit to probit, adding interaction controls, and applying ℓ₁-penalization.

## Theory / model

The paper considers any model in which the true parameter θ₀ ∈ Θ ⊆ ℝ^p solves a population optimization problem (eq. 1.1, p. 2):

$$
\boldsymbol{\theta}_0 = \operatorname*{argmin}_{\boldsymbol{\theta} \in \Theta} \mathbb{E}\!\left[m\!\left(\boldsymbol{X}^\top \boldsymbol{\theta},\, \boldsymbol{Y}\right)\right] \tag{1.1}
$$

where $$m : \mathbb{R} \times \mathcal{Y} \to \mathbb{R}$$ is a known (potentially non-smooth) loss function convex in its first argument, $$\boldsymbol{X} = (X_1,\ldots,X_p)^\top \in \mathbb{R}^p$$ is the regressor vector, and $$\boldsymbol{Y} \in \mathcal{Y}$$ is the outcome. Examples covered (§2, pp. 7–9): the logit loss $$m(t,y) = \ln(1 + e^t) - yt$$, the probit loss $$m(t,y) = -y \ln \Phi(t) - (1-y)\ln(1-\Phi(t))$$, the ordered response loss, the expectile (asymmetric least squares) loss, and the trimmed LAD/LS loss for censored panel models. The framework deliberately includes non-Lipschitz losses (e.g., probit), ruling out the self-normalized moderate deviation (SNMD) method of Belloni et al. (2012) and Belloni et al. (2016).

The true parameter is assumed to lie in an ℓ_q-ball of radius $$s_q^{1/q}$$ for some $$q \in [0,1]$$ (Assumption 3.6, p. 11): $$\sum_{j=1}^p |\theta_{0,j}|^q \leq s_q$$. The case $$q = 0$$ is exact sparsity with at most $$s_0$$ non-zeros; $$q > 0$$ allows approximately sparse coefficients.

The key theoretical object motivating the penalty choice is the score at the truth (eq. 3.3, p. 11):

$$
\boldsymbol{S}_n := \mathbb{E}_n\!\left[m'_1\!\left(\boldsymbol{X}_i^\top \boldsymbol{\theta}_0,\, Y_i\right) \boldsymbol{X}_i\right] \tag{3.3}
$$

Theorem 3.1 (p. 12) shows that for the ℓ₁-penalized M-estimator (ℓ₁-ME) to achieve good ℓ₁ and ℓ₂ estimation error, λ must dominate $$c_0 \|\boldsymbol{S}_n\|_\infty$$ with high probability; this requires estimating the $$(1-\alpha)$$-quantile $$q_n(1-\alpha)$$ of $$\|\boldsymbol{S}_n\|_\infty$$. The bound itself builds on and extends the penalty bound derived for ℓ₁-quantile regression in Belloni and Chernozhukov (2011a) to the general M-estimation setting of equation (1.2).

The paper has no formal economic model: the framework is a statistical model for estimation and inference. The identification logic is that θ₀ is identified by the population FOC $$\nabla \mathcal{E}(\boldsymbol{\theta}_0) = \boldsymbol{0}$$ under a quadratic margin condition (Assumption 3.4), ensuring $$\mathcal{E}(\boldsymbol{\theta}) \geq c_M \|\boldsymbol{\theta} - \boldsymbol{\theta}_0\|_2^2$$ near the truth.

## Method

BCV combines K-fold cross-validation to estimate score residuals with a Gaussian multiplier bootstrap to estimate their quantile, and then applies debiasing for inference. It builds on `lasso` (ℓ₁ penalization), `k-fold-cross-validation`, and `gaussian-multiplier-bootstrap`.

**ℓ₁-ME (eq. 1.2, p. 3).** The penalized estimator is:

$$
\widehat{\boldsymbol{\theta}}(\lambda) \in \widehat{\Theta}(\lambda) := \operatorname*{argmin}_{\boldsymbol{\theta} \in \Theta} \left\{ \frac{1}{n}\sum_{i=1}^n m\!\left(\boldsymbol{X}_i^\top \boldsymbol{\theta},\, Y_i\right) + \lambda \|\boldsymbol{\theta}\|_1 \right\} \tag{1.2}
$$

The penalty parameter λ must be chosen; it governs the bias-variance tradeoff and size control.

**Step 1: cross-validating residuals.** The CV procedure partitions observations into K folds $$\{I_k\}_{k=1}^K$$, selects a preliminary penalty level $$\widehat{\lambda}^{\text{cv}}$$ by minimizing out-of-fold prediction loss (eq. 4.9, p. 16):

$$
\widehat{\lambda}^{\text{cv}} \in \operatorname*{argmin}_{\lambda \in \Lambda_n} \sum_{k=1}^K \sum_{i \in I_k} m\!\left(\boldsymbol{X}_i^\top \widehat{\boldsymbol{\theta}}_{I_k^c}(\lambda),\, Y_i\right) \tag{4.9}
$$

and extracts out-of-fold residuals (eq. 4.10, p. 16):

$$
\widehat{U}_i^{\text{cv}} := m'_1\!\left(\boldsymbol{X}_i^\top \widehat{\boldsymbol{\theta}}_{I_k^c}\!\left(\widehat{\lambda}^{\text{cv}}\right),\, Y_i\right), \quad i \in I_k,\; k \in [K] \tag{4.10}
$$

Because observation $$i$$ is held out from fold $$I_k$$, the derivative exists almost surely for kinked losses.

**Step 2: bootstrap penalty (BCV).** Given the residual estimates, the bootstrap quantile estimate and the BCV penalty level are (eqs. 4.11–4.12, p. 17):

$$
\widehat{q}^{\text{bcv}}(1-\alpha) := (1-\alpha)\text{-quantile of } \max_{1 \leq j \leq p} \left| \mathbb{E}_n\!\left[e_i \widehat{U}_i^{\text{cv}} X_{i,j}\right] \right| \text{ given } \{(\boldsymbol{X}_i, Y_i)\}_{i=1}^n, \tag{4.11}
$$

$$
\widehat{\lambda}_\alpha^{\text{bcv}} := c_0 \,\widehat{q}^{\text{bcv}}(1-\alpha) \tag{4.12}
$$

where $$e_1,\ldots,e_n \overset{\text{iid}}{\sim} \mathrm{N}(0,1)$$ are independent of the data. This Gaussian multiplier bootstrap approximates the distribution of $$\|\boldsymbol{S}_n\|_\infty$$ by replacing unobservable true residuals $$U_i = m'_1(\boldsymbol{X}_i^\top \boldsymbol{\theta}_0, Y_i)$$ with $$\widehat{U}_i^{\text{cv}}$$. The approximation is justified by the Chernozhukov et al. (2013, 2017) Gaussian approximation and multiplier bootstrap theorems for maxima of sums of high-dimensional random vectors (eqs. 4.1–4.2, p. 14). Adding BCV to CV carries essentially zero additional computational cost because glmnet already stores the out-of-fold linear forms.

**Theorem 4.1 (BCV convergence rates, p. 17–18).** Under Assumptions 3.1–3.6 and 4.1–4.3, the ℓ₁-ME with BCV penalty $$\widehat{\lambda}_\alpha^{\text{bcv}}$$ achieves (eq. 4.14):

$$
\sup_{\widehat{\boldsymbol{\theta}} \in \widehat{\Theta}(\widehat{\lambda}_\alpha^{\text{bcv}})} \|\widehat{\boldsymbol{\theta}} - \boldsymbol{\theta}_0\|_2 \lesssim_{\mathrm{P}} \sqrt{s_q \eta_n^{2-q}} \quad \text{and} \quad \sup_{\widehat{\boldsymbol{\theta}} \in \widehat{\Theta}(\widehat{\lambda}_\alpha^{\text{bcv}})} \|\widehat{\boldsymbol{\theta}} - \boldsymbol{\theta}_0\|_1 \lesssim_{\mathrm{P}} s_q \eta_n^{1-q} \tag{4.14}
$$

where $$\eta_n = \sqrt{\ln(pn)/n}$$. Under exact sparsity $$q = 0$$, the ℓ₂ rate is $$\sqrt{s_0 \ln(pn)/n}$$, matching the minimax-optimal LASSO rate in Negahban et al. (2012). Crucially, both the ℓ₁ and ℓ₂ rates are established; cross-validation as analyzed in the existing literature yields only the ℓ₂ rate.

**Debiased inference (Algorithm 5.1, p. 21–22).** For inference on a scalar component $$\beta_0$$, three-step debiasing (building on Chernozhukov et al. (2018)) proceeds:
- Step 1: compute the ℓ₁-ME $$\widetilde{\boldsymbol{\theta}} = (\widetilde{\beta}, \widetilde{\boldsymbol{\gamma}}^\top)^\top$$ of $$\boldsymbol{\theta}_0 = (\beta_0, \boldsymbol{\gamma}_0^\top)^\top$$ using BCV penalty $$\lambda_1$$ (with optional refitting).
- Step 2: compute the debiasing coefficient estimate $$\widetilde{\boldsymbol{\mu}}$$ of $$\boldsymbol{\mu}_0$$ by solving a weighted ℓ₁-penalized regression (eq. 5.3, p. 22) using BCV penalty $$\lambda_2$$.
- Step 3: form the debiased estimate by a one-step update (eq. 5.5, p. 22):

$$
\widehat{\beta} := \widetilde{\beta} - \frac{\mathbb{E}_n\!\left[m'_1\!\left(\boldsymbol{X}_i^\top \widetilde{\boldsymbol{\theta}},\, Y_i\right)(D_i - \boldsymbol{W}_i^\top \widetilde{\boldsymbol{\mu}})\right]}{\mathbb{E}_n\!\left[m''_{11}\!\left(\boldsymbol{X}_i^\top \widetilde{\boldsymbol{\theta}},\, Y_i\right)(D_i - \boldsymbol{W}_i^\top \widetilde{\boldsymbol{\mu}}) D_i\right]} \tag{5.5}
$$

Theorem 5.1 (p. 24) establishes $$\sqrt{n}(\widehat{\beta} - \beta_0)/\sigma_0 \xrightarrow{d} \mathrm{N}(0,1)$$ where:

$$
\sigma_0^2 := \frac{\mathbb{E}\!\left[\!\left(m'_1\!\left(\boldsymbol{X}^\top \boldsymbol{\theta}_0, Y\right)(D - \boldsymbol{W}^\top \boldsymbol{\mu}_0)\right)^2\right]}{\left(\mathbb{E}\!\left[m''_{11}\!\left(\boldsymbol{X}^\top \boldsymbol{\theta}_0, Y\right)(D - \boldsymbol{W}^\top \boldsymbol{\mu}_0) D\right]\right)^2} \tag{5.6}
$$

and $$\sigma_0^2$$ is consistently estimated by a plug-in formula (eq. 5.7 or 5.8, p. 24). Neyman orthogonality ensures $$\widehat{\beta}$$ is first-order insensitive to estimation error in $$\boldsymbol{\gamma}_0$$ and $$\boldsymbol{\mu}_0$$, enabling $$\sqrt{n}$$-consistent inference despite high-dimensional nuisance.

## Empirical specifications

**Simulation DGP (§6.1, p. 25).** The study uses a binary probit model (Example 1, §2) with:

$$
Y_i = \mathbf{1}\!\left(\beta_0 D_i + \sum_{j=1}^{p-1} \gamma_{0j} W_{i,j} + \varepsilon_i > 0\right), \quad \varepsilon_i \mid D_i, \boldsymbol{W}_i \sim \mathrm{N}(0,1), \quad i \in [n],
$$

and jointly centered Gaussian regressors $$\boldsymbol{X} = (D, \boldsymbol{W}^\top)^\top \sim \mathrm{N}(\boldsymbol{0}, \boldsymbol{\Sigma}(\rho))$$ with Toeplitz covariance $$\Sigma_{j,k}(\rho) = \rho^{|j-k|}$$, $$\rho \in \{0, 0.2, 0.4, 0.6, 0.8\}$$. Three coefficient patterns: exactly sparse ($$\boldsymbol{\theta}_0 = (1,1,0,\ldots,0)^\top$$, $$s_0 = 2$$), intermediate ($$\theta_{0,j} = (1/\sqrt{2})^{j-1}\mathbf{1}(j \leq 5)$$), and approximately sparse ($$\theta_{0,j} = (1/\sqrt{2})^{j-1}$$ for all $$j$$). Sample sizes $$n = p \in \{100, 200, 400\}$$ fix the problem in the high-dimensional regime. All runs use $$K = 3$$ folds, 2,000 simulation draws, and 1,000 Gaussian bootstrap draws per draw per estimation step. Four estimators are compared: BCV, post-BCV, and CV (post-CV is dropped after failing in ~15% of cases; see R3). The score markup is $$c_0 = 1.1$$ and tolerance is $$\alpha_n = 0.1/\ln(p \vee n)$$, following Belloni et al. (2012). Implemented in R 4.2.2 using `glmnet::cv.glmnet` (CV and BCV) and `stats::glm` (refitting).

**Empirical application (§7, p. 35).** Chetverikov and Sørensen (2025) re-examine Fryer Jr (2019)'s analysis of racial differences in police use of force using the Police-Public Contact Survey (PPCS). The estimating model is (eq. 7.1, p. 35):

$$
\mathrm{P}(\text{Force} = 1 \mid \textbf{Race},\, \boldsymbol{W}) = F\!\left(\textbf{Race}^\top \boldsymbol{\alpha}_0 + \boldsymbol{W}^\top \boldsymbol{\gamma}_0\right) \tag{7.1}
$$

where Force is an indicator for any police use of force conditional on a civilian-officer encounter, $$\textbf{Race} = (\text{Black}, \text{Hisp}, \text{Other})^\top$$ are race dummies (white is reference), $$\boldsymbol{W}$$ is the control vector, and $$F$$ is either the logistic CDF (logit) or the standard normal CDF (probit). Two regressor sets: Basic Controls (p = 30, matching Fryer's largest set, with categorical variables expanded to dummies) and Basic Controls + Interactions (p = 327, adding all first-order pairwise interactions among the original non-race controls). Complete separation occurs at p = 327, so unpenalized ML does not exist; ℓ₁-penalization via post-BCV with K = 10 folds is required.

Three-step post-BCV debiasing (Algorithm 5.1) is applied, testing $$\beta_{\text{Black}} = 0$$ and computing the average partial effect (APE):

$$
\widehat{\text{APE}}_{\text{Black}} := \mathbb{E}_n\!\left[F\!\left(\widehat{\beta}_{\text{Black}} + \boldsymbol{W}_i^\top \widehat{\boldsymbol{\gamma}}\right) - F\!\left(\boldsymbol{W}_i^\top \widehat{\boldsymbol{\gamma}}\right)\right] \tag{7.2}
$$

Fryer Jr (2019) Table 2.B (logit) is replicated to all reported digits using data from his supplementary files, confirming dataset identity (n = 9,930 complete cases for 2002 and 2011 surveys out of 59,668 total encounters).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Police-Public Contact Survey (PPCS), BJS | Empirical illustration: n = 9,930 civilian-police encounters from 2002 and 2011 PPCS surveys; binary outcome (force used); 30 basic controls and 327 controls + interactions | no page yet |
| Synthetic simulation data | Binary probit DGP with Toeplitz covariance, n = p = 100, 200, 400, three sparsity patterns; 2,000 draws per design | n/a (simulated) |

## When to read the full paper

Read the source at [doi.org/10.1086/736770](https://doi.org/10.1086/736770) if you are: implementing ℓ₁-penalized M-estimators with a non-Lipschitz loss (probit, ordered response, expectile) and need a penalty selection method with valid ℓ₁ and ℓ₂ error bounds; performing debiased inference on individual components in a high-dimensional M-estimation setting; evaluating the finite-sample size properties of cross-validation versus BCV penalty selection; or extending the Fryer Jr (2019) analysis with richer control sets or alternative loss functions. Online Appendices A–H contain all proofs, verification of assumptions for each example class, and additional simulation results.

## Attribution and rights

Source: peer-reviewed, *Journal of Political Economy* 133(10), October 2025, pp. 3208–3248. DOI: 10.1086/736770. Publisher: University of Chicago Press. Article is paywalled; no CC license found in Crossref metadata. This distillation was extracted by an LLM on 2026-06-26 and is **not human-verified or independently reproduced**. Extract-only: reproduction of the publisher's text is not permitted.

> Chetverikov, Denis, and Jesper Riis-Vestergaard Sørensen. "Selecting Penalty Parameters of High-Dimensional M-Estimators Using Bootstrapping after Cross Validation." *Journal of Political Economy* 133, no. 10 (October 2025): 3208–3248. DOI: 10.1086/736770.


==============================================================================
# Trade with Nominal Rigidities: Rodriguez-Clare, Ulate & Vasquez (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jpe/2025/clare-et-al-trade-nominal-rigidities-understanding-2025/
# Distilled: A dynamic quantitative trade and migration model with downward nominal wage rigidity shows that the China shock generates temporary unemployment reducing U.S. aggregate welfare gains by roughly two-thirds (from 31 to 12 basis points) and turning 18 additional states from net gainers into net losers. Journal of Political Economy 2025, CC BY 4.0 (accepted version). Eight core results with source locators, model equations, and calibration method.
# Tags: paper-summary, international-trade, labor-markets, nominal-rigidities
==============================================================================

**What this is.** The paper's core results, the dynamic trade model with downward nominal wage rigidity, and the key equations for the production structure, labor supply, DNWR constraint, and welfare calculation: enough to know what it found and how, without reading all 42 pages. To replicate or extend it, read the full source at [doi:10.1086/738344](https://doi.org/10.1086/738344) or the [open-access accepted version](https://researchonline.lse.ac.uk/id/eprint/127629/).

## TL;DR

Rodríguez-Clare, Ulate, and Vasquez build a dynamic quantitative trade and migration model with downward nominal wage rigidity (DNWR) and use it to evaluate the China shock. DNWR prevents nominal wages from falling more than roughly 1% per year, generating temporary unemployment when a negative productivity shock demands a larger wage cut. Calibrated to match Autor, Dorn, and Hanson (2013) cross-sectional regressions, the model generates aggregate U.S. unemployment peaking at 1.25% in 2007, which then fades to near zero by 2016. DNWR reduces aggregate U.S. welfare gains from the China shock by roughly two-thirds (from 31 to 12 basis points). In the longer-shock variant (shock lasting until 2011), the welfare gains nearly disappear entirely.

## Core results

Magnitudes are as reported; locators point into the accepted-version PDF. Column (2) of Table 1 refers to the baseline specification.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | DNWR reduces aggregate U.S. welfare gain from the China shock by roughly **two-thirds** | Table 1 col. 2, p. 26; §6.3, p. 29 | 12 bp with DNWR vs. 31 bp without (flexible-wage delta=0 counterfactual) |
| R2 | Aggregate U.S. **unemployment peaks at 1.25%** in 2007 due to the China shock and declines to near zero by 2016 | Figure 3, p. 27 | Cumulative 6 year-points of unemployment over 2001-2010 (§8.2, p. 36) |
| R3 | More-exposed states face **lower welfare gains**: -9.1 bp per $1,000/worker China exposure | Table 1 row "Welfare vs exposure" col. 2, p. 26; §6.3, p. 28 | Coefficient on exposure = -0.091 pp |
| R4 | With DNWR: **20 states lose welfare**; without DNWR: only 2 states lose | Figure 4, p. 29; §6.3 | 30 states gain and 20 lose with DNWR; 48 gain and 2 lose without DNWR |
| R5 | CZs in **high-DNWR states experience 0.17 pp larger unemployment increase** per $1,000 exposure in 2007 | Figure 2 panel a, p. 11; §2.3, p. 11 | Coefficient beta\_{3h} at h=2007; large relative to ADH average of 0.22 pp |
| R6 | **Unemployment effect is transitory** (non-significant by 2011); NILF effect persists to 2020 | Figure 1 panels b-c, pp. 9-10; §2.2 | NILF effect in 2020 still about half the 2007 magnitude |
| R7 | **Longer shock (to 2011) nearly eliminates welfare gains**: 1.1 bp vs. 12.6 bp baseline | Table 1 col. 3 "Longer", p. 26; §7.1, pp. 30-32 | Mean welfare change 0.011 (col. 3) vs. 0.126 (col. 2) |
| R8 | **Sacrifice ratio** near baseline: 1.63 year-points of inflation per year-point of unemployment reduction | Figure 8, p. 37; §8.2, p. 36 | Ratio rises sharply (toward infinity) as unemployment is pushed 6 year-points below baseline |

**Overall (paper's conclusion).** The China shock is a positive terms-of-trade shock for the U.S. as a whole, but DNWR converts a large fraction of that gain into temporary unemployment, sharply reducing aggregate welfare. The baseline welfare gain without DNWR (31 bp) is quantitatively similar to models by Caliendo, Dvorkin, and Parro (2019) and Galle, Rodriguez-Clare, and Yi (2023); DNWR is the source of the large divergence between this paper's welfare estimates and those benchmarks. Under the baseline calibration the U.S. still gains on net; under the longer-shock calibration (which better matches the dynamic pattern of cross-sectional evidence) the gains nearly vanish. The results imply that nominal frictions and aggregate demand management are first-order considerations in evaluating trade shocks, not a side issue.

## Theory / model

The model is a dynamic, multi-sector, multi-region quantitative trade and migration model building on Caliendo, Dvorkin, and Parro (2019) (CDP), extended with two features: DNWR and a nested-Gumbel labor supply that allows different elasticities of sectoral versus regional mobility.

**Production and trade.** There are $$I$$ regions and $$S+1$$ sectors (S productive market sectors plus a home-production sector indexed 0). Each region $$j$$ produces in each sector $$s$$ using labor and intermediates under a Cobb-Douglas production function. With perfect competition and iceberg trade costs $$\tau_{ij,s,t} \geq 1$$, the price of region $$i$$'s good $$s$$ in region $$j$$ at time $$t$$ is (eq. 3, p. 13):

$$
p_{ij,s,t} = \tau_{ij,s,t} A_{i,s,t}^{-1} W_{i,s,t}^{\phi_{i,s}} \prod_k P_{i,k,t}^{\phi_{i,sk}} \tag{3}
$$

where $$A_{i,s,t}$$ is TFP, $$W_{i,s,t}$$ the wage, $$\phi_{i,s}$$ the labor share, $$\phi_{i,sk}$$ the intermediate-input share from sector $$k$$, and $$P_{i,k,t}$$ the sector-$$k$$ price index. The CES price index satisfies (eq. 4):

$$
P_{j,s,t}^{1-\sigma_s} = \sum_{i=1}^{I} p_{ij,s,t}^{1-\sigma_s} \tag{4}
$$

with elasticity of substitution $$\sigma_s > 1$$. Trade shares are (eq. 6):

$$
\lambda_{ij,s,t} = \frac{p_{ij,s,t}^{1-\sigma_s}}{\sum_{r=1}^{I} p_{rj,s,t}^{1-\sigma_s}} \tag{6}
$$

Labor demand equates wage bill to revenue share (eq. 7): $$W_{i,s,t} L_{i,s,t} = \phi_{i,s} R_{i,s,t}$$.

**Labor supply and migration.** Workers are forward-looking with discount factor $$\beta$$. An agent in region $$j$$, sector $$s$$ at time $$t$$ chooses a destination $$(i,k)$$ by solving (eq. in §3.2, p. 14):

$$
V_{j,s,t} = U_{j,s,t} + \max_{\{i,k\}} \left\{ \beta \mathbb{E}(V_{i,k,t+1}) - \varphi_{ji,sk} + \epsilon_{i,k,t} \right\}
$$

Idiosyncratic shocks $$\epsilon$$ follow a nested Gumbel distribution with nesting parameter $$\kappa > \nu > 0$$, allowing the elasticity of inter-regional mobility (governed by $$1/\kappa$$) to differ from the elasticity of inter-sectoral mobility (governed by $$1/\nu$$). This yields closed-form migration shares (eqs. 9-10, p. 14). The expected lifetime utility and labor-supply evolution satisfy eqs. (8) and (11) in the paper.

**DNWR.** Following Schmitt-Grohe and Uribe (2016), the key departure from CDP is that employment $$L_{i,k,t}$$ can fall below labor supply $$\ell_{i,k,t}$$:

$$
L_{i,k,t} \leq \ell_{i,k,t} \tag{16}
$$

Nominal wages in local currency units cannot fall by more than a factor $$\delta_k$$:

$$
W_{i,k,t} \geq \delta_{i,k} W_{i,k,t-1}, \qquad \delta_{i,k} \geq 0 \tag{17}
$$

Both constraints hold with complementary slackness (eq. 18, p. 16). In the baseline calibration, $$\delta_{i,k} = \delta \approx 0.99$$ for all U.S. manufacturing sectors, and $$\delta_{i,k} = 0$$ (flexible wages) elsewhere.

**Nominal anchor.** To close the nominal model the paper assumes world nominal GDP grows at a constant gross rate $$\gamma$$ (eq. 19, p. 17):

$$
\sum_{i=1}^{I} \sum_{s=1}^{S} W_{i,s,t} L_{i,s,t} = \gamma \sum_{i=1}^{I} \sum_{s=1}^{S} W_{i,s,t-1} L_{i,s,t-1} \tag{19}
$$

This anchor is set so that the ratio $$\delta/\gamma$$ determines the bite of DNWR; in the baseline $$\gamma = 1$$ so the full burden of adjustment falls on $$\delta$$.

**Welfare.** In the dynamic hat-algebra (ratio-form) representation, the welfare change for sector-region $$(j,s)$$ due to the China shock is (§3.6, p. 19):

$$
\mathcal{V}_{j,s} = \sum_{t=1}^{\infty} \beta^t \ln \left( \frac{\hat{\Delta}_{j,s,t} \hat{\omega}_{j,s,t}}{(\hat{\mu}_{jj,ss|j,t})^{\nu} (\hat{\mu}_{jj,s\#,t})^{\kappa}} \right)
$$

where hats denote counterfactual-to-baseline ratios, $$\hat{\Delta}_{j,s,t}$$ is the risk-adjustment factor, $$\hat{\omega}_{j,s,t}$$ is the real wage ratio, and $$\hat{\mu}$$ terms capture mobility gains. This is a permanent equivalent variation in real income.

## Method

**Dynamic hat algebra.** Following Dekle et al. (2007) and CDP, the model is solved in ratio form so that counterfactual exercises require only initial-period observables (revenues, trade shares, labor supply, migration matrices) and parameters $$(\delta, \nu, \kappa, \sigma_s, \alpha_{j,s}, \phi_{j,s}, \phi_{j,sk})$$, without data on TFP levels or wages per efficiency unit. This is the `dynamic-general-equilibrium` and `dynamic-hat-algebra` technique. The contraction-mapping algorithm adapted from Alvarez and Lucas (2007) handles the complementary-slackness conditions (eqs. 16-18) in Appendix B.4-B.7.

**Calibration.** Parameters $$(\delta, \nu, \kappa)$$ are calibrated by `method-of-simulated-moments`-style matching: the model is simulated at each candidate parameter vector and OLS regressions on simulated data are compared to three ADH-style targets (pp. 23-24):
- Unemployment-to-population effect: +0.22 pp per $1,000 exposure
- NILF-to-population effect: +0.55 pp per $1,000 exposure
- Population effect: -0.05 pp per $1,000 exposure

This yields $$\delta \approx 0.99$$, $$\nu = 0.54$$, $$\kappa = 6.55$$ (Table 1, p. 26). The China productivity shocks $$\{\hat{A}_{\text{China},s,t}\}$$ are calibrated to match U.S. import growth from China in each sector using a gravity regression and the other-high-income-country import instrument from ADH (§5, pp. 21-23).

**Welfare counterfactual.** For any set of parameters and shocks the equilibrium is solved forward from 2001 using dynamic hat algebra, and the welfare expression above is evaluated at discount rate $$\beta = 0.95$$. The counterfactual (with China shock) is compared to the baseline (no shock). For the no-DNWR comparison, $$\delta$$ is set to zero without recalibrating $$\nu$$ and $$\kappa$$.

## Empirical specifications

Two empirical exercises motivate and validate the model.

**ADH-style dynamic regressions (Section 2).** The paper estimates the following specification in the spirit of Autor, Dorn, and Hanson (2021), stacking the 1990-2000 and 2000+h changes for $$h = 6, \ldots, 20$$ (eq. 1, p. 8):

$$
\Delta Y_{i,t+h} = \alpha_t + \beta_{1h} \Delta IP^{cu}_{i,\tau} + X'_{i,t} \beta_2 + \varepsilon_{i,t+h} \tag{1}
$$

where $$\Delta Y_{i,t+h}$$ is the ten-year-equivalent change in outcome $$Y$$ for commuting zone $$i$$, $$\Delta IP^{cu}_{i,\tau}$$ is the growth in Chinese import competition in interval $$\tau$$, and $$X_{i,t}$$ are controls. The endogenous import exposure is instrumented with $$\Delta IP^{cu}_{0i,\tau}$$, the analogous import growth in other high-income countries (ADH's instrument). Standard errors are not clustered (OLS/2SLS on stacked CZ data with year fixed effects). Data source: ACS employment data and ADH replication files (§2.2, p. 8).

**DNWR heterogeneity regressions (Section 2.3).** To link DNWR intensity to the unemployment response, the paper augments eq. (1) with a state-level DNWR proxy and its interaction with exposure (eq. 2, p. 10):

$$
\Delta U_{i,t+h} = \gamma_t + \beta_{1h} \Delta IP^{cu}_{i,\tau} + \beta_{2h} \text{Rig}_{s(i),\tau} + \beta_{3h} \text{Rig}_{s(i),\tau} \times \Delta IP^{cu}_{i,\tau} + X'_{i,t} \beta_4 + \varepsilon_{i,t+h} \tag{2}
$$

where $$\text{Rig}_{s(i),\tau}$$ is a state-level dummy for high DNWR (below-median share of workers with negative wage changes, drawn from CPS data following Jo and Zubairy 2023). The interaction $$\text{Rig} \times \Delta IP^{cu}$$ is instrumented with $$\text{Rig} \times \Delta IP^{cu}_{0i,\tau}$$. Results: $$\hat{\beta}_{3h}$$ at $$h = 2007$$ equals 0.17 pp (Figure 2 panel a, p. 11), statistically significant and large relative to the average ADH effect.

**Model validation.** The calibrated model is used to generate simulated state-level changes in employment, NILF, unemployment, wages, and population. OLS regressions of simulated changes on the ADH exposure measure are then compared to the empirical estimates in column (1) of Table 1 (p. 25-26). The model matches the targeted moments by construction but also closely replicates the non-targeted manufacturing and non-manufacturing employment effects (-0.605 vs. ADH's -0.596 and -0.169 vs. ADH's -0.178, respectively).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| American Community Survey (ACS) | Employment (manufacturing and non-manufacturing), NILF, and unemployment data for commuting zones and states, 2000-2020 | [ACS](/wiki/datasets/acs/) |
| Bureau of Labor Statistics (BLS) sector-state employment | Initial labor-supply distribution and migration matrix construction for U.S. sector-state pairs | [BLS](/wiki/datasets/bls/) |
| U.S. Bureau of Economic Analysis regional accounts | Share of labor in production and value-added in gross output for U.S. states; scaling of state relative importance in U.S. total | [BEA I-O Accounts](/wiki/datasets/bea-io/) |
| U.S. Census Bureau trade statistics | Import and Export Merchandise Trade Statistics for state-country bilateral flows in manufacturing and agriculture | [Census public data](/wiki/datasets/census/) |
| World Input-Output Database (WIOD, 2013 release) | Bilateral trade flows, I-O coefficients, and production data for 36 countries; labor and intermediate input shares | no page yet |
| Commodity Flow Survey (CFS) | Intra-U.S. bilateral manufacturing trade flows between states | no page yet |
| IRS Statistics of Income (SOI) Tax Stats | State-to-state migration flows used to construct the initial migration matrix | no page yet |
| Current Population Survey (CPS) | State-level DNWR proxies (share of workers with negative year-over-year wage changes); intra-state migration and labor-flow data | no page yet |
| ADH replication files (Autor, Dorn, and Hanson 2013) | Controls $$X_{i,t}$$ for the cross-sectional regressions; CZ-level import exposure definition | no page yet |

Sample: 87 regions (50 U.S. states, 36 countries, rest of world), 15 sectors (12 manufacturing, services, agriculture, home production), annual, 2000-2007 baseline.

## When to read the full paper

Read the [original source](https://doi.org/10.1086/738344) if you are: (i) building or extending a quantitative trade model with DNWR or nominal frictions; (ii) studying welfare distributional effects of the China shock across U.S. states (Figure 4 and Appendix A.9 are the key outputs); (iii) calibrating mobility elasticities in a spatial labor market model (the nested-Gumbel structure in eqs. 8-11 with $$\nu \neq \kappa$$ is the key methodological contribution to labor supply); (iv) interested in the sacrifice ratio between unemployment and inflation in a trade context (Figure 8); or (v) replicating the ADH dynamic evidence (Figure 1 updates ADH to 2020). The model code and calibration details are in online Appendices B-C.

## Attribution and rights

Source: peer-reviewed, *Journal of Political Economy* 134(2), February 2026. This distillation was extracted by an LLM on 2026-06-26 and is **not human-verified or independently reproduced**. The accepted version is CC BY 4.0; the version of record is paywalled. Extract-only; PDF not hosted in this batch.

> **Attribution (CC BY 4.0, accepted version).** Rodríguez-Clare, Andrés, Mauricio Ulate, and Jose P. Vasquez.
> "Trade with Nominal Rigidities: Understanding the Unemployment and Welfare Effects of the China Shock."
> *Journal of Political Economy* 134, no. 2 (February 2026): 626-664.
> DOI: 10.1086/738344.
> Accepted version: LSE Research Online, eprint 127629, CC BY 4.0.
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Colluding against Workers: Delabastita & Rubens (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jpe/2025/delabastita-rubens-colluding-against-workers-2025/
# Distilled: proposes a new identification approach for employer collusion in labor markets using production and cost data, applied to 227 Belgian coal firms 1845-1913. The 1897 coal cartel explains the entire post-1900 surge in wage markdowns and depressed wages and employment by 6%-17% relative to pre-cartel conduct. Journal of Political Economy 2025, paywalled. Seven core results with source locators, datasets used, the structural model, and the method with its defining equations.
# Tags: paper-summary, labor-markets, monopsony, wage-markdowns, employer-collusion
==============================================================================

**What this is.** The paper's core results, the structural model it proposes, and the identification method with its defining equations: enough to understand what was found and how, without reading all 44 pages. To replicate or extend, read the full source at [https://doi.org/10.1086/734780](https://doi.org/10.1086/734780).

## TL;DR

The paper develops an empirical method to detect and quantify employer collusion in labor markets using firm-level production and cost data. The idea is to estimate wage markdowns (the ratio of labor's marginal revenue product to the wage) from a production model that imposes no conduct assumptions, then compare those estimates to the markdown bounds that would arise under Cournot (no-collusion) and fully collusive behavior. Applied to 227 Belgian coal mining firms in the Liege and Namur provinces between 1845 and 1913, the paper finds: (i) wages were roughly 40% below labor's marginal revenue product at the median firm; (ii) wage markdowns were higher at employers' association members throughout the nineteenth century but that premium disappeared after the 1897 coal cartel; (iii) markdowns jumped to the fully collusive upper bound right after the cartel began, an increase the authors can detect without ex ante knowledge of the cartel's timing; and (iv) the cartel reduced equilibrium wages and employment by 6%-17% depending on assumptions about coal market competition.

## Core results

Results extracted from Tables 1-4 and Figures 3-4 of the source PDF. Magnitudes are as reported; the PDF was electronically published before final pagination (all pages show "000"), so locators reference equation, table, and figure numbers. Standard errors (SE) are block-bootstrapped with 200 iterations.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Median wage markdown (MRPL/wage) estimated at 1.680, implying miners were paid ~40% below their marginal revenue product; the coal price markup is below 1, suggesting firms derived profits mainly from labor market power | Table 1, panels A-B; markup equation (13) | Median markdown 1.680 (SE 0.450); average 1.828 (SE 0.491); median markup 0.714 (SE 0.494); firm-level labor supply elasticity 10.172 |
| R2 | Wage markdowns were 11.2% higher at employers' association members, consistent with wage-fixing collusion through these associations | Table 2, panel A, col. 1 | 0.112 (SE 0.052) |
| R3 | After the 1897 coal cartel, the employers' association premium entirely disappears, indicating informal association-based collusion was replaced by formal cartel collusion | Table 2, panel B | Pre-1897 association coefficient 0.132 (SE 0.042); post-1897 coefficient -0.058 (SE 0.091) |
| R4 | The collusion null cannot be rejected before 1900; from 1901 onward, it is rejected at the 10% level for every year except 1903, identifying cartel collusion without requiring prior knowledge of its existence | Figure 4B | Median collusion index near zero pre-1900; statistically positive from 1901 at 10% |
| R5 | The cartel reduced wages and employment by about 6% relative to the observed pre-1898 conduct (which was itself partially collusive), under exogenous coal prices | Table 4, panel A, column "Pre-1898 Conduct" | Wage change -0.059; employment change -0.059 |
| R6 | Compared to Cournot (no collusion) competition, the cartel reduced wages and employment by about 10%, under exogenous coal prices | Table 4, panel A, column "Cournot" | Wage change -0.103; employment change -0.102 |
| R7 | Under endogenous coal prices, the cartel reduced wages and employment by ~17% relative to pre-1898 conduct and ~25% relative to Cournot; coal output fell ~20%-28% and coal prices rose ~10%-17% | Table 4, panel B | vs pre-1898: wages -0.167, employment -0.166, output -0.195, price +0.100; vs Cournot: wages -0.251, employment -0.249, output -0.283, price +0.174 |

**Overall (paper's conclusion).** Wage markdowns in Belgian coal mining were stable during the first half of the nineteenth century but increased sharply around 1900. Decomposing markdowns into collusive and noncollusive components shows the pre-1900 growth was driven by noncollusive sources (concentration, productivity), while the post-1900 surge was entirely driven by the 1897 coal cartel. The method can identify this collusion without ex ante information about the cartel, and counterfactuals show it caused substantial losses in wages and employment for workers.

## Theory / model

The paper builds a model of the labor market with three components: a production technology, a labor supply function, and an employer behavior model with a conduct parameter.

**Production function.** Output $$Q_{ft}$$ at firm $$f$$ in year $$t$$ follows a Cobb-Douglas specification in log labor $$l_{ft}$$, log materials $$m_{ft}$$, and log capital $$k_{ft}$$, with log total factor productivity $$\omega_{ft}$$ (eq. 1, Section III.A):

$$
q_{ft} = \beta^l l_{ft} + \beta^m m_{ft} + \beta^k k_{ft} + \omega_{ft} \tag{1}
$$

TFP follows an AR(1) Markov process with serial correlation $$\rho$$ and innovation $$v_{ft}$$ (eq. 2):

$$
\omega_{ft} = \rho \, \omega_{f,t-1} + v_{ft} \tag{2}
$$

**Labor supply.** Firms face an upward-sloping market-level labor supply curve with inverse elasticity $$\Psi^l$$. The log-linear supply function at market $$i$$ in year $$t$$ is (eq. 3, Section III.B):

$$
W^l_{it} = L^{\Psi^l}_{it} \, \nu_{it} \tag{3}
$$

where $$L_{it}$$ is market-level employment and $$\nu_{it}$$ is a labor supply shifter. If firms are wage takers, $$\Psi^l = 0$$; labor market power implies $$\Psi^l > 0$$. The firm-level inverse elasticity $$\psi^l_{ft} \equiv (\partial W^l_{ft}/\partial L_{ft})(L_{ft}/W^l_{ft})$$ is related to the market-level elasticity by the firm's labor market share.

**Wage markdown and markup.** The wage markdown is defined as the ratio of labor's marginal revenue product to the wage:

$$
\mu^l_{ft} \equiv \frac{\text{MRPL}_{ft}}{W^l_{ft}}, \qquad \text{where} \quad \text{MRPL}_{ft} \equiv \frac{\partial(P_{ft} Q_{ft})}{\partial L_{ft}}
$$

A percentage wage wedge $$\delta^l_{ft} = (\mu^l_{ft} - 1)/\mu^l_{ft}$$ measures how far below MRPL workers are paid. The product market markup is $$\mu_{ft} \equiv P_{ft}/\text{MC}_{ft}$$.

**Employer behavior and conduct.** Firms minimize a weighted combination of their own and rivals' input costs (eq. 4, Section III.C), with collusion weights $$\lambda_{fgt}$$ parameterizing the degree to which firm $$f$$ internalizes firm $$g$$'s costs. When $$\lambda_{fgt} = 0$$ for $$f \ne g$$, firms minimize only their own costs (Cournot). When $$\lambda_{fgt} = 1$$, firms jointly minimize the cartel's total costs.

**Markdown bounds.** Under no collusion (Cournot), the first-order condition for labor yields the no-collusion markdown lower bound (eq. 7):

$$
\underline{\mu}^l_{ft} = 1 + s^l_{ft} \, \Psi^l \tag{7}
$$

where $$s^l_{ft} = L_{ft}/L_{it}$$ is the firm's labor market share. Under full collusion, all firms within market $$i$$ minimize joint costs, treating the market-level supply curve as endogenous; the fully collusive markdown upper bound is (eq. 10):

$$
\bar{\mu}^l_{ft} = 1 + \Psi^l \tag{10}
$$

Nesting both cases through a scalar conduct parameter $$\tilde{\lambda}_{ft} \in [s^l_{ft}, 1]$$ (eq. 12):

$$
\mu^l_{ft} = 1 + \tilde{\lambda}_{ft} \, \Psi^l \tag{12}
$$

## Method

The central methodological contribution is a way to identify employer conduct without imposing a conduct assumption. The approach builds on the demand-side conduct-identification tradition of Bresnahan (1987) but uses the supply side: the identification comes from combining two independently estimable quantities: (i) a cost-side markdown estimate that does not depend on conduct, derived from the production function; and (ii) the conduct-dependent markdown bounds from the labor supply model. De Loecker and Scott (2016) applied a similar comparison for goods market price markups without imposing conduct; the present paper extends that logic to the factor market and allows for collusive behavior.

**Key identification equation.** Following De Loecker and Warzynski (2012), the product market markup is $$\mu_{ft} = \beta^m / \alpha^m_{ft}$$, where $$\alpha^m_{ft} = W^m_{ft} M_{ft} / (P_{ft} Q_{ft})$$ is the revenue share of materials. Substituting the production function output elasticity of labor $$\beta^l$$ and the revenue share of labor $$\alpha^l_{ft} = W^l_{ft} L_{ft} / (P_{ft} Q_{ft})$$ into the general first-order condition (eq. 11) yields the key markdown expression (eq. 14, Section III.D):

$$
\mu^l_{ft} = 1 + \tilde{\lambda}_{ft} \, \Psi^l = \frac{\beta^l \alpha^m_{ft}}{\beta^m \alpha^l_{ft}} \tag{14}
$$

The right-hand side is the cost-side markdown estimate: it depends only on the production function parameters $$\beta^l$$ and $$\beta^m$$ and the observable cost shares, not on $$\tilde{\lambda}_{ft}$$. This separates the cost-side estimate from the conduct-side model; equating them identifies the conduct parameter.

**Collusion index.** The paper rescales the conduct parameter to the unit interval (eq. 15), where 0 denotes no collusion (Cournot) and 1 denotes full collusion:

$$
\hat{\lambda}_{ft} \equiv \frac{\mu^l_{ft} - \underline{\mu}^l_{ft}}{\bar{\mu}^l_{ft} - \underline{\mu}^l_{ft}} = \frac{\mu^l_{ft} - (1 + s^l_{ft}\Psi^l)}{\Psi^l(1 - s^l_{ft})} \tag{15}
$$

**Production function estimation.** The paper builds on Olley and Pakes (1996) timing assumptions (capital fixed and dynamic; labor and materials static), combined with Blundell and Bond (2000) AR(1) differencing to avoid inverting the input demand function. The GMM moment conditions (eq. 17) are:

$$
\mathbb{E}\!\left[\left(q_{ft} - \rho q_{f,t-1} - \beta^0(1-\rho) - \beta^l(l_{ft} - \rho l_{f,t-1}) - \beta^m(m_{ft} - \rho m_{f,t-1}) - \beta^k(k_{ft} - \rho k_{f,t-1})\right)\Big|\, l_{f,t-1}, m_{f,t-1}, k_{ft}, k_{f,t-1}, w^{\text{agr}}_{t-1}\right] = 0 \tag{17}
$$

The instruments include lagged inputs plus lagged agricultural wages $$w^{\text{agr}}_{t-1}$$, which shift labor supply to coal mines (Walloon coal mines drew agricultural surplus labor from Flanders) but are assumed not to affect mining productivity directly.

**Labor supply estimation.** The market-level inverse labor supply elasticity $$\Psi^l$$ is estimated by IV on the market-year panel. Two instruments shift labor demand without shifting supply: (i) an indicator for 1871-1875, the coal demand surge from the Franco-Prussian War and Lorraine annexation; and (ii) cartel membership interacted with the post-cartel period, which reduced coal output and hence labor demand for cartel participants.

**Counterfactual equilibrium.** Under exogenous coal prices, closed-form equilibrium wages and employment as a function of the conduct parameter $$\tilde{\lambda}_{it}$$ and revenue $$R_{it} = P_{it}Q_{it}$$ are (Section IV.D):

$$
W^l_{it} = \left(\frac{\beta^l R_{it} \, \nu^{1/\Psi^l}_{it}}{1 + \Psi^l \tilde{\lambda}_{it}}\right)^{\Psi^l/(1+\Psi^l)}, \qquad L_{it} = \left(\frac{\beta^l R_{it}}{(1 + \Psi^l \tilde{\lambda}_{it})\,\nu_{it}}\right)^{1/(1+\Psi^l)}
$$

The cartel effects are computed by setting $$\tilde{\lambda}_{it}$$ to the Cournot value ($$1/N_{it}$$) or the pre-1898 average collusion level and comparing to the observed post-cartel fully collusive state.

## Empirical specifications

**Production function (Table 1, panel A).** Estimated by GMM on 4,005 firm-year observations (GMM sample, after conditioning on all variables being observed) using the moment conditions in eq. (17). Block-bootstrap with 200 iterations. The preferred specification (column 2, free RTS) gives $$\hat{\beta}^l = 0.699$$ (SE 0.327), $$\hat{\beta}^m = 0.222$$ (SE 0.138), $$\hat{\beta}^k = 0.153$$ (SE 0.075), serial correlation $$\hat{\rho} = 0.866$$ (SE 0.198). The model is overidentified; the Hansen J-test gives p = 0.126. A version with RTS restricted to 1.05 (column 3) yields tighter standard errors: $$\hat{\beta}^l = 0.661$$, $$\hat{\beta}^m = 0.237$$, $$\hat{\beta}^k = 0.102$$.

**Labor supply (Table 1, panel C).** The market-level inverse labor supply elasticity is estimated by IV on 1,990 market-year observations, regressing log wage on log employment with the two demand shifters as instruments. The IV estimate is $$\hat{\Psi}^l = 1.009$$ (SE 0.265), implying that at a monopsonistic firm the MRPL is twice the wage. The first-stage F-statistic is 462. The firm-level elasticity implied by the model is 10.172.

**Markdown correlations (Table 2).** OLS regressions of log markdown $$\mu^l_{ft}$$ on employer association and cartel membership indicators. Panel A (all years) on 4,432 observations with year fixed effects: employers' association coefficient 0.112 (SE 0.052); cartel coefficient 0.080 (SE 0.041). Panel B splits by pre- vs. post-1897: association coefficient pre-1897 is 0.132 (SE 0.042) and post-1897 is -0.058 (SE 0.091).

**Size-markdown correlations (Table 3).** Regressions of log markdown on log labor market share, separately for cartel and noncartel firms, with no, market, and market-by-year fixed effects. For noncartel firms, market-by-year FE explain 56% of markdown variation and the size-markdown gradient is positive (0.065, SE 0.005), consistent with the Cournot model. For cartel firms, conditioning on market-by-year FE makes the size-markdown gradient near zero (-0.004, SE 0.002), consistent with equalized markdowns under collusion.

**Collusion test (Figure 4B).** Year-by-year estimation of the median collusion index $$\hat{\lambda}_{ft}$$ with 10%-90% confidence intervals (200 bootstrap iterations). The collusion index fluctuates around 0-50% of the collusive range before 1900. From 1901 onward, the null of zero collusion is rejected at the 10% level for every year except 1903.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Administration des Mines annual inspection reports (Liege and Namur, 1845-1913) | Firm-level coal output (tons by type), employment (days worked, underground vs. surface), intermediate input expenditure, extraordinary expenses (capital investment), horsepower of machines | no page yet |
| Union des Charbonnages Ligeois monthly Bulletin (digitized) | Employer association membership per firm, per year | no page yet |
| Cartel lists from De Leener (1904) | Coal cartel membership 1897 onward | no page yet |
| Municipality-level railroad and tramway station opening dates | Control for transport network access; instrument validity check | no page yet |
| Consumer price index (Segers 2003; extended to 1845 using Scholliers 1995) | Price deflation of all nominal variables | no page yet |
| Agricultural wages in Belgium (Segers 2003) | Instrument for labor supply to coal mines; reflects labor supply shocks from the agricultural sector | no page yet |

Sample: 227 coal mining concessions in Liege and Namur provinces, annual observations 1845-1913. The GMM production function sample has 4,005 observations (requiring lagged variables to be observed); the full markdown-estimation sample has up to 8,779 firm-year observations. The market-level labor supply sample has 1,990 municipality-year observations.

## When to read the full paper

Read the [original](https://doi.org/10.1086/734780) if you are:
working on identification of employer conduct or monopsony in labor markets (Section III gives the complete model and identification logic, including the generalization to heterogeneous employers in appendix A.1);
studying the labor market effects of cartels historically or in contemporary antitrust contexts (Section IV.D and Table 4 give the counterfactual framework and parameter estimates);
interested in production function estimation with labor supply instruments or factor market power (the GMM approach with agricultural wage instruments is fully developed in Section IV.A and appendices B-C); or
working on economic history of the Industrial Revolution and employer associations (Sections II and V cover the Belgian coal setting and robustness checks including unionization, factor-biased technical change, and democratization); or concerned with antitrust policy toward labor markets (the results bear on arguments in Naidu, Posner, and Weyl (2018) that antitrust should address labor-market collusion, not only product-market collusion).

## Attribution and rights

Source: peer-reviewed, *Journal of Political Economy* 133(6), June 2025. Copyright 2025 The University of Chicago. All rights reserved. Published by The University of Chicago Press. This distillation was extracted by an LLM on 2026-06-26 and is **not human-verified or independently reproduced**. The journal version is paywalled; an accepted-manuscript copy is available at the Radboud University repository. Replication code: [Harvard Dataverse, https://doi.org/10.7910/DVN/FG1JSE](https://doi.org/10.7910/DVN/FG1JSE) (Delabastita and Rubens 2024).

> Delabastita, Vincent, and Michael Rubens. "Colluding against Workers."
> *Journal of Political Economy* 133, no. 6 (June 2025): 1796-1839.
> DOI: 10.1086/734780. Extract-only; all rights reserved.


==============================================================================
# Asset-Price Redistribution: Fagereng et al. (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jpe/2025/fagereng-et-al-asset-price-redistribution-2025/
# Distilled: Rising asset valuations redistribute welfare from buyers to sellers, not from non-holders to holders. Individual welfare gains range from -$185,000 (p1) to +$273,000 (p99) in Norway 1994-2019, with redistribution from young cohorts to old and from the poor to the wealthy. Journal of Political Economy 2025, paywalled. Six core results with source locators, datasets used, the model (envelope-theorem sufficient statistic), and the empirical implementation (NPV of net asset sales weighted by price-dividend deviation).
# Tags: paper-summary, household-finance, wealth-inequality, redistribution, life-cycle
==============================================================================

**What this is.** The paper's core results, the model it builds on (the envelope-theorem welfare formula for asset-price changes), and the empirical implementation: enough to know what it found and how, without reading the full 56-page manuscript. To replicate or extend it, read the full source at the [original](https://doi.org/10.1086/736769).

## TL;DR

The paper asks who wins and who loses from rising asset valuations. It derives a sufficient statistic for the individual money-metric welfare effect of a price-dividend deviation: the net present value of an individual's net asset sales, weighted by the percentage price deviation from a balanced-growth baseline. Applied to Norwegian administrative data on all financial transactions from 1994 to 2019 (four asset classes: housing, debt, deposits, equity), the paper finds that the rise in asset valuations had large redistributive effects, redistributing from the young to the old and from the poor to the wealthy. Housing is the dominant driver. Welfare gains differ sharply from revaluation gains (the Spearman rank correlation is 0.19), because revaluation gains accrue to holders while welfare gains accrue only to net sellers.

## Core results

Magnitudes and signs are as reported in the paper (2011 US dollars unless stated). Locators point to the accepted manuscript.

| \# | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Distribution of individual welfare gains is wide and centered near zero | Figure 4a, p. 24 | avg +$10,000; p1 = -$185,000; p99 = +$273,000; interquartile range $31,000 (2011 USD) |
| R2 | Welfare gains as a fraction of total wealth: near zero on average with fat tails | Figure 4b, p. 24 | avg 0.0%; p1 = -30%; p99 = +27%; interquartile range 5.0% |
| R3 | Intergenerational redistribution: old gain, young lose | Figure 7c, p. 27 | Millennials (<=15 in 1993): avg -$13,000; Baby Boomers (>50 in 1993): avg +$22,000 |
| R4 | Redistribution across wealth percentiles: gains concentrated at the top | Figure 8c-d, pp. 29-30 | top 1%: avg +$73,000 (3.1% of total wealth); bottom 1%: avg +$8,000; population avg normalized 1.5% |
| R5 | Welfare gains and revaluation gains are weakly correlated | Figure 6, p. 26 | Spearman rank correlation = 0.19; revaluation gains avg 16.4% of total wealth vs. welfare gains avg 0.0% |
| R6 | Housing and debt dominate the welfare-gain decomposition | Table 2, p. 24 | housing: -$4,700; debt: +$16,900; deposits: -$2,400; equity: +$200; total: +$10,000 (avg, 2011 USD) |

**Overall (paper's conclusion).** Asset-price changes redistribute welfare rather than creating it. The rise in Norwegian asset valuations from 1994 to 2019 benefited the old and the wealthy because they are net sellers of assets, while the young and the poor are net buyers. This contrasts with the view in Piketty and Zucman (2014) that rising asset prices benefit existing asset holders generally. Revaluation gains, which are positive for all asset holders, overstate welfare gains by a factor of more than ten on average and are at best weakly correlated with actual welfare gains at the individual level. The government (via the sovereign wealth fund, a net saver) bore losses that offset household gains in the aggregate. Compared to Doepke and Schneider (2006), who estimate redistributive effects of inflation using Survey of Consumer Finances data for the US, this paper uses admin microdata and covers all four major asset classes over a longer horizon.

## Theory / model

The paper builds a framework for measuring the welfare effect of a deviation in asset prices. The key insight, illustrated first in a two-period model (Section 1.1, pp. 5-7), is that the welfare gain depends on net asset *sales*, not net asset *holdings*. An individual who holds an asset but never sells it gains nothing from a price increase at first order.

**Two-period intuition (Section 1.1).** Individual $$i$$ maximizes utility over consumption paths subject to budget constraints (eqs. 3-4, p. 5):

$$
V_{i,0} \equiv \max_{\{C_{i,0},\, C_{i,1},\, N_{i,0}\}} U(C_{i,0}) + \beta U(C_{i,1}) \tag{3}
$$

$$
C_{i,0} + (N_{i,0} - N_{i,-1}) P_0 = Y_{i,0}, \qquad C_{i,1} = N_{i,0} D_1 + Y_{i,1}. \tag{4}
$$

Applying the envelope theorem to evaluate the welfare effect of an infinitesimal change in the asset price $$P_0$$, holding dividends $$D_1$$ fixed, yields (eq. 5, p. 6):

$$
\frac{dV_{i,0}}{U'(C_{i,0})} = (N_{i,-1} - N_{i,0}) \, dP_0. \tag{5}
$$

The welfare gain is the net asset sale $$(N_{i,-1} - N_{i,0})$$ multiplied by the price change. A pure holder who plans never to trade ($$N_{i,0} = N_{i,-1}$$) is unaffected; sellers gain; buyers lose. This two-period result was also derived by Davila and Korinek (2018) and Moll (2020) in similar two- and three-period environments; the present paper generalizes it to multiple assets, infinite horizon, and the empirical Norwegian setting.

**Baseline infinite-horizon model (Section 1.2, pp. 8-10).** There are $$K$$ long-lived assets and a sequence of one-period bonds. Individual $$i$$ maximizes (eq. 8, p. 9):

$$
V_{i,0} \equiv \max_{\{C_{i,t},\, B_{i,t},\, \{N_{i,k,t}\}_k\}_{t=0}^{\infty}} \sum_{t=0}^{\infty} \beta^t U(C_{i,t}), \tag{8}
$$

subject to budget constraints with adjustment costs $$\chi_k$$ (eq. 9, p. 9):

$$
C_{i,t} + \sum_{k=1}^{K}(N_{i,k,t} - N_{i,k,t-1}) P_{k,t} + B_{i,t} Q_t + \sum_{k=1}^{K} \chi_k(N_{i,k,t} - N_{i,k,t-1}) = \sum_{k=1}^{K} N_{i,k,t-1} D_{k,t} + B_{i,t-1} + Y_{i,t}. \tag{9}
$$

**Proposition 1** (welfare gain, eq. 10, p. 9): the welfare gain of a price deviation $$\{dQ_t, \{dP_{k,t}\}_k\}_{t=0}^{\infty}$$ is:

$$
\frac{dV_{i,0}}{U'(C_{i,0})} = \sum_{t=0}^{\infty} R_{0 \to t}^{-1} \left( \sum_{k=1}^{K} (N_{i,k,t-1} - N_{i,k,t}) \, dP_{k,t} - B_{i,t} \, dQ_t \right). \tag{10}
$$

The formula sums the NPV of net asset sales weighted by the price deviation. The adjustment-cost function $$\chi_k$$ drops out via the envelope theorem: what matters is observed transactions, not why they happen. For bonds, holdings rather than sales matter (because bonds must be rolled over continuously). Welfare gains aggregate to zero across all individuals trading with each other (but not across sectors, since Norwegian households also trade with the government and foreigners).

**Extensions (Section 4):** (i) With uninsurable idiosyncratic income risk (Proposition 2, eq. 18), the formula gains a covariance term between the growth in marginal utility and future net sales, causing welfare gains to no longer aggregate to zero across the population. (ii) With borrowing constraints indexed to loan-to-value following Kiyotaki and Moore (1997) (Proposition 3 / Corollary 4, eqs. 24-27), asset holdings also matter via a collateral channel, and individual discount rates rise with the LTV ratio. (iii) Second-order effects require accounting for portfolio responses to price changes (eq. 29) but are quantitatively small given stable observed transaction patterns across cohorts. A related welfare decomposition for interest-rate changes is provided by Greenwald et al. (2021); that duration-mismatch formula is a special case of the general sufficient statistic here.

## Method

The empirical method has two parts: defining the counterfactual price deviation and implementing the welfare formula with administrative data.

**Price-dividend deviation (eq. 14, p. 16).** The counterfactual is a balanced growth path where asset prices grow at the same rate as dividends (constant price-dividend ratio). The percentage price deviation for asset class $$k$$ is:

$$
\frac{\Delta P_{k,t}}{P_{k,t}} = \frac{PD_{k,t} - \overline{PD}_k}{PD_{k,t}}, \tag{14}
$$

where $$\overline{PD}_k$$ is the 1992-1996 average price-dividend ratio for asset class $$k$$. For housing, the price-dividend ratio is the price-to-rent ratio; for debt and deposits, it is the inverse of the real interest rate; for equity, it is the enterprise-value-to-cash-flow ratio.

**First-order sufficient statistic (eq. 15, p. 17).** Combining Proposition 1 with the percentage deviation formula and truncating the sum at $$T = 25$$ years (1994-2019) gives the implementable statistic:

$$
\text{Welfare Gain}_i = \sum_{t=0}^{T} R^{-t} \left( \sum_{k=1}^{K} (N_{i,k,t-1} - N_{i,k,t}) P_{k,t} \times \frac{PD_{k,t} - \overline{PD}_k}{PD_{k,t}} - B_{i,t} Q_t \times \frac{Q_t - \overline{Q}}{Q_t} \right), \tag{15}
$$

with discount rate $$R = 1.05$$ (the approximate average deposit/mortgage rate at the start of the sample), baseline values from 1992 to 1996, and $$T = 25$$.

**Asset-class decomposition (eq. 16, p. 17).** The statistic is computed separately for four asset classes and summed:

$$
\text{Welfare Gain}_{i,\text{housing}} = \sum_{t=0}^{25} R^{-t} \, (N_{i,H,t-1} - N_{i,H,t}) P_{H,t} \times \frac{PD_{H,t} - \overline{PD}_H}{PD_{H,t}},
$$

$$
\text{Welfare Gain}_{i,\text{debt}} = \sum_{t=0}^{25} R^{-t} \, (-B_{i,M,t} Q_{M,t}) \times \frac{Q_{M,t} - \overline{Q}_M}{Q_{M,t}},
$$

with analogous terms for deposits and equity. The formula builds on the `life-cycle-model` portfolio-choice framework and relies on `panel-regression` to estimate the loan-to-value elasticity of interest rates ($$\hat{\xi} \approx 0.0025$$ to $$0.004$$, from panel regressions in Appendix D.2.2) used in the Section 4.2 collateral-channel extension.

## Empirical specifications

**Data construction (Section 2.3, pp. 20-22).** Individual-level financial transactions are constructed from Norwegian administrative registries. Net transactions in housing are directly observed from housing transaction registries. For equity, net transactions are imputed as the change in market value minus capital gains (using individual stock ownership from VPS from 2005 onward; Financial Accounts aggregate capital gains before 2005). Indirect asset positions via privately owned businesses are consolidated through ten layers of ownership, allocating firm-level transactions to ultimate owners. Holdings and transactions are aggregated at the household level and distributed equally across adult household members.

**Baseline sample (Section 2.1, p. 17).** The sample covers the universe of Norwegians aged at least 18 at any point between 1994 and 2019, observed annually ($$T = 25$$ years). All monetary values are expressed in real 2011 Norwegian krone using the CPI, then converted to 2011 USD at a fixed rate of 5.607.

**Revaluation gains comparison (Section 3.1, eq. 17, p. 26).** The revaluation gain is the NPV of the price-deviation effect on asset *holdings* rather than sales:

$$
\text{Revaluation Gain}_i = \sum_{t=0}^{T} R_{0 \to t}^{-1} \sum_{k=1}^{K} N_{i,k,t-1} P_{k,t-1} \, \Delta\!\left(\frac{P_{k,t}}{P_{k,t-1}}\right), \tag{17}
$$

where $$\Delta(P_{k,t}/P_{k,t-1})$$ is the deviation in the capital-gains component caused by the price deviation. Revaluation gains are positive for virtually all asset holders (average 16.4% of total wealth), while welfare gains average 0.0% and can be negative for buyers. The Spearman rank correlation between the two is 0.19 across individuals.

**Robustness and extensions (Section 4, pp. 31-42).** The baseline welfare gains are shown to be robust to (i) uninsurable income risk (which moderately dampens the welfare loss of the young), (ii) borrowing constraints with collateral effects (small average effect), (iii) second-order effects (small, since observed transaction patterns are roughly stable across cohorts), and (iv) extrapolation beyond 2019 (higher-persistence scenarios shift gains toward younger cohorts). Table 3 (p. 31) reports welfare gains across cohorts for each generalization; the combined-extension mean rises from $11,900 to $23,700.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Norwegian administrative registries (tax records, private business balance sheets, housing transaction registries, equity ownership registers) | Individual-level asset holdings, net transactions, income, and household identifiers, 1993-2019 | [Statistics Norway](/wiki/confidential/statistics-norway/) (licensed/confidential) |
| VPS (Norwegian Securities Registry, Verdipapirsentralen) | Listed equity holdings at the individual level from 2005 onward | [VPS Norway](/wiki/confidential/vps-norway/) (licensed/confidential) |
| Statistics Norway interest-rate and CPI databases | Real interest rates on mortgages and deposits for price-dividend ratio of debt/deposit asset classes | [Statistics Norway](/wiki/confidential/statistics-norway/) (licensed/confidential) |
| Norges Bank Historical Monetary Statistics (Eitrheim and Erlandsen 2005) | House price index (combined with Stats Norway rental index for price-to-rent series) | No page yet |
| Worldscope | Enterprise value and cash flows for listed Norwegian non-financial firms; equity valuation ratio | [Worldscope](/wiki/commercial/worldscope/) (licensed) |

Sample: Norway, universe of adults aged >=18, 1994-2019 (25 years, annual). Asset classes: housing (principal residence, secondary homes, recreational estates), mortgage debt, bank deposits, public and private equity.

## When to read the full paper

Read the [original](https://doi.org/10.1086/736769) if you are: (i) extending the sufficient statistic approach to other countries or asset classes (the method is general, the Norway calibration is country-specific); (ii) implementing the incomplete-markets or borrowing-constraint corrections (Propositions 2-3, Appendix D.1-D.2); (iii) replicating any of the distributional figures using the Harvard Dataverse replication package; or (iv) studying optimal capital gains and wealth taxation in environments with changing asset prices. The locators above point to the exact tables, figures, and equations.

## Attribution and rights

Source: peer-reviewed, *Journal of Political Economy* 133(11), November 2025. This distillation was extracted by an LLM on 2026-06-26 and is **not human-verified or independently reproduced**. The version of record is paywalled (University of Chicago Press). The LSE-hosted accepted manuscript is available under CC BY 4.0 at [https://researchonline.lse.ac.uk/id/eprint/129496/](https://researchonline.lse.ac.uk/id/eprint/129496/); this page extracts and re-expresses core results only.

> **Citation.** Fagereng, Andreas, Matthieu Gomez, Emilien Gouin-Bonenfant, Martin Holm, Benjamin Moll, and Gisle Natvik.
> "Asset-Price Redistribution."
> *Journal of Political Economy* 133, no. 11 (November 2025): 3494-3549.
> DOI: [10.1086/736769](https://doi.org/10.1086/736769).


==============================================================================
# Optimal Fiscal Policy with Heterogeneous Agents: Le Grand & Ragot (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jpe/2025/grand-ragot-optimal-fiscal-policy-heterogeneous-2025/
# Distilled: Le Grand and Ragot (2025) show that positive capital taxes and public debt can both be optimal in a heterogeneous-agent model when credit constraints occasionally bind and utility is non-CRRA (GHH or DRRA), overturning the Chamley-Judd zero-capital-tax result. Optimal public debt rises after a low-persistence public spending shock but falls after a high-persistence shock. Journal of Political Economy 133(7), 2025, paywalled. Six core results with source locators, the structural model equations, and the solution method.
# Tags: paper-summary, optimal-taxation, heterogeneous-agents, macro, fiscal-policy, structural
==============================================================================

**What this is.** The paper's core results, the structural model it builds on, and the solution
method: enough to know what it found and how, without reading all 50 pages. To replicate or
extend, read the full source at [doi.org/10.1086/734877](https://doi.org/10.1086/734877) or
the open preprint at [hal.science/hal-05547657](https://hal.science/hal-05547657).

## TL;DR

Le Grand and Ragot (2025) analyze optimal fiscal policy in a Bewley-Huggett-Aiyagari
heterogeneous-agent model with capital accumulation, progressive labor taxation, a linear
capital tax, and public debt. The government finances exogenous public spending via taxes
and new debt. Three contributions:

First, in a simple analytical model, the steady-state optimal capital tax is positive when
credit constraints occasionally bind AND the utility function deviates from Constant Relative
Risk Aversion (CRRA): for GHH or Decreasing RRA (DRRA) preferences, an externality of savings
on post-tax factor prices creates a rationale for a positive capital tax. With CRRA utility,
the Chamley (1986) and Judd (1985) zero-capital-tax result generalizes exactly (Corollary 1).

Second, the existence of a Stationary Ramsey Equilibrium (SRE) with positive capital tax and
positive public debt requires three independent conditions: a non-first-best condition, the
Straub and Werning (2020) stationarity condition, and a Laffer condition.

Third, for a given net present value (NPV) of a public spending shock, optimal public debt
rises when shock persistence is low (the government borrows to smooth taxes) and falls when
persistence is high (the cost of future tax increases to retire debt is too large). A
quantitative model calibrated to the US via an inverse optimal approach confirms these results.

## Core results

Magnitudes are as reported; all results are from the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Positive optimal capital tax when credit constraints bind for unemployed agents (GHH/DRRA utility) | Proposition 1 (eq. 29, p. 15); GHH case eqs. (33)-(34), p. 19; tractable example p. 24 | Simple GHH example: $$\tau^K = 6\%$$, $$\tau^L = 3\%$$, $$B > 0$$ (parameters: $$\alpha=0.3, \beta=0.7, \varphi=0.3, \delta=1, G=0.01$$) |
| R2 | Zero capital tax with CRRA separable utility, even with binding credit constraints | Corollary 1, p. 18 | $$\tau^K = 0$$ for $$U(c,l) = u(c) - v(l)$$ with CRRA $$u$$; generalizes Chamley-Judd to incomplete markets with occasionally binding constraints |
| R3 | Positive public debt is optimal when savings motive dominates public spending needs | Result 1 (eq. 39), p. 23 | $$B \geq 0$$ iff $$\bar{g}_1 \leq 0$$ and $$G \leq \bar{g}_{\text{pos}} Y_{FB}$$; tractable example: $$B > 0$$ with $$\alpha=0.3, \beta=0.7$$ |
| R4 | Optimal public debt response to spending shock decreasing in shock persistence at fixed NPV | Proposition 5 (p. 26); Figure 3 (p. 43) | $$\partial \hat{B}_0/\partial \rho_G < 0$$; quantitatively: debt rises for $$\rho_G=0.1$$ (1% of GDP shock), falls for $$\rho_G=0.99$$ (0.02% of GDP shock) |
| R5 | Capital tax rises significantly at impact after a public spending shock (both persistence levels) | Figure 1 (panel 4, p. 40-41) | Change in $$\tau^K$$ at impact is an order of magnitude larger than the change in the labor tax level; capital tax increases for both high- and low-persistence shocks |
| R6 | Labor tax progressivity rises and level falls at impact after a spending shock | Figure 1 (panels 2-3, p. 41) | Progressivity $$\tau$$ increases and labor tax level $$\kappa$$ decreases; both changes are much smaller than the capital tax response; public debt path differs markedly by persistence level (panel 5) |

**Overall (paper's conclusion).** The key friction for positive optimal capital taxation is
an occasionally binding credit constraint: it introduces a price externality of savings that
the planner corrects with a positive capital tax. The result fails for CRRA utility because
the externality cancels exactly. For public debt dynamics, shock persistence is the key driver
of the optimal financing structure: transitory shocks call for borrowing (lower future taxes
via smoothing) while persistent shocks call for front-loading adjustment (raising taxes now to
avoid a highly distortionary persistent increase later).

## Theory / model

The economy runs in discrete time. A continuum of $$F$$ ex-ante types of heterogeneous agents
face idiosyncratic productivity risk. A representative firm produces using Cobb-Douglas
technology. The government has access to a linear capital tax, a nonlinear (HSV) labor tax,
and public debt. Aggregate uncertainty enters only through an exogenous public spending path
(an MIT shock), so the economy is otherwise deterministic at the aggregate level.

**Production.** The net-of-depreciation production function is (p. 7, eq. 1):

$$
Y_t = F(K_{t-1}, L_t) = K_{t-1}^\alpha L_t^{1-\alpha} - \delta K_{t-1}. \tag{1}
$$

Factor prices satisfy $$\tilde{w}_t = F_{L,t}$$ and $$\tilde{r}_t = F_{K,t}$$.

**Tax instruments.** The labor tax follows Heathcote, Storesletten, and Violante (2017) (HSV).
An agent earning pre-tax labor income $$\tilde{w} y l$$ pays (p. 8, eq. 2):

$$
T_t(\tilde{w}yl) := \tilde{w}yl - \kappa_t(\tilde{w}yl)^{1-\tau_t}, \tag{2}
$$

where $$\kappa_t$$ governs the level of labor taxation and $$\tau_t \in [0,1]$$ governs progressivity
($$\tau_t = 0$$ is a linear tax; $$\tau_t = 1$$ is full income redistribution). The capital tax
$$\tau_t^K$$ is linear and applied to all interest-bearing assets. Defining post-tax factor
prices (p. 9, eqs. 4-5):

$$
w_t := \kappa_t(\tilde{w}_t)^{1-\tau_t}, \qquad R_t := 1 + r_t = 1 + (1-\tau_t^K)\tilde{r}_t, \tag{4,5}
$$

the government budget constraint in post-tax prices simplifies to (p. 9, eq. 6):

$$
G_t + R_t B_{t-1} + w_t \sum_{f=1}^F m^f \int_i (y_{i,t}^f l_{i,t}^f)^{1-\tau_t} \ell^f(di)
\leq F(K_{t-1}, L_t) - (R_t - 1)K_{t-1} + B_t. \tag{6}
$$

**Agents.** Each agent $$i$$ of type $$f$$ maximizes expected discounted utility (p. 10, eq. 7):

$$
\max_{\{c,l,a\}} \; \mathbb{E}_0 \sum_{t=0}^\infty \beta^t U(c_{i,t}^f, l_{i,t}^f), \tag{7}
$$

subject to the budget constraint (eq. 8) and a borrowing limit $$a_{i,t}^f \geq -\underline{a}$$:

$$
c_{i,t}^f + a_{i,t}^f = R_t a_{i,t-1}^f + w_t(y_{i,t}^f l_{i,t}^f)^{1-\tau_t}. \tag{8}
$$

Denoting by $$\beta^t \nu_{i,t}^f \geq 0$$ the multiplier on the credit constraint, the consumption
Euler equation is (eq. 10):

$$
U_c(c_{i,t}^f, l_{i,t}^f) = \beta \mathbb{E}_t\!\left[ R_{t+1} U_c(c_{i,t+1}^f, l_{i,t+1}^f) \right] + \nu_{i,t}^f. \tag{10}
$$

The labor supply first-order condition is (eq. 11):

$$
-U_l(c_{i,t}^f, l_{i,t}^f) = (1-\tau_t) w_t y_{i,t}^f (y_{i,t}^f l_{i,t}^f)^{-\tau_t} U_c(c_{i,t}^f, l_{i,t}^f). \tag{11}
$$

**Social welfare and Ramsey problem.** The government is a utilitarian planner with type-specific
Pareto weights $$\omega^f$$. Aggregate social welfare is (eq. 14):

$$
W_0 = \sum_{f=1}^F m^f \omega^f \left( \mathbb{E}_0 \sum_{t=0}^\infty \beta^t \int_{i \in I^f} U(c_{i,t}^f, l_{i,t}^f) \ell^f(di) \right). \tag{14}
$$

A Ramsey Equilibrium (RE) is the competitive equilibrium with the highest $$W_0$$ over all
fiscal policies satisfying the government budget constraint. A Stationary Ramsey Equilibrium
(SRE) is an RE in which aggregate quantities, prices, fiscal policy, and public spending are
all constant. At the SRE the planner's FOC for public debt implies the modified golden rule
$$1 + F_K = 1/\beta$$, a condition first derived in the context of optimal capital taxation
under incomplete markets by Aiyagari (1995); this pins down the long-run capital stock
independently of the SWF weights $$\omega^f$$.

**Simple model and the capital tax condition.** Section 3 studies a simplified environment
with deterministic productivity fluctuations (Woodford 1990): two agent types (employed and
unemployed) alternating each period, a linear labor tax $$\tau^L$$, and a zero borrowing limit.
Binding credit constraints affect only the unemployed at the SRE.

The planner's first-order condition (FOC) linking post-tax interest and wage rates at the SRE is
Proposition 1 (p. 15, eq. 29):

$$
\underbrace{1 - \beta R}_{\text{smoothing wedge}}
= \underbrace{\frac{F_L - w}{w}}_{\text{labor wedge}}
\cdot \underbrace{\frac{\sigma_u - \sigma_e + \varsigma^l_{c,e}}{\sigma_e + \frac{1}{\varphi_e} - \varsigma^l_{c,e} + \varsigma^c_{l,e}}}_{\text{net distributional gain}}, \tag{29}
$$

where $$\sigma_e, \sigma_u$$ are the inverses of the intertemporal elasticity of substitution (IES)
for employed and unemployed agents, $$\varphi_e$$ is the Frisch elasticity of labor supply, and
$$\varsigma^l_{c,e}$$, $$\varsigma^c_{l,e}$$ are cross-derivative terms that vanish for separable utility.
The smoothing wedge equals $$\beta(1+F_K-R) = (1-\beta)\tau^K$$, so a positive smoothing wedge
is equivalent to a positive capital tax.

For separable CRRA utility, $$\sigma_u = \sigma_e$$, the numerator vanishes, and hence $$\tau^K = 0$$
(Corollary 1). The capital tax is positive when the IES differs between employed and
unemployed agents (DRRA utility, so $$\sigma_u > \sigma_e$$) or when the utility is non-separable
in a suitable way (GHH, KPR).

**GHH utility.** For the Greenwood-Hercowitz-Huffman utility function (p. 19, eq. 32):

$$
U(c,l) := u\!\left(c - \chi^{-1} \frac{l^{1+1/\varphi}}{1+1/\varphi}\right), \tag{32}
$$

where $$u$$ has constant IES $$1/\sigma$$ and $$\varphi > 0$$ is the Frisch elasticity. In the
log-GHH case (IES = 1), Proposition 1 reduces to a simple relation between the capital and
labor taxes (eq. 34):

$$
(1-\beta)\tau^K = \frac{\tau^L}{1-\tau^L} \varphi(1+\beta). \tag{34}
$$

The capital tax is thus positive whenever the labor tax is positive, and increases with the
discount factor $$\beta$$ and the Frisch elasticity $$\varphi$$.

## Method

The paper applies two computational methods and an identification strategy.

**Factorization approach.** In the general model of Section 4, the Ramsey program is solved
using the factorization method of Marcet and Marimon (2019). This writes the Lagrangian of
the sequential Ramsey problem so that the discounted sum collapses to a single-period term
embedding forward-looking constraints (the agents' Euler equations). The resulting
first-order conditions (FOCs) are derived in Appendix A.6. The capital tax FOC (eq. 60,
p. 32) equates the net distributive gain of a capital tax to the cost imposed on savings
incentives:

$$
\sum_{f=1}^F m^f \int_i \hat{\psi}_{i,t}^f a_{i,t-1}^f \ell^f(di)
= \sum_{f=1}^F m^f \int_i \lambda_{i,t-1}^f u'(x_{i,t}^f) \ell^f(di), \tag{60}
$$

where $$\hat{\psi}_{i,t}^f := \mu_t - \psi_{i,t}^f$$ is the net value to the planner of reallocating
one unit from agent $$(i,f)$$ to public funds, $$\mu_t$$ is the shadow value of government
resources, and $$\lambda_{i,t}^f$$ is the Lagrange multiplier on agent $$(i,f)$$'s Euler equation.
The public-debt FOC implies the modified golden rule at the steady state:

$$
\mu_t = \beta(1+\tilde{r}_{t+1})\mu_{t+1}, \qquad \Rightarrow \quad 1 + F_K = \frac{1}{\beta}. \tag{59}
$$

**Truncation method.** For the quantitative model, the paper uses the truncation approach of
LeGrand and Ragot (2022a) with the refinement of LeGrand and Ragot (2022b), both building on
heterogeneous-agent-bewley-model traditions. The method aggregates agents by their recent
idiosyncratic productivity histories of length $$N$$, replacing the full infinite-dimensional
distribution with a finite number of "representative histories." Histories that are more
frequently visited are given longer truncation lengths (refined truncation), reducing the
state space from exponential to linear in the maximum truncation length.

**Inverse optimal approach.** The Social Welfare Function (SWF) weights $$(\omega^f)_{f=1,\ldots,F}$$
are identified from the observed US fiscal system via an inverse optimal approach following
Bourguignon and Amadeo (2015) and Heathcote and Tsujiyama (2021). Given the calibrated
steady-state fiscal parameters $$(\tau^K, \kappa, \tau, B)$$, the model's FOCs at the SRE are
solved for the unique $$(\omega^f)_{f=1,\ldots,F}$$ consistent with optimality. With $$F=3$$ agent
types, the identification reduces to inverting a 3x3 matrix of FOC constraints.

**Public debt dynamics.** In the simple log-GHH model, capital is the unique state variable
in the linearized dynamics. The optimal capital path after a public spending shock of initial
size $$\hat{G}_0$$ and persistence $$\rho_G$$ is (Result 2, eq. 41):

$$
\hat{K}_t = \rho_K \hat{K}_{t-1} + \sigma_K \hat{G}_t, \qquad \rho_K > 0, \; \sigma_K < 0, \tag{41}
$$

from which the closed-form public debt impulse response function follows (eq. 44):

$$
\hat{B}_t = \hat{G}_0 \!\left(\Theta^K \rho_K^t - \Theta^G \rho_G^t\right). \tag{44}
$$

The impact response $$\hat{B}_0 = \hat{G}_0(\Theta^K - \Theta^G)$$ can be positive or negative
depending on $$\rho_G$$. Proposition 5 (p. 26) proves that $$\partial \hat{B}_0/\partial \rho_G < 0$$
at fixed $$\hat{G}_0$$ and, more importantly, also at fixed NPV of public spending. The
intuition: when persistence is low, the planner borrows to smooth the large transitory shock
and retires debt with a small future tax increase; when persistence is high, the capital stock
falls persistently, making future tax increases very costly, so the planner front-loads fiscal
adjustment without issuing new debt.

## Empirical specifications

The quantitative model (Section 5) is calibrated to the US and solved numerically.

**Parameters and calibration targets.** The period is a quarter. Technology is Cobb-Douglas:
$$F(K,L) = K^\alpha L^{1-\alpha} - \delta K$$ with $$\alpha = 0.36$$ (capital share) and $$\delta = 0.025$$
(depreciation, corresponding to 10% annually), following Krueger, Mittman, and Perri (2018).
The discount factor $$\beta$$ is set to match an annual capital-to-output ratio of 2.7. The GHH
utility has Frisch elasticity $$\varphi = 0.5$$ (recommended by Chetty et al. (2011) for the
intensive margin in heterogeneous-agent models) and scaling $$\chi = 0.05$$ to generate a
steady-state labor supply of roughly $$1/3$$.

**Ex-ante heterogeneity.** Three agent types ($$F=3$$) are distinguished by their ex-ante
productivity processes, corresponding to educational attainment: high-school or less, some
college, and at least a bachelor's degree, with average productivity levels of 0.8, 1, and 2
and population shares of $$1/3$$ each (2022 CPS data, footnote 28, p. 38). Each type follows
an AR(1) log-productivity process:

$$
\log y_t^f = \rho_y^f \log y_{t-1}^f + \varepsilon_t^f, \qquad \varepsilon_t^f \overset{\text{iid}}{\sim} \mathcal{N}(0,(\sigma_y^f)^2),
$$

discretized with five idiosyncratic states per type using Rouwenhorst (1995), yielding 15
productivity levels and 455 truncated histories total after refinement.

**Fiscal calibration.** The capital tax rate $$\tau^K = 36\%$$ is taken from Trabandt and Uhlig
(2011), using the Mendoza, Razin, and Tesar (1994) methodology on US data prior to 2008.
The labor tax progressivity $$\tau = 0.18$$ is from Heathcote, Storesletten, and Violante
(2017). The level $$\kappa$$ is chosen to match a public-spending-to-GDP ratio of $$G/Y = 17\%$$.

**Inverse optimal identification.** At the calibrated steady state, the planner's FOCs
identify the SWF weights as $$\omega^1 = 13.1\%$$, $$\omega^2 = 81.6\%$$, $$\omega^3 = 5.3\%$$ for
the three types (p. 40). These weights are positive, consistent with a sensible SWF, and
sum to 100% by normalization.

**MIT shock specification.** The public spending shock enters as (eq. 40):

$$
\hat{G}_t = \begin{cases} \hat{G}_0 & t = 0 \\ \rho_G \hat{G}_{t-1} & t > 0, \end{cases} \tag{40}
$$

with $$\rho_G \in (-1,1)$$. Two persistence values are studied: $$\rho_G = 0.1$$ (low persistence,
initial shock = 1% of GDP) and $$\rho_G = 0.99$$ (high persistence, initial shock = 0.02%
of GDP), calibrated to the same NPV of public spending (Panel 1, Figure 1, p. 40).

**Robustness.** Results hold under an affine tax system (Appendix A.9, linear labor tax plus
lump-sum transfer as in Dyrda and Pedroni (2022)) and under a productivity-dependent SWF
that assigns weights to instantaneous rather than intertemporal utility (Appendix A.10).
Results for TFP shocks and discount factor shocks are reported in Appendix A.11 and are
qualitatively similar to the public spending shock results.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| US Current Population Survey (CPS) 2022 | Calibration of average earnings for three education groups (high-school, some college, bachelor+); sets relative productivity levels 0.8, 1, 2 | [no page yet] |
| Trabandt and Uhlig (2011) capital tax estimates | Sets steady-state capital tax target $$\tau^K = 36\%$$ (Mendoza-Razin-Tesar 1994 methodology, US pre-2008) | no page yet |
| Heathcote, Storesletten, and Violante (2017) estimates | Sets labor tax progressivity target $$\tau = 0.18$$ | no page yet |

Sample: calibrated to US steady-state fiscal data circa 2007; productivity AR(1) processes estimated to target US income risk moments. Dynamics are first-order perturbations around the calibrated SRE; not estimated from time-series data.

## When to read the full paper

Read the [source](https://doi.org/10.1086/734877) (or [HAL preprint](https://hal.science/hal-05547657)) if you are:
studying the analytical conditions for existence of a stationary Ramsey equilibrium in
heterogeneous-agent models (Propositions 2-3 and Appendices A.3-A.5); building or comparing
quantitative Ramsey optimal policy models for the US (the calibration and truncation method
details are in Sections 5 and Appendix A.7-A.8); or working on the question of whether
capital taxes should rise or fall in response to public spending shocks (the key quantitative
IRFs are in Figures 1-3). The replication code at [doi.org/10.7910/DVN/ZMIFAZ](https://doi.org/10.7910/DVN/ZMIFAZ)
reproduces all tables and figures.

## Attribution and rights

Source: peer-reviewed, *Journal of Political Economy* 133(7), July 2025, pp. 2320-2369.
Published by the University of Chicago Press; paywalled. An open preprint is available at
[hal.science/hal-05547657](https://hal.science/hal-05547657) under CC BY-NC-ND 4.0.

This page was extracted by an LLM (claude-sonnet-4-6) on 2026-06-26 and is **not human-verified
or independently reproduced**. Redistribution of the VOR is not permitted (paywalled); this
page contains extracted summaries only.

> Le Grand, François, and Xavier Ragot. "Optimal Fiscal Policy with Heterogeneous Agents
> and Capital: Should We Increase or Decrease Public Debt and Capital Taxes?"
> *Journal of Political Economy* 133, no. 7 (2025): 2320-2369.
> DOI: [10.1086/734877](https://doi.org/10.1086/734877).


==============================================================================
# Opening Up Military Innovation: Howell, Rathje, Van Reenen & Wong (2025)
# https://instituteforautomatedresearch.org/wiki/papers/jpe/2025/howell-et-al-opening-up-military-innovation-2025/
# Distilled: Using a sharp regression discontinuity design in the U.S. Air Force SBIR program, the paper shows that Open (bottom-up, unspecified) awards increase military technology adoption by 11.4 pp, VC investment by 12 pp, and patenting by 7-9 pp, while Conventional (top-down, specified) awards have no such effects and create program lock-in. Journal of Political Economy 2025, VOR paywalled. Six core results with source locators, datasets used, and the estimating equation.
# Tags: paper-summary, innovation-policy, defense-rd, procurement, sbir
==============================================================================

**What this is.** The paper's core results, the identification strategy, and the estimating
equation: enough to understand what the Air Force SBIR reform found and how it was
identified, without reading all 34 pages. To replicate or extend it, read the full source
at [doi.org/10.1086/737235](https://doi.org/10.1086/737235) or the accepted manuscript at
[LSE Research Online](https://researchonline.lse.ac.uk/id/eprint/128343).

## TL;DR

Should governments procuring innovation specify desired products (a "Conventional" approach)
or allow firms to propose their own ideas (an "Open" approach)? The paper studies a 2018
reform at the U.S. Air Force SBIR program that introduced an Open competition alongside the
existing Conventional one. Using a sharp RDD that exploits the rank-based award rule within
each competition topic, the paper finds that winning an Open award increases military
technology adoption (subsequent non-SBIR DoD contracts) by 11.4 pp, VC investment by 12 pp,
and high-originality patenting by 7 pp. Winning a Conventional award has no positive effects
on any of these outcomes and instead creates program lock-in: it raises the probability of
winning another SBIR award by roughly three times the mean. Three complementary designs
(firm-characteristic controls, specificity variation within Conventional, and firms applying
to both programs) rule out differential firm selection as the explanation. Openness matters
independently from applicant composition.

## Core results

Magnitudes and significance are as reported in the text and tables; `\*` = 10%, `\*\*` = 5%,
`\*\*\*` = 1% (standard errors clustered by topic). Locators refer to the accepted manuscript
pagination.

| # | Result | Locator | Magnitude as reported |
|---|---|---|---|
| R1 | Open award raises probability of subsequent non-SBIR DoD contract (technology adoption) | Table 2 Panel A col 1, pp. 19-20 | Open: +11.4 pp = +0.200-0.086, p = 0.019, 69% of mean; Conventional: -8.6 pp (insig.) |
| R2 | Open award raises probability of VC investment | Table 2 Panel A col 2, p. 19 | Open: +12 pp, >sample mean of 9.2%, sig. at 5%; Conventional: no effect |
| R3 | Open award raises probability of any patent grant | Table 2 Panel A col 3, pp. 19-20 | Open: +8.9 pp, 79% of mean, sig. at 5%`\*\*`; Conventional: negative, weakly sig. |
| R4 | Open award raises probability of high-originality patent | Table 2 Panel A col 4, p. 20 | Open: +7 pp, 194% of mean, sig. at 1%`\*\*\*`; Conventional: no effect |
| R5 | Conventional award creates lock-in (future SBIR); Open does not | Table 2 Panel A col 5, p. 20 | Conventional: positive, ~3x mean (weakly sig.); Open: no effect on future SBIR |
| R6 | Open effects hold for firms that applied to both programs (selection ruled out) | Table 6 panels A-B, pp. 24-25 | Among 507 cross-applicant firms: Open raises DoD contracts by 15.1 pp and high-originality patents by 10.7 pp (controls model); no Conventional effects |

**Overall (paper's conclusion).** The Air Force Open SBIR program succeeded in its stated
objectives: it increased commercial technology adoption by the military, expanded the
nontraditional industrial base (via VC investment), and raised commercial innovation intent
(via patenting). Conventional awards have zero effect on these outcomes and instead create
SBIR-mill incumbency. Openness matters as a program design feature, independently from the
type of firm it attracts: three complementary research designs all point to the same
conclusion. Among non-defense-sector firms, the Open effect on DoD contracts is even larger
(16.4 pp vs. 9.4 pp in the full sample, per Appendix Table E.13).

## Theory / model

The paper has no formal economic model. It frames the Open vs. Conventional comparison
as a principal-agent information problem: the government (DoD) holds a need but imperfect
knowledge of the technological landscape, while firms hold private knowledge of potentially
useful technologies. A Conventional approach forces the government to specify ex ante what
it wants, limiting the signal it sends to firms with unrecognized solutions. An Open
approach delegates identification of solutions to the private sector, allowing firms to
reveal technologies DoD did not know it needed.

The paper draws on two theoretical benchmarks. Belenzon and Cioaca (2021) show that
government R&D contracts (which carry an implicit promise of future downstream procurement)
crowd in private R&D investment; Open appears to activate this channel more effectively
because winning firms can credibly signal to venture capitalists that a large customer
exists for their commercially-oriented technology. Che et al. (2021) show that bundled
approaches in which the innovating firm receives the follow-on contract are optimal for
unsolicited proposals; the Open program's structure matches this prediction and the
positive VC and patent results are consistent with it.

On the SBIR program specifically, Bhattacharya (2021) develops a structural model of
R&D procurement contests in the Navy SBIR; the paper here complements that work with a
causal RDD design at the Air Force focused on program design rather than selection dynamics.

The paper presents three identification arguments that openness matters beyond selection.
First, adding lifecycle and technology fixed effects to Equation (1) leaves the Open
coefficients unchanged (Table 2, Panel B). Second, within the Conventional program,
less-specific topics (measured by cosine-similarity dispersion of proposal text) yield
larger positive effects on patenting, with specific topics yielding significantly negative
effects (Table 5). Third, among firms that applied to both Open and Conventional and thus
share unobservables by construction, only Open awards generate positive outcomes (Table 6).

## Method

The identification exploits a sharp RDD: within each SBIR competition topic, applicants
receive an aggregate evaluation score (sum of three independent sub-scores on Technology,
Team, and Commercialization), and winners are exactly those above a rank threshold
determined by the available budget. Because the cutoff is set independently of the
evaluation process and no single evaluator can manipulate position around it, the running
variable (rank) is as-good-as-randomly assigned near the threshold.

Ranks are normalized within topic so that rank 1 is the lowest-scoring winner and rank -1
is the highest-scoring loser. A triangular kernel weights observations closest to the
cutoff more heavily (p. 17):

$$
\text{Kernel}_{iT} = 1 - \frac{|\text{Rank}_{iT}|}{\max_j |\text{Rank}_{jT}| + 0.01}
$$

The main estimating equation pools Open and Conventional topics and estimates the Open
effect as the interaction of winning with an Open indicator (Equation 1, p. 18):

$$
Y_i = \alpha + \beta_1 \text{Award}_{iT} + \beta_2 \text{Award}_{iT} \cdot \text{Open}_T
$$

$$
+ \gamma_1 [\text{Rank}_{iT} \mid \text{Rank}_{iT} > 0] + \gamma_2 [\text{Rank}_{iT} \mid \text{Rank}_{iT} > 0] \cdot \text{Open}_T
$$

$$
+ \gamma_3 [\text{Rank}_{iT} \mid \text{Rank}_{iT} < 0] + \gamma_4 [\text{Rank}_{iT} \mid \text{Rank}_{iT} < 0] \cdot \text{Open}_T
+ \delta \text{Score}_{iT} + \mathbf{X}_i' \theta + \alpha_T + \varepsilon_{iT} \tag{1}
$$

Here $$\beta_1$$ is the Conventional award effect, and $$\beta_1 + \beta_2$$ is the Open award
effect. $$\text{Open}_T$$ is an indicator for the topic being Open. Topic fixed effects
$$\alpha_T$$ absorb all time-invariant topic characteristics (including the date of award and
the program type per se). Standard errors $$\varepsilon_{iT}$$ are clustered by topic.
Optional controls $$\mathbf{X}_i$$ include firm age, firm size (employees), and 25 narrow
technology-class fixed effects constructed from k-means clustering of proposal abstract
text (Forgy 1965; Bonhomme and Manresa 2015), using a 25-cluster model on word embeddings.

The paper also uses topic-level specificity scores based on `k-means-clustering` of proposal
texts: for each topic, the standard deviation of cosine similarities between individual
proposals and the topic centroid measures how much the topic allows diverse technology
proposals (high SD = open, low SD = specific). This variable is used in Table 5 to show
that more open-style Conventional topics also have larger positive effects on patenting
(§6.2, pp. 22-24).

## Empirical specifications

**Main RDD specification (R1-R5).** The specification is Equation (1) above, estimated
by weighted OLS with the triangular kernel. The main sample is 2,283 unique firms from
the 2017-19 SBIR solicitation periods (restricting to first-time winners for homogeneity).
Outcomes are binary ever-after indicators measured through January 2023, at least 37 months
after the last award. Standard errors are clustered by topic. The coefficient of interest
is $$\beta_2$$ (Award $\times$ Open interaction), which captures the incremental Open effect
over and above the Conventional effect $$\beta_1$$.

Panel A of Table 2 runs Equation (1) with no firm-level controls. Panel B adds firm
lifecycle controls (age, employees) and 25 technology fixed effects. Panel C expands to
all applications from 2003 onwards (firms may appear more than once). Results are similar
across all panels.

**Narrow-bandwidth robustness (Table 7, Panel B).** The sample is restricted to ranks
$$\pm 2$$ around the cutoff (two ranks on each side), so no control for rank is needed.
Results remain significant, supporting the local randomization interpretation and
suggesting the results are not confined to the immediate neighborhood.

**Specificity test within Conventional (Table 5, §6.2).** The sample is restricted to
Conventional topics using data from 2003 onwards. The regression interacts winning with
an indicator for being in a non-specific topic (above the 66th percentile of the
topic specificity distribution). Columns 3-5 show that more open-style Conventional
topics have significantly higher patent effects; highly specified Conventional topics
have a significantly negative patent effect, suggesting over-specification deters
commercialization.

**Cross-applicant design (Table 6, §6.3).** The sample is restricted to 507 firms that
applied to both the Open and Conventional programs. These firms are observationally
similar by construction. The Open award effect on DoD contracts (+15.1 pp) and
high-originality patents (+10.7 pp) is robust within this sample (Table 6 Panel B with
full controls), while Conventional effects remain zero, ruling out selection as the sole
explanation.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Air Force SBIR administrative microdata (proposals, evaluation sub-scores, award decisions, 2003-2019) | Running variable (rank/score), treatment (award), sample frame; obtained via research collaboration (Howell as Special Government Employee) | no page yet |
| Federal Procurement Data System (FPDS) | Non-SBIR DoD contract outcomes (technology adoption); linked to SBIR firms by firm identifier | no page yet |
| Pitchbook, CB Insights, SDC VentureXpert, Crunchbase (VC databases) | Venture capital investment outcome; VC deals matched to SBIR firms | [PitchBook](/wiki/commercial/pitchbook/) (licensed); [Crunchbase](/wiki/commercial/crunchbase/) (licensed) |
| USPTO patent data (granted patents, originality, citations) | Patent and high-originality patent outcomes; patent originality scored per Jaffe and Trajtenberg (2002) | no page yet (`data:uspto`) |
| SBA SBIR award data (all agencies) | Future SBIR outcome (lock-in measure) | no page yet (`data:sbir`) |

Sample: 2,283 unique firms applying 2017-2019 for the first time (main analysis).
Outcomes measured through January 2023 (at least 37 months after the last award).
The extended sample uses 21,365 proposals from 6,701 unique firms, 2003-2019.

## When to read the full paper

Use the [original](https://doi.org/10.1086/737235) or the
[accepted manuscript](https://researchonline.lse.ac.uk/id/eprint/128343) if you are:
designing or evaluating open vs. specified procurement programs in the public or private
sector; studying the SBIR program's innovation effects; comparing the Air Force results
to the DoE SBIR positive results in Howell (2017); examining the theoretical mechanisms
(downstream procurement signaling per Belenzon and Cioaca (2021), bundled follow-on
contracts per Che et al. (2021)); or using the replication data at Harvard Dataverse
([doi.org/10.7910/DVN/78W8M6](https://doi.org/10.7910/DVN/78W8M6)) to extend the analysis.

## Attribution and rights

Source: peer-reviewed, *Journal of Political Economy* 133(11), November 2025.
DOI: [10.1086/737235](https://doi.org/10.1086/737235).

This distillation was extracted by an LLM (paper-distiller, claude-sonnet-4-6) on
2026-06-26 and is **not human-verified or independently reproduced**. The VOR is
paywalled (University of Chicago Press). An accepted manuscript is available under
CC BY 4.0 at LSE Research Online; the PDF mirror is not hosted in this batch.

> Howell, Sabrina T., Jason Rathje, John Van Reenen, and Jun Wong.
> "Opening Up Military Innovation: Causal Effects of Reforms to US Defense Research."
> *Journal of Political Economy* 133, no. 11 (November 2025): 3605-3651.
> DOI: 10.1086/737235. Extract only: reproduction rights not granted for the VOR.


==============================================================================
# Revolutionary Transition: Gay, Gobbi & Goñi (2026)
# https://instituteforautomatedresearch.org/wiki/papers/jpe/2026/gay-et-al-revolutionary-transition-inheritance-change-2026/
# Distilled: The 1793 French inheritance reforms, which abolished impartible inheritance and imposed equal asset partition among all children, reduced completed fertility by 0.60-0.70 children per woman in affected areas, providing the first empirical support for Le Play's (1875) hypothesis that inheritance law drove France's early demographic transition. Journal of Political Economy 2026, paywalled. Eight core results with source locators, datasets used, the theoretical model with equations, and the estimating specifications.
# Tags: paper-summary, demographic-transition, fertility, french-revolution
==============================================================================

**What this is.** A distilled skeleton of Gay, Gobbi, and Goñi (2026), "Revolutionary Transition: Inheritance Change and Fertility Decline," Journal of Political Economy 134(6): 1666-1713. Read the [original paper](https://doi.org/10.1086/739821) to replicate or extend; a freely available preprint is at <https://hal.science/hal-04285818v1/document>. Extracted by LLM; not human-verified; not reproduced.

## TL;DR

The 1793 French inheritance reforms unexpectedly abolished impartible inheritance (primogeniture and unigeniture) and imposed equal partition of assets among all children, including women. Gay, Gobbi, and Goñi (2026) test Le Play's (1875) 150-year-old hypothesis that these reforms drove France's early fertility transition. They build the first complete atlas of pre-Revolutionary inheritance customs at the judicial-district level (141 customs, 435 districts), link it to individual fertility data from two independent sources (the Enquête Louis Henry and Geni.com genealogies), and use difference-in-differences (DD) and regression-discontinuity difference-in-differences (RD-DD) designs. The DD results show that each additional fertile year of exposure to the 1793 reforms reduced completed fertility by roughly 1 percent (0.024-0.028 children per year), with a cumulative effect of 0.60-0.70 children over a full fertile cycle. This effect is almost identical to the pre-reform fertility gap between impartible and partible inheritance areas (0.68-0.75 children), confirming that the reforms brought high-fertility impartible regions to the same low-fertility equilibrium as partible regions, and thereby sharply accelerating France's demographic transition.

## Core results

| \# | Result | Locator | Magnitude as reported |
|---|---|---|---|
| R1 | DD: completed fertility response to 1793 reforms per fertile year of exposure | Table 2, Panel A, p. 33 | -0.024 to -0.028 (p<0.001); over 25-year cycle: -0.60 to -0.70 children |
| R2 | Pre-reform fertility gap: impartible vs partible areas | Table 2, Panel A, p. 33 | +0.682 to +0.748 more children in impartible areas (p<0.001) |
| R3 | DD: completed fertility of mothers (intensive margin) | Table 2, Panel B, p. 33 | -0.020 to -0.022 per fertile year (p<0.01); cumulative: -0.50 to -0.55 children |
| R4 | DD: childlessness rate (extensive margin) | Table 2, Panel C, p. 33 | +0.003 per fertile year (p<0.001); +7.5 pp over full cycle |
| R5 | Flexible-trend DD: completed fertility (controls for economic, religious, political, geography trends) | Table 3, cols (1)-(4), p. 35 | -0.026 to -0.031 per fertile year (p<0.001); cumulative: -0.65 to -0.78 children |
| R6 | Flexible-trend DD: age at first marriage | Table 3, col (7), p. 35 | +0.073 years per fertile year (p<0.01); +1.8 years over full cycle |
| R7 | RD-DD (Geni data): completed fertility of mothers at spatial inheritance border | Table 4, p. 47 | -0.032 to -0.054 per fertile year (p<0.01); cumulative: -0.75 to -1.25 children |
| R8 | RD pre-reform gap (Geni data): fertility discontinuity at inheritance border | Figure 9, Panel A, p. 46 | b = 0.78 (se 0.21, p<0.001): ~0.78 more children in impartible areas before reform |

**Overall (paper's conclusion).** The 1793 inheritance reforms contributed to France's early fertility decline by eliminating the economic incentives for high fertility in formerly impartible inheritance areas. The reform effect is robust across two independent datasets (Henry family-reconstitution and Geni crowdsourced genealogies), two identification strategies (nationwide DD and spatial RD-DD), and extensive controls for economic conditions, religiosity, political factors, and geography. The convergence in fertility across regions was large: the reforms brought roughly half of France to the low-fertility regime that already prevailed under partible inheritance.

## Theory / model

The paper develops a parsimonious model of endogenous fertility under different inheritance rules, formalizing Le Play's (1875) hypothesis. The model features households that care about own consumption and total child endowments, following de la Croix and Doepke (2003) warm-glow altruism (Section 3, p. 12). Unlike Becker and Lewis (1973) quantity-quality tradeoff models where human capital is the key margin, this model focuses on the land indivisibility constraint as the mechanism linking inheritance rules to fertility.

**Utility function** (Eq. 1, p. 12):

$$u(c_i, n_i) = \ln c_i + \beta \ln(n_i y'_i) \tag{1}$$

where $$n \geq 1$$ is the number of children, $$y'$$ is each child's income, and $$\beta > 0$$ weights utility from the next generation's endowment.

**Budget constraint** (Eq. 2, p. 12):

$$c_i = (1 - \phi n_i) y_i \tag{2}$$

where $$\phi \in (0,1)$$ is the fixed cost of raising each child as a share of household income.

**Production function** (Eq. 3, p. 13): Household income is determined by land $$L$$ and labor $$N$$ via a Stone-Geary function with a minimum land threshold $$\bar{L}$$:

$$f(L, N_i) = \begin{cases} 0 & \text{for } L \leq \bar{L} \\ (L - \bar{L})^{1-\alpha} N_i^{\alpha} & \text{otherwise} \end{cases} \tag{3}$$

The threshold $$\bar{L} > 0$$ captures land indivisibility: below it, production is zero. The parameter $$\alpha \in (0,1)$$ is the relative importance of labor.

**Inheritance rules** (Eq. 4, p. 14): Under partible inheritance ($$i = P$$), land is divided equally among $$n_P$$ children, each forming their own household. Under impartible inheritance ($$i = I$$), land is kept whole and the extended family (heir plus $$n_I$$ siblings) works it together:

$$L'_P = \frac{L}{n_p}, \quad N_P = 1, \quad y'_P = f\!\left(\frac{L}{n_p}, 1\right), \quad L'_I = L, \quad N'_I = n_I, \quad y'_I = \frac{f(L, n_I)}{n_I} \tag{4}$$

**Assumption 1** (Eq. 5, p. 14) ensures pre-industrial fertility above replacement:

$$\phi < \frac{\alpha\beta}{\alpha\beta + 1} \tag{5}$$

**Equilibrium fertility** (Eqs. 6-8, p. 14): Optimal fertility differs by inheritance regime when land is in the intermediate range $$\bar{L} < L < \tilde{L}$$, where indivisibility binds under partible but not impartible inheritance. Let $$\tilde{L} \equiv \frac{((1+\beta)\phi - \beta)\bar{L}}{\phi - \alpha\beta(1-\phi)}$$ and $$\Delta \equiv (\beta\bar{L} + (1+\alpha\beta)\phi L)^2 - 4\alpha\beta(1+\beta)\phi\bar{L}L$$:

$$\text{If } L \leq \bar{L}: \quad n^*_I = n^*_P = 1 \tag{6}$$

$$\text{If } \bar{L} < L < \tilde{L}: \quad n^*_I = \frac{\alpha\beta}{(1+\alpha\beta)\phi} \text{ and } n^*_P = 1 \tag{7}$$

$$\text{If } L \geq \tilde{L}: \quad n^*_I = \frac{\alpha\beta}{(1+\alpha\beta)\phi} \text{ and } n^*_P = \frac{\beta\bar{L} + (1+\alpha\beta)\phi L - \sqrt{\Delta}}{2(1+\beta)\phi\bar{L}} \tag{8}$$

**Proposition 1** (p. 15): Fertility is strictly higher under impartible inheritance than under partible inheritance (for $$L > \bar{L}$$).

The intuition: under partible inheritance, dividing a plot of productive land among many heirs can push per-heir land below $$\bar{L}$$, making production zero. Parents internalize this and limit family size. Under impartible inheritance, land is kept whole, remaining productive regardless of family size. The reform abolishing impartible inheritance in 1793 sets $$\theta = 0$$ (the share of households under impartible rules), moving the economy to the low-fertility partible equilibrium.

The model also rationalizes why similar reforms did not reduce fertility in England or Prussia, where large concentrated landholdings meant the indivisibility constraint would not bind for most heirs.

## Method

The empirical strategy combines two complementary identification designs:

**1. Difference-in-differences (Section 5, pp. 24-28).** The primary design exploits regional variation in pre-reform inheritance systems (treated: formerly impartible areas where the 1793 reforms changed the system; control: already-partible areas where the reforms were immaterial) interacted with cohort exposure to the reforms. Exposure is continuous: $$F_c$$ is the remaining number of fertile years after 1793 for a woman in cohort $$c$$, rising linearly from 0 for women aged 40+ in 1793 (who completed their fertile cycle before the reforms) to 25 for women aged 15 in 1793 (whose entire fertile cycle lay after the reforms). This "treatment intensity" design, building on difference-in-differences, avoids the binary classification of exposure and tests whether effects grow proportionally with reform exposure.

**2. Regression-discontinuity difference-in-differences (Section 7, pp. 41-47).** The supplementary design (using Geni data) restricts to women born near borders between judicial districts that had different pre-reform inheritance systems. It compares the fertility gap at the border for cohorts fertile before the reforms versus after, following Avdic and Karimi (2018). MSE-optimal bandwidths and triangular kernel functions are used; robustness checks span bandwidths of 15-30 km and both linear and quadratic polynomial fits.

**Standard errors** are clustered at the municipality level throughout (39 clusters for Henry; hundreds to over 1,000 clusters for Geni). The religiosity index $$R_m$$ (Eq. 9, p. 23) used as a control variable is constructed from lent and advent marriages after 1792:

$$R_m = \frac{\text{Lent and advent marriages}}{\text{All marriages}} \times \frac{365.25}{46 + \text{days advent}} \tag{9}$$

## Empirical specifications

**Baseline DD** (Eq. 10, p. 24): individual-level regression on completed fertility with cohort fixed effects:

$$Y_{icm} = \alpha + \beta \, I_m \times F_c + \gamma \, I_m + \mu_c + \mathbf{X}'_i \theta + \varepsilon_{icm} \tag{10}$$

where $$Y_{icm}$$ is completed net fertility of woman $$i$$ in municipality $$m$$ born in cohort $$c$$; $$I_m$$ is an indicator for municipalities that had impartible inheritance before 1793 (the treatment group); $$F_c$$ is the number of fertile years remaining after 1793 for cohort $$c$$; $$\mu_c$$ are birth-cohort fixed effects; and $$\mathbf{X}_i$$ includes individual controls (literacy indicators, accuracy of Henry form, parents-in-law alive at marriage). The coefficient of interest is $$\beta$$, capturing the effect of exposure to the 1793 reforms. Standard errors are clustered by municipality (39 clusters). Results shown in Table 2 (p. 33).

**Extended DD with flexible trends** (Eq. 11, p. 27): adds municipality-level wheat prices by decade ($$p_{mc}$$) and allows fertility trends to differ across municipalities by religious, political, and economic-geography characteristics ($$\mathbf{Z}_m$$ interacted with cohort dummies):

$$Y_{icm} = \alpha + \beta \, I_m \times F_c + \gamma \, I_m + \mu_c + p_{mc} + \mathbf{X}'_i \theta + \sum_t \mathbf{1}[c=t] \times \mathbf{Z}'_m \delta_t + \varepsilon_{icm} \tag{11}$$

$$\mathbf{Z}_m$$ includes distance to religious centers, political societies, rebellions 1779-89, legal centers, fiscal centers, territorial administrative centers, paved roads, and horse posts. Results shown in Table 3 (p. 35).

**RD design** (Eq. 13, p. 42): restricts the Geni sample to mothers born within an MSE-optimal bandwidth of the partible-impartible inheritance border; $$d_m$$ is signed distance to the border (positive in impartible areas); $$\phi_b$$ are border-segment fixed effects:

$$Y_{icm} = \alpha + \beta \mathbf{1}[d_m \geq 0] + \phi_b + \mu_c + \mathbf{1}[d_m \geq 0] \times f_I(d_m, B_I) + \mathbf{1}[d_m < 0] \times f_P(-d_m, B_P) + \varepsilon_{icm} \tag{13}$$

**Combined RD-DD** (Eq. 14, p. 44): interacts the spatial RD with treatment intensity $$F_c$$ and allows polynomial fits to differ across pre- and post-reform sub-samples ($$S_c = 1$$ for cohorts completing the fertile cycle before 1793; $$S_c = 2$$ for those fertile after):

$$Y_{icm} = \alpha + \beta \mathbf{1}[d_m \geq 0] \times F_c + \gamma \mathbf{1}[d_m \geq 0] + \phi_b + \mu_c + \mathbf{Z}'_{mc} \delta_c + \sum_{s=1}^{2} \mathbf{1}[S_c=s] \times \!\left\{ \mathbf{1}[d_m \geq 0] \times f_I(d_m, B_{Is}) + \mathbf{1}[d_m < 0] \times f_P(-d_m, B_{Ps}) \right\} + \varepsilon_{icm} \tag{14}$$

Identification relies on: (1) regional variation in pre-reform inheritance systems (rooted in Germanic legal traditions, historically unrelated to economic conditions); (2) rapid take-up of the 1793 reforms enforced by family tribunals; and (3) exogeneity of the reforms to fertility concerns (inheritance was not among the grievances in the 1789 Estates General and fertility was not an objective of the reformers). Parallel trends are confirmed in Figure 6 (p. 26) for the Henry data and Appendix Figure B11 for the Geni data.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Enquête Louis Henry (Henry database) | Primary fertility data: 20,332 women born 1700-1803 in 39 rural French municipalities; baseline for DiD estimates | no page yet |
| Geni.com (crowdsourced genealogies) | Supplementary fertility data: 11,649 women born 1700-1810, 2,966 locations across France; used for the RD-DD design covering broader territory | no page yet |
| Atlas of inheritance customs in Ancien Regime France (hand-collected by authors) | Treatment variable: complete map of 141 inheritance customs across 435 judicial districts at the eve of the Revolution, coded partible/impartible and whether women were included; available on Harvard Dataverse (Gay, Gobbi, and Goni 2023b, 2023d) | no page yet |

The Henry data cover 39 rural municipalities with 100-2,000 individual observations each. The sample average of net completed fertility is 2.35 children. The Geni data use the horizontal restriction of Blanc (2023a) to correct for genealogy over-representation of single-child families. Both datasets record net fertility (children surviving to age 6). Wheat prices from Ridolfi (2019) provide municipality-decade proxies for local economic conditions.

## When to read the full paper

Read Gay, Gobbi, and Goni (2026) if you are studying the causes of the demographic transition, especially for France or Europe's early fertility decline. The paper is essential for researchers interested in: legal institutions as drivers of household behavior; inheritance law and its economic consequences (Table 1 balance tests, Table 2-3 DD estimates, Table 4 RD-DD); the quantity-quality tradeoff as a framework contrasted with institutional explanations; and historical demography using the Henry database or Geni genealogies. The inheritance atlas (Harvard Dataverse) is independently useful for studying economic legacies of Ancien Regime France.

## Attribution and rights

Gay, Victor, Paula E. Gobbi, and Marc Goñi. 2026. "Revolutionary Transition: Inheritance Change and Fertility Decline." *Journal of Political Economy* 134(6): 1666-1713. <https://doi.org/10.1086/739821>

Published by University of Chicago Press. The journal article is paywalled; no open license in Crossref metadata. A preprint is freely available under CC BY 4.0 at <https://hal.science/hal-04285818v1/document> (HAL open science, hal-04285818v4). The inheritance atlas data are available on Harvard Dataverse under open licenses (Gay, Gobbi, and Goni 2023b, 2023d).

This page is an LLM-distilled summary (extract-only); it is not human-verified and the results have not been reproduced. Read the original paper before relying on specific numbers.


==============================================================================
# Enlightenment Ideals and Belief in Progress: Almelhem et al. (2026)
# https://instituteforautomatedresearch.org/wiki/papers/qje/2026/almelhem-enlightenment-ideals-belief-progress-2026/
# Distilled: Using LDA topic modeling and sentiment analysis on 264,443 English volumes printed 1500-1900, this paper documents that science-language volumes secularized by the mid-eighteenth century, that those at the nexus of science and political economy became the most progress-oriented during the Enlightenment, and that industrial volumes at this nexus were the most progress-oriented from the mid-eighteenth century onward. QJE 2026, CC BY 4.0. Five core results with source locators, datasets used, the classification and sentiment methods with equations, and the estimating specifications.
# Tags: paper-summary, text-as-data, economic-history, panel-data, panel-regression
==============================================================================

**What this is.** The paper's core results, the LDA-based classification method it applies, and the two estimating regressions with real equations: enough to know what it found and how, without reading the full 52 pages. To replicate or extend, read the original at [https://doi.org/10.1093/qje/qjaf054](https://doi.org/10.1093/qje/qjaf054).

## TL;DR

The paper applies Latent Dirichlet Allocation to 264,443 English volumes from the HathiTrust Digital Library (printed in England, 1500-1900) to trace how the languages of science, religion, and political economy evolved in the centuries leading to the British Industrial Revolution. Three findings emerge. First, the languages of science and religion diverged in the mid-eighteenth century: science volumes that had used roughly 30% religious language in the early eighteenth century used only about 10% by 1850. Second, regression analysis shows that volumes using language at the nexus of science and political economy became the most progress-oriented beginning in the late seventeenth century, while volumes using purely scientific language were largely neutral. Third, within this nexus, those that also used the language of industrialization were the most progress-oriented from the mid-eighteenth century onward. The findings support Mokyr (2016)'s Industrial Enlightenment thesis: it was pragmatic, industrially oriented scientific writing aimed at a broad literate audience, not elite scientific discourse, that carried progress-oriented culture into Britain's economic take-off.

## Core results

Magnitudes and descriptions are as reported; locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Languages of science and religion became **distinct in the mid-eighteenth century**; science volumes ceased to use religious language | Figure II, p. 284; Figure III, p. 286 | Science volumes used ~30% religious language in the early 18th century, declining to ~10% by 1850; by 1750 essentially no volumes sit at the science-religion vertex of the language simplex |
| R2 | **Average progress sentiment rose** from the mid-seventeenth century and persisted through the period | Figure V, p. 291 | Progress score (percentile) rose from ~10th-15th percentile pre-1650 to ~50th-60th percentile by 1800-1850 |
| R3 | **Science-political economy nexus** volumes were the most progress-oriented from ~1700 onward | Figure VII, p. 295; Online Appendix Table B.1 | Predicted progress score for 50%/50% science-political economy mix is highest among all language combinations beginning late 17th century; volumes using purely scientific or purely religious language score lower |
| R4 | **Industrial language** at the science-political economy nexus amplified progress orientation from the mid-eighteenth century | Figure XI, p. 305; Online Appendix Table B.4 | Within the 50%/50% science-political economy nexus, volumes at the 75th percentile of industrial language had approximately 2x the predicted progress score of zero-industry volumes by 1800 |
| R5 | Pattern is **robust to alternative category definitions**; replacing political economy with law or economics yields the same 18th-century rise | Figure VIII, p. 298 | Science-law and science-economics nexus volumes both show a sharp 18th-century rise in predicted progress sentiment, mirroring the science-political economy finding; the pattern does not emerge for arts and literature |

**Overall (paper's conclusion).** The results are consistent with Mokyr (2016)'s claim that the Industrial Enlightenment diffused progress-oriented views of science into industry and political economy. It was the literate artisan and applied-science audience, not the elite scientific community, whose language became most progress-oriented in the run-up to Britain's industrialization.

## Theory / model

The paper has no formal economic model. The empirical analysis tests three subsidiary hypotheses derived from Mokyr (2016) and Mokyr (2009)'s Industrial Enlightenment and Culture of Growth theses:

1. The language of science and religion became increasingly distinct during the Enlightenment (secularization of science).
2. The language of science became more progress-oriented during the Enlightenment, with the effect concentrated at the nexus of science and political economy rather than in pure scientific discourse.
3. Volumes using the language of industrialization at the science-political economy nexus were particularly progress-oriented in the period before and during Britain's Industrial Revolution.

**Identification strategy.** The analysis is descriptive: the regressions are accounting exercises documenting how progress-oriented language correlates with category weights and industrial language over time. The authors explicitly state that the regressions "are not meant to imply a causal relationship, as omitted variable biases and reverse causation may be present" (p. 293, p. 305). The evidence is structural in the sense of testing whether the pattern predicted by Mokyr (2016) is present in the data, but causality is not claimed.

## Method

The method builds on the `lda-topic-model` technique applied to historical text corpora. Related work using this approach includes Erikson (2021), who applies LDA and sentiment analysis to political and economic tracts from England, 1550-1720, and Grajzl and Murrell (2024), who study English print culture across 1530-1700; both cover shorter time windows than the 400-year corpus here.

### LDA topic model

The corpus is a document-term matrix $$D \times V$$, where $$D$$ is the number of volumes and $$V$$ is the vocabulary size. LDA (Blei, Ng, and Jordan (2003)) models each volume as a mixture over $$T = 60$$ topics and each topic as a multinomial distribution over words. The optimal $$T$$ is chosen by 4-fold cross-validation on perplexity (Section II.D, p. 276-277). The output is, for each volume $$v$$ and topic $$t$$, a weight $$\alpha_{t,v}$$ representing how strongly the topic appears in that volume, with $$\sum_{t=1}^{60} \alpha_{t,v} = 1$$ for each volume.

### Topic categorization and volume classification

Topics are grouped into three categories (science, religion, political economy) based on topic-pair co-occurrence. For each topic pair $$i \in \{1, \ldots, 1{,}770\}$$ and each volume $$v$$, let $$w_{iv}$$ be the product of the two topics' weights. The corpus-wide share of topic-pair $$i$$ is (equation 1, p. 279):

$$
\text{Share}_i = \frac{\sum_{v=1}^{V} w_{iv}}{\sum_{i=1}^{I} \sum_{v=1}^{V} w_{iv}} \tag{1}
$$

Categories are identified as the triplets of topics with the highest total share (Incidence) that are sufficiently distinct from each other. The three resulting categories are science (topics 3, 41, 43), religion (topics 10, 34, 38), and political economy (topics 13, 35, 36); see Table I, p. 282.

Each volume is assigned weights for all three categories by weighting the LDA topic weights by each topic's time-varying category coefficient (equations 2-4, p. 285):

$$
\text{Science}_v = \sum_{t=1}^{60} \alpha_{t,v}\,\beta_{t,\text{Science}} \tag{2}
$$

$$
\text{PolitEcon}_v = \sum_{t=1}^{60} \alpha_{t,v}\,\beta_{t,\text{PolitEcon}} \tag{3}
$$

$$
\text{Religion}_v = \sum_{t=1}^{60} \alpha_{t,v}\,\beta_{t,\text{Religion}} \tag{4}
$$

where $$\beta_{t,c}$$ is the category coefficient of topic $$t$$ for category $$c \in \{\text{Science, PolitEcon, Religion}\}$$, computed from the time-varying topic-pair shares over 20-year moving bins. By construction $$\text{Science}_v + \text{PolitEcon}_v + \text{Religion}_v = 1$$.

### Sentiment (progress-oriented score)

A progress dictionary (Table II, p. 289) lists 7 modern English synonyms of "progress" (progress, improvement, stride, betterment, advance, rise, amelioration), all in use before 1643 per the Oxford English Dictionary. The progress score for volume $$i$$ is the share of dictionary words in the volume (equation 5, p. 289):

$$
\text{Sentiment}_i = \frac{\sum_{\ell \in L} w_{i,\ell}}{W_i} \tag{5}
$$

where $$w_{i,\ell}$$ is the count of word $$\ell$$ from the progress dictionary $$L$$ in volume $$i$$, and $$W_i$$ is the total word count of volume $$i$$. Scores are converted to percentile ranks over the full corpus for comparability.

### Industrial score

An industrial score is constructed from the weighted index of machine-related root words transcribed from the five volumes of *Appleby's Illustrated Handbook of Machinery* (Appleby 1877-1903). The top 10 industrial words (by index frequency) include: crane (51), electr (42), weight (37), rope (27), cost (27); see Table IV, p. 301. Each volume's industrial score is the normalized sum of industrial word counts weighted by each word's Appleby index frequency.

## Empirical specifications

### Regression 1: Progress sentiment and language category weights (eq. 6, p. 294)

Volumes are placed into 20-year bins by publication date. The baseline estimating equation is:

$$
\text{Sentiment}_{v,t} = \alpha_1 + \alpha_2\,\text{Science}_v + \alpha_3\,\text{PolitEcon}_v + \alpha_4\,\text{Science}_v \times \text{PolitEcon}_v + \alpha_5\,\text{Science}_v \times \text{Religion}_v + \alpha_6\,\text{Religion}_v \times \text{PolitEcon}_v + \lambda_t + \lambda_t\,\mathbf{A}_{v,t}\,\boldsymbol{\alpha} + \varepsilon_{v,t} \tag{6}
$$

where $$\text{Sentiment}_{v,t}$$ is the progress score (percentile) of volume $$v$$ in bin $$t$$; $$\text{Science}_v$$, $$\text{Religion}_v$$, $$\text{PolitEcon}_v$$ are the volume's category weights from equations (2)-(4); $$\text{Religion}_v$$ is excluded as the reference category; $$\lambda_t$$ are 20-year bin fixed effects; and $$\mathbf{A}_{v,t}$$ is the vector of all variables and interactions in equation (6), with time-varying slope $$\lambda_t\,\boldsymbol{\alpha}$$ allowing all coefficients to change across bins. Standard errors are clustered by year of publication. Full results are in Online Appendix Table B.1; predicted values for key language mixes are plotted in Figure VII (p. 295).

### Regression 2: Progress sentiment, category weights, and industrial language (eq. 7, p. 303)

The industrial score is added as an additional regressor with all two-way and three-way interactions:

$$
\text{Sentiment}_{v,t} = \beta_1 + \beta_2\,\text{Science}_v + \beta_3\,\text{PolitEcon}_v + \beta_4\,\text{Industry}_v + \beta_5\,\text{Science}_v \times \text{PolitEcon}_v + \beta_6\,\text{Science}_v \times \text{Religion}_v + \beta_7\,\text{Religion}_v \times \text{PolitEcon}_v + \beta_8\,\text{Science}_v \times \text{Industry}_v + \beta_9\,\text{PolitEcon}_v \times \text{Industry}_v + \beta_{10}\,\text{Science}_v \times \text{Religion}_v \times \text{Industry}_v + \beta_{11}\,\text{Science}_v \times \text{PolitEcon}_v \times \text{Industry}_v + \beta_{12}\,\text{Religion}_v \times \text{PolitEcon}_v \times \text{Industry}_v + \lambda_t + \lambda_t\,\mathbf{B}_{v,t}\,\boldsymbol{\beta} + \varepsilon_{v,t} \tag{7}
$$

where $$\text{Industry}_v$$ is the normalized industrial language score of volume $$v$$ and all other notation follows equation (6). As before, all slope coefficients are interacted with bin fixed effects to allow time variation. Full results are in Online Appendix Table B.4; predicted values for the 50%/50% science-political economy location at varying industry percentiles are plotted in Figure XI (p. 305).

Both regressions are run on pre-1650 data excluded in robustness checks (Online Appendix Figure B.19, B.27); results are similar. Alternative dictionaries using 1708 *Dictionarium Anglo-Britannicum* progress words and a ChatGPT-generated Enlightenment-era synonym list also yield similar patterns (Online Appendix Figures B.10-B.13).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| HathiTrust Digital Library (HDL) extracted features, 264,443 volumes (Almelhem et al. 2025, Harvard Dataverse) | Main corpus: bag-of-words representation of all English volumes printed in England, 1500-1900; used for LDA and sentiment analysis | [no page yet](/wiki/) |
| *Appleby's Illustrated Handbook of Machinery*, vols. 1-5 (Appleby 1877-1903) | Source of industrial root-word index used to construct per-volume industrial scores | no page yet |

Sample: 264,443 unique English volumes printed in England, 1500-1900, after removing duplicates and non-English volumes from an initial set of 420,081. Data available at Harvard Dataverse: [https://doi.org/10.7910/DVN/DQRO8L](https://doi.org/10.7910/DVN/DQRO8L) (Almelhem et al. 2025).

## When to read the full paper

Use the [original](https://doi.org/10.1093/qje/qjaf054) if you are: replicating the LDA estimation or the sentiment regressions (Online Appendix Sections A-H contain the full data-cleaning protocol, all 60 topic definitions, robustness checks with alternative dictionaries and unbinned data, and author fixed-effect specifications); extending the corpus to other European languages to test the McCloskey (2006) thesis; or using the qualitative volume examples in Section VI (Clare 1735, Saul 1735, Stephenson 1831) to understand what progress-oriented industrial language looked like in practice.

## Attribution and rights

Source: peer-reviewed, *The Quarterly Journal of Economics* 141(1), 2026. This distillation was extracted by an LLM on 2026-06-28 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch. Replication data are available at Harvard Dataverse (Almelhem et al. 2025), DOI: [10.7910/DVN/DQRO8L](https://doi.org/10.7910/DVN/DQRO8L).

> **Attribution (CC BY 4.0).** Almelhem, Ali, Murat Iyigun, Austin Kennedy, and Jared Rubin.
> "Enlightenment Ideals and Belief in Progress in the Run-up to the Industrial Revolution:
> A Textual Analysis."
> *The Quarterly Journal of Economics* 141, no. 1 (2026): 263-314.
> DOI: 10.1093/qje/qjaf054. (c) The Author(s) 2025.
> Published by Oxford University Press on behalf of President and Fellows of Harvard College.
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Vanguard: Ang & Chinoy (2026)
# https://instituteforautomatedresearch.org/wiki/papers/qje/2026/ang-vanguard-black-veterans-civil-2026/
# Distilled: using random variation from the WWI draft lottery and millions of digitized military and NAACP records, Ang and Chinoy provide the first causal evidence that military service nearly tripled Black veterans' likelihood of joining the NAACP, driven by institutional discrimination rather than socioeconomic gains. The Quarterly Journal of Economics 141(1), 2026, paywalled. Seven core results with source locators, datasets used, the IV design, and the estimating equations.
# Tags: paper-summary, civil-rights, race-discrimination, economic-history, political-economy
==============================================================================

**What this is.** The paper's core results, the identification strategy (WWI draft lottery as instrument for veteran status), the estimating equations, and the mechanism tests: enough to know what the paper found and how, without reading all 50 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1093/qje/qjaf046).

## TL;DR

Nearly 400,000 Black men were drafted into the National Army during World War I, assigned primarily to racially segregated labor units under white supervisors. Ang and Chinoy (2026) leverage random variation from the WWI draft lottery and millions of digitized military records, NAACP membership rosters, and census data to estimate the causal effect of military service on postwar civil rights activism. Two-stage least squares (2SLS) estimates show that Black men randomly induced to enlist were nearly three times more likely to join the nascent NAACP than comparable nonveterans from the same draft board (baseline mean 1.6% vs implied veteran rate 4.4%), with a first-stage F-statistic of 565. A back-of-the-envelope calculation implies the draft induced approximately 10,000 Black men to join the NAACP, about 20% of the organization's 1940 membership.

These effects are not explained by residential mobility or improved socioeconomic status: near-zero effects on literacy, income, and employment are consistent with the limited training and scant postwar benefits provided to Black WWI soldiers (in contrast to the human capital gains documented for Vietnam-era veterans by Angrist and Chen (2011) and for modern volunteers by Greenberg et al. (2022)). Rather, effects increase monotonically with measures of discriminatory treatment in the draft and in military camps, consistent with the historical account of Williams (2007) that perceptions of injustice catalyzed Black postwar activism in the New Negro era.

## Core results

Magnitudes and significance are as reported; `\*\*` = 5%. All 2SLS specifications instrument veteran status with the scaled draft lottery order number and include draft board fixed effects; standard errors clustered by serial number.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Military service nearly **tripled NAACP membership** among Black draftees relative to comparable nondraftees from the same board | Table III, col. 4, p. 817 | 2SLS: 0.0282\*\* (SE 0.0118); baseline nonveteran mean 1.6%, implied veteran rate 4.4%; first-stage F = 565.1 |
| R2 | Draftees were induced into **lasting NAACP participation**: 1 in 100 Black draftees participated for more than 7 years | Figure IV, p. 820 | 2SLS coefficient ~0.01 for 7+ years of membership; 95% CI excludes zero through at least 7 years |
| R3 | Military service nearly **tripled the probability of becoming a historically prominent community leader** in the New Negro era | Online Appendix Table A.III, p. 820 | 2SLS point estimate comparable to NAACP membership effect; similar magnitude across both biographical databases |
| R4 | **Near-zero socioeconomic effects**: serving in WWI did not improve Black soldiers' literacy, home ownership, employment, or income | Figure V, Panel A, p. 823 | All 2SLS estimates statistically insignificant; CIs rule out occupational income gains above 0.10 SD (~$88 in 1950 dollars) |
| R5 | NAACP effects increase **monotonically with draft board racial bias**: near-zero for equitable boards, large for the most discriminatory | Figure VIII, p. 830 | 2SLS: ~0 for Q1 (smallest Black-white induction rate gap); ~0.08 for Q4 (largest gap) |
| R6 | Large NAACP effects for men at **most discriminatory camps** (denying Black soldiers training and promotion); near-zero for less discriminatory camps | Figure IX, p. 832 | 2SLS ~0.10 for camps with 0-25% Black NCO share or no Black troop training; ~0 when majority of NCOs are Black or training was provided |
| R7 | **Injustice is the sole survey theme** significantly predictive of voter participation: veterans citing injustice were 24 percentage points more likely to vote | Figure XI, p. 836 | OLS coefficient ~0.24 on the injustice theme; all other themes statistically insignificant or borderline; robust to controls for birth year, birth state, county, and prior military experience |

**Overall (paper's conclusion).** Military service causally increased Black Americans' participation in the early civil rights movement. The effect is not driven by skill acquisition or economic improvement, but by soldiers' encounters with systematic racial discrimination in the draft and in segregated military camps, which catalyzed organized political resistance in the New Negro era. The draft likely induced approximately 10,000 Black men to join the NAACP, accounting for roughly 20% of its 1940 membership.

## Theory / model

The paper has no formal economic model. It tests two competing mechanisms for how military service could affect civil rights activism:

**Mechanism 1 (Human capital / socioeconomic mobility).** Military service provides formal training, organizational skills, income, and social networks that lower the cost of civic participation. This predicts positive effects on income and education alongside NAACP membership, and effects concentrated among soldiers who received more training.

**Mechanism 2 (Discrimination-induced radicalization).** Experiences of institutional racism in the draft process (racial bias in induction rates, denial of exemptions) and in segregated military camps (exclusion from training, leadership, and equal treatment) catalyze perceptions of injustice and political resistance. This predicts effects concentrated among soldiers who experienced the most discrimination, with no corresponding socioeconomic gains.

The paper's evidence supports Mechanism 2 and rules out Mechanism 1 throughout:
- Near-zero effects on all socioeconomic outcomes (R4);
- Effects increasing with draft board racial bias in induction rates (R5);
- Large effects at camps denying Black soldiers training and promotion; near-zero at less discriminatory camps (R6);
- Injustice as the sole survey theme predictive of postwar civic engagement (R7).

This also rules out a policy-feedback channel of the kind Mettler (2002) documents for the WWII GI Bill, since no comparable postwar benefits extended to Black WWI soldiers.

These results are consistent with the historical account of Williams (2007, p. 351) that "systematic discrimination during military service politicized black soldiers" and their "postwar disillusionment" drove racial militancy in the New Negro era.

**Identification.** The WWI draft lottery supplies quasi-random variation in military service. Six weeks after the first registration (June 5, 1917), paper slips containing serial numbers 1-10,500 were drawn in a national lottery; the rank of each man's serial number among those at his draft board determined his induction order. The authors infer order numbers from serial numbers, board size, and the national drawing sequence. The instrument is exogenous by construction: Figure III, Panel B (p. 816) shows all prewar registrant characteristics (exemption claim, marital status, occupation) have coefficients below 0.01 standard deviations on the scaled order number, all statistically insignificant.

Unlike the Vietnam and WWII draft lotteries examined in Angrist and Krueger (1994) and related work, the WWI lottery was based on registration serial numbers rather than birth dates, generating variation within draft boards rather than within birth cohorts.

## Method

The paper applies 2SLS using the scaled draft lottery order number as an instrument for veteran status. It builds on `instrumental-variables` for the causal identification strategy and `panel-regression` for the fixed-effects specification with clustered standard errors. Historical record linkage across administrative databases follows Abramitzky, Boustan, and Eriksson (2012), henceforth ABE, requiring exact matches on birth state and name initials with close alignment on full names and birth years. For the heterogeneity analysis, county-level racial prejudice is measured using a version of the Confederate Culture Index adapted from Bazzi et al. (2023).

**First stage** (equation 2, p. 815):

$$\text{Veteran}_i = \delta_b + \gamma \, \text{Order}_i + X_i'\Lambda + v_i \tag{2}$$

Here $$\text{Order}_i$$ is registrant $$i$$'s inferred order number scaled by the total number of registrants in his draft board, so $$\text{Order}_i \in [0,1]$$ with lower values indicating earlier lottery draws and greater induction likelihood. $$\delta_b$$ are draft board fixed effects that restrict the variation to within-board comparisons; $$X_i$$ are prewar individual controls (birth year, state of birth, prewar occupation, exemption claim interacted with marital status). Standard errors are clustered by serial number throughout.

Figure III, Panel A (p. 816) shows the first-stage binned scatterplot: slope = -0.088 (SE 0.0038). The gap in veteran likelihood between men with the lowest and highest order numbers is approximately 10 percentage points. First-stage F-statistics range from 534 to 565 across the main specifications (Table III, p. 817).

## Empirical specifications

**Main specification** (equation 1, p. 813):

$$\text{NAACP}_i = \lambda_b + \beta \, \text{Veteran}_i + X_i'\Gamma + u_i \tag{1}$$

where $$\text{NAACP}_i$$ indicates whether individual $$i$$ appears in NAACP membership rosters between the end of WWI and 1939; $$\text{Veteran}_i$$ is an indicator combining the 1930 census self-report and VAMI/ATS military record matches; $$\lambda_b$$ are draft board fixed effects; and $$X_i$$ are prewar controls. Veteran status is instrumented with $$\text{Order}_i$$ from equation (2). Analysis sample: 204,923 draft registrants linked to the 1930 census.

The preferred specification (Table III, column 4) adds 1930 county fixed effects to rule out residential mobility during the Great Migration as a confound; the 2SLS coefficient is 0.0282\*\* (SE 0.0118, first-stage F = 565.1). The OLS coefficient (0.0034-0.0044) is appreciably smaller, consistent with measurement error in veteran status attenuating the first-stage slope and with civic-minded volunteers being less inclined to join organizations that challenged the government.

**Heterogeneity by discrimination (R5, R6).** Equation (1) is estimated separately for quartiles of the board-level Black-white induction rate gap (Figure VIII, p. 830) and for bins of camp discrimination measures derived from Military Intelligence Division reports: share of Black noncommissioned officers (NCOs) in Black units and share of Black soldiers receiving military training (Figure IX, p. 832). Camp assignment is predicted for all registrants, including non-draftees, using a crosswalk linking draft boards to camp recruitment areas via National Geographic maps and Southern draft call lists. Bazzi et al. (2023)'s approach for measuring historical racial prejudice is adapted to the county level to measure prewar racial context (Figure VI, p. 826).

**Socioeconomic outcomes (R4).** The same 2SLS specification is applied to standardized outcomes from the 1930 and 1940 censuses: literacy, home ownership, employment, occupational income score, wage and salary income in 1940, and college attendance. All estimates are near zero and statistically insignificant (Figure V, p. 823); confidence intervals rule out occupational income gains above 0.10 standard deviations.

**Survey analysis (R7).** Multivariate OLS regression of self-reported voter participation on indicators for narrative themes from Virginia and Connecticut WWI veterans' questionnaires, using approximately 1,300 Black army veterans who returned surveys with non-missing responses. Standard errors are heteroskedasticity-robust. Injustice mentions yield a 24 percentage point coefficient, robust to controls for birth year, birth state, prior military experience, and county of residence; all other themes are statistically insignificant or borderline.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| WWI Draft Registration Cards, First Registration June 5, 1917 (NARA) | Primary sample: ~936K Black male registrants; serial numbers, birth date/state, address, occupation, exemption claim; serial numbers used to reconstruct lottery order | No page yet |
| NAACP Membership Rosters, ProQuest History Vault (NAACP Papers collection) | Main outcome: 233,517 member observations across 227 branches, 1912-1940; linked to census via name and address | No page yet |
| Veterans Administration Master Index (VAMI), 1917-1919 | Veteran status measure: ~4M veterans who claimed VA benefits, linked to draft cards by name/county/birth date; 120,412 cards linked | No page yet |
| Army Transport Service (ATS) Passenger Lists | Veteran status and unit assignment for soldiers deployed to or returning from Europe; ~6M passengers; Black units identified via *Directory of Troops* | No page yet |
| 1930 Full-Count Census (IPUMS USA, Ruggles et al. 2024) | Primary linking bridge: ~6M Black men; name, address, birth data, veteran self-report; socioeconomic outcomes (literacy, income, employment) | [U.S. Census Bureau data](/wiki/datasets/census/) |
| 1940 Census (subset linked via Abramitzky, Boustan, and Rashid 2020) | Extended outcomes for wage income and college attendance | [U.S. Census Bureau data](/wiki/datasets/census/) |
| Virginia and Connecticut WWI Veterans' Questionnaires | Survey mechanism analysis: ~1,300 Black army veterans' responses to "effects upon yourself" of draft, camp, and overseas experience | No page yet |
| African American National Biography / ProQuest African American Biographical Database | Community leadership outcome (R3): 7,554 Black men born 1860-1900 identified as historically significant by modern historians and contemporaneous publications | No page yet |

Main analysis sample: 204,923 draft registrants linked to the 1930 census, of whom 35% are identified as veterans using the union measure (census self-report OR VAMI/ATS match) and 1.8% are identified as NAACP members. The NAACP membership data spans 1912-1940, with the vast majority of observations from after 1925 (roughly 75% of all NAACP members from the mid-1920s to late 1930s).

## When to read the full paper

Use the [original](https://doi.org/10.1093/qje/qjaf046) if you are:
examining the full set of robustness checks (alternative linking strategies, alternative veteran status definitions, alternative first-stage functional forms including cubic polynomial and nonparametric specifications, alternative instruments based on actual vs. inferred order numbers) in the Online Appendix;
studying the evolution of NAACP membership effects across the 1920s and 1930s (Online Appendix Figure A.IX) or the distribution of membership duration (Online Appendix Figure A.X);
analyzing spillover effects of veterans on undrafted Black community members and the macrolevel relationship between local veteran shares and NAACP branch formation (Figure XII, p. 838);
or replicating from the publicly available dataset at Harvard Dataverse (https://doi.org/10.7910/DVN/LSPICA).

## Attribution and rights

Source: peer-reviewed, *The Quarterly Journal of Economics* 141(1), 2026, pp. 795-844. This distillation was extracted by an LLM on 2026-06-28 and is **not human-verified or independently reproduced**. The article is paywalled; extract-only applies.

> Ang, Desmond, and Sahil Chinoy. "Vanguard: Black Veterans and Civil Rights After World War I."
> *The Quarterly Journal of Economics* 141, no. 1 (2026): 795-844.
> DOI: 10.1093/qje/qjaf046. © The Author(s) 2025. All rights reserved.
> Published by Oxford University Press on behalf of President and Fellows of Harvard College.
> This page is an **extract** by the Institute for Automated Research: core results and
> equations re-expressed; not a reproduction.


==============================================================================
# Republican Support and Economic Hardship: Arteaga & Barone (2026)
# https://instituteforautomatedresearch.org/wiki/papers/qje/2026/arteaga-republican-support-economic-hardship-2026/
# Distilled: Using quasi-exogenous variation in Purdue Pharma's OxyContin cancer-market targeting, this paper establishes a causal link between opioid epidemic exposure and a 4.5 percentage-point increase in Republican House vote share by 2022, operating through economic hardship and conservative media framing. QJE 2026, paywalled. Ten core results with source locators, datasets used, the empirical specification with equations, and identification strategy.
# Tags: paper-summary, political-economy, health-economics, opioids, elections, partisan-realignment
==============================================================================

**What this is.** A distilled skeleton of Arteaga and Barone (2026). Read the
[original paper](https://doi.org/10.1093/qje/qjaf051) to replicate or extend; this
page records headline results with source locators, the estimating equations, and
the data used, extracted by an LLM and not human-verified.

## TL;DR

Arteaga and Barone use unsealed Purdue Pharma litigation records to show that the
pharmaceutical company targeted OxyContin marketing at communities with high cancer
mortality in 1996, creating quasi-exogenous geographic variation in opioid exposure.
This exposure caused persistent increases in drug-induced mortality, disability, and
food-stamp receipt. By 2022, a one standard deviation higher 1996 cancer mortality
rate raised the Republican two-party House vote share by 4.5 percentage points. The
political shift was not anti-incumbent, was not confined to any demographic group,
and was concentrated in communities where economic hardship was greatest. Conservative
media covered the epidemic more extensively and framed it around crime and economic
distress, themes aligned with the Republican Party's messaging to white working-class
America.

## Core results

| # | Result | Locator | Magnitude as reported |
|---|--------|---------|----------------------|
| R1 | Drug-induced mortality: 1 SD higher 1996 cancer mortality rate causes persistent rise in drug deaths | Fig. III, p. 523 | 46% above pre-epidemic average by 2017 |
| R2 | Prescription opioid deaths: 1 SD increase causes 61% rise by 2012 | p. 523, Fig. III | 61% higher than pre-epidemic average at 2012 peak |
| R3 | SSDI applications: 1 SD increase causes 12% rise by 2012 | Fig. IV, p. 524 | +12% SSDI applications by 2012 |
| R4 | SSI applications: 1 SD increase causes 7.6% rise by 2012 | Fig. IV, p. 524 | +7.6% SSI applications by 2012 |
| R5 | SNAP receipt: 1 SD increase predicts 8% more recipients by 2022 | Fig. IV, p. 524 | +8% SNAP recipients (0.14 SD) by 2022 |
| R6 | Republican House vote share: 1 SD increase raises GOP share 4.5 pp in 2022 | Fig. V, p. 526 | +4.5 pp GOP two-party vote share in 2022 midterms |
| R7 | Republican presidential vote share: 1 SD increase raises GOP share 4.6 pp | Fig. VII, p. 530 | +4.6 pp GOP presidential vote share |
| R8 | Republican gubernatorial vote share: 1 SD increase raises GOP share 4.3 pp after six elections | Fig. VIII, p. 532 | +4.3 pp GOP gubernatorial vote share |
| R9 | Policy preferences: exposure predicts support for law enforcement and opposition to marijuana legalization | Table V, p. 538 | +0.0393\*\*\* (more police), +0.125\*\*\* (safety around police), -0.0179\*\* (marijuana legalization) |
| R10 | Media: exposure predicts higher Fox News viewership | Table V col. 4, p. 538 | +0.0475\*\*\* Fox News viewership share (CCES 2020) |

**Overall (paper's conclusion).** The opioid epidemic causally reshaped the
political landscape of affected communities, generating sustained gains for
Republicans in House, presidential, and gubernatorial elections from 2006 onward.
These gains were mediated by economic hardship (higher disability and SNAP
enrollment) and amplified by conservative media coverage that framed the epidemic
around crime and economic distress, resonating with the Republican Party's
messaging on working-class economic decline.

## Theory / model

The paper has no formal structural model. It tests a set of empirical hypotheses
derived from the political economy of economic distress and partisan realignment.

The core identification assumption is that in the absence of OxyContin marketing,
commuting zones with higher 1996 cancer mortality would have followed the same
trends in health, economic, and political outcomes as zones with lower cancer
mortality (parallel trends). The validity of this assumption is supported by
pre-trend tests, out-of-sample placebo exercises, and controls for alternative
economic and political shocks (NAFTA exposure, China import shock, recessions,
robot adoption, Fox News rollout, Southern partisan realignment).

The paper rules out three alternative explanations:

1. **Anti-incumbent voting**: the GOP vote share increase does not depend on the
   party of the incumbent at the time of each election (Online Appendix Figure A9,
   p. 533).
2. **Voter turnout**: no meaningful changes in turnout rates are estimated across
   the distribution of opioid exposure (Online Appendix Figure A6, p. 527).
3. **Direct mortality effect**: a counterfactual calculation shows that missing
   votes attributable to opioid-related deaths shift the observed vote share by at
   most 0.22 pp relative to 2020 (p. 527), far smaller than the 4.5 pp House effect.

The mechanisms the paper tests are:
- Economic hardship (SNAP, disability) predicts subsequent Republican gains with a
  lag: state-level SNAP effects in 2006 are the strongest predictor of GOP vote
  share shifts in 2022 (Figure IX, p. 536).
- Conservative media engaged more extensively with the epidemic and framed it
  around crime and economic hardship, themes favoring the Republican Party's
  platform (Online Appendix Figure A10, Table V, pp. 538-541).
- Residents preferred Republican law enforcement policy solutions (more police,
  opposition to harm-reduction policies) over Democratic harm-reduction approaches
  (Table V, p. 538).

The paper is related to three strands of literature. First, the causal
identification design extends Autor, Dorn, Hanson, and Majlesi (2020), who use the China
trade shock as an instrument for local economic distress and document downstream
political polarization. Second, the finding that acute epidemics can shift politics
rightward complements Voigtländer and Voth (2012), who document that the Black Death
increased anti-Semitic persecution in medieval Germany. Third, the paper extends
Goodwin et al. (2018), who document an observational association between chronic
opioid use and presidential voting patterns in US counties, to a causal
quasi-experimental framework. Evans, Lieber, and Power (2019) provide the evidence
that OxyContin's 2010 abuse-deterrent reformulation pushed users toward heroin,
used here to contextualize the epidemic's transition from prescription to illicit
drugs.

## Method

The main estimating equation is an event-study specification on a panel of
commuting zones (CZs), interacting a pre-determined exposure proxy with year
indicators (p. 515):

$$y_{ct} = \sum_{\tau=1982}^{2022} \phi_\tau \, \text{CancerMR}_{c,1996} \cdot \mathbf{1}(\text{Year} = \tau) + \alpha X_{ct} + \gamma_c + \gamma_{st} + \upsilon_{ct} \tag{1}$$

where $$c$$ indexes the CZ, $$s$$ the state, and $$t$$ the year.
$$\text{CancerMR}_{c,1996}$$ is the 1996 cancer mortality rate per 1,000 in CZ
$$c$$, standardized so that a one-unit change equals one standard deviation.
The sum runs from 1982 (or the earliest available year for each outcome) to 2022.
Coefficients $$\phi_{1982}$$ to $$\phi_{1995}$$ are pre-trend estimates that test
for differential trends before OxyContin's 1996 launch; $$\phi_{1997}$$ to
$$\phi_{2022}$$ trace the cumulative effect of opioid exposure over time. The
interaction for 1996 is omitted, so all coefficients are relative to that base year.

$$\gamma_c$$ are CZ fixed effects capturing time-invariant unobserved heterogeneity.
$$\gamma_{st}$$ are state-year fixed effects controlling for state-level policy
changes (prescription drug monitoring programs, pill-mill regulations, naloxone
access laws). $$X_{ct}$$ are time-varying controls at the CZ level: contemporaneous
cancer mortality, white and female population shares, and age shares (18-29, 30-49,
50-64, 65+, under 1). Standard errors are clustered at the CZ level.

The instrument for opioid exposure is 1996 county-level cancer mortality, which
proxies the cancer pain market that Purdue targeted for OxyContin at launch. The
validity rests on Purdue's documented strategy: internal records state "OxyContin
will be marketed at the cancer pain market" (Purdue Pharma 1994, cited p. 509) and
cancer-market penetration was used as a springboard into the far larger noncancer
pain market. Columns (2) and (3) of Table I (p. 512) confirm that 1994 MS Contin
prescription rates (the pre-OxyContin cancer opioid) and 1996 cancer mortality both
strongly predict 1996-1998 OxyContin prescription rates.

To identify geographic concentration of political effects and test whether
communities with the largest economic effects saw the largest political shifts, the
paper also estimates a state-level in-differences model (p. 535):

$$\Delta y_{ct} = \sum_{s=1}^{50} \sum_{\tau=1982}^{2022} \phi_\tau^s \, \text{CancerMR}_{c,1996} \cdot \mathbf{1}(\text{Year} = \tau \text{ and State} = s) + \alpha \Delta X_{ct} + \gamma_t + \upsilon_{ct} \tag{2}$$

The state-by-year interactions $$\phi_\tau^s$$ capture state-specific exposure
effects at each point in time, allowing a within-state comparison that is equivalent
to the level specification augmented with state-by-year dummies.

## Empirical specifications

**First stage (opioid supply, Figure II, p. 514).** The estimating equation is
equation (1) above, with $$y_{ct}$$ replaced by DEA ARCOS prescription opioid doses
per capita (1997-2020). The peak coefficient is in 2012, when a one standard
deviation higher cancer mortality corresponded to 0.97 additional opioid doses
prescribed per capita (65% above the baseline mean), consistent with the magnitude
of Alpert, Evans, Lieber, and Powell (2022).

**Mortality (Figure III, p. 522-523).** Equation (1) with $$y_{ct}$$ = drug-induced
mortality per 1,000. Drug-induced mortality is the broadest measure, covering
poisoning and medical conditions caused by legal or illegal drugs (ICD-9/10 codes
linked per CDC 2013 guidance). The coefficient rises steadily from 1996 onward,
peaks around 2015-2017, then begins to decline as fentanyl-related deaths shift the
geographic distribution. Pre-trend test: $$p$$-value for joint test of
$$\phi_{1982}, \ldots, \phi_{1995} = 0$$ is 0.3462 (Figure III caption, p. 522).
Deaths are concentrated among individuals under 55 (Online Appendix Figure A4).

**Economic outcomes (Figure IV, p. 524-525).** Same specification with outcomes:
(i) share of working-age population (18-65) applying for SSDI, (ii) share of
population applying for SSI (blindness/disability), and (iii) share of population
receiving SNAP benefits. SSDI and SSI data available for 1990-2015 (438 CZs);
SNAP data from USDA Food and Nutrition Service, January counts, 1989-2022.
Unemployment does not respond significantly (Online Appendix Section C.3, p. 524n).

**Political outcomes: House elections (Figure V, Table IV, pp. 526-528).**
Equation (1) with $$y_{ct}$$ = two-party Republican vote share in House elections.
Outcome is the ratio of Republican votes to total Republican plus Democratic votes.
CZs cover 616 House election observations. Pre-trend: $$p$$-value for joint zero
of $$\phi_{1982}, \ldots, \phi_{1994}$$ is 0.7510 (Figure V caption). Cross-demographic
stability: Table IV shows coefficients of 0.0477-0.0721 (standard errors 0.0131-0.0165)
across race (white vs. non-white), age (under vs. over 50), gender, and education
groups, all statistically indistinguishable. The effect on Republican seat wins appears
from 2012 onward (Figure VI, Panel A, p. 529).

**Policy preferences (Table V, p. 538).** Cross-sectional specification:

$$y_{it} = \alpha_i + \beta \, \text{CancerMR}_{c,1996} + \alpha X_i + \gamma_s + \varepsilon_{it}$$

where $$i$$ indexes CCES respondents in 2020 and $$\gamma_s$$ are state fixed
effects. Columns (1)-(2) and (4) use 2020 CCES data; column (3) uses secretary of
state vote records for marijuana ballot initiatives in 18 states (2012-2023).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---------|--------------|-----------|
| DEA ARCOS (Automation of Reports and Consolidated Orders System), 1997-2020 | CZ-level opioid prescription doses per capita (first stage); digitized from DEA records | no page yet |
| CDC/NCHS National Vital Statistics System (NVSS) Detailed Multiple Cause of Death files, 1976-2022; restricted-use county identifiers | Drug-induced, prescription opioid, and all-opioid mortality rates; cancer mortality rate (instrument) | no page yet |
| SSA SSDI and SSI application and receipt data, 1990-2015 (applications), 1998-2020 (receipts); 438 CZs | SSDI and SSI application and receipt rates (economic hardship measures) | no page yet |
| USDA Food and Nutrition Service SNAP county-level participation, January 1989-2022 | SNAP receipt rate (economic hardship measure) | no page yet |
| Dave Leip's Atlas of U.S. Elections + ICPSR U.S. Historical Election Returns, 1976-2022 | Two-party Republican vote shares for House, presidential, and gubernatorial elections; voter turnout | no page yet |
| Cooperative Congressional Election Study (CCES), 2006-2020 | Individual-level survey data: vote choice, demographic heterogeneity analysis, policy preferences (police, marijuana), Fox News viewership | no page yet |
| Purdue Pharma unsealed litigation documents | OxyContin marketing strategy; source of identification rationale (cancer market targeting strategy) | no page yet |
| State Drug Utilization Data (SDUD), 1994-1998 | MS Contin and OxyContin prescription rates at state level for validating instrument (Table I) | no page yet |
| Newspapers.com local newspaper archive, 1995-2020; TV News Internet Archive (Fox News/CNN/MSNBC), 2009-2020 | Media coverage content and framing analysis; newspaper political affiliation from Gentzkow and Shapiro (2011) | no page yet |

Sample: 621 CZs (restricted to those with more than 20,000 residents, covering more than 99% of all opioid deaths and 99% of total US population). Panel period 1982-2022 for political and mortality outcomes; 1989-2022 for SNAP; 1990-2015/1998-2020 for disability.

## When to read the full paper

Read Arteaga and Barone (2026) if you need:

- A credible causal estimate of the opioid epidemic's effect on partisan realignment
  (Table I, Figures V-VIII are the key evidence).
- A detailed treatment of how economic hardship mediates the link from a public health
  crisis to political preferences (Figure IX, Online Appendix Table A2).
- Evidence on how media framing of a crisis shapes political outcomes, with a useful
  contrast between conservative and liberal coverage (Online Appendix Figure A10,
  Table V columns 1-4).
- A worked example of using industry marketing documents (unsealed litigation records)
  as a source of quasi-exogenous variation (Sections III.A-B, pp. 509-516).

## Attribution and rights

Arteaga, Carolina, and Victoria Barone. "Republican Support and Economic Hardship:
The Enduring Effects of the Opioid Epidemic." *The Quarterly Journal of Economics*
141, no. 1 (2026): 499-558. https://doi.org/10.1093/qje/qjaf051

Copyright The Author(s) 2025. Published by Oxford University Press on behalf of
President and Fellows of Harvard College. All rights reserved. This page contains
only extracts for academic reference (titles, locators, magnitudes, descriptions).
Replication data are available at Harvard Dataverse:
https://doi.org/10.7910/DVN/R5PKQL.

LLM-distilled by paper-distiller (claude-sonnet-4-6), 2026-06-28. Not
human-verified. Not reproduced.


==============================================================================
# Ideas Have Consequences: Ash, Chen & Naidu (2026)
# https://instituteforautomatedresearch.org/wiki/papers/qje/2026/ash-ideas-have-consequences-impact-2026/
# Distilled: The Manne Economics Institute for Federal Judges shifted judicial behavior: trained judges used more economics language in their opinions, voted more often against federal regulatory agencies, and imposed stricter criminal sentences. The Quarterly Journal of Economics (2026), CC BY-NC 4.0. Seven core results with source locators, datasets used, and the DiD estimating equations.
# Tags: paper-summary, law-economics, judicial-decision-making, criminal-sentencing
==============================================================================

**What this is.** The paper's core results, identification design, and estimating equations: enough to know what it found and how, without reading the full 43 pages. To replicate or extend, read the original at [https://doi.org/10.1093/qje/qjaf042](https://doi.org/10.1093/qje/qjaf042). Replication data are available at [Harvard Dataverse](https://doi.org/10.7910/DVN/XATYFX).

## TL;DR

The paper estimates the causal effect of economics training on U.S. federal judges. The treatment is attendance at the Manne Economics Institute for Federal Judges, an intensive two-week program run by Henry Manne and the Law and Economics Center (LEC) starting in 1976. The program was oversubscribed from its second year onward, with admission on a first-come, first-served basis, which the authors exploit as a source of quasi-random variation in the timing of attendance. The sample covers the universe of published opinions in U.S. Circuit Courts (1970-2005) and approximately 1 million District Court criminal sentencing decisions (1992-2011; main event-study analysis uses 1992-2003). After attending the program, judges used more economics language in their opinions (roughly 0.36 standard deviations more word-embedding similarity to the economics lexicon), voted more often against federal labor and environmental regulatory agencies (roughly 16 percentage points more), and imposed prison sentences more frequently (roughly 6 percentage points more in the short run). Effects are concentrated in economics-related cases; no significant shift appears in non-economics cases.

## Core results

Significance is as reported; standard errors clustered by judge. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Manne attendance **increased economics language use** by ~0.36 SD short-run (all years) | Table I col. (1), p. 868; Figure III, p. 869 | beta=0.363 (se=0.154), N=5,261; significant at 5% |
| R2 | Effect on economics language is larger in the **early period** (pre-1987): ~0.43 SD | Table I col. (2), p. 868 | beta=0.430 (se=0.1305), N=3,214; significant at 1% |
| R3 | Manne attendance **raised probability of voting against labor/environmental regulators** by ~16 pp | Table I col. (5), p. 868; Figure IV, p. 872 | beta=0.162 (se=0.0675), N=2,408; significant at 5% |
| R4 | **Conservative voting in economics cases** increased by ~27 pp short-run (long-run conservative estimate: ~4 pp) | Table I col. (9), p. 868; col. (11) discussed p. 877 | beta=0.265 (se=0.1316), N=804 short-run; beta=0.043 long-run conservative estimate |
| R5 | **No significant effect** on conservative voting in non-economics cases | Table I cols. (13)-(16), p. 868 | beta=0.02 to 0.06 across specs, all insignificant; consistent with economics-specific channel |
| R6 | Manne attendance **increased probability of imposing a prison sentence** by ~6.2 pp short-run | Table II col. (1), p. 869; Figure V, p. 875 | beta=0.0617 (se=0.0202), N=70,784; significant at 1% |
| R7 | Long-run effect on **prison sentence probability**: ~3.5 pp | Table II col. (2), p. 869 | beta=0.035 (se=0.0135), N=260,516; significant at 5% |

**Overall (paper's conclusion).** The U.S. law and economics movement, disseminated through the Manne program, shifted legal outcomes across the federal judiciary. The finding that neither the legalist view (judges apply law as written) nor the attitudinalist view (judges follow party) can explain the results points to a third channel: within-judge shifts in the intellectual framework used to decide cases. The persuasion rate for conservative voting in economics cases is approximately 8%, comparable in magnitude to major media-effects estimates such as the Fox News effect on Republican voting estimated by DellaVigna and Kaplan (2007).

## Theory / model

The paper has no formal structural model. The theoretical framework proceeds in three steps.

**Step 1: Two null benchmarks.** A legal formalist would predict no effect: judges apply statutes and precedents as written, leaving no room for a training program to shift outcomes. An attitudinalist would likewise predict no effect: judges follow their appointing party platform, and training within the window of a sitting judge's tenure cannot move a party affiliation. Both predict a zero treatment effect. The paper argues that neither model fits because judges in a common-law system have substantial discretion, and the empirical results confirm it.

**Step 2: Why economics?** The best prior evidence on ideological influence in courts came from Bonica et al. (2019), who found that changes in clerk ideology sometimes shift a Supreme Court justice's votes; the Manne program extended and broadened that channel. The paper invokes the Bayesian persuasion framework of Gentzkow and Kamenica (2011) to explain why economics, as a rigorous social science, is especially persuasive with professional judges (p. 880). The Manne curriculum corresponds to a signal structure with commitment: the instructor, bound by academic or scientific norms, reveals the results of economic analysis even when the findings may conflict with the preferred conclusion of funders. In the Gentzkow-Kamenica (2011) framework, the principal will choose either an informative signal or none at all; thus, even if the judge knows the economist is biased for a particular outcome, the economist can still influence the judge to vote in the preferred direction some of the time, precisely because the economist is committed to revealing the signal generated by the economic analysis. The credibility of academic economic analysis means the signal is informative even to skeptical judges, raising the persuasion probability.

**Step 3: Persuasion rate.** The paper benchmarks its effect estimates against media-effects studies using the persuasion rate formula from DellaVigna and Gentzkow (2010) (p. 877):

$$p = 100 \times \frac{\Delta y}{\Delta e} \cdot \frac{1}{1 - y_0},$$

where $\Delta y$ is the change in the outcome, $\Delta e = 1$ (all attenders exposed), and $y_0 = 0.46$ is the pre-treatment mean conservative vote rate in economics cases. Setting $\Delta y = 0.043$ (Table I, col. (11), the most conservative long-run estimate) gives $p \approx 8\%$, not that different from the Fox News persuasion rate of 11.6% estimated by DellaVigna and Kaplan (2007) (p. 877).

## Method

**Economics language measurement.** The first outcome measure is constructed using word embeddings (p. 858-860). Judicial opinions are preprocessed (lowercased, punctuation removed) and represented as word lists. The paper uses the word2vec algorithm from Mikolov et al. (2013) to map each word to a dense vector. The starting point for the economics lexicon is the Ellickson (2000) index of 11 economics-related terms used in legal scholarship (externalit\*, transaction\_cost, efficien\*, deterr\*, cost\_benefit, capital, game\_theo, chicago\_school, marketplace, law1economic, law2economic). The economics language measure for opinion $i$ is the cosine similarity in word-embedding space between the opinion vector and the Ellickson lexicon vector:

$$\text{EconLang}_i = \cos(\text{emb}(\text{opinion}_i),\, \text{emb}(\text{lexicon})),$$

where emb(·) maps a document to its embedding-space centroid. The measure is approximately normally distributed and captures contextual use of economics reasoning, not just raw word counts. The paper also constructs an alternative supervised-learning measure (predicting whether an opinion is on an economics topic), which gives consistent results (Online Appendix D).

**Identification.** The identification strategy exploits the quasi-random timing of Manne program attendance. From 1977 through the late 1980s, the program was oversubscribed, with admission on a first-come, first-served basis. This means that among judges who applied, the specific year of attendance was largely determined by luck of timing rather than by judge characteristics. The paper's baseline sample restricts to ever-attenders (excludes never-attenders, who differ on pretreatment levels and trends) and exploits within-judge variation in the years before versus after attendance.

**Peer-share controls.** To address SUTVA concerns (judges communicate within circuits), the paper augments the main specification with a judge-specific peer attendance variable: $\bar{Z}^{-j}_{ct}$ = share of peer judges (weighted by caseload) on the same court who have attended Manne (p. 866-867). This is included interacted with circuit-judge fixed effects, allowing a judge-specific spillover effect.

## Empirical specifications

The outcome $Y_{ijct}$ is a decision, vote, or text metric for case $i$ by judge $j$ in court $c$ during year $t$. The baseline difference-in-differences specification is (equation (1), p. 863):

$$Y_{ijct} = \alpha_j + \alpha_{ct} + \gamma Z^{\text{post}}_{jt} + \mathbf{X}'_{ijct}\beta + \varepsilon_{ijct}, \tag{1}$$

where $\alpha_j$ is a judge fixed effect, $\alpha_{ct}$ is a court-year fixed effect, $Z^{\text{post}}_{jt}$ is an indicator for years after judge $j$ attended the Manne program, $\mathbf{X}_{ijct}$ includes optional elastic-net-selected judge characteristics interacted with five-year time effects (predicting timing of attendance), and $\varepsilon_{ijct}$ is an error term clustered by judge. The court-year fixed effect absorbs time-varying court-level confounders including case-mix changes, since cases are quasi-randomly assigned to judges within court-year cells.

The design differs from the standard Callaway and Sant'Anna (2021) staggered DiD in that parallel trends are required in the already-treated group (not only the not-yet-treated), because many judges attend early and restricting to not-yet-treated comparators would cost too much statistical power (p. 862).

The event-study specification replaces the post-indicator with leads and lags (equation (2), p. 863):

$$Y_{ijct} = \alpha_j + \alpha_{ct} + \sum_{k \in K} \gamma_k Z^k_{jt} + \mathbf{X}'_{ijct}\beta + \varepsilon_{ijct}, \tag{2}$$

where $Z^k_{jt}$ are indicators for $k$ years before/after Manne attendance ($k \in K$, with the year before attendance and the first year of the window excluded following Borusyak, Jaravel, and Spiess (2024)). The event-study plots for each of the three main outcomes (Figures III, IV, V) show minimal pre-trends and a jump at the time of attendance.

Specific implementations by outcome:

- **Economics language (R1-R2):** Sample limited to majority-opinion authors on economics cases; circuit court level (N approx 5,261 in the short-run sample). The early-period (pre-1987) subsample isolates the heyday when classes were most oversubscribed (N=3,214).
- **Voting against regulators (R3):** Binary outcome: whether judge voted against the government in a labor or environmental agency case. Sample: ever-attenders in the event-study window (N=2,408 short run; 4,244 long run).
- **Conservative voting (R4-R5):** Binary vote direction from the Songer-Auburn database (hand-coded 5% sample of circuit court cases through 2002). Smaller N (N=804 short-run economics cases) due to the limited Songer-Auburn coverage.
- **Criminal sentencing (R6-R7):** Binary outcome: any prison sentence given. OLS. District court cases; event-study (short-run) sample 1992-2003, N=70,784; long-run sample uses all ever-attenders across all years (1992-2011), N=260,516. Sentence length analyzed separately via Poisson regression; no significant effect found (Table II cols. (3)-(4)), consistent with limited judicial discretion over length under mandatory guidelines in this period.

Observations are weighted so that judge-years count equally (reweighting by judge-year to prevent courts and years with more cases from having disproportionate influence).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Bloomberg Law circuit court opinions | ~200,000 cases (1970-2005); source of judicial opinions, three-judge panel composition, vote direction, and case topic codes | No page yet |
| TRAC district court criminal sentencing records | ~1 million cases (1992-2011); prison sentence given, sentence length in months, crime type, judge identity; event-study analysis uses 1992-2003 window | No page yet |
| Federal Judicial Center (FJC) judge biographies | Judge biographical data: appointment party, birth cohort/region, education, career history | No page yet |
| Manne Program attendance records | Hand-compiled from Butler (1999) and FOIA requests to the LEC at George Mason University; attendance year for each of ~840 ever-attending federal judges, 1976-1998 | No page yet |
| Songer-Auburn database | Hand-coded 5% sample of circuit court cases through 2002; liberal/conservative/neutral vote coding for 5% of cases | No page yet |

Sample scope: circuit courts 1970-2005 (economics language and regulatory voting); district courts 1992-2011 (criminal sentencing; event-study window 1992-2003). Judge-year is the effective unit of observation. Observations weighted to treat judge-years equally.

## When to read the full paper

Use the [original](https://doi.org/10.1093/qje/qjaf042) if you are:
- Studying how ideas and intellectual frameworks influence policy decisions;
- Interested in the history and effects of the law and economics movement on the U.S. judiciary;
- Replicating (the Online Appendix covers the complete robustness battery including never-attenders, alternative clustering, unweighted regressions, and SUTVA checks);
- Extending the analysis to other ideological training programs, other policy domains, or other countries with analogous judge training programs;
- Benchmarking text-as-data measures of ideological style against supervised alternatives.
- Extending the decision-quality evidence from Baye and Wright (2011), who show Manne-trained judges are less likely to have antitrust decisions reversed on appeal, to the full breadth of judicial outcomes studied here.

## Attribution and rights

Source: peer-reviewed, *The Quarterly Journal of Economics* (2026), pp. 845-887. This distillation was extracted by an LLM on 2026-06-28 and is **not human-verified or independently reproduced**.

The paper is distributed under the Creative Commons Attribution-NonCommercial License (CC BY-NC 4.0). Commercial re-use requires permission from the publisher.

> **Attribution (CC BY-NC 4.0).** Ash, Elliott, Daniel L. Chen, and Suresh Naidu.
> "Ideas Have Consequences: The Impact of Law and Economics on American Justice."
> *The Quarterly Journal of Economics* (2026), 845-887.
> DOI: 10.1093/qje/qjaf042. © 2025 The Author(s).
> Published by Oxford University Press on behalf of President and Fellows of Harvard College.
> Licensed under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Permanent Capital Losses after Banking Crises: Baron et al. (2026)
# https://instituteforautomatedresearch.org/wiki/papers/qje/2026/baron-permanent-capital-losses-banking-2026/
# Distilled: Studying 76 bank equity crises across 46 economies since 1870, this paper documents that banking crises produce large, permanent declines in bank capital driven by asset write-downs rather than temporary price dislocations, and that forceful liquidity interventions restore only a transient fraction of bank value. Historical government recapitalizations have been too small, delayed, and narrow to restore banking sector capitalization. The Quarterly Journal of Economics, 2026, paywalled. Eight core results with source locators, datasets used, and empirical specifications.
# Tags: paper-summary, banking-crises, financial-crises, bank-capital, bank-equity
==============================================================================

**What this is.** The paper's core results, identification strategy, datasets, and empirical
specifications: enough to know what it found and how. To replicate or extend, read the full
source at the [original](https://doi.org/10.1093/qje/qjaf052).

## TL;DR

The paper studies the mechanisms driving bank losses across historical banking crises
in 46 economies and the effectiveness of policy interventions in restoring bank
capitalization. It constructs several new historical datasets: a country-level panel of
bank and nonfinancial equity returns (extending Baron, Verner and Xiong (2021), henceforth
BVX), individual-bank-level balance sheet and income data for the 10 largest banks in 17
economies around each crisis (building on Jordà, Schularick and Taylor (2017), henceforth
JST), and a comprehensive database of policy interventions extending Laeven and Valencia (2020).
The central finding is that bank stock prices experience large, permanent declines at banking
crisis onset, predicting commensurate long-run declines in banks' earnings and dividends rather
than elevated future equity returns. This earnings-driven pattern is inconsistent with models
that posit bank losses are primarily due to temporary price dislocations or liquidity strains.
Write-downs on nonperforming assets account for the bulk of realized bank losses; asset
sales during panics contribute little on average. Liquidity-based interventions (central bank
support, blanket liability guarantees) provide only a transient rebound in bank equity that
reverses within 12-36 months. Historical government recapitalizations have been too small,
delayed, and narrow to restore banking sector capitalization.

## Core results

Magnitudes and significance as reported; \* = 10%, \*\* = 5%, \*\*\* = 1%. Locators
point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Bank equity has large, permanent abnormal declines at crisis onset; no elevated returns in years t+1 to t+5** | Figure I, p. 685; Table II Panel A, p. 687 | Average abnormal return: -68 log points (banks), -36 log points (nonfinancials) at crisis onset; bank cumulative abnormal return at t+3 = -0.263\*\*\* (s.e. 0.047); at t+5 = -0.043 (s.e. 0.079); nonfinancials at t+5 = +0.313\*\*\* (s.e. 0.081) |
| R2 | **Initial bank equity declines predict long-run earnings and dividend declines** (earnings-driven, not discount-rate-driven) | Table III, p. 695 | Country-level: coeff. on 5-yr log-change in earnings per share = 1.750\*\*\* (s.e. 0.408); dividends = 2.483\*\*\* (s.e. 0.545); bank-level: earnings = 0.981\*\*\* (s.e. 0.218); with crisis FEs = 0.792\*\*\* (s.e. 0.213) |
| R3 | **Even with perfect trough timing, bounce-back averages only ~30% of initial decline; gains reverse after ~12 months** | Figure II, pp. 691-692 | Peak bounce-back from trough = ~30% of initial bank equity decline; gains begin reversing after approximately one year; banks do not recover to precrisis levels by t+5 |
| R4 | **Write-downs account for nearly all bank losses; trading losses from asset sales during panics are small on average** | Figure IV, Table IV, pp. 703-706 | Cumul. write-downs by t+5 = -0.338\*\*\* of precrisis book equity; write-downs account for approx. 100% of market-value losses by t+5; trading income = small fraction; banks with large securities portfolios show larger immediate trading losses |
| R5 | **Countries with larger bank equity declines exhibit higher subsequent NPL rates** (consistent with asset quality as primary mechanism) | Figure V, p. 709 | Adj. R² = 0.232 (full sample); adj. R² = 0.529 excluding USA 1990 and SWE 1991 outliers; relationship statistically significant across advanced economies |
| R6 | **Liquidity-based interventions yield only a transient ~20% bank equity rebound that reverses by months 12-36** | Figure VI, pp. 711-712 | Within two months of intervention, bank equity continues to decline; bank equity then rises by about ~20%, but this gain is short-lived; gains reverse between months 12 and 36; no persistent or large increase in bank capitalization from liquidity support or liability guarantees |
| R7 | **Government recapitalizations are small (~24% of book equity, ~43% of losses), delayed, and narrow** | Table V, Figure VII, pp. 715-722 | Mean recap size: ~24% of precrisis book equity; ~43% of market-value losses; 65% of programs narrow (few banks targeted); median delay ~8-9 months from panic start; bank market cap remains persistently lower 5 years after crisis |
| R8 | **Early liquidity interventions can avert incipient panics only before large equity declines occur**; 75% of bank equity crises feature equity decline strictly preceding the panic | Section VI, pp. 725-728 | Of 183 incipient liquidity shocks without prior bank equity decline: ~92 (~50%) averted by early intervention; of 76 bank equity crises: 57 (75%) show equity decline before any panic; essentially no bank equity crises averted by liquidity intervention after large equity decline has occurred |

**Overall (paper's conclusion).** Bank equity crises produce permanent capital losses driven by
deteriorating asset quality and eventual write-downs, not temporary price dislocations. Forceful
liquidity policy can avert crises when invoked before fundamental bank equity weakness occurs,
but after large equity declines have materialized, neither liquidity support nor historical
recapitalization programs have reliably restored banking sector capitalization.

## Theory / model

The paper has no formal structural model. It organizes the analysis around two competing
theoretical views and tests their empirical predictions:

**Temporary-loss view.** Several prominent models posit that bank losses during crises are
primarily temporary. Under this view, crises are times when assets held by financial
intermediaries trade at sharp discounts due to binding borrowing constraints or temporary
illiquidity (e.g., Gertler and Kiyotaki (2015); He and Krishnamurthy (2012); Dang, Gorton,
and Holmstrom (2020)). The temporary-loss view predicts: (i) elevated bank equity returns in
the years following a crisis, as discount rates normalize; (ii) limited long-run declines in
bank dividends and earnings; and (iii) large and lasting increases in bank equity following
liquidity interventions by central banks.

**Permanent-loss view.** An alternative view holds that banking crises give rise to permanent
bank losses through deterioration in asset quality (e.g., Kaminsky and Reinhart (1999);
Calomiris and Mason (2003); Schularick and Taylor (2012)). Nonperforming assets and borrower
defaults lead to asset impairments, permanently lowering bank equity and earnings. Banks are
slow to recognize these losses in accounting statements, but equity markets price them at crisis
onset. This view predicts: (i) no elevated future bank equity returns; (ii) commensurate
long-run declines in earnings and dividends; (iii) persistent rise in nonperforming loan rates;
and (iv) limited effectiveness of liquidity-based interventions.

The paper defines a "bank equity crisis" as the starting year in which both: (i) the bank
equity index declines by more than 30% in any year within the past five years; and (ii) a
top-20 bank (ranked by total assets) fails within a 0-5 year window around that decline.
This is a real-time, objective indicator based entirely on public information available to
market participants, designed to avoid the retrospective look-ahead bias in narrative crisis
chronologies. The sample contains 76 bank equity crises across 46 economies over 1870-2019.

## Method

The primary estimator is a panel regression with country (or bank) fixed effects and
Driscoll-Kraay standard errors, which allow for arbitrary serial correlation and
cross-sectional dependence across countries. This builds on `driscoll-kraay-regression`.
The event-study-style analysis (`event-study`) traces cumulative coefficients for each
horizon $$h$$ relative to the crisis onset.

**Cumulative abnormal returns.** For each crisis, the paper traces cumulative buy-and-hold
abnormal returns using equation (1) (p. 684):

$$
r_{i,t-k,t+h} = \alpha_i + \beta^h \text{BankEqCrisis}_{i,t} + \varepsilon_{i,t+h} \tag{1}
$$

where $$r_{i,t-k,t+h}$$ is the cumulative log excess total return from year $$t-k$$ to
$$t+h$$ for either the bank or nonfinancial equity index in country $$i$$;
$$\text{BankEqCrisis}_{i,t}$$ equals one if country $$i$$ enters a crisis in year $$t$$;
and $$\alpha_i$$ are country fixed effects. Setting $$k=1$$ normalizes cumulative returns
relative to the year before crisis onset. The coefficient $$\beta^h$$ measures abnormal
returns at horizon $$h$$. Standard errors are Driscoll-Kraay; 95% confidence intervals
for $$h \in [-5, 5]$$ are plotted in Figure I (p. 685) and tabulated in Table II (p. 687).
The key test: if $$\beta^h \leq 0$$ for $$h > 0$$, the temporary-loss view's prediction
of elevated post-crisis returns is rejected.

**Earnings and dividends predictability.** Equations (2a) and (2b) (p. 694) regress the
log-change in real dividends or earnings per share from year $$t-1$$ to $$t+5$$ on the
bank log excess total return in the crisis year, estimated conditional on the start of a
bank equity crisis. At the country level:

$$
\Delta y_{i,t-1,t+5} = \alpha_i + \beta r_{i,t-1,t} + \varepsilon_{i,t} \tag{2a}
$$

At the individual-bank level (banks indexed by $$b$$):

$$
\Delta y_{i,b,t-1,t+5} = \alpha_b + \beta r_{i,b,t-1,t} + \varepsilon_{i,b,t} \tag{2b}
$$

A coefficient $$\beta \approx 1$$ means a 1 log-point initial equity decline predicts an
approximately equal 1 log-point long-run decline in earnings or dividends, supporting the
permanent-loss view. Standard errors are Huber-White at the country level (Panel A) and
clustered by crisis episode at the bank level (Panel B), reported in Table III (p. 695).

**Heterogeneity by market-to-book ratio.** Equation (3) (p. 697) sorts banks into five bins
by their market-to-book (M/B) ratio at crisis onset to assess cross-bank heterogeneity:

$$
\Delta y_{i,b,t,t+5} = \alpha_i + \sum_k \beta_k \left(\frac{M}{B}\right)_{(\text{lower}_k,\, \text{upper}_k),\, i,b,t} + \varepsilon_{i,b,t} \tag{3}
$$

The five bins are M/B ratios 0-0.2, 0.2-0.4, 0.4-0.6, 0.6-0.8, and above 0.8. Country
fixed effects $$\alpha_i$$ are included. Results in Online Appendix Table A.10 show
monotonically worse five-year outcomes for more distressed (low M/B) banks.

**Book income decomposition.** Cumulative abnormal book income is computed relative to each
bank's average precrisis income (years $$t-4$$ to $$t-1$$), normalized by aggregate
precrisis book equity, then averaged across crises in 17 advanced economies. Income is
decomposed into: (i) write-downs (revaluations of balance-sheet assets: loan loss
provisions, impairments, goodwill write-downs); (ii) trading income (realized gains and
losses from securities trading and all asset sales); and (iii) all other book income.
Results in Table IV (p. 705) and Figure IV (p. 703).

## Empirical specifications

All main regressions use annual data conditional on the start of a bank equity crisis.
The primary country-level sample is 76 bank equity crises across 46 economies, 1870-2019;
individual-bank analyses cover the 17 JST economies.

**Returns analysis (R1).** Equation (1) is estimated with $$k=0$$ (cumulative returns
after crisis onset) and $$k=1$$ (normalized to the year before onset). Table II (p. 687)
Panel A reports bank returns and Panel B nonfinancial returns for $$h \in [1, 5]$$, with
panels C-D repeating the analysis using real total returns in place of excess returns.
Driscoll-Kraay standard errors are used throughout; the within-R² is reported as a
measure of explanatory power.

**Earnings and dividends predictability (R2).** Specifications (2a) and (2b) are estimated
by OLS, with the sample restricted to country-years (or bank-years) at the onset of bank
equity crises. The dependent variable is the log-change from year $$t-1$$ to $$t+5$$ in
real earnings per share (columns 1-2 in Table III) or real dividends per share (columns 3-4).
The independent variable is the past one-year bank log excess total return. For bank-level
regressions (Panel B), even-numbered columns add crisis fixed effects, so that the coefficient
captures within-crisis cross-bank heterogeneity (banks harder hit in a given crisis also show
larger long-run earnings declines).

**Income decomposition (R4).** Cumulative abnormal income components are averaged across
37 crisis episodes for the write-down vs. all-other split (Panel A) and 21 crisis episodes
for the three-way split (Panel B). Banks are sorted by the ratio of securities to total assets
in year $$t-1$$ to form top and bottom quartile subsamples (Panels C and D of Figure IV,
p. 703), examining whether banks with large tradeable-securities holdings experience more
immediate trading losses during panics.

**NPL cross-section (R5).** Figure V (p. 709) plots the unlevered cumulative log excess
total return of the bank equity index (from $$t-1$$ to $$t+5$$) against the peak NPL rate
(maximum from $$t$$ to $$t+5$$) across crises in advanced JST economies with available NPL
data. Unlevered returns (returns divided by banking sector book leverage) measure the implied
market value of asset losses. A line of best fit with adjusted R² is shown.

**Liquidity intervention event study (R6).** Using monthly BVX equity index returns, equation
(1) is re-estimated with the event month $$t$$ set to the first month of extraordinary central
bank liquidity support or blanket bank liability guarantee announcement (whichever occurs first)
for each crisis. Cumulative abnormal monthly returns are traced over a $$\pm 60$$ month window
with Driscoll-Kraay standard errors (Figure VI, p. 711).

**Crisis aversion probit (R8).** A probit regression on the full sample of historical banking
events (Online Appendix Table A.17) estimates predictors of crisis aversion using four
binary variables: (i) small bank equity decline preceding the panic (below 30%), (ii) early
liquidity intervention within one month of the panic, (iii) outbreak of war, and (iv) run
initially focused on a single institution (Online Appendix Table A.18).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| BVX equity index data (Baron, Verner and Xiong (2021)) | Country-level bank and nonfinancial equity index total returns for 46 economies, annual 1870-2016 and monthly; primary source for abnormal return analysis and crisis crisis dating | No page yet |
| JST Macrohistory Database (Jordà, Schularick and Taylor (2017)) | Individual-bank balance sheets, income statements, and equity returns for the 10 largest banks in each of 17 economies around each crisis; source for write-down decomposition and bank-level regressions | [JST Macrohistory](/wiki/datasets/jst-macrohistory/) |
| BSZ database (Baron, Schularick and Zimmermann (2024)) | Identities and annual balance sheets of the top-20 banks in 17 economies since 1870; used to identify bank failures and implement the real-time crisis definition | No page yet |
| Policy interventions database (new, this paper) | Monthly starting dates of extraordinary central bank liquidity support, blanket liability guarantees, and government recapitalizations across all 76 bank equity crises; extends Laeven and Valencia (2020) and Metrick and Schmelzing (2024) | No page yet |
| Ari, Chen and Ratnovski (2021) NPL data | Peak nonperforming loan rates for advanced economies used in the cross-crisis NPL scatter (Figure V) | No page yet |

Sample: 46 economies, annual, 1870-2019 for the main returns analysis; 17 JST economies
for individual-bank income decompositions and policy-effectiveness analyses. BVX monthly
data are used for short-horizon bounce-back (Figure II) and intervention event studies
(Figure VI). Replication data are available at Harvard Dataverse:
<https://doi.org/10.7910/DVN/NCUHLW>.

## When to read the full paper

Read the original if you are:

- **Building or calibrating a structural model of banking crises**: Tables II-IV provide empirical
  targets for bank equity dynamics, earnings declines, and write-down timing at both the country
  and individual-bank levels.
- **Assessing the effectiveness of lender-of-last-resort policy vs. recapitalization programs**:
  Sections V and VI cover both types of intervention with historical evidence; Table V documents
  recapitalization size, speed, and breadth across all 17 JST economies since 1870.
- **Studying heterogeneity in crisis outcomes**: Sections III.D-E compare banks sorted by M/B
  ratio; Section VI contrasts bank equity crises to panic-only crises and averted crises.
- **Using the historical banking crisis or policy intervention database**: Online Appendix
  Table A.1 lists all 76 bank equity crisis episodes with policy intervention dates; Table V
  lists individual government recapitalization programs with size, timing, and breadth statistics.

## Attribution and rights

Source: peer-reviewed, *The Quarterly Journal of Economics* (2026), pp. 667-732. This
distillation was extracted by an LLM on 2026-06-28 and is **not human-verified or
independently reproduced**. The work is paywalled; reproduction is extract-only.

> Baron, Matthew, Luc Laeven, Julien Pénasse, and Yevhenii Usenko. "Permanent Capital
> Losses after Banking Crises." *The Quarterly Journal of Economics* (2026): 667-732.
> DOI: 10.1093/qje/qjaf052.


==============================================================================
# Digital Distractions with Peer Influence: Barwick, Chen, Fu & Li (2026)
# https://instituteforautomatedresearch.org/wiki/papers/qje/2026/barwick-digital-distractions-peer-influence-2026/
# Distilled: Mobile app usage is contagious among college roommates and causally harms academic performance, physical health, and labor market outcomes. The Quarterly Journal of Economics 2026, paywalled. Nine core results with source locators, datasets used, the linear-in-means peer effects model, and shift-share IV identification.
# Tags: paper-summary, peer-reviewed, unreplicated, peer-effects, education
==============================================================================

**What this is.** This page distils the paper's core results, the linear-in-means peer effects model it tests, and the shift-share IV identification strategy, with exact table and page locators. To replicate or extend it, read the full source at the [original](https://doi.org/10.1093/qje/qjaf048). Replication data are available on Harvard Dataverse at [doi.org/10.7910/DVN/PAOKUU](https://doi.org/10.7910/DVN/PAOKUU).

## TL;DR

Using administrative records for 7,479 college students at a Chinese university linked to detailed mobile phone usage data from a major telecom carrier, the paper estimates causal effects of individual and peer app usage on academic performance, physical health, and early labor market outcomes. Three identification strategies handle endogeneity: (i) the university's random dormitory assignment, (ii) a shift-share IV interacting the September 2020 launch of blockbuster game *Yuanshen* (Genshin Impact) with students' precollege app usage, and (iii) a shift-share IV interacting China's October 2019 minors' game restriction policy with the evolving count of each student's underage precollege friends. App usage is contagious: a one standard deviation (SD) increase in roommates' in-college app usage raises a student's own usage by 5.8%, driven by behavioral spillover rather than shared contextual traits. A one SD increase in own app usage reduces GPA for required courses by 36.2% of a within-cohort-major SD and initial wages by 2.3%. The total peer effect on GPA (combining contagion and the direct disruption channel) reaches 22.7% of a GPA SD, more than half the own-usage effect. High-frequency GPS data show that app usage displaces time from study halls, increases class lateness and absences, and disrupts sleep. Extending China's gaming restriction to college students would boost initial wages by 0.9%, equivalent to roughly half the return to an additional year of work experience. The paper extends Stinebrickner and Stinebrickner (2008), who found roommates' video game console use harms GPA, by covering mobile apps across all categories, separating behavioral from contextual peer effects, and tracing consequences to wages and physical health.

## Core results

Magnitudes are as reported in the paper; `\*` / `\*\*` / `\*\*\*` = 10% / 5% / 1%. App usage measured in log hours; GPA on a 0-100 scale; wages in log RMB. All regressions control for class-by-gender (cohort-major-administrative-unit) and dorm-size fixed effects at minimum.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Behavioral peer effect on app usage**: 1 SD increase in roommates' in-college total app usage raises own usage by 5.8% | Table III Panel A col (4), p. 24 | IV coefficient 0.050 (s.e. 0.030)\*; F-stat 34.5; contextual effect 0.024 (s.e. 0.032, insig.), Table IV |
| R2 | **Own app usage reduces GPA** (required courses, IV): 1 SD increase reduces GPA by 36.2% of within-cohort-major SD | Table V IV model col (1), p. 28 | IV coefficient -0.613\*\*\* (s.e. 0.214); KP F-stat 16.9; OLS -0.546\*\*\* |
| R3 | **Roommates' app usage reduces GPA** (direct channel, IV): 1 SD increase in roommates' usage reduces GPA by 20.6% of within-cohort-major SD | Table V IV model col (1), p. 28 | IV coefficient -0.349\*\* (s.e. 0.155) |
| R4 | **Total peer effect on GPA** combining contagion and direct channels: 22.7% of a GPA SD reduction per 1 SD roommate increase | p. 32 (derived from R1 and R3) | -0.450 GPA points; exceeds 60% of own-usage effect |
| R5 | **Own app usage reduces PE scores** (IV): 1 SD increase reduces PE grade by 2.74 points, roughly four times the GPA effect; no direct roommate PE effect | Table V IV model col (4), p. 30-32 | IV coefficient -2.350\*\*\* (s.e. 0.854); SD-normalized -2.74 points; roommates' IV coefficient 0.140 (insig.) |
| R6 | **Own app usage reduces initial wages** (IV): 1 SD increase reduces graduation wages by 2.3%, or 12.1% of within-cohort-major wage SD | Table VI IV model col (4), p. 34-35 | IV coefficient -0.020\*\*\* (s.e. 0.006); KP F-stat 317.3 |
| R7 | **Roommates' app usage reduces wages**: direct IV -0.9% (4.8% SD); total effect including contagion -1.0% (5.3% SD) | Table VI IV model col (4), p. 35 | IV coefficient -0.008\* (s.e. 0.005); total ~half of own-usage wage effect |
| R8 | **Policy counterfactual**: extending China's three-hour weekly gaming cap to college students would raise initial wages by 0.9%, equivalent to roughly half the return to one extra year of work experience | pp. 36-37 (back-of-envelope) | Restriction binds 34.3% of student-month observations; reduces average monthly gaming 12.1 to 7.65 hours at steady state |
| R9 | **Time-allocation mechanism**: *Yuanshen* release causes students to arrive at study halls 18.2 minutes later and return to dorms 23.4 minutes earlier; minors' restriction has the opposite sign | Table VII cols (1)-(2), p. 39 | Study-hall arrival: +18.2 min (Yuanshen); dorm return: -23.4 min (Yuanshen); effects evident in lateness and class absences |

**Overall (paper's conclusion).** Mobile app usage imposes economically significant costs on both users and their peers. Behavioral peer spillovers dominate contextual peer effects: it is what roommates *do* (their app usage), not who they are (their prior characteristics), that drives the contagion. The negative consequences extend from academic performance to physical health and early wages, with time displacement from study and sleep as the primary mechanism. A gaming restriction targeted at college students would meaningfully offset these costs.

## Theory / model

The paper does not build a structural model. It tests a linear-in-means peer effects framework, originally formalized by Manski (1993), in which an individual's outcome depends on both their own predetermined characteristics and on the contemporaneous behavior and characteristics of their peer group.

**Linear-in-means specification.** Let $$y_{it}$$ be individual $$i$$'s in-college app usage in month $$t$$, $$x_i$$ their predetermined characteristic (precollege app usage), and $$N_i$$ the set of their roommates. The model (equation (1), p. 18) is:

$$
y_{it} = \alpha + \gamma x_i + \beta \frac{1}{|N_i|} \sum_{j \in N_i} y_{jt} + \delta \frac{1}{|N_i|} \sum_{j \in N_i} x_j + \epsilon_{it} \tag{1}
$$

where $$\beta$$ is the behavioral peer effect (how peers' contemporaneous app usage affects own usage) and $$\delta$$ is the contextual peer effect (how peers' prior characteristics affect own usage, independent of their behavior). The reflection problem, first formalized by Manski (1993), means $$\beta$$ and $$\delta$$ are not separately identified in a cross-section. Bramoullé, Djebbari, and Fortin (2020) survey conditions under which panel variation and exclusion restrictions restore identification.

**Hypothesis tested.** The paper's central hypothesis is that app usage is *contagious*: $$\beta > 0$$. Beyond this, it tests whether $$\beta$$ dominates $$\delta$$ (behavioral versus contextual), and whether the combined peer channel transmits harmful effects on GPA and wages. The paper also tests whether the displacement of study time and sleep (mechanism) and the direct disruption of the study environment (direct channel) account for the academic penalty.

**Identification logic.** Three sources of quasi-random variation are used. (i) The university randomly assigns freshmen to single-gender dorm rooms within administrative classes (cohort-major-administrative-unit), yielding exogenous peer groups; the random assignment is verified by a balance table (Online Appendix Table C.1). (ii) The September 2020 launch of *Yuanshen* (Genshin Impact) differentially increased app usage for students with higher precollege gaming intensity; interacting the launch indicator with precollege usage produces a shift-share instrument. (iii) China's October 2019 National Press and Publication Administration policy prohibiting individuals under 18 from gaming between 10 p.m. and 8 a.m. (and capping gaming at 90 minutes on weekdays) differentially reduced app usage for students with more underage precollege friends; interacting the policy with the evolving underage-friend count produces a second shift-share instrument for peer usage.

## Method

The paper estimates three nested models using these identification strategies: the reduced-form peer effect (which mixes behavioral and contextual channels), the behavioral peer effect (isolated via the minors' restriction IV), and the GPA / wage effects (via the Yuanshen and restriction IVs). It builds on `panel-regression` with student and class-semester fixed effects, `instrumental-variables` (2SLS) for endogenous app usage, and `event-study` graphs to validate the exclusion restrictions.

**Reduced-form peer effect.** Substituting the linear-in-means model into itself and using the random assignment of roommates, the estimating equation (equation (2), p. 18) is:

$$
y_{it} = \theta_a + \theta_{\gamma_1} x_i + \theta_{\gamma_2} \frac{1}{|N_i|} \sum_{j \in N_i} x_j + \mathbf{z}'_{it} \rho + \eta_{cg} + \eta_m + \eta_t + \varepsilon_{it} \tag{2}
$$

where $$\theta_{\gamma_2}$$ is the reduced-form peer effect (a function of both $$\beta$$ and $$\delta$$), $$\mathbf{z}_{it}$$ is a vector of demographic controls, $$\eta_{cg}$$ are class-by-gender fixed effects, $$\eta_m$$ are dorm-size fixed effects, and $$\eta_t$$ are month-of-sample fixed effects. Standard errors are clustered at the class level. OLS on this equation yields the causal reduced-form estimate because roommate precollege usage is predetermined (set before any college interaction) and randomly assigned. Sacerdote (2001) established this strategy for peer effects at Dartmouth; the paper extends it to a Chinese university context with mobile app data.

**Behavioral peer effect isolation.** Adding student fixed effects absorbs time-invariant individual characteristics (including $$x_i$$) to estimate equation (3) via 2SLS:

$$
y_{it} = \eta_i + \beta \frac{1}{|N_i|} \sum_{j \in N_i} y_{jt} + \epsilon_{it} \tag{3}
$$

The IV for average roommates' in-college usage $$\frac{1}{|N_i|} \sum y_{jt}$$ is the interaction of the minors' game restriction policy timing with the evolving count of underage precollege friends among roommates. Because roommate friend networks are predetermined and do not overlap with the focal student's, this instrument affects peers but not the focal student directly, satisfying the exclusion restriction.

**GPA and wage estimation.** OLS regresses GPA on own and roommates' current app usage (equation (4), p. 27):

$$
\text{GPA}_{is} = \alpha_1 \text{Phone}_{is} + \alpha_2 \frac{1}{|N_i|} \sum_{j \in N_i} \text{Phone}_{js} + \alpha_3 \text{CEE}_i \times \eta_s + \eta_i + \eta_{cs} + \epsilon_{is} \tag{4}
$$

where $$\text{Phone}_{is}$$ is log app usage in semester $$s$$, $$\text{CEE}_i \times \eta_s$$ is an interaction between the student's college-entrance exam score and a semester trend (absorbing differential GPA trends by incoming ability), and $$\eta_{cs}$$ are class-by-semester fixed effects. The IV first stage (equation (5), p. 31) instruments both own and roommates' usage using the Yuanshen and restriction shocks:

$$
y_{is} = \lambda_1 \text{YS}_s \times \text{PrePhone}_i + \lambda_2 \text{YS}_s \times \frac{1}{|N_i|} \sum_{j \in N_i} \text{PrePhone}_j + \lambda_3 \text{Policy}_s \times \text{Minor}_{is} + \lambda_4 \text{Policy}_s \times \frac{1}{|N_i|} \sum_{j \in N_i} \text{Minor}_{js} + \lambda_5 \text{Minor}_{is} + \lambda_6 \frac{1}{|N_i|} \sum_{j \in N_i} \text{Minor}_{js} + \text{CEE} \times \eta_s + \eta_i + \eta_{cs} + \epsilon_{is} \tag{5}
$$

where YS is the Yuanshen release indicator, PrePhone is precollege app usage, Policy is the minors' restriction indicator, and Minor is the evolving count of underage precollege friends.

For wages, because outcomes are measured once per student, student fixed effects are infeasible; the paper instead controls for a rich set of student attributes and the estimated GPA fixed effect $$\hat{\eta}_i$$ (equation (6), p. 33):

$$
y_i = \gamma_1 \text{Phone}_i + \gamma_2 \frac{1}{|N_i|} \sum_{j \in N_i} \text{Phone}_j + X'_i \gamma_X + \eta_{cg} + \eta_m + \hat{\eta}_i + \varepsilon_i \tag{6}
$$

where $$\text{Phone}_i$$ is the student's average in-college app usage across all semesters (predicted from the IV first stage), $$X_i$$ includes demographic controls and hometown fixed effects, and $$\hat{\eta}_i$$ captures time-invariant ability and effort.

## Empirical specifications

**Peer effects on app usage (Section III).** Reduced-form estimates use equation (2) with OLS; the random assignment delivers causal identification without instruments. Standard errors are clustered at the class level (a cohort-major-administrative-unit triplet of 20-50 students). Behavioral peer effects in equation (3) are estimated by 2SLS using the minors' restriction shift-share instrument; Kleibergen-Paap F-statistics for total app usage is 34.5 (Table III col (4)) and for game apps 31.2 (col (2)). Event studies (Figure II) confirm pre-trend flatness for both the Yuanshen and policy shocks.

**GPA and PE effects (Section IV.A).** OLS on equation (4) and 2SLS using four instruments from equation (5): the Yuanshen interaction with own and roommates' precollege usage, and the restriction-policy interaction with own and roommates' underage friend count. Sample is student-semester cells (excluding spring 2020, COVID). Kleibergen-Paap F-statistics: 16.9 for total app usage (Table V col (1)), 14.3 for games (col (2)), 19.6 for games + video (col (3)). Hansen J p-values exceed 0.29 in all columns (Table V), supporting instrument validity. Standard errors clustered at the class level.

**Wage effects (Section IV.B).** OLS and 2SLS on equation (6). Sample: 2,812 students from 2018 and 2019 cohorts who had a job one month after graduation. Phone is predicted in-college usage averaged across semesters. Kleibergen-Paap F-statistics are very large (317.3 for total app time, up to 4,625 for game apps, Table VI). Standard errors clustered at the class level.

**Time allocation (Section V.A).** Specification (5) with the dependent variable replaced by one of six on-time performance measures (time of study-hall arrival, dorm return, duration at study hall, duration at dorm, class lateness, class absence). Controls include week-of-sample, day-of-week, class-semester, and student fixed effects, plus interactions between week-of-year and precollege app usage.

**Sleep patterns (Section V.B).** OLS regressions for the 2020 cohort (November 2023 to June 2024 with hourly app usage data) regressing sleep duration or late-sleep / late-wakeup indicators on total nighttime and daytime app usage, controlling for student, class-by-semester, week-of-sample, and day-of-week fixed effects.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Chinese university administrative records (2018-2020 cohorts, 7,479 students) | Roommate assignments, CEE scores, demographics, college transcripts (GPA per course per semester), postgraduation employment status and initial wages | No page yet (proprietary-confidential institutional data) |
| Mobile phone usage data (major Chinese telecom carrier, province-level, 2018-2021) | Monthly in-college app usage in log hours by category (social media, video, games, other) for 6,430 matched students; GPS location data at five-minute intervals; hourly usage for 2020 cohort | No page yet (proprietary-confidential telecom data) |
| Precollege friend network (phone call records) | Identifies predetermined "private" friends: bilateral calls from 2 months before college start; used as instrument-construction input (underage friend count) | No page yet |
| University annual survey (2 waves: 2022 for 2018 cohort; 2023 for 2019-2020 cohorts) | Personality, health, job search behaviors, attitudes toward gaming; 1,798 respondents (24% response rate); reweighted for representativeness | No page yet (university internal survey) |

Sample coverage: September 2018 to June 2021 (spring 2020 excluded for COVID). Wage data: graduates of 2018 cohort (June 2022) and 2019 cohort (June 2023). Average total monthly app usage: 92.9 hours (s.d. 108.5); average GPA: 78 (s.d. 6.6). The anonymized replication dataset is publicly available at Harvard Dataverse ([doi.org/10.7910/DVN/PAOKUU](https://doi.org/10.7910/DVN/PAOKUU)).

## When to read the full paper

Read the [original](https://doi.org/10.1093/qje/qjaf048) if you are: (i) replicating with the Harvard Dataverse data and code; (ii) extending the peer-effects decomposition framework (behavioral vs. contextual via equation (3)) to other technology use or addiction contexts; (iii) designing or evaluating screen-time restriction policies for students; (iv) studying mechanisms in detail via the GPS time-allocation or sleep-pattern analyses (Section V, Tables VII-VIII). The locators above (Table III, V, VI, VII) point to the exact panels.

## Attribution and rights

Source: peer-reviewed, *The Quarterly Journal of Economics* 141(1), 2026. © The Author(s) 2025. Published by Oxford University Press on behalf of President and Fellows of Harvard College. All rights reserved. This page is an extract-only LLM distillation by the Institute for Automated Research (2026-06-28). It is **not human-verified and not independently reproduced**.

> Barwick, Panle Jia, Siyu Chen, Chao Fu, and Teng Li. "Digital Distractions with Peer Influence: The Impact of Mobile App Usage on Academic and Labor Market Outcomes." *The Quarterly Journal of Economics* 141, no. 1 (2026): 1–49. DOI: [10.1093/qje/qjaf048](https://doi.org/10.1093/qje/qjaf048).


==============================================================================
# Traditional Institutions in Modern Times: Bau, Khanna, Low & Voena (2026)
# https://instituteforautomatedresearch.org/wiki/papers/qje/2026/bau-traditional-institutions-modern-times-2026/
# Distilled: Using two new surveys on dowry property rights and a natural experiment from India's highway expansion, this paper shows that grooms' parents commonly retain dowry from migrant sons and that stronger historical dowry traditions predict higher male out-migration rates and larger migration responses to falling migration costs. The Quarterly Journal of Economics 141(1), 2026, paywalled. Seven core results with source locators, datasets, the theoretical model, and empirical specifications.
# Tags: paper-summary, development-economics, migration, marriage-markets
==============================================================================

**What this is.** The paper's core results, the household model, and the empirical specifications with the actual equations: enough to know what it found and how, without reading all 58 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1093/qje/qjaf041).

## TL;DR

The paper introduces and tests the hypothesis that dowry in modern India functions as a pension for grooms' parents. When sons migrate for work, they can no longer provide traditional in-person old-age support; the liquidity from the bride's dowry lets families make upfront transfers to parents at the time of marriage, easing this intergenerational friction. Using two new surveys (Destination Survey 2018, Origin Survey 2020) with first-ever quantitative data on who owns dowry within the household, the paper documents that 27-45% of grooms' parents are net takers from the dowry, especially when sons are migrants and earn more. In nationally representative NSS data, male migration rates are higher in districts with stronger historical dowry traditions (Giuliano and Nunn (2018) methodology). Exploiting the staggered construction of the Golden Quadrilateral (GQ) and North-South/East-West highway corridors as a natural experiment (Borusyak, Jaravel, and Spiess (2024) event-study estimator), young men (15-30) from dowry districts show significantly larger migration increases than those from non-dowry districts when migration costs fall.

## Core results

Magnitudes and significance as reported; `\*\*`/`\*\*\*` = 5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Net dowry taking by grooms' parents is **common but heterogeneous**: some parents take, many give, distribution roughly centered at zero | Online Appendix Tables A.1, A.2; Figure III, p. 231 | 45% net takers (Destination Survey); 27% (Origin Survey); distribution of Net Transfers to Grooms' Parents spans both positive and negative, unlike gross/net dowry measures in the literature |
| R2 | **Migrant sons' parents are significantly more likely to take from the dowry** than non-migrant sons' parents | Table I cols (1) and (3), p. 232 | Origin Survey: migrant coefficient = 0.076\*\* (se=0.038); Destination Survey: 0.218\*\* (se=0.086); controlling for coresidence, marriage year, age, son/father education FE, and net dowry |
| R3 | **Higher-earning migrant sons' parents are more likely to take**: interaction of migration and son's occupational score is positive and significant | Table I col (2), p. 232 | Origin Survey: Migrant x ln(son occ. score) = 0.199\*\* (se=0.090); non-migrant x score = 0.014 (ns) |
| R4 | **Greater parental bargaining power (Pareto weight) increases net taking** for both migrant and non-migrant sons | Table I col (4), p. 232 | Migrant x veto power = 0.229\*\*\* (se=0.076); Non-migrant x veto power = 0.221\*\* (se=0.086); consistent with dowry redistributing resources according to Pareto weights regardless of migration |
| R5 | **Sons who remit are more likely to have parents who took from the dowry** (Prediction 4: remittances signal high-Pareto-weight households that exhaust all transfer channels) | Table II col (2), p. 238 | Son transfers x migrant son = 0.176\*\* (se=0.075); robust to income controls |
| R6 | **Historical dowry traditions positively predict male out-migration rates** in NSS data | Table III cols (1)-(2), p. 246 | Continuous dowry = 0.0257\*\* (se=0.0101) to 0.0451\*\* (se=0.0199); remains positive controlling for state and birth-year FE, caste, ethnographic and geographic controls |
| R7 | **GQ highway construction raises out-migration significantly more for young men (15-30) in dowry districts** than in non-dowry districts; no effect for older men (31-45) | Figure VI, p. 252 | Large, significant post-construction increase in all and employment-based migration for men aged 15-30 in dowry districts (0.1% cutoff); coefficient near zero for non-dowry districts; Panel B effect for ages 31-45 is indistinguishable from zero |

**Overall (paper's conclusion).** All six model predictions are confirmed. Dowry appears to have adapted in modern India: rather than functioning purely as a bequest to daughters as in its traditional form described by Botticini and Siow (2003), it increasingly operates as a mechanism for sons' parents to secure resources at the time of marriage when sons migrate and can no longer co-reside and provide in-kind old-age support. Dowry traditions may therefore help explain why rural-urban migration in India remains below its aggregate-productivity-maximizing level as identified by Munshi and Rosenzweig (2016): areas with weaker dowry traditions face higher effective barriers to migration.

## Theory / model

The paper (Section III, pp. 215-223) develops a two-stage collective household model (Chiappori 1988) in which a family with parents and one son decides jointly over marriage gifts, old-age savings, and migration.

**Setup.** Parents have Pareto weight $$\theta(\mathbf{z}) \in (0,1)$$, where $$\mathbf{z}$$ are distribution factors (e.g., whether parents have veto power over the son's marriage). The son has Pareto weight $$1 - \theta(\mathbf{z})$$. Parents earn $$y_{1P}$$ in stage 1 (working age) and zero in stage 2 (retirement). The son earns zero in stage 1 and $$y_{2K}$$ in stage 2, plus a net return to migration $$R$$ if $$m = 1$$. The bride's endowment is $$E$$, of which a liquid fraction $$d \in [0,1]$$ is available as dowry at marriage (stage 1); the illiquid fraction $$(1-d) \cdot E$$ represents future inheritance or human capital. Migration introduces a friction $$\gamma > 0$$: transferring one unit $$\alpha$$ from the son to the parents costs $$(1 + \gamma \cdot m) \cdot \alpha$$ when $$m = 1$$.

**Household optimization.** The family solves (equation (1), p. 217):

$$
\max_{\substack{G \geq 0,\; S_1 \geq 0,\; \alpha \geq 0 \\ m \in \{0,1\},\; c_{2P} \geq 0,\; c_{2K} \geq 0}}
\theta(\mathbf{z}) \ln(c_{2P}) + (1 - \theta(\mathbf{z})) \ln(c_{2K}) \tag{1}
$$

subject to:

$$
S_1 + G \leq y_{1P} + d \cdot E
$$

$$
c_{2P} \leq S_1 + \alpha
$$

$$
c_{2K} \leq y_{2K} + (1-d) \cdot E + R \cdot m + G - (1 + \gamma \cdot m) \cdot \alpha
$$

Here $$G \geq 0$$ is the net marriage gift to the son (the son is a net taker when $$G < d \cdot E$$, i.e., parents retain $$d \cdot E - G > 0$$), $$S_1$$ is parents' savings, and $$\alpha \geq 0$$ is the son's stage-2 transfer to parents. Grooms' parents are **net takers** ($$d \cdot E > G$$) when the net transfer to them is positive.

**Solutions (pp. 218-219).** Three regimes arise when $$m = 1$$:

(i) *Marriage-gifts solution*: $$\alpha^* = 0$$, $$G^* > 0$$. Stage-1 resources are sufficient for parents to achieve their first-best consumption without costly stage-2 remittances. Migration does not distort allocation.

(ii) *Autarky solution*: $$\alpha^* = 0$$, $$G^* = 0$$. Stage-1 resources are in an intermediate range; parents exhaust them and the son sends nothing in stage 2.

(iii) *Remittances solution*: $$\alpha^* > 0$$, $$G^* = 0$$. Stage-2 resources are high enough relative to stage-1 to warrant costly remittances despite $$\gamma > 0$$.

In regimes (ii) and (iii), migration only occurs when $$R$$ exceeds a strictly positive threshold (the consumption distortion cost of migration). A larger $$d$$ (stronger dowry practice) lowers this threshold by pre-funding the stage-2 consumption shortfall at the cheaper stage-1 price, reducing the migration friction. This is the core mechanism: dowry practices enable migration by front-loading intergenerational transfers to the pre-migration stage, where the transfer cost is zero.

**Six testable predictions** (Section III.C, pp. 220-223):
1. Net transfers to grooms' parents can be positive or negative (heterogeneous across families).
2. Parents are more likely to be net takers when sons migrate.
3. Net taking is increasing in the migrant son's income and in the parental Pareto weight.
4. Parents who receive remittances from migrant sons are more likely to be net dowry takers.
5. Families in areas with stronger dowry practices (higher $$d$$) are more likely to have a migrant son.
6. A decline in migration cost increases migration more in areas with higher $$d$$ (when migration rates are low).

## Method

**Survey design (Section IV, pp. 223-230).** The paper introduces two original data sets on dowry property rights:

- *Destination Survey (2018)*: in-person interviews with 557 men aged 21-41 in Gurugram, stratified 20% Delhi natives and 80% migrants. Detailed gift-by-gift ownership questions for every item transferred at marriage, including who gave it and who holds property rights today, allowing construction of gross and net transfers to each family member.
- *Origin Survey (2020)*: phone interviews with 2,541 households in 34 districts of six Indian states (Rajasthan, Uttar Pradesh, Bihar, Jharkhand, Madhya Pradesh, Maharashtra). Random sample of one married son per household, oversampling migrants (70%/30%); yielded data on 3,069 sons.

Net transfers to grooms' parents are defined as: gross transfers received by grooms' parents minus gross transfers made by grooms' parents to other parties (p. 225). Grooms' parents are net takers when this quantity is positive.

**Historical dowry traditions (Section V.A, pp. 239-243).** Following Giuliano and Nunn (2018), the paper constructs a district-level measure of the share of the current population belonging to linguistic groups with historical dowry practices, drawing on ethnographic data from the Murdock (1967) *Ethnographic Atlas* combined with current language group maps from the *Ethnologue* and LandScan population weights. Districts are coded as high-dowry when the measure exceeds 0.1% (368 of 582 districts).

**Highway natural experiment (Section V.D.2, pp. 248-251).** The paper exploits the staggered construction of the Golden Quadrilateral and NS-EW highway corridors (1999-2016, ~5,846 km) as a quasi-random reduction in migration costs across districts and years. Highway project timing and district location are matched from the NHDP project list to the CMIE CapEx database. The staggered event-study estimator of Borusyak, Jaravel, and Spiess (2024) is used (doubly-robust estimator; Callaway and Sant'Anna (2021) as robustness). Standard errors are wild-bootstrapped and clustered at the district level.

## Empirical specifications

**Predictions 2 and 3: OLS on net-taker indicator (Table I, p. 232).** For each son $$h$$ in survey $$s$$:

$$
\text{NetTaker}_{hs} = \alpha_c + \beta_1 \text{Migrant}_{hs} + \mathbf{X}_{hs}\boldsymbol{\gamma} + \varepsilon_{hs} \tag{T1}
$$

where $$\alpha_c$$ denotes coresidence status fixed effects and marriage year and age fixed effects (two separate additive groups as listed in Table I, not a three-way interaction cell); $$\mathbf{X}_{hs}$$ includes son education dummies, father education dummies, and a quadratic in net dowry. For Prediction 3 (income interaction), the specification adds $$\text{Migrant}_{hs} \times \ln(\text{son occ. score})_{hs}$$ and $$\text{Non-migrant}_{hs} \times \ln(\text{son occ. score})_{hs}$$. For the Pareto weight test, the specification adds $$\text{Migrant}_{hs} \times \text{VetoPower}_{hs}$$ and $$\text{Non-migrant}_{hs} \times \text{VetoPower}_{hs}$$.

**Prediction 4: remittances and net-taking (Table II, p. 238).** Using Origin Survey data:

$$
\text{NetTaker}_{hs} = \alpha + \beta_1 \text{SonTransfers}_{hs} + \beta_2 \text{Migrant}_{hs} + \beta_3 (\text{SonTransfers} \times \text{Migrant})_{hs} + \mathbf{X}_{hs}\boldsymbol{\gamma} + \varepsilon_{hs} \tag{T2}
$$

where SonTransfers is an indicator for net financial transfers from son to parents in the year before the survey.

**Prediction 5: dowry traditions and migration (Table III, p. 246).** NSS Round 64 cross-section of men aged 15-45:

$$
\text{Migrated}_{ids} = \beta \cdot \text{Dowry}_d + \delta_s + \delta_{\text{year}} + \mathbf{X}_{ids}\boldsymbol{\gamma} + \varepsilon_{ids} \tag{T3}
$$

where $$d$$ indexes districts, $$s$$ indexes states, year is birth year. $$\text{Dowry}_d$$ is the continuous population-share measure or a discrete indicator. $$\mathbf{X}_{ids}$$ includes caste FE, ethnographic controls (percent patrilineal, percent plow), geographic controls (latitude, longitude, coastal distance, city distance), and education controls.

**Prediction 6: GQ highway event study (Equation (2), p. 249).** Individual $$i$$ of age $$a$$, state $$j$$, district $$d$$, year $$t$$:

$$
y_{iajdt} = \alpha_i + \theta_{jt} + \delta_a + \sum_s \beta_s \cdot GQ_{dts} + \mathbf{X}_{iajdt}\boldsymbol{\gamma} + \varepsilon_{iajdt} \tag{2}
$$

where $$y_{iajdt}$$ is an indicator for having migrated before year $$t$$; $$\alpha_i$$ are individual FE; $$\theta_{jt}$$ are state-by-year FE; $$\delta_a$$ are age FE; $$GQ_{dts}$$ is an indicator equal to one if a highway segment was constructed $$s$$ years ago in district $$d$$ by year $$t$$. The equation is estimated separately for dowry districts (above 0.1% cutoff) and non-dowry districts, with $$t = 0$$ normalized to the year of first highway construction in the district. The sample is restricted to men aged 13-45 at the time of receiving a GQ project to avoid capturing age-dependent migration (dependent migration for under-15s, old-age migration for over-45s).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Author Destination Survey 2018 (Gurugram) | Primary survey: property rights over dowry items for 557 male migrants and non-migrants; net transfers to grooms' parents (Predictions 1-4) | No page yet (hand-collected) |
| Author Origin Survey 2020 (6 Indian states) | Primary survey: phone interviews with 2,541 households, 3,069 married sons; dowry allocation and migration by same-origin comparison (Predictions 1-4) | No page yet (hand-collected) |
| NSS Round 64 (2007-08 migration module) | Nationally representative individual-level migration data; ~188,192 men aged 15-45 in 588 districts; tests Predictions 5 and 6 | No page yet ([data:nss-india](/wiki/tags/)) |
| Giuliano-Nunn Ancestral Characteristics | District-level share of population with historical dowry tradition; constructed from Murdock (1967) Ethnographic Atlas + Ethnologue language maps | No page yet ([data:giuliano-nunn-ancestral](/wiki/tags/)) |
| REDS 1999 (NCAER) | Validation of historical dowry measure against contemporary dowry payment size | No page yet ([data:reds-india](/wiki/tags/)) |
| IHDS 2005 and 2011-12 | Validation of historical dowry measure (gold payment likelihood); robustness tests for migration predictions | No page yet ([data:ihds-india](/wiki/tags/)) |
| CapEx (CMIE, 2023) | Infrastructure project timing and district location for GQ and NS-EW highway segments; matched to NHDP project list | No page yet ([data:capex-cmie](/wiki/tags/)) |

Sample periods: Destination Survey 2018, Origin Survey 2020, NSS migration 1996-2007, CapEx highway projects 1999-2016.

## When to read the full paper

Use the [original](https://doi.org/10.1093/qje/qjaf041) if you are: (i) studying how traditional institutions adapt to economic development (the mechanism literature building on Botticini and Siow (2003) and Anderson and Bidner (2015)); (ii) working on migration frictions in low-income countries, particularly the role of old-age support constraints identified by Munshi and Rosenzweig (2016); (iii) applying staggered event-study methods to infrastructure programs in developing countries; (iv) using the Giuliano and Nunn (2018) Ancestral Characteristics methodology for India; or (v) replicating the dowry property-rights surveys (replication data at Harvard Dataverse, DOI above). Tables I-III and Figure VI contain the headline numbers; Online Appendix B.5 details survey construction.

## Attribution and rights

Source: peer-reviewed, *The Quarterly Journal of Economics* 141(1), 2026. This distillation was extracted by an LLM on 2026-06-28 and is **not human-verified or independently reproduced**. The paper is paywalled; all rights reserved by Oxford University Press on behalf of President and Fellows of Harvard College. No PDF is hosted here.

> Bau, Natalie, Gaurav Khanna, Corinne Low, and Alessandra Voena. "Traditional Institutions in Modern Times: Dowries as Pensions When Sons Migrate." *The Quarterly Journal of Economics* 141, no. 1 (2026): 205-262. DOI: 10.1093/qje/qjaf041.


==============================================================================
# Bargaining and Inequality in the Labor Market: Caldwell, Haegele & Heining (2026)
# https://instituteforautomatedresearch.org/wiki/papers/qje/2026/caldwell-bargaining-inequality-labor-market-2026/
# Distilled: A novel matched firm-worker survey linked to German administrative data documents that individual wage bargaining is pervasive (78% of workers exposed), that labor market factors predict firms' bargaining strategies better than firm productivity, that workers with better outside options negotiate more successfully, and that gender wage gaps are 3-5 percentage points larger at bargaining firms. The Quarterly Journal of Economics (2026), paywalled. Eight core results with source locators, datasets used, the empirical framework, and the estimating equations.
# Tags: paper-summary, labor-economics, wage-inequality, bargaining, gender-gap
==============================================================================

**What this is.** The paper's core results, the empirical framework, and the estimating equations: enough to understand what it found and how, without reading all 57 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1093/qje/qjaf049).

## TL;DR

The paper introduces and validates a novel survey measure of firm wage-bargaining strategies for 772 German firms (fielded through the ifo Institute), linked to German Social Security records (IEB) and a complementary worker survey of nearly 10,000 full-time workers. Most workers (78%) are employed in positions where the firm reports that individual bargaining is possible, and the typical firm is willing to differentiate pay by 6%-12% depending on employee group. Firm productivity does not predict which firms bargain; labor market factors such as job tightness and employee group replaceability do. Worker-firm bargaining events typically begin with workers providing salary expectations; most outside offers are ultimately rejected and used to renegotiate at the incumbent firm. Workers with better outside options and higher risk tolerance ask for more and receive more, while women are systematically less likely to ask and to succeed. At firms that bargain, residual gender wage gaps are 3-5 percentage points larger than at posting firms, and a worker's prior-firm pay premium (AKM firm effect) predicts current pay only under individual bargaining.

## Core results

Magnitudes and significance as reported; `\*\*`/`\*\*\*` = 5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Individual bargaining is pervasive**: 78% of workers at surveyed firms are in positions where the firm can differentiate pay by individual bargaining | Figure I (Panels A-B), p. 333-334 | 95% of firms can differentiate pay for managers; 85% for experienced non-managers; 57% for recent entrants with outside offers; 50% for new labor market entrants |
| R2 | **Firms expect substantial initial-offer variation**: typical expected spread between highest and lowest initial offers to identically qualified candidates is 3%-10% depending on group | Figure II, p. 335; Figure III, p. 336 | 3% for recent entrants, 5% for experienced non-managers, 10% for managers (conditional on nonzero: 6%, 10%, 12%); final-offer gap is similar |
| R3 | **Labor market factors beat firm characteristics** in explaining bargaining strategies; employee-group dummies explain as much as 500+ firm fixed effects | Table III, p. 343 | Group dummies alone: R² = 0.33, adj. R² = 0.33; all firm FE: R² = 0.40, adj. R² = 0.19; firm controls without industry dummies (cols 4-7) keep adj. R² ≤ 0.35; 4-digit industry dummies (cols 8-9) reach adj. R² = 0.44 |
| R4 | **Most outside offers are rejected**; workers use them to renegotiate at the incumbent firm | Table IV, p. 346 | 91% of workers who received outside offers stayed; 33% attempted renegotiation with incumbent; 46% of renegotiation attempts succeeded |
| R5 | **Outside options drive bargaining success**: workers with better outside options are 9 pp more likely to ask for a wage increase at the start of an employment spell | Table V Panel A, p. 351 | Outside options (binary): +8.7 pp`\*\*\*` asked, +6.7 pp`\*` successfully negotiated upward; level: +0.056 ask, +0.487 pp negotiated; consistent effects in previous 6 months |
| R6 | **Women ask for and receive less**: women are 6 pp less likely to successfully negotiate pay up at the start of a spell and during employment; gap cannot be fully explained by worse outside options | Table V Panels A-B, pp. 351-352 | Start of spell: -5.8 pp ask, -6.4 pp`\*\*` succeeded; in previous 6 months: -5.8 pp`\*\*\*` asked, -6.4 pp`\*\*\*` received raise |
| R7 | **Gender pay gaps are 3-5 pp larger at bargaining firms** after controlling for occupation-establishment fixed effects; 44% of residual gender gap at surveyed firms attributed to bargaining | Table VI Panel A, p. 359; Figure V, p. 361 | Female coeff. at bargaining firms: -0.053`\*\*` (s.e. 0.023, occ-est FE, col 5); at posting firms: 0.008 (s.e. 0.032, col 2); ~6 pp difference robust to controlling for hours and excluding special pay |
| R8 | **Prior-firm pay premium predicts current pay only under bargaining**: 10 pp higher prior-firm AKM effect associated with 0.5% higher current pay, but only at bargaining firms | Table VII Panel A, p. 363 | Bargaining firms: prior-firm effect = 0.049`\*\*\*` (s.e. 0.010); posting firms: 0.006 (s.e. 0.018); p-value of equality = 0.016 |

**Overall (paper's conclusion).** Individual wage bargaining is empirically pervasive in Germany, is driven by labor market factors rather than firm productivity, and generates meaningful wage dispersion. Providing workers with pay information (a common policy proposal) would not suffice to close gender gaps in bargaining behavior; residual differences persist even in hypothetical scenarios with equalized information. The prior-firm pay persistence result implies that when workers move jobs, their starting wage partly reflects the bargaining regime at their previous employer, not just their productivity.

## Theory / model

The paper has no formal model of its own. It situates itself within the theoretical literature on imperfect competition in the labor market (Manning (2011)): when workers face search frictions and firms earn monopsonistic rents, wages can be set by individual bargaining rather than wage posting, and the distribution of rents varies with outside options and bargaining power.

The paper tests two classes of theoretical predictions from models of firm wage-setting strategy:

1. **Productivity-based theories** (Postel-Vinay and Robin (2004); Doniger (2015); Flinn and Mullins (2021)): more productive firms will be more likely to bargain with workers to capture a larger share of the surplus. The paper finds these predictions are **rejected**: firm age, size, and assets per employee do not predict whether a firm bargains.

2. **Labor market factor theories** (Ellingsen and Rosén (2003); Michelacci and Suarez (2006)): firms bargain when it is difficult to replace workers or when vacancy tightness is high. The paper finds these predictions are **confirmed**: market tightness (bottleneck occupations) and employee replaceability (experienced workers, managers) predict bargaining strategies, and employee-group dummies explain as much variation as all firm fixed effects combined.

For the inequality analysis, the identification logic is:
- **Gender pay gap (Section VI.A)**: the paper extends Biasi and Sarsons (2022) to a broader multi-sector German sample, comparing conditional gender gaps within occupation-establishment cells across firms with and without bargaining.
- **Prior-firm pay persistence (Section VI.B)**: the paper compares the relation between a worker's prior-firm AKM wage premium (Abowd, Kramarz, and Margolis (1999); estimated from 2010-2017 population data by Bellmann et al. (2020)) and current pay at bargaining vs. posting firms. Card, Heining, and Kline (2013) documented the rise of firm heterogeneity in these AKM effects in Germany; the present paper identifies individual bargaining as a contributing mechanism.

Both exercises are descriptive/comparative; neither makes a causal identification claim beyond selection-on-observables within occupation-establishment cells.

## Method

The key methodological contribution is the design and validation of a survey instrument to measure firm wage-bargaining strategies, following the management-practices survey approach of Bloom and Van Reenen (2007).

**Protocol question (main measure of firm bargaining strategy).** Firms were asked separately for four employee groups:

> How much more could a person maximally receive compared to the fixed compensation you would have offered based on the person's qualification/fit for the position alone?

Response options: 0% (no adjustment), 1%-10%, 11%-20%, 21%-30%, 31%-40%, more than 40%. A firm is classified as having a bargaining strategy if it reported any nonzero adjustment. For incumbent workers facing outside offers:

> Suppose an employee at your company receives an external offer from another company and requests a salary increase. What is the maximum percentage by which your firm could possibly increase the fixed compensation (without changing the person's tasks) in order to retain the person?

**Incidence question (intensive margin).** Firms were asked to imagine 10 candidates with identical qualifications but differing salary expectations and outside offers, and report the expected spread between the highest and lowest initial and final offers (p. 330).

**Validation.** The paper conducts three validation exercises:
1. *Stability across respondents within the same firm* (following Bloom and Van Reenen (2007)): independent responses from 37 multi-respondent firms show significant overlap, confirming firm-level determination (Online Appendix Table A2).
2. *External validity with published data*: answers on observable firm practices (e.g., CBA coverage) align with publicly available sources (Online Appendix C.3).
3. *Correlation with worker survey*: elicited firm strategies are positively and significantly correlated with worker reports at those firms (Online Appendix Table A5).

The survey-based approach builds on Hall and Krueger (2012), who provided early evidence on the incidence of wage posting vs. bargaining using worker surveys. This paper adds the firm side and links both to administrative records. The results on back-and-forth negotiation dynamics are analogous to Backus et al. (2020a), who documented sequential bargaining in eBay product markets; this paper provides the labor-market counterpart.

**Worker bargaining outcomes regression (p. 350, equation 1).**

$$
y_i = \beta X_i + \delta \text{age}_i + \alpha \exp_i + \gamma \exp_i^2 + \zeta_{\text{educ}(i)} + \lambda_{o(i),\text{est}(i)} + \epsilon_i \tag{1}
$$

where $$y_i$$ is a bargaining outcome (probability of asking for a raise, successfully negotiating, etc.); $$X_i$$ is the heterogeneity dimension of interest (outside options, risk tolerance, gender, or AKM person effect); $$\lambda_{o(i),\text{est}(i)}$$ are three-digit occupation-establishment fixed effects. Standard errors are clustered at the firm level.

**Gender pay gap regression (p. 358, equation 2).**

$$
\log w_i = \beta \, \text{Female}_i + \delta \, \text{age}_i + \alpha \exp_i + \gamma \exp_i^2 + \zeta_{\text{educ}(i)} + \lambda_{o(i),\text{est}(i)} + \epsilon_i \tag{2}
$$

where $$\log w_i$$ is log daily pay. Estimated separately for workers exposed to individual bargaining and those whose wages are set by posting (based on the firm's reported strategy for the worker's group). Standard errors are clustered at the firm level.

**Prior-firm pay persistence regression (p. 362).**

$$
\log w_i = \beta \, \psi_{i,j^{\text{prev}}(i)} + \delta \, \text{age}_i + \alpha \exp_i + \gamma \exp_i^2 + \zeta_{\text{educ}(i)} + \lambda_{o(i),\text{est}(i)} + \epsilon_i \tag{3}
$$

where $$\psi_{i,j^{\text{prev}}(i)}$$ is the AKM wage premium of individual $$i$$'s previous employer (from population regressions using log daily pay 2010-2017; Bellmann et al. (2020)). Estimated separately by bargaining exposure.

## Empirical specifications

**Table II (Section IV.B): predicting firm bargaining strategies.** Equality-of-means tests comparing posting vs. bargaining firms on financial characteristics (total assets per employee, fixed assets per employee), firm size, firm age, other characteristics (CBA coverage, East Germany HQ, stock corporation). Tests run separately for three employee groups and two bargaining measures. Key finding: productivity proxies have p-values uniformly above 0.10; CBA coverage and East Germany HQ are significantly correlated with bargaining.

**Table III (Section IV.B): variance decomposition of bargaining strategies.** OLS of the firm-group indicator $$b_{ig}$$ (firm $$i$$ bargains with group $$g$$) on sets of covariates. The four employee-group dummies (column 1) explain R² = 0.33, comparable to all firm fixed effects (column 2, R² = 0.40, adjusted R² = 0.19). Columns (4)-(7) add firm size, productivity, norms, or 1-digit industry dummies; adjusted R² stays in the 0.34-0.35 range, barely above the group-only 0.33. Columns (8)-(9) add 4-digit industry dummies, reaching adj. R² = 0.44 (Panel A). This is shown for new hires (Panel A), incumbents (Panel B), and on the intensive margin (Panel C).

**Tables V-VI (Sections V, VI): bargaining behavior and the gender gap.** Main regression is equation (1) for bargaining outcomes and equation (2) for log wages, with three-digit occupation-establishment fixed effects and firm-level clustering. For the gender gap, Table VI runs columns (1)/(4) without FE, columns (2)/(5) with occ-est FE, and columns (3)/(6) with level-occ-est FE, separately for posting and bargaining firms. The main estimate is columns (2) vs. (5): the gender gap narrows to near zero at posting firms but remains at -5 pp at bargaining firms (Panel A). Figure V shows robustness across pay measures (daily pay, daily base pay, hourly wages), samples (surveyed workers, all workers), and FE specifications.

**Table VII (Section VI.B): prior-firm pay persistence.** Regression (3) estimated for all workers at surveyed firms (columns 1-2) and surveyed workers only (columns 3-4), separately by bargaining exposure. Panel A uses current daily pay; Panel B uses starting daily pay at current firm. The p-value of equality of prior-firm effect across bargaining and posting regimes is 0.016 (Panel A, all workers).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| IAB Integrated Employment Biographies (IEB) | Administrative employer-employee records; daily pay, demographics, occupation codes, employer IDs; linked to firm survey (553/772 firms consented); 416,821 full-time employees at matched firms in 2020; AKM firm and worker effects from 2010-2017 population data | [IEB Germany](/wiki/confidential/ieb-germany/) (confidential) |
| ifo HR Survey Panel (firm survey) | Novel survey of 772 German private-sector firms on wage-bargaining strategies; elicited for four employee groups and two bargaining contexts (new hires, incumbents with outside offers); fielded 2021-2022 | No page yet (new data, introduced by this paper) |
| IAB Worker Survey (HOPP, worker survey) | Novel survey of 9,756 full-time German workers on bargaining behavior, outside options, and risk tolerance; linked to IEB; fielded 2022-2024 to a sample drawn from Social Security records | No page yet (new data, introduced by this paper) |
| Orbis (Bureau van Dijk) | Balance sheet characteristics for surveyed firms (firm age, total assets, fixed assets, stock corporation status); matched to 99% of surveyed firms | [Orbis BvD](/wiki/commercial/orbis-bvd/) (licensed) |

Sample: 772 firms, workers ages 25-50 employed in 2020, IEB data from 1975 onward (main analysis uses 2010-2020), AKM effects estimated on 2010-2017 population.

## When to read the full paper

Read the [original](https://doi.org/10.1093/qje/qjaf049) if you are:
- Building or calibrating a model of individual wage bargaining in the labor market (the survey statistics on firm willingness to differentiate pay and the share of workers exposed to bargaining are key inputs);
- Studying the sources of the gender wage gap (the paper provides unusually direct evidence that bargaining, not productivity, drives residual gaps);
- Using AKM firm effects to interpret wage dispersion (the result that prior-firm AKM effects carry over to current pay only under bargaining has implications for how firm effects should be interpreted);
- Interested in survey-based measurement of firm practices (the validation exercises in the Online Appendix provide a detailed template).

The locators above point to the exact tables and figures in the source.

## Attribution and rights

Source: peer-reviewed, *The Quarterly Journal of Economics* (2026), pp. 315-371. Published by Oxford University Press on behalf of Harvard University. All rights reserved; no CC licence. This distillation was extracted by an LLM on 2026-06-28 and is **not human-verified or independently reproduced**.

> Caldwell, Sydnee, Ingrid Haegele, and Jörg Heining.
> "Bargaining and Inequality in the Labor Market."
> *The Quarterly Journal of Economics* (2026): 315-371.
> DOI: [10.1093/qje/qjaf049](https://doi.org/10.1093/qje/qjaf049).
> © The Author(s) 2025. Published by Oxford University Press on behalf of President and Fellows of Harvard College. All rights reserved.
> Extracted under fair-use / extract-only policy; no verbatim reproduction of substantial portions.


==============================================================================
# Diversifying Society's Leaders: Chetty, Deming & Friedman (2026)
# https://instituteforautomatedresearch.org/wiki/papers/qje/2026/chetty-diversifying-society-leaders-determinants-2026/
# Distilled: Using anonymized admissions data linked to federal tax records, Chetty, Deming, and Friedman show that top-0.1% income families are 2.5x more likely than middle-class applicants to gain admission to Ivy-Plus colleges with identical test scores, driven by legacy preferences (46%), nonacademic credentials (31%), and athletic recruitment (24%), none of which predict postcollege success. Attending an Ivy-Plus college instead of an average flagship public college causally increases the probability of reaching the top 1% of earnings by 5 pp and triples chances of working at an elite firm. Quarterly Journal of Economics 141(1), 2026, paywalled. Eight core results with source locators, the statistical model, and both research designs. LLM-distilled; not human-verified.
# Tags: paper-summary, higher-education, inequality, social-mobility, income-mobility
==============================================================================

**What this is.** A distilled skeleton of Chetty, Deming, and Friedman (2026). Read the [original article](https://doi.org/10.1093/qje/qjaf050) to replicate or extend. Equations, tables, and figures referenced below are from that source. This page is LLM-extracted and has not been human-verified.

## TL;DR

The paper uses a newly linked panel dataset combining federal income tax records, college attendance records, SAT/ACT scores, and internal applications data from Ivy-Plus and flagship public colleges to study two questions: (i) why children from top-income families disproportionately attend Ivy-Plus colleges (Harvard, Yale, Princeton, and the other eight Ivy League colleges, Chicago, Duke, MIT, and Stanford), and (ii) whether attending those colleges causally improves students' postcollege outcomes. The analysis proceeds in four parts: characterizing the pipeline from application through matriculation, identifying the mechanisms driving the high-income admissions advantage, estimating causal effects using two quasi-experimental designs, and predicting the effects of counterfactual admissions policies on socioeconomic diversity.

The headline findings are that (i) Ivy-Plus attendance does causally improve upper-tail outcomes by substantial magnitudes relative to attending a flagship public college, and (ii) the credentials that give high-income applicants their admissions advantage (legacy status, nonacademic ratings, athletic recruitment) do not predict better postcollege outcomes once college quality is held fixed, while academic credentials do. Contrary to the well-known findings of Dale and Krueger (2002) on mean log earnings, large causal effects of Ivy-Plus attendance on upper-tail income and nonmonetary outcomes emerge once richer data allow college value-added to be measured directly rather than through test-score proxies. The paper reconciles with Dale and Krueger (2002) and Dale and Krueger (2014): both papers agree on effects on mean log earnings; the divergence arises entirely on upper-tail outcomes where Ivy-Plus colleges have disproportionate effects.

## Core results

| # | Result | Locator | Magnitude as reported |
|---|--------|---------|----------------------|
| R1 | Top 0.1% income families are 2.5x more likely to gain admission to Ivy-Plus than middle-class applicants (70th-80th pctile) with the same test scores; 99th-99.9th pctile are 44% more likely; flagship public admissions rates are uncorrelated with parental income conditional on test scores | Figure III Panel B, p. 80 | Relative admission rate: 2.5x for top 0.1%; 1.44x for 99th-99.9th pctile; roughly 1.0x at flagship publics |
| R2 | 68% of the income gap in Ivy-Plus attendance conditional on test scores arises from admissions rather than applications or matriculation; decomposed into legacy preferences (31%), nonacademic credentials (21%), and athletic recruitment (16%), accounting for 114 of 168 extra top-1% students | Table II, pp. 77, 82; Figures V, VI | 52 extra top-1% students from legacy preferences; 35 from nonacademic credentials; 27 from athletic recruitment; 141 of 168 from admissions-related factors combined including athletes |
| R3 | Attending an Ivy-Plus college instead of the average flagship public college causally increases the predicted probability of reaching the top 1% of income at age 33 by 5 pp (+42%) | Table IV col. 1 and 6, p. 114 | TOT = 5.01 pp (SE 1.31), $$p < .001$$; from 11.8% to 16.8% |
| R4 | Ivy-Plus attendance nearly doubles the probability of attending an elite graduate school | Table IV Panel B, p. 114 | TOT = 5.64 pp (SE 2.79); from 6.1% to 11.7%, +92% |
| R5 | Ivy-Plus attendance more than triples the probability of working at an elite firm at age 25 | Table IV Panel B, p. 114 | TOT = 16.96 pp (SE 4.01); from 8.5% to 25.5%, +199% |
| R6 | Ivy-Plus attendance nearly quadruples the probability of working at a prestigious firm at age 25 | Table IV Panel B, p. 114 | TOT = 17.51 pp (SE 4.26); from 7.2% to 24.7%, +245% |
| R7 | The credentials underlying the high-income admissions advantage (legacy status, nonacademic ratings, athletic recruitment) have zero or negative association with postcollege success after adjusting for college quality; high academic ratings have a +4.8 pp effect on top-1% probability | Figure XV Panel B, p. 129 | Legacy: negative (negatively associated, p. 129); nonacademic rating: approx. 0 (no significant association); athlete: approx. 0 (no significant association); high academic rating: +4.8 pp on top-1% probability |
| R8 | Eliminating all three high-income admissions advantages (legacy preferences, nonacademic-credentials boost, athletic recruitment income gradient) would increase the share of Ivy-Plus students from the bottom 95% of parental income by 8.8 pp, with no reduction in average student outcomes | Table V Panel A rows 1-4, p. 133 | Top-1% parental income share falls from 15.8% to 9.9%; bottom-60% share rises from 15.7% to 20.0%; average predicted outcomes unchanged or improved |

**Overall (paper's conclusion).** Ivy-Plus colleges have large causal effects on students' chances of achieving upper-tail earnings and nonmonetary leadership outcomes, but they also substantially over-admit students from high-income families relative to what academic credentials alone would predict. The three factors driving this admissions advantage (legacies, nonacademic credentials, athletes) are uncorrelated with, or negatively predictive of, postcollege success, meaning that admissions policy changes that eliminate these advantages would increase socioeconomic diversity by an amount comparable to race-based affirmative action without reducing student-body quality. Because Ivy-Plus colleges account for a relatively small share of all Americans, changes in admissions policy have small effects on the share of top-1% earners from low-income families but could meaningfully diversify the socioeconomic backgrounds of people in nonmonetary leadership positions (senators, Supreme Court justices, Nobel laureates).

## Theory / model

The paper presents a formal statistical model (Section IV.A, pp. 92-97) to clarify what each research design identifies.

**Admissions ratings.** College $$j$$ assigns applicant $$i$$ a composite rating

$$Z_{ij} = \gamma_{1j} X_{1i} + \gamma_{2j} X_{2i} + \eta_i + \epsilon_{ij}, \tag{3}$$

where $$X_{1i}$$ is observable (e.g., SAT/ACT score), $$X_{2i}$$ is unobservable but correlated with long-term outcomes $$Y_i$$ (e.g., intrinsic ability or motivation), $$\eta_i$$ is a common idiosyncratic component uncorrelated with $$Y_i$$ (e.g., a strong guidance counselor letter that helps at all colleges), and $$\epsilon_{ij}$$ is pure noise at college $$j$$ uncorrelated with $$Y_i$$ across all colleges (e.g., whether the student plays an instrument needed for the college's orchestra in the application year). Colleges admit student $$i$$ to college $$j$$ if $$Z_{ij} > C_j$$, where $$C_j$$ is a college-specific cutoff. Colleges are assumed to decide independently.

**Postcollege outcomes.** The student's outcome (e.g., earnings or one of the leadership proxies in Figure I) follows

$$Y_i = \sum_{j \in J_i} D_{ij} \phi_j + \beta_1 X_{1i} + \beta_2 X_{2i} + \epsilon_i^Y, \tag{4}$$

where $$D_{ij}$$ is an enrollment indicator, $$\phi_j$$ is college $$j$$'s causal value added (normalized to zero for the average flagship public, the outside option $$O$$), and $$\epsilon_i^Y$$ is an outcome error orthogonal to $$\eta_i$$ and $$\epsilon_{ij}$$. The goal is to estimate $$\phi_A$$, the causal effect of attending an Ivy-Plus college $$A$$ instead of college $$O$$.

**Identification.** OLS on admitted students is biased because $$X_{2i}$$ affects both admission and outcomes. The paper offers two designs to remove this bias, both exploiting data on admissions decisions at multiple colleges.

**Research Design 1 (idiosyncratic-admissions IV, Section IV.A.2, p. 93).** Among students on the waitlist at college $$A$$, the paper uses being admitted off the waitlist as a quasi-instrument. The rescaled waitlist estimator is

$$r_A = \frac{E[Y_i | P_{iA}=1, X_{1i}, \tilde{X}_{2i}] - E[Y_i | P_{iA}=0, X_{1i}, \tilde{X}_{2i}]}{E[D_{iA} | P_{iA}=1, X_{1i}, \tilde{X}_{2i}]}, \tag{5}$$

where $$\tilde{X}_{2i}$$ is a proxy for $$X_{2i}$$ (e.g., whether the student was placed on the waitlist, itself a signal of near-marginal quality). Under the correlated-admissions-criteria assumption (Assumption 1, p. 95) that $$\gamma_{2A} > 0 \Rightarrow \gamma_{2B} > 0$$ for colleges $$B$$ with similar holistic admissions processes, the estimator identifies $$\phi_A$$ if and only if the test statistic $$T_{B|A} = 0$$, where $$T_{B|A}$$ measures whether being admitted vs. rejected from college $$A$$'s waitlist predicts admission at college $$B$$. This multiple-rater test (Figure VII, pp. 102-104) passes empirically: waitlisted students' admission outcomes at other Ivy-Plus colleges are statistically indistinguishable from each other, regardless of whether they are admitted from the waitlist at the reference college.

**Research Design 2 (matriculation design, Section IV.A.3, p. 96).** Following Mountjoy and Hickman (2021) and Dale and Krueger (2002), the paper compares outcomes for students admitted to the same portfolio of colleges $$J_i = \{A, O\}$$ who choose to attend different colleges:

$$r_M = E[Y_i | D_{iA}=1, X_{1i}, J_i = \{A, O\}] - E[Y_i | D_{iO}=1, X_{1i}, J_i = \{A, O\}], \tag{6}$$

under Assumption 2 that conditional on the admissions portfolio and $$X_{1i}$$, the unobservable $$X_{2i}$$ is orthogonal to the matriculation choice (p. 96). Both designs yield consistent estimates (Table IV, p. 114), which strengthens the credibility of both.

## Method

**Surrogate index for early-career outcomes (Section II.C.4, p. 67; Section IV.B, pp. 97-100).** Because income ranks at age 33 are not observed for recent cohorts, the paper constructs a surrogate index (Athey et al., forthcoming) using employers and graduate schools at ages 22-25 to predict the probability of reaching the top 1% of income at age 33. This is motivated by the finding that firms' employment composition at ages 22-25 strongly predicts income at 33 (Figure IX and Online Appendix Figure A.23a, pp. 107-108). The paper verifies that early-career employers and graduate schools capture the income dynamics that produce age-33 outcomes, with a near-zero treatment effect at age 25 that grows steadily to ~5 pp by age 33 (Figure IX Panel A, p. 107).

**Treatment-effect heterogeneity by outside options (Section IV.C.5, pp. 110-113).** To identify $$\phi_{Ivy}$$ (the causal effect relative to the average flagship public as outside option), the paper exploits variation in the value added of each applicant's outside option. Applicants are grouped by home state, parental income, and race; the outside-option quality is measured as the average observational value added of colleges attended by rejected non-waitlisted applicants in each group. The paper then estimates:

$$\hat{\phi}_{Ivy} \approx 5.01 \text{ pp (SE 1.31), rescaling the reduced-form waitlist estimate by the ratio of observational VA differences}$$

(Table IV col. 1 and Figure X Panel A, pp. 112-113). The slope of the heterogeneity relationship equals -0.79 (Figure X Panel A), indicating that most variation in outcomes between colleges is driven by genuine causal effects rather than selection, with students facing weaker outside options gaining most from Ivy-Plus attendance.

**Multiple-rater admissions test (Section IV.C.1, pp. 100-104; Figure VII, p. 102).** The paper develops a new validation test for the quasi-random variation assumption in Research Design 1. The test compares admission rates at lower-ranked Ivy-Plus colleges $$B$$ (ranked by revealed student preferences) for students who are admitted vs. rejected from the waitlist at college $$A$$. Under the correlated-admissions-criteria assumption, $$T_{B|A} = 0$$ if and only if the residual variation in $$A$$'s admissions decisions among waitlisted students is orthogonal to $$X_{2i}$$. The paper implements this test for three specifications (no controls, with controls, dropping legacies/athletes/top-1%) and finds $$T_{B|A}$$ statistically indistinguishable from zero across all three (Figure VII, p. 102), supporting the identification assumption.

## Empirical specifications

**Pipeline analysis: counterfactual attendance rate under income-neutral admissions (Section III.A.1, p. 75, Equation 1).** To quantify how many extra top-1% students are in the Ivy-Plus class conditional on test scores, the paper computes:

$$\text{Counterfactual Attendance Rate}_c = \sum_a N_{Top 1\%, a} \times \text{Attendance Rate}_{P70-80, ac}, \tag{1}$$

where $$N_{Top 1\%, a}$$ is the number of test takers with score $$a$$ from families in the top 1% and $$\text{Attendance Rate}_{P70-80, ac}$$ is the fraction attending college $$c$$ among students with score $$a$$ from the 70th-80th percentile. Scaling to a class of 1,650 students, this counterfactual implies 93 students from the top 1% rather than the observed 261, a gap of 168 "extra" top-1% students (10.2% of enrollment).

**Pipeline decomposition (Section III.B.4, p. 81, Equation 2).** For non-athletes, the paper decomposes the 168-student gap by sequentially equalizing applications, admissions, and matriculation rates across income groups:

$$\text{Equal Admit CF}_c = \sum_a N_{Top 1\%, a} \times \text{Application Rate}_{Top 1\%, ac} \times \text{Admission Rate}_{P70-80, ac} \times \text{Matriculation Rate}_{Top 1\%, ac}. \tag{2}$$

Setting application, admission, and matriculation rates to those of the middle class, and averaging across orderings, admissions account for 58% (96 students) of the 168-student gap (Online Appendix Table A.6, p. 82). Including athletes, 114 of 168 extra students (68%) are from admissions-related factors.

**Causal-effect regression (Section IV.C.3, pp. 105-106).** The paper estimates the treatment-on-the-treated (TOT) effect from Research Design 1 by regressing an outcome indicator on a waitlist-admission indicator $$P_{iA}$$ plus college-by-cohort fixed effects, clustering standard errors by student (to account for students on multiple waitlists), and dividing the reduced-form coefficient by the first-stage effect (the probability of attending college $$A$$ conditional on admission). In the primary specification (Table IV col. 1, p. 114), the outcome is the predicted top-1% income probability based on age-22-25 employers. Controls include a quintic in SAT/ACT scores, parent income bin dummies, race/ethnicity indicators, gender, home-state indicators, recruited-athlete and legacy status indicators, and college-by-cohort fixed effects. Robustness to dropping legacies, athletes, and top-1% applicants (the third bar set in Figures VIII and XI) confirms the estimates are not driven by the same characteristics that generate the admissions imbalance.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---------|--------------|-----------|
| Federal income tax records (IRS, 1996-2021) | Parental income (1040, W-2), children's individual income (W-2, Form SE, 1040), employer identification (W-2) | [no page yet](/wiki/confidential/) |
| 1098-T college attendance forms (Dept of Education via NSLDS) | College attendance indicator for all US colleges; linked to tax records at individual level | [no page yet](/wiki/confidential/) |
| SAT scores (College Board, 2001-2005 and odd years 2007-2015) | Standardized academic qualification measure; composite score (math + critical reading) | [no page yet](/wiki/confidential/) |
| ACT scores (ACT, 2001-2015) | Standardized test scores converted to SAT equivalents via ACT 2016 concordance tables | [no page yet](/wiki/confidential/) |
| Pell Grant records (NSLDS, 1999-2013) | Low-income student identification; supplementary college attendance signal | [no page yet](/wiki/confidential/) |
| Application and admissions records (Ivy-Plus and flagship public colleges, 1998-2015) | Admission indicators, legacy/athlete/faculty-child flags, admissions-office ratings (academic and nonacademic), application round, GPA; from several Ivy-Plus colleges plus 9 flagship public systems | no page yet |

**Sample note.** The pipeline analysis sample covers 5,063,263 students on pace to graduate high school in 2011, 2013, or 2015 (Table I col. 1, p. 70). The college-specific analysis sample covers 486,150 Ivy-Plus applicants and 1,877,770 flagship public applicants for whom internal admissions records are available (Table I cols. 3-4, p. 72). All data were linked at the individual level using Social Security numbers and stripped of personally identifiable information before analysis; the IRS component was accessed under IRS contract TIRNO-16-E-00013. The dataset construction and income variable definitions build on the earlier linked administrative dataset of Chetty et al. (2020) on income segregation across US colleges; the current paper extends that work with internal admissions records and later cohorts. Application-rate differences as a driver of access gaps were documented by Hoxby and Avery (2013) using geographic imputations of family income; this paper finds admissions rates are the primary driver at Ivy-Plus colleges in the more recent period studied, after private colleges expanded low-income recruitment programs.

## When to read the full paper

Read Chetty, Deming, and Friedman (2026) if you need:

- The exact statistical model and identification assumptions (Equations 3-6, Online Appendix H with formal proofs), especially the multiple-rater test logic and how the two designs nest into a unified framework (Section IV.A, pp. 92-97).
- Heterogeneity in treatment effects by parental income, race, and outside-option quality (Figure XII Panel D, p. 120; Table IV, p. 114; Online Appendix Table A.10).
- The admissions-ratings analysis showing how legacy preferences and nonacademic credentials mechanically generate the high-income admissions advantage and how it is measured via counterfactual admissions predictions (Figures V and VI, pp. 85, 89; Section III.C, pp. 84-91).
- The counterfactual admissions simulations and their predicted effects on leadership outcomes across specific categories (senators, Supreme Court justices, Nobel laureates) under alternative admissions policies (Table V, pp. 133-134; Section VI, pp. 131-141).
- The quantile treatment effects analysis showing why Ivy-Plus effects are concentrated at the very top of the income distribution rather than distributed proportionally (Figure XIII, p. 124; Section IV.E, pp. 122-125).
- The college-level data on parental income distributions at each stage of the application process, publicly released at www.opportunityinsights.org/data (Online Appendix O, p. 142).

## Attribution and rights

Chetty, Raj, David J. Deming, and John N. Friedman. "Diversifying Society's Leaders? The Determinants and Causal Effects of Admission to Highly Selective Private Colleges." *The Quarterly Journal of Economics* 141(1), 2026, 51-145. DOI: [10.1093/qje/qjaf050](https://doi.org/10.1093/qje/qjaf050).

Copyright (c) The Author(s) 2025. Published by Oxford University Press on behalf of President and Fellows of Harvard College. All rights reserved. Replication code and data available at Harvard Dataverse: [https://doi.org/10.7910/DVN/YMVK4K](https://doi.org/10.7910/DVN/YMVK4K).

This page is a machine-generated distillation (LLM-extracted). It has not been human-verified and reproduces no paywalled content, only brief quotations of results as permitted for commentary and research purposes (extract-only).


==============================================================================
# Failing Banks: Correia, Luck & Verner (2026)
# https://instituteforautomatedresearch.org/wiki/papers/qje/2026/correia-failing-banks-2026/
# Distilled: Using a new panel covering 37,000 US commercial banks from 1863 to 2024, Correia, Luck, and Verner show that bank failures across the full history of the US banking system are strongly predicted by deteriorating fundamentals, that failures with bank runs are as predictable as other failures, and that most pre-FDIC failures involved fundamentally insolvent banks. The Quarterly Journal of Economics 2026, public domain (US Government employee work). Nine core results with source locators, datasets, the insolvency condition, and the failure prediction specification.
# Tags: paper-summary, banking, bank-failures, bank-runs, financial-crises, economic-history
==============================================================================

**What this is.** This is a machine-extracted skeleton of the paper. Read the
[original at QJE](https://doi.org/10.1093/qje/qjaf044) (or the [arXiv preprint](https://arxiv.org/pdf/2506.06082))
to replicate or extend the results.

## TL;DR

Correia, Luck, and Verner build a new panel of virtually all US commercial banks from 1863
to 2024 and study the history of bank failures. The central finding is that bank failures are
almost always preceded by deteriorating fundamentals: rising asset losses, declining solvency,
and increasing reliance on expensive noncore funding. These patterns make failures highly
predictable from public accounting data. Failures that involve large deposit outflows (bank
runs) are just as predictable as other failures, contradicting the view that non-fundamental
panic runs commonly topple otherwise healthy banks. Low recovery rates on failed banks' assets
further suggest that most pre-FDIC banks that ran were already fundamentally insolvent. The
aggregate failure rate during systemic banking crises is also largely forecast by deteriorating
micro-level fundamentals, with an out-of-sample R-squared of 40% for the full sample and 81%
for the modern era. The result extends, across 160 years of data, the cross-bank evidence of
Calomiris and Mason (2003) on the Great Depression and reinforces the aggregate-data findings
of Gorton (1988) and Baron, Verner, and Xiong (2021) that banking crises follow bad
macroeconomic news and declining bank equity rather than purely self-fulfilling panics.

## Core results

| # | Result | Locator | Magnitude as reported |
|---|--------|---------|----------------------|
| R1 | AUC for predicting bank failure within 1 year, historical pre-FDIC sample (1863-1934), full specification | Table I, Panel A, col. 4, p. 172 | In-sample AUC = 0.864; OOS AUC = 0.851 |
| R2 | AUC for predicting bank failure within 1 year, modern sample (1959-2024), full specification | Table I, Panel B, col. 4, p. 172 | In-sample AUC = 0.953; OOS AUC = 0.945 |
| R3 | 3-year failure probability at top-5th-percentile insolvency and noncore funding (both samples) | Figure IV, p. 168 | 27% (historical and modern); unconditional = 2.5% (historical), 1% (modern); 10-27x higher |
| R4 | Average deposit growth immediately before failure | Table II, Panel A, p. 178 | -14% pre-FDIC (1880-1934); -2.5% post-FDIC (1993-2024); 25% of pre-FDIC failures had outflows exceeding 20% |
| R5 | AUC for failures with large deposit outflows (runs), historical | Table I, Panel A, col. 5, p. 172; text p. 180 | In-sample AUC = 0.855; OOS AUC = 0.839; same as all-failures AUC (0.864) |
| R6 | Out-of-sample R-squared, aggregate bank failure rate on predicted aggregate failure rate | Table III, col. 1 (full) and col. 3 (modern), p. 184 | R² = 0.40 full sample (1874-2024); R² = 0.81 modern era (1970-2024); slope coefficient ≈ 1.0 in modern era |
| R7 | Average asset recovery rate in bank receiverships, pre-FDIC sample (1863-1934) | Table IV, p. 186 | Average R = 0.52; 43% of failures have R < 0.50; OCC assessed 47% of assets as doubtful and 18% as worthless (Table V, p. 187) |
| R8 | Share of failed banks that were fundamentally insolvent, conditional on rho and v | Table VII, p. 193 | 0.81 (rho = 0, v = 0 baseline); 0.60 (rho = 0.1, v = 0.05); under extreme assumptions (rho = v = 0.2), still 0.31 |
| R9 | OCC-classified cause of failure attributed to bank runs (1863-1937) | Figure IX, p. 195-196 | <2% of failures; economic conditions most common (>30%); losses second (~25%) |

**Overall (paper's conclusion, p. 196-198).** Bank failures are almost always and everywhere
a deterioration of bank fundamentals. Runs are a frequent mechanical trigger but typically
close insolvent banks rather than causing solvent banks to fail. The predictability of failures,
including failures with runs, suggests that non-fundamental, self-fulfilling runs on healthy
banks are a rare cause of US bank failures both before and after deposit insurance.

## Theory / model

The paper organizes its empirical analysis around two competing theoretical frameworks for
why banks fail (Section II.A, pp. 154-156).

**The solvency view.** Banks fail when realized credit losses, interest rate losses, or fraud
erode asset values below debt claims, making the bank insolvent regardless of whether a run
occurs. Morris and Shin (2016) formalize solvency risk as the probability of failure in a
counterfactual with no withdrawals. Under this view, the runnable nature of bank liabilities
is not the root cause.

**The bank runs view (Diamond and Dybvig (1983)).** Banks finance illiquid assets with
demandable deposits. A coordination failure among depositors can produce a self-fulfilling
panic run on an otherwise solvent bank, forcing it to liquidate assets at a loss.

**Fundamental-based panic runs (Goldstein and Pauzner (2005)).** Bank fundamentals
$$\theta$$ are stochastic. Three regions determine equilibrium behavior (p. 155):

- $$\theta > \bar{\theta}$$: fundamentals are strong, no depositor has incentive to withdraw.
- $$\theta \leq \underline{\theta}$$: the bank is insolvent; all depositors withdraw regardless of others' actions (a "fundamental run").
- $$\underline{\theta} < \theta \leq \theta^*$$: a panic region in which a coordinated run can cause failure even though absent the run the bank could pay all creditors.

This model predicts that failures with runs occur randomly within the panic region and should
therefore be harder to predict from fundamentals than failures outside it. The paper tests
this implication and finds it is rejected.

**The insolvency condition (Section VIII.B, p. 190).** The paper develops a simple
framework to gauge what fraction of failed banks were fundamentally insolvent absent a run.
A bank with book assets $$A$$, debt $$D$$, unrealized asset losses $$\lambda$$ before failure,
and additional receivership losses $$\rho$$ has observed recovery rate $$R = (1-\lambda)(1-\rho)$$.
Let $$v$$ be the franchise value as a share of current book assets. The bank is fundamentally
insolvent irrespective of any run if:

$$
(1-\lambda)(1+v)A < D
$$

Rewriting in terms of leverage $$\ell = D/A$$ and the observed recovery rate:

$$
\frac{1+v}{1-\rho} < \frac{\ell}{R} \tag{3}
$$

A bank satisfying this condition was insolvent even if the run had not occurred. When $$\rho$$
and $$v$$ are low (baseline: both zero), the condition simplifies to $$R < \ell$$, which holds
for 81% of pre-FDIC bank failures in the sample.

## Method

**Measures of bank fundamentals (Section IV.A, pp. 161-162).** Three observable proxies
for financial health are constructed:

- *Insolvency risk.* Pre-1934: surplus profit relative to total equity (surplus profit = sum
  of surplus fund and undivided profits; proxies profitability and capitalization). Post-1959:
  net income / total assets.
- *Noncore funding.* Pre-1934: total assets net of total deposits, equity, and national bank
  notes, all scaled by assets (captures expensive nondeposit wholesale funding). Post-1959:
  (time deposits + wholesale funding) / total assets.
- *Asset growth.* Change in log real bank assets, used in quintile buckets to capture the
  nonlinear boom-bust relation.

**Failure prediction model (Section V, p. 170, eq. 2).** Bank failures are predicted via
linear probability models (with logit as robustness):

$$
\text{Failure}_{b,t+1\to t+h} = \alpha + \beta_1\,\text{Insolvency}_{bt}
+ \beta_2\,\text{Noncore Funding}_{bt}
+ \beta_3\,\text{Insolvency}_{bt}\times\text{Noncore Funding}_{bt}
$$

$$
+ \beta_4\,\text{Asset Growth}_{bt}
+ \beta_5\,\text{Aggregate Conditions}_t
+ \epsilon_{b,t+1\to t+h} \tag{2}
$$

where $$\text{Failure}_{b,t+1\to t+h}$$ is an indicator equal to one if bank $$b$$ fails
within $$h$$ years of call report date $$t$$. Only real-time observables enter; no bank or
time fixed effects are included. Predictive performance is evaluated via the area under the
receiver operating characteristic curve (AUC), computed both in-sample and pseudo-out-of-sample
using an expanding training window (first 10 years of data as initial training sample).

**Aggregate failure rate prediction (Section VII, p. 184).** The bank-level predicted
probabilities are aggregated into a predicted aggregate failure rate:

$$
\bar{p}_{t|t-1} = \sum_{b \in B_{t-1}} w_{bt-1}\,\hat{p}_{b,t|t-1}
$$

and regressed on the realized failure rate:

$$
\text{FailureRate}_t = \alpha + \beta\,\bar{p}_{t|t-1} + u_t
$$

using Newey-West standard errors (truncation parameter $$S = 1.3T^{1/2}$$).

**Event-study dynamics (Section IV.B, p. 163, eq. 1).** To characterize how fundamentals
evolve in failing banks, the paper estimates:

$$
y_{b,t} = \alpha_b + \sum_{j=-9}^{0} \beta_j \times \mathbf{1}[\text{YearsToFail}_{b,t} = j] + \epsilon_{b,t} \tag{1}
$$

restricted to failing banks within 10 years of failure; the omitted period is $$j = -10$$.
Coefficients $$\{\beta_j\}$$ trace the pre-failure dynamics of solvency, funding, and assets.

## Empirical specifications

**Predictability of bank failures (Section V.B, pp. 170-174).** Equation (2) is estimated
separately for the historical pre-FDIC sample (1863-1934) and the modern sample (1959-2024)
at one-, three-, and five-year horizons. Standard errors are not clustered (real-time
observable specification; no fixed effects). The in-sample AUC for the full specification
(insolvency, noncore funding, their interaction, asset growth quintiles, and aggregate
conditions) ranges from 0.864 to 0.739 across horizons in the historical sample and from
0.953 to 0.831 in the modern sample (Table I, p. 172). The pseudo-OOS performance is nearly
as strong, with OOS AUC = 0.851 (historical, 1-year) and 0.945 (modern, 1-year).

**Failures with bank runs (Section VI.B, pp. 178-181).** Failures with large deposit outflows
are defined as those where deposits decline by more than 7.5% between the last call report and
failure (data available for 1880-1934 historically; 1993-2024 for the modern sample). Equation
(2) is re-estimated restricting to this subsample. The AUC for failures with large deposit
outflows is 0.855 (in-sample, historical, col. 5 vs. 0.864 for all failures in col. 4, Table I),
confirming that fundamentals predict run-failures equally well. On average, banks in the
historical pre-FDIC sample saw deposits decline by 14% before failure (Table II, Panel A),
with 25% experiencing outflows exceeding 20%. Post-FDIC, average outflows are only 2.5%.

**Aggregate waves of bank failures (Section VII, pp. 182-185).** The predicted aggregate
failure rate $$\bar{p}_{t|t-1}$$ is constructed pseudo-out-of-sample using only data up to
year $$t-1$$. Regressing the actual aggregate failure rate on $$\bar{p}_{t|t-1}$$ yields
$$R^2 = 0.40$$ for the full sample and $$R^2 = 0.81$$ for the modern era 1970-2024 (Table III,
p. 184). The estimated coefficient $$\hat\beta$$ is close to one in the modern era, indicating
that predicted and actual failure rates move in proportion. The Great Depression years
(1929-1934) are underpredicted, consistent with excess failures beyond what deteriorating
micro-level fundamentals alone forecast.

**Recovery rates and fundamental insolvency (Section VIII, pp. 185-194).** Recovery rate
$$R$$ is defined as total cash collected by the OCC receiver divided by book assets at
suspension. Regressing realized $$R$$ on OCC asset-quality categories (good, doubtful,
worthless) yields coefficients of 0.89, 0.54, and 0.08 respectively, with $$R^2 = 0.936$$
(Table VI, col. 1, p. 189), validating the OCC assessments. The insolvency condition
(equation 3) is applied bank by bank using observed $$\ell$$ and $$R$$ for a grid of $$(\rho, v)$$
values. Under the baseline ($$\rho = v = 0$$), 81% of pre-FDIC failures satisfy the
insolvency condition (Table VII, p. 193). Even under the generous assumption that
receiverships destroy 10% of asset value and franchise value equals 5% of assets
($$\rho = 0.1$$, $$v = 0.05$$), 60% of failures were insolvent.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---------|--------------|-----------|
| OCC Annual Reports to Congress (1863-1941) | Historical bank balance sheets (assets, loans, deposits, equity), failure events, causes of failure, receiver postmortem reports (deposits and assets at suspension, funds collected); digitized via OCR using methods of Correia and Luck (2023) | no page yet |
| FFIEC Call Reports via Federal Reserve (1959-2024) | Modern quarterly bank balance sheets (FFIEC 031/041/051 from 1976; FFIEC 010/011 extended back to 1959); income statements; foundation of the modern-sample analysis | no page yet |
| FDIC Failure Transaction Database (1934-2024) | Failure dates, deposits and assets at resolution for post-FDIC failures (1993-2024 subset used for deposit outflows); defines bank failure as closure involving purchase-and-assumption or liquidating receivership | no page yet |
| National Information Center (NIC) tables | Bank charter and founding dates for de novo bank identification | no page yet |

**Sample scope.** Historical sample: 14,152 national banks, 1863-1941 (annual; national bank
notes era and early Fed era). Modern sample: 23,209 FDIC member banks, 1959-2024 (annual
data used for comparability). Combined: 37,361 unique bank entities; 5,120 bank failures
(2,887 pre-1935; 2,233 post-1959). Recovery rate analysis: 2,917 receiverships with complete
data, 1863-1934.

## When to read the full paper

- **Banking history and crises**: Section IV documents the fundamental dynamics in failing
  banks (Figures II-IV); Section VII shows these dynamics forecast systemic crises including
  the Great Depression and 2008 (Figure VII, Table III).
- **Bank runs and solvency**: Sections VI and VIII contain the main tests of whether runs
  caused failures (Table I col. 5; Table VII); the framework in Section VIII.B (equation 3)
  is the cleanest tool for the fundamental insolvency calculation.
- **Early warning / stress testing**: Table I provides benchmark AUC statistics (0.86-0.95)
  for bank failure prediction from simple accounting ratios; the full regression coefficients
  are in Online Appendix Tables B.4 and B.5.
- **Historical database users**: Section III and Online Appendix C describe the OCR digitization
  of OCC Annual Reports and the construction of the new 1863-1941 balance-sheet panel; replication
  data are on Harvard Dataverse (Correia, Luck, and Verner 2025a).

## Attribution and rights

This paper is in the public domain in the United States. The PDF footer (p. 204) states:
"Published by Oxford University Press on behalf of President and Fellows of Harvard College
2025. This work is written by (a) US Government employee(s) and is in the public domain in
the US."

LLM-distilled summary, not human-verified, not reproduced. For the canonical text see:

> Correia, Sergio, Stephan Luck, and Emil Verner. "Failing Banks."
> *The Quarterly Journal of Economics* 141(1), 2026, 147-204.
> https://doi.org/10.1093/qje/qjaf044

Replication data: Correia, Luck, and Verner (2025a), Harvard Dataverse,
https://doi.org/10.7910/DVN/Q22XR1.


==============================================================================
# The Price of Housing in the United States: Lyons, Shertzer, Gray & Agorastos (2026)
# https://instituteforautomatedresearch.org/wiki/papers/qje/2026/lyons-price-housing-united-states-2026/
# Distilled: Lyons, Shertzer, Gray, and Agorastos construct the first annual market rent and home sales price series for 30 U.S. cities over 1890-2006 from 2.7 million newspaper real estate listings. Real rents rose 60% rather than fell over the postwar period; real sales prices reached four times their 1890 level by 2006; and the average annual real return to housing was 9% (rental 7.7%, capital gain 1.3%). Q.J. Econ. 2026, paywalled. Seven core results with source locators, datasets used, the rolling-window hedonic method with its equations, and the user cost framework.
# Tags: paper-summary, macro, real-estate, housing-markets, measurement
==============================================================================

**What this is.** The core results, the user cost framework, and the rolling-window hedonic method with its defining equations: enough to understand what the paper found and how, without reading the full 45 pages. To replicate or extend, read the original at the [DOI](https://doi.org/10.1093/qje/qjaf047). The underlying data are in the [Harvard Dataverse](https://doi.org/10.7910/DVN/CYX1PQ).

## TL;DR

The paper constructs the first annual market rent index (RI) and housing price index (HPI) for 30 U.S. cities from 1890 to 2006, drawing on over 2.7 million archival newspaper real estate listings. The authors apply a rolling-window hedonic regression to remove unobserved quality trends and aggregate by population weight to a national series. The headline finding is that the standard sources misrepresent the history of U.S. housing prices: real market rents rose 60% nationally rather than fell (as the BLS Rent of Primary Residence series implies), and real sales prices grew 142% from 1890 to 1987 rather than the 37% implied by the Shiller (2015) national index. The paper also revises upward the U.S. capital gain estimate of Jordà et al. (2019) (who report 1.0% AGR for housing capital gains) and provides city-level evidence extending the national long-run series of Knoll, Schularick, and Steger (2017). The average annual real total return to housing over the full period was 9%, dominated by rental income at 7.7%; capital gains contributed just 1.3% but were near-zero before 1940 and rose to 1.7% per year thereafter, concentrated in cities with more restrictive land use regulation.

## Core results

Magnitudes and significance are as reported in the paper. Locators point to the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | HHP real market rents rose 60% nationally from 1890 to 2006, contrasting with the BLS Rent of Primary Residence (RoPR) series, which implies rents fell roughly 50% from 1914 to 2006 | Figure II, p.572; p.573 | HHP national RI: +60% relative to 1890; +36% relative to 1914; BLS RoPR series: approximately -0.6% per year from 1914 to 2006 |
| R2 | HHP real sales prices grew 142% from 1890 to 1987, nearly four times the 37% implied by the Shiller index, and reached approximately four times the 1890 level by 2006 | Figure III, p.575; p.576 | HHP: +142% real growth 1890-1987 (Shiller: +37%); HHP real HPI in 2006 ~4x the 1890 level; the two series diverge mainly in 1953-1987 |
| R3 | The rent-to-price ratio fell from approximately 8-9% in 1890-1920 to 3% in 2006, driven by the emergence of the modern long-term amortized mortgage and falling user costs | Figure VII Panel A, p.589; p.590 | Rent-to-price: ~8-9% in 1890-1920; 9.4% in 1935; 7.4% in 1970; ~3% in 2006; tracks declines in user cost component from LTV expansion and mortgage term extension |
| R4 | The average annual real total return to housing over 1890-2006 was 9.0%, with rental returns accounting for 7.7% and real capital gains just 1.3%; returns were negative in only 5 of 116 years | Figure VIII, p.595; pp.595-596 | Total return 9.0%; rental return 7.7% (SD 1.9%, CV 0.25); capital gain 1.3% (SD 5.0%, CV 3.8); capital gain close to zero before 1940 (+0.04% pa) |
| R5 | HHP capital gains averaged +0.04% pa before 1940 and +1.7% pa from 1940 to 2006, revising the Shiller long-run capital gain estimate of 0.7% upward by approximately 0.6 percentage points | pp.595-596 | HHP overall capital gain AGR 1.3% vs Shiller 0.7%; the revision is driven by the post-World War II decades where the Shiller data relied on appraisals and excluded non-conforming loans |
| R6 | Over 1890-2006, housing offered a real return of 9% vs 11.5% for equities, but with standard deviation 5.4% vs 17.6% (coefficient of variation 0.6 vs 1.5) | p.597; Figure IX, p.598 | Housing: mean 9.0%, SD 5.4%, CV 0.6; equities: mean 11.5%, SD 17.6%, CV 1.5; equity capital gain ~3.6% pa vs housing 1.3% pa; housing outperformed equities only in the early 1920s, early 1940s, and late 1970s to early 1980s |
| R7 | Cities with higher zoning restrictiveness had substantially higher average annual real capital gains, with the OLS coefficient on the Saiz (2010) Wharton zoning index stable across both 1940-1969 and 1970-2006 | Figure VI, p.586; p.585 | OLS slope on zoning index: 0.63 (1940-1969), 0.79 (1970-2006); the positive relationship between regulatory constraint and capital gains is statistically similar across the two periods |

**Overall (paper's conclusion).** The HHP series revise several standard facts about U.S. housing markets. Postwar real rents rose rather than fell: adjustments to the BLS series by Crone, Nakamura, and Voith (2010) and by Gordon and van Goethem (2007) move in the same direction as HHP but do not reach the same level. Real sales prices grew substantially more before 1987 than the Shiller index records, consistent with limitations in the Grebler, Blank, and Winnick retrospective survey and the exclusion of non-conforming loans identified by Fishback and Kollmann (2014). Over the full 116 years, housing offered a stable annual return of 9% driven primarily by rental income (7.7%), with capital gains (1.3%) more volatile but rising in importance after 1940, particularly in cities with binding land use regulation.

## Theory / model

The paper has no formal equilibrium model. It applies two accounting frameworks to interpret the HHP data.

**User cost of housing (Section VI, p.587).** The no-arbitrage condition equating annual rent $$R_t$$ to the user cost of owning one unit of housing at price $$P_t$$ is (equations 4 and 5):

$$
R_t = P_t \, u_t, \tag{4}
$$

$$
u_t = i_t + \tau_t + \delta_t - g_{t+1} + \gamma_t, \tag{5}
$$

where $$i_t$$ is the real risk-free interest rate, $$\tau_t$$ the tax benefit or cost associated with homeownership, $$\delta_t$$ maintenance costs, $$g_{t+1}$$ expected appreciation over the coming year, and $$\gamma_t$$ the risk premium for owning relative to renting. The user cost framework motivates the rent-to-price ratio series (R3) and is used to interpret the three-phase narrative of mortgage market development (Section VI).

**Housing production cost identity (Section V, p.581).** Following Glaeser and Gyourko (2018), the production cost of housing in period $$t$$ is

$$
\text{PC}_t = (L_t + \text{CC}_t) \times \text{EP}_t, \tag{3}
$$

where $$L_t$$ is land cost, $$\text{CC}_t$$ construction costs, and $$\text{EP}_t$$ entrepreneurial profit. The paper focuses on $$\text{CC}_t$$ using RSMeans indices; land value estimation is deferred to future work.

**Total return to housing (Section VII, pp.593-594).** For city $$c$$ in year $$t$$, the total real return $$r_{c,t}$$ decomposes into a real capital gain $$h_{c,t}$$ and a rental return $$y_{c,t}$$:

$$
R_{c,t} = H_{c,t} + Y_{c,t} \;=\; \frac{\text{HPI}_{c,t} - \text{HPI}_{c,t-1}}{\text{HPI}_{c,t-1}} + \frac{\text{RI}_{c,t}}{\text{HPI}_{c,t-1}}, \tag{6-7}
$$

where $$\pi_t = (\text{CPI}_t - \text{CPI}_{t-1})/\text{CPI}_{t-1}$$ is CPI inflation. The inflation-adjusted capital gain is

$$
h_{c,t} = \frac{1 + H_{c,t}}{1 + \pi_t} - 1,
$$

and the average return over $$T = \{1891, \ldots, 2006\}$$ is defined as the arithmetic mean (equation 8, p.594):

$$
\bar{h}_c = \frac{1}{|T|} \sum_{t=1891}^{2006} h_{c,t}. \tag{8}
$$

National averages are population-weighted Paasche aggregates.

## Method

The core method is a hedonic rolling-window (RW) regression, following the approach formalized by Silver (2016), implemented here for the first time at city scale over a century-long horizon. The method builds on `panel-regression` (hedonic OLS) but avoids imposing fixed quality coefficients across the full sample.

**Rolling-window hedonic regression (Section III, p.568).** For a rolling window of size $$s$$ with base year $$b$$, and city $$c$$, the estimating equation is (equation 1):

$$
\ln(\text{price})_{ict} = \alpha_{bc} + \sum_{y=b+1}^{b+s-1} \beta_{cy} \cdot \mathbf{1}_{\{y=t\}} + \mathbf{X}_{ict} \boldsymbol{\Gamma}_{bc} + \varepsilon_{ict}, \tag{1}
$$

where the regression uses only observations $$i$$ from years $$t \in \{b, \ldots, b+s-1\}$$, $$\beta_{cy}$$ captures the log price change from year $$y$$ to $$y+1$$ for city $$c$$, $$\mathbf{1}_{\{y=t\}}$$ is an indicator for year $$t$$, and $$\mathbf{X}_{ict}$$ is a vector of property characteristics. The baseline is $$s = 2$$ (two-year windows), giving one $$\beta$$ per window. Allowing coefficients $$\boldsymbol{\Gamma}_{bc}$$ to vary by window means the relative price of a bathroom or an extra room can change over time, addressing unobserved quality drift that a single pooled regression cannot accommodate.

**Chain-linking.** The city-level index at time $$t$$ is the product of all preceding window-specific price changes (equation 2, p.569):

$$
\iota_{ct} = \prod_{y=1891}^{t} \exp(\beta_{cy}). \tag{2}
$$

**National aggregation.** The national index aggregates city-level percentage changes each year using the city's population share as a weight (effectively a Paasche price index). Population counts come from U.S. Census metropolitan area data, interpolated between census years.

**Robustness variants** include three-year and five-year rolling windows; results are qualitatively similar with the main difference in the wartime rental segment (1944-1947), where the two-year specification with rent-control adjustments is preferred.

## Empirical specifications

**HHP rent and sales price indices (Sections II-IV).** Each city-window regression (equation 1) is estimated on the cross-section of newspaper listings for that city in the relevant two-year window. Controls in $$\mathbf{X}_{ict}$$:

- **Location:** 20 standardized geographic areas per city, defined by machine-learning geocoding of address or intersection information; allowed to expand as the metro area grows over time.
- **Size:** Dummies for total rooms (or bedrooms post-WWII) and bathrooms (rounded to nearest half); stories; missing-size indicators included.
- **Type:** House vs. apartment indicator.
- **Rental frequency (rents only):** Rental payment period dummies (weekly/monthly/annual), with frequency imputed for approximately 33,000 listings without a stated period using city-year percentile comparisons.

When the two-year window yields too few listings in a city-year (mainly rental listings during WWII and sales during the Great Depression), the window length is extended to three or five years; cases are documented in Online Appendix Table B2.

**Rent-to-price ratio (Section VI).** The ratio is constructed from the national HPI and RI series, benchmarked to a 2006 value of 3.16% taken from the Davis land-price indicators dataset. The user cost decomposition (equation 5) is used descriptively: real mortgage interest rates from Drehmann, Juselius, and Quincy (2024) and loan-to-value ratios from Fetter (2013) and the Historical Statistics of the United States are plotted against the ratio to identify the three credit-condition phases (1890-1935; 1935-1970; 1970-2006).

**Total return decomposition (Section VII).** For each city and year, the real capital gain and rental return are computed from the baseline HPI and RI indices and the Officer-Williamson CPI. National returns are population-weighted arithmetic means. The volatility comparison uses Shiller (1992) S&P 500 earnings and price data from Robert Shiller's website for equities.

**Housing supply and zoning (Sections V-VI, R7).** The zoning result uses a city-level OLS regression of average annual real capital gains (from HHP HPI) over 1940-1969 and 1970-2006 on the Saiz (2010) zoning restrictiveness measure (Wharton Residential Urban Land Regulation Index from Gyourko, Saiz, and Summers 2008). Reported coefficients are 0.63 and 0.79 respectively (Figure VI, p.586); the difference is not statistically significant.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| HHP Newspaper Real Estate Listings | Primary source: 2.7 million archival newspaper listings for 30 U.S. cities, 1890-2006; supports rent and sales price index construction | No page (hand-collected; data in [Harvard Dataverse](https://doi.org/10.7910/DVN/CYX1PQ)) |
| BLS Rent of Primary Residence (RoPR / CUUR0000SEHA) | Benchmark comparison for the rental price series (Sections III-IV) | [FRED](/wiki/datasets/fred/) |
| Shiller national housing price index (Yale website) | Benchmark comparison for the sales price series (Sections III-IV, VII) | No page yet |
| FHFA House Price Index | City-level benchmark for sales prices from 1975 onward (Online Appendix) | No page yet |
| RSMeans construction cost index | City-level construction cost series (five-year intervals 1940-1980; annual thereafter) used in housing supply analysis (Section V) | No page yet |
| BLS city building permit surveys (1920-1950) and Census Bureau Building Permit Survey (1959-2006) | Population-adjusted housing permit series for supply analysis (Section V) | No page yet |
| Officer-Williamson extended CPI | Deflating all nominal series to real terms | No page yet |
| Wharton Residential Urban Land Regulation Index (WRLURI) | City-level zoning stringency from Gyourko, Saiz, and Summers (2008) / Saiz (2010); used in zoning and capital gains analysis (Section V) | No page yet |
| Drehmann, Juselius, and Quincy (2024) mortgage data | Real mortgage interest rates and loan-to-value ratios for the user cost analysis (Section VI) | No page yet |

Sample scope: 30 U.S. cities from 1890 to 2006 (annual frequency). Most cities enter in 1890; Las Vegas enters later. Approximately 1.23 million rental listings and 1.47 million sales listings.

## When to read the full paper

Read the source at the [original DOI](https://doi.org/10.1093/qje/qjaf047) if you are: constructing city-level housing models and need the annual HPI or RI series (available in Harvard Dataverse); examining the history of U.S. housing returns over the twentieth century; studying the relationship between mortgage market development and asset pricing; or investigating whether the standard Shiller index understates historical housing price growth for a specific period or city. The locators above point to the exact figures and tables.

## Attribution and rights

Source: peer-reviewed, *The Quarterly Journal of Economics* (2026), 559-603. Advance Access published October 10, 2025. This distillation was extracted by an LLM on 2026-06-28 and is **not human-verified or independently reproduced**. The paper is paywalled; redistribution is extract-only.

> Lyons, Ronan C., Allison Shertzer, Rowena Gray, and David Agorastos.
> "The Price of Housing in the United States, 1890-2006."
> *The Quarterly Journal of Economics* (2026), 559-603.
> DOI: 10.1093/qje/qjaf047.
> © The Author(s) 2025. Published by Oxford University Press on behalf of Harvard University.
> All rights reserved. This page is an extract by the Institute for Automated Research.


==============================================================================
# Dollar Dominance and the Transmission of Monetary Policy: McLeay & Tenreyro (2026)
# https://instituteforautomatedresearch.org/wiki/papers/qje/2026/mcleay-dollar-dominance-transmission-monetary-2026/
# Distilled: The MCP model shows monetary easing can still strongly boost exports even under dollar pricing, with export quantities rising 0.95% vs. only 0.14% in sticky-price DCP models, because the binding constraint is supply capacity not demand. Panel evidence from 37 emerging economies and case studies of Canada, Chile, and three large Latin American devaluations confirm significant export responses to monetary-policy-induced exchange rate changes. The Quarterly Journal of Economics 2026, CC BY 4.0. Seven core results with source locators, datasets used, the model, and the method.
# Tags: paper-summary, monetary-policy, open-economy-macro, exchange-rates, international-trade, local-projections, panel-regression, structural, open-access, cc-by, peer-reviewed, unreplicated, data:un-comtrade
==============================================================================

**What this is.** The paper's core results, the model, and the empirical specifications: enough to know what it found and how, without reading all 62 pages. To replicate or extend, read the full source at the [original](https://doi.org/10.1093/qje/qjaf043).

## TL;DR

McLeay and Tenreyro challenge the dominant-currency pricing (DCP) view that dollar invoicing undermines exchange-rate-based monetary policy transmission. They build a mixed currency pricing (MCP) framework in which competitive, homogeneous-good exporters price in dollars with flexible prices, while differentiated-good exporters retain sticky monopoly-power pricing. In the MCP model, a monetary loosening that depreciates the currency lowers domestic production costs expressed in dollars, allowing competitive exporters to expand supply substantially. The binding constraint is export supply capacity (upward-sloping marginal cost from decreasing returns to scale), not demand. The model replicates the empirical fact of limited exchange rate pass-through to dollar export prices (as in DCP), yet delivers a strong export quantity response (as in the classic PCP framework of Obstfeld and Rogoff (1995)), and challenges the optimal-DCP-policy conclusions of Egorov and Mukhin (2023) by showing price flexibility relaxes the binding dollar-pricing constraint. Three empirical exercises confirm the mechanism: monetary policy-induced depreciations raise exports significantly in a panel of 37 emerging and developing economies, in Canada and Chile (where dollar-priced commodity exports dominate), and large devaluations in Argentina, Brazil, and Mexico were followed by visible export expansions relative to trend.

## Core results

Magnitudes as reported; `\*\*\*` = 1%. Locators point to the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | MCP model: export quantity response to monetary easing is 7x the sticky-price DCP response | Table III, p. 639 | Year-1 avg: MCP 0.95%, DCP 0.14%, PCP 0.69%; impact (Figure VI, p. 637): MCP ~1.34%, DCP ~0.05% |
| R2 | MCP model: aggregate output response is 2.5x the DCP response | Table III, p. 639 | Year-1 avg: MCP 0.81%, DCP 0.32%; common exchange rate depreciation: 0.52% |
| R3 | Dollar invoicing is strongly positively associated with homogeneous-goods export share | Table I, p. 620 | OLS coefficient 0.712-0.799\*\*\* (1,173 obs, R² 0.29-0.37); 10 pp more homogeneous goods → 7-8 pp more dollar invoicing |
| R4 | Panel LP: monetary tightening causes significant fall in exports in 37 emerging and developing economies | Figure X, p. 652 | Dollar exports peak fall ~1.5% at 11 months; year-1 avg ~0.99%; 68% CI excludes zero at peak |
| R5 | Canada: monetary tightening causes large export falls consistent with MCP predictions | Figure XI, p. 656 | Energy exports peak -1.5% after 3 months; chemicals -1% after 7 months per 1 pp policy rate increase |
| R6 | Chile: monetary tightening causes large mining export falls consistent with MCP | Figure XII, p. 657 | Mining exports fall ~10% on impact; manufacturing exports avg -1.25% first 6 months per 1 pp tightening |
| R7 | MCP dollar export price pass-through is small (-0.06%), matching DCP, but arises from equilibrium not stickiness | Table III, p. 639 | Year-1 avg: MCP -0.06%, PCP -0.34%, DCP -0.07% (100 bps easing) |

**Overall (paper's conclusion).** The pass-through of monetary policy via the export channel is strong even when goods are priced in dollars, as long as dollar-pricing exporters face competitive markets with flexible prices. The standard interpretation of low exchange rate pass-through to dollar export prices as evidence of nominal rigidities is misleading: in the MCP model, low pass-through is an equilibrium result of high demand elasticity and rising marginal costs, not a friction. Monetary policy and the exchange rate remain effective stabilization tools in a world of dollar dominance. The policy implications of dollar pricing may need to be reassessed.

## Theory / model

The model economy consists of households who consume domestic and imported goods and provide labor for firms that produce for home consumption and exports. A monetary authority sets domestic interest rates via a Taylor rule. The key structural innovation is a nested CES demand system that reverses the standard open-economy nesting, placing intra-sector variety competition at the inner level and cross-industry substitution at the outer level.

**Household preferences.** Each household in country $$j$$ maximizes lifetime expected utility (equation 1, p. 622):

$$
\mathbb{E}_0 \sum_{t=0}^\infty \beta^t \left( \frac{C_{j,t}^{1-\sigma_c}}{1-\sigma_c} - \frac{N_{j,t}(h)^{1+\varphi}}{1+\varphi} \right), \tag{1}
$$

where $$C_{j,t}$$ is total consumption, $$N_{j,t}(h)$$ is labor supply, $$\sigma_c$$ is the coefficient of relative risk aversion (equal to the inverse of the intertemporal elasticity of substitution), and $$\varphi$$ is the reciprocal of the labor supply elasticity.

**Demand structure.** Total consumption aggregates across goods $$g$$ (equation 2, p. 622):

$$
C_{j,t} \equiv \left( \int_0^1 C_{j,t}(g)^{\frac{\sigma-1}{\sigma}} dg \right)^{\frac{\sigma}{\sigma-1}}, \tag{2}
$$

where $$\sigma$$ is the cross-industry elasticity of substitution. Within each good $$g$$, consumption aggregates varieties from all countries (equation 3, p. 623):

$$
C_{j,t}(g) \equiv \left( \sum_i \left( \frac{\gamma^g_{ij}}{|\Omega^g_i|} \right)^{\frac{1}{\eta^g}} \int_{\omega \in \Omega^g_i} C^g_{ij,t}(\omega)^{\frac{\eta^g-1}{\eta^g}} d\omega \right)^{\frac{\eta^g}{\eta^g-1}}, \tag{3}
$$

where $$\eta^g$$ is the within-good cross-variety elasticity (which may vary across goods) and $$\gamma^g_{ij}$$ captures preference for varieties from country $$i$$, arising from home bias and trade costs. Setting $$\eta^g \gg \sigma$$ for homogeneous goods means the relevant price for export demand is the variety price relative to competing foreign varieties, not the aggregate price index. This makes demand highly elastic at the variety level, enabling large quantity adjustments in response to small price changes.

**Firms and production.** A firm in country $$j$$ producing variety $$\omega$$ of good $$g$$ uses labor and intermediate inputs (equation 15, p. 626):

$$
Y^g_{j,t}(\omega) = A^g_{j,t} (L^g_{j,t}(\omega))^{1-\alpha} (X^g_{j,t}(\omega))^\alpha \left[ (L^g_{j,t})^{1-\alpha} (X^g_{j,t})^\alpha \right]^{v_g - 1}, \tag{15}
$$

where $$\alpha$$ is the intermediate input share, $$1-\alpha$$ is the labor share, and $$v_g \leq 1$$ governs returns to scale at the industry level. The term $$[(L^g_{j,t})^{1-\alpha}(X^g_{j,t})^\alpha]^{v_g-1}$$ generates decreasing returns when $$v_g < 1$$, capturing fixed good-specific factors such as structures. The resulting industry-level marginal cost (equation 24, p. 629) is:

$$
MC^g_{j,t} = \frac{1}{(1-\alpha)^{1-\alpha} \alpha^\alpha} \frac{W^{1-\alpha}_{j,t} P^\alpha_{j,t} \left[ L^{1-\alpha}_{j,t} X^\alpha_{j,t} \right]^{1-v_g}}{A^g_{j,t}}, \tag{24}
$$

which rises with industry output when $$v_g < 1$$, generating an upward-sloping marginal cost curve. This supply-side constraint, not demand, limits the export expansion after a depreciation.

**Pricing.** Each firm resets its price with good-specific Calvo probability $$1 - \delta^g_p$$ each period. Dollar-pricing firms solve (equation 21, p. 628):

$$
\mathbb{E}_t \left[ \sum_{s=0}^\infty (\beta \delta^g_p)^s \frac{C^{-\sigma_c}_{j,t} P_{j,t}}{C^{-\sigma_c}_{j,t+s} P_{j,t+s}} Y^g_{ji,t+s}(\omega) \left( \bar{P}^{g,\$}_{ji,t}(\omega) - \frac{\eta^g}{\eta^g - 1} \frac{MC_{j,t+s}(\omega)}{\mathcal{E}_{\$j,t+s}} \right) \right] = 0, \tag{21}
$$

setting the dollar reset price as a markup $$\eta^g / (\eta^g - 1)$$ over the weighted average of future dollar marginal costs. When prices are flexible ($$\delta^g_p \to 0$$), the optimal dollar price depends only on current dollar marginal costs and the invoicing currency is irrelevant: a depreciation that lowers home costs in dollar terms leads to a small equilibrium price cut (smaller when demand is more elastic) and a large quantity increase.

**Monetary policy.** The central bank sets domestic nominal interest rates via a Taylor rule (equation 25, p. 629):

$$
\frac{1 + i_{j,t}}{1 + \bar{i}_j} = \left( \frac{1 + i_{j,t-1}}{1 + \bar{i}_j} \right)^\rho (1 + \pi_{j,t})^{(1-\rho)\phi_\pi} \zeta^M_{j,t}, \tag{25}
$$

where $$\rho$$ is policy smoothing, $$\phi_\pi > 1$$ is the inflation response coefficient, $$\bar{i}_j$$ is the steady-state nominal rate, and $$\zeta^M_{j,t}$$ is an AR(1) monetary policy shock. A negative shock (easing) reduces the policy rate, leading to a nominal exchange rate depreciation.

**The central finding.** The depreciation lowers domestic dollar costs (wages expressed in dollars fall). For a monopolistic sticky-price DCP exporter, the price cannot adjust so markups rise but quantities stay flat. For a competitive flexible-price exporter with high $$\eta^g$$, the optimal reset price falls only slightly (elastic demand means profits respond more to volume than to margin). The quantity adjustment is large, continuing until rising marginal cost from expanding production offsets the improved profitability. The supply constraint parameter $$v_g$$ determines the size of the export response: under constant returns ($$v_g = 1$$), the quantity response is very large; under decreasing returns ($$v_g = 0.85$$), it is still substantially larger than in the DCP model. The MCP model nests sticky-price DCP (set $$\eta^g = \sigma$$, $$\delta^g_p = 0.75$$) and PCP (set $$\delta^g_p = 0$$, $$\eta^g = \sigma$$) as special cases.

## Method

**Model calibration and simulation.** The model is linearized around a deterministic steady state and simulated using impulse response functions (Figure VI, p. 637). The baseline calibration for households and policy follows Gopinath et al. (2020): cross-product elasticity $$\sigma = 2$$, labor demand elasticity $$\vartheta = 4$$, Calvo price rigidity $$\delta_p = 0.75$$ (four-quarter mean duration), Calvo wage rigidity $$\delta_w = 0.75$$ (Table II, pp. 634-635). The key departures for the homogeneous export sector: fully flexible prices ($$\delta^{g_H}_p = 0$$), cross-variety elasticity $$\eta^{g_H} = 17$$ (from Broda and Weinstein (2006) for crude oil 1972-1988), and decreasing returns $$v_{g_H} = 0.85$$ (calibrated from the share of structures in Canadian mining value-added). Country-specific calibrations for Canada and Chile are in Table IV (p. 654).

The method builds on `local-projections` (Jordà 2005) for the empirical tests and `panel-regression` for the motivating cross-country facts, with the proposed `nk-soe-dsge` framework as the structural basis.

**Monetary policy shock identification.** Because the exchange rate is endogenous, the empirical strategy uses monetary policy shocks identified by purging the interest rate of its response to current macroeconomic conditions. Shocks are obtained as residuals $$\hat{\epsilon}_{i,t}$$ from a forward-looking interest rate rule (equation 38, p. 650):

$$
\Delta i_{i,t} = \alpha + \phi_\pi E_t \pi^f_{i,t+12} + \phi_y E_t \Delta y^f_{i,t+12} + \sum_{j=1}^2 \phi_\pi \pi_{i,t-j} + \sum_{j=1}^2 \phi_y \Delta y_{i,t-j} + \sum_{j=1}^2 \phi_e \Delta NEER_{i,t-j} + \sum_{j=1}^2 \phi_i i_{i,t-j} + \epsilon_{i,t}, \tag{38}
$$

where $$E_t \pi^f_{i,t+12}$$ and $$E_t \Delta y^f_{i,t+12}$$ are 12-month-ahead forecasts of inflation and output growth. The residual $$\hat{\epsilon}_{i,t}$$ is by construction uncorrelated with past macro conditions and current forecasts, providing an exogenous driver of exchange rate changes.

## Empirical specifications

**Section III: Fact 3 invoicing regression (Table I, p. 620).** Cross-country OLS establishes that dollar invoicing is concentrated in homogeneous-good sectors. The regression uses four-digit SITC data from UN Comtrade (Rauch (1999) homogeneous-goods classification) and invoicing data from Boz et al. (2022):

$$
\text{Dollar share}_{i,t} = \beta_0 + \beta_1 \text{Homogeneous share}_{i,t} + \mu_t + \epsilon_{i,t},
$$

with specifications adding year fixed effects and GDP weighting. Estimated on 1,173 observations across 101 countries (1990-2019) with robust standard errors. Coefficient $$\hat{\beta}_1 = 0.712$$ to $$0.799$$ (all significant at 1%).

**Section V.B: Panel local projections in 37 EMEs (equation 39, p. 651).** Macroeconomic effects of identified monetary shocks on exports and activity are estimated using Jordà (2005)'s LP method with country fixed effects:

$$
z_{i,t+h} = \mu^h_i + \sum_{j=0}^2 \gamma^h_j \hat{\epsilon}_{i,t-j} + \delta^h_0 \Delta NEER_{i,t} \times \hat{\epsilon}_{i,t} + \sum_{j=0}^2 \beta^h_j \times \text{controls}_{i,t-j} + \omega^h_{i,t}, \tag{39}
$$

where $$h$$ is the horizon in months, the interaction term $$\Delta NEER_{i,t} \times \hat{\epsilon}_{i,t}$$ captures the differential effect through the exchange rate, and $$\omega^h_{i,t}$$ is the residual. Impulse responses are normalized to a 1 percentage point interest rate increase on impact (so all results correspond to a monetary tightening). The panel database is from Brandao-Marques et al. (2021), covering 37 countries.

**Section V.C: Country VARs for Canada and Chile (equation 40, p. 655).** For each economy, a hybrid VAR with six lags (Canada) or four lags (Chile) is estimated:

$$
\mathbf{X}_t = \mathbf{c} + \delta t + B(\mathbf{L}) \mathbf{X}_{t-1} + C(\mathbf{L}) \mathbf{W}_{t-1} + \boldsymbol{\epsilon}_t, \tag{40}
$$

where $$\mathbf{X}_t$$ contains the monetary policy shock series (ordered first for recursive identification), exchange rate, CPI, GDP, and sectoral exports; $$\mathbf{W}_t$$ includes the U.S. dollar price of Canadian commodities (Canada only). Identification is recursive (Cholesky), with the cumulative monetary shock ordered first. For Canada: Champagne and Sekkel (2018) narrative shocks, monthly, 1981-2015. For Chile: Brandao-Marques et al. (2021) shocks, monthly, 2003-2017. Model impulse responses (solid red lines in Figures XI-XII) are scaled to match the average estimated exchange rate response over the first six months.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| UN Comtrade (four-digit SITC) | Share of homogeneous goods in total goods exports, 1985-2023 (Figure IV); base data for invoicing regression (Table I) | no page yet |
| Boz et al. (2022) invoicing database | Share of exports invoiced in dollars, 1990-2019, used in Table I regression | no page yet |
| Brandao-Marques et al. (2021) panel | Monetary policy shocks and macro data for 37 emerging and developing economies; used in Section V.B LP estimation (equation 39, Figure X) | no page yet |
| Champagne and Sekkel (2018) shock series | Narrative monetary policy shocks for Canada, 1974-2015; used in Canada VAR (equation 40, Figure XI) | no page yet |
| Canadian national statistics (Bank of Canada / Statistics Canada) | Monthly interest rate, CPI, GDP, energy and chemicals exports, exchange rate, 1981-2015 | no page yet |
| Chilean national statistics (Banco Central de Chile) | Monthly IMACEC (output), policy rate, CPI, mining and manufacturing exports, exchange rate, 2003-2017 | no page yet |
| Harvard Dataverse replication files | Assembled replication dataset (McLeay and Tenreyro 2025, doi:10.7910/DVN/SASVME) | no page yet |

Sample: panel LP covers quarterly data for 37 countries (1990-2019); country VARs use monthly data (Canada: T ≈ 418 months; Chile: T ≈ 172 months); invoicing regression covers 1,173 country-year observations across 101 countries.

## When to read the full paper

Read the [original](https://doi.org/10.1093/qje/qjaf043) if you are: building or evaluating open-economy monetary models where the invoicing currency choice matters; empirically studying exchange rate pass-through and its interpretation for monetary policy; assessing whether monetary policy transmission through exports remains effective in highly dollar-invoiced developing and emerging economies; or using local projections to identify monetary policy effects on trade flows in a panel with heterogeneous export structures.

## Attribution and rights

Source: peer-reviewed, *The Quarterly Journal of Economics* 141(1), 2026. This distillation was extracted by an LLM on 2026-06-28 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** McLeay, Michael, and Silvana Tenreyro.
> "Dollar Dominance and the Transmission of Monetary Policy."
> *The Quarterly Journal of Economics* 141, no. 1 (2026): 605-666.
> DOI: 10.1093/qje/qjaf043. © 2025 The Author(s).
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Marginal Returns to Public Universities: Mountjoy (2026)
# https://instituteforautomatedresearch.org/wiki/papers/qje/2026/mountjoy-marginal-returns-public-universities-2026/
# Distilled: Using a fuzzy regression discontinuity design across hundreds of SAT/ACT admission cutoffs at all 35 Texas public universities, this paper establishes that marginal admission raises four-year credits by one year, BA completion by 12 percentage points, and earnings by 8.6%; internal rates of return are 26% for students and 16% for society. QJE 2026, CC BY 4.0. Nine core results with source locators, datasets used, the RD design with equations, and the intensive/extensive margin bounding method.
# Tags: paper-summary, education-economics, returns-to-education, higher-education
==============================================================================

**What this is.** The paper's core results, the identification design with its key equations, and the intensive/extensive margin bounding method: enough to know what it found and how without reading all 69 pages. To replicate or extend it, read the full source at [https://doi.org/10.1093/qje/qjaf055](https://doi.org/10.1093/qje/qjaf055).

## TL;DR

Barrow and Malamud (2015) documented the scarcity of credible causal evidence on the returns to college, and Carneiro, Heckman, and Vytlacil (2011) showed that marginal returns to education may be heterogeneous across individuals. This paper addresses both gaps. It assembles a large collection of decentralized SAT/ACT admission score cutoffs at all 35 Texas public universities and links them to administrative records on high school preparation, college trajectories, and long-run quarterly earnings. A fuzzy regression discontinuity (RD) design compares barely admitted versus barely rejected applicants at each cutoff. Marginally admitted students gain roughly one additional year of four-year college credits, become 12 percentage points more likely to ever earn a bachelor's degree, and earn 8.6% more than their marginally rejected counterparts eight to twelve years out. Grant aid nearly fully offsets the additional tuition cost, so marginally admitted students pay essentially no additional net tuition. A formal cost-benefit analysis finds internal rates of return of 26% for students, 16% for society, and 7% for the government. Earnings gains are uniform across institutions of widely varying selectivity, while students from low-income families reap smaller earnings gains, driven by lower degree completion, more time enrolled, and majoring in less lucrative fields. Finally, a novel bounding method reveals that the pooled gains are driven primarily by students who would not have enrolled in any four-year college if rejected (the extensive margin), with a smaller contribution from students who would have attended a less selective four-year institution (the intensive margin).

## Core results

Magnitudes are as reported; all estimates use the main fuzzy RD specification of Section II.D. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Marginal admission increases cumulative four-year credits by 28.4 (roughly 1 full year of four-year education) | Table II, p. 463; Figure VI, p. 456 | LATE = 28.4 credits (se = 3.0); untreated complier mean = 64.7 credits; net gain in two-year credits = -15.5 (se = 1.9) |
| R2 | Marginal admission raises the probability of ever earning a BA by 12 pp | Table II, p. 463; Figure VI, p. 456 | LATE = 0.119 (se = 0.026); untreated complier mean = 0.406; effect is entirely in non-STEM fields (Figure VII, p. 458) |
| R3 | Log earnings gain of 8.6% (pooled 8-12 years after application) | Table II, p. 463; Figure VIII, p. 460 | LATE = 0.086 (se = 0.032); crossover from negative to positive earnings effects at year 6 |
| R4 | Annual dollar earnings premium of $3,339 (8-12 years out) | Table II, p. 463; Figure VIII, p. 460 | LATE = $3,339 (se = 1,288); untreated mean = $40,829; 8.2% of baseline mean |
| R5 | Earnings rank gain of 4.1 percentile points in statewide distribution | Table II, p. 463; Figure VIII, p. 460 | LATE = 4.1 pp (se = 1.4); untreated mean = 52.4; from about the 50th to the 54th percentile |
| R6 | Zero additional net tuition; private IRR of 26%, social IRR of 16%, taxpayer IRR of 7% | Online Appendix Figure A.13; Figure X, p. 469 | Net tuition LATE ≈ $0 (grant aid offsets $4,600 gross tuition increase); NPV at 3%: $70,000 student, $80,000 social; society payoff horizon: 11 years, taxpayer: 19 years |
| R7 | No systematic earnings difference across institution selectivity | Figure XII, p. 477 | Log-earnings slope per 100 SAT points of selectivity = -0.000 (se = 0.019); dollar-earnings slope weakly positive but not significant; selectivity predicts peer earnings gain (\~$3,000 per 100 SAT) but not complier own-earnings gain (\~$900 per 100 SAT, insignificant) |
| R8 | Low-income (FRPL-eligible) compliers experience near-zero earnings gains 8-12 years out | Figure XIII, p. 480 | Point estimate for FRPL group indistinguishable from zero; non-FRPL gains are positive and significant; difference driven by fewer degrees, more time enrolled, and less lucrative majors |
| R9 | Extensive-margin BA completion effect is 20-25 pp; intensive-margin effect is only 4-7 pp | Figure XIV, p. 486; Figure XV, p. 489 | A₀=0 group (93% extensive-margin) LATE bounds: [20, 25] pp; A₀=1 group (71% intensive-margin) LATE bounds: [4, 7] pp; extensive margin drives the majority of the pooled 12 pp effect |

**Overall (paper's conclusion).** Marginal admission to U.S. public universities generates large and positive returns for students, society, and taxpayers, even after accounting for the costs of additional education and the long delay before earnings gains materialize. The returns are broadly similar across institutions of varying selectivity, but smaller for students from low-income families. The predominance of extensive-margin gains suggests that the returns are primarily from expanding access to any four-year institution rather than from upgrading to a more selective one.

## Theory / model

The paper has no structural economic model. The key parameter of interest is the local average treatment effect (LATE) of enrolling at a target public university for marginal applicants who enroll if and only if they barely cross the admission cutoff.

**Potential outcomes setup (Section II.D, p. 449).** Let D indicate whether applicant i enrolls at the target university. Define $$Y_1$$ as the potential outcome if enrolled and $$Y_0$$ if not. The applicant's concorded test score serves as the running variable R, with c the admission cutoff for a given application cell. The marginal applicants are those with $$R = c$$; among them, $$D_1(c) = 1$$ (would enroll if admitted) and $$D_0(c) = 0$$ (would not enroll if rejected) defines the complier population. The parameter of interest is:

$$
\text{LATE} = \mathbb{E}[Y_1 - Y_0 \mid R = c,\; D_1(c) = 1,\; D_0(c) = 0]
$$

This is identified by the fuzzy RD estimand (equation 1, p. 449):

$$
\text{LATE} = \frac{\lim_{r \downarrow c} \mathbb{E}[Y \mid R=r] - \lim_{r \uparrow c} \mathbb{E}[Y \mid R=r]}{\lim_{r \downarrow c} \mathbb{E}[D \mid R=r] - \lim_{r \uparrow c} \mathbb{E}[D \mid R=r]} \tag{1}
$$

Four standard assumptions justify this (Hahn, Todd, and Van der Klaauw (2001); Dong (2018)): (i) first-stage relevance (Figure III bottom panel, p. 446: first-stage LATE = 0.150, F = 2,024); (ii) continuity of conditional expectations of unobservables through the cutoff; (iii) monotonicity (cutoff-crossing weakly increases enrollment); and (iv) the exclusion restriction (cutoff-crossing affects outcomes only through initial enrollment at the target university). The smooth density of applicants and their pre-college characteristics through the cutoffs supports assumption (ii) (Figure IV, p. 447).

The pooled estimand identified by pooling all application cells is a weighted average of cell-specific LATEs, with more weight on cells where applicants at the cutoff are more numerous and more likely to be compliers (Cattaneo et al. (2016), p. 450).

**Identified hypotheses.** The paper tests three conceptually distinct questions: (1) Does marginal admission increase four-year educational attainment? (2) Does it generate long-run earnings gains that exceed the costs of additional education? (3) Are the effects driven by the extensive margin (first access to four-year college) or the intensive margin (upgrading selectivity)? These are tested sequentially across Sections III, IV, and VII respectively.

## Method

**Cutoff inference (Section II.B-C, pp. 441-443).** Because publicly posted admissions criteria may differ from operative criteria and invite strategic manipulation, cutoffs are inferred from the data. For each of roughly 700 application cells (university × year × high-school GPA quartile × test type), local linear regressions are estimated at each distinct test-score value and the cutoff is defined as the value producing the largest discontinuity in admission and enrollment. Porter and Yu (2015) show this yields a superconsistent estimator of the true cutoff, leaving the asymptotic distribution of the second-stage RD estimator unaffected. The procedure identifies cutoffs distributed across the full SAT/ACT score range (Figure I, p. 444).

**Intensive/extensive margin bounding method (Section VII, p. 484-489).** This builds on `partial-identification-bounds` from Manski and Pepper (2000) and the complier-describing logic of Abadie (2002). Let $$A \in \{0, 1\}$$ indicate whether a given applicant has any other Texas public university admission offer. The pooled LATE decomposes as (equation 2, p. 485):

$$
\mathbb{E}[Y_1 - Y_0 \mid D_0=0, D_1=1] = \omega\, \mathbb{E}[Y_1-Y_0 \mid D_0=0, D_1=1, A_0=1] + (1-\omega)\,\mathbb{E}[Y_1-Y_0 \mid D_0=0, D_1=1, A_0=0] \tag{2}
$$

where $$\omega$$ is the share of cutoff compliers with $$A_0 = 1$$ (another four-year option available) and $$(1-\omega)$$ is the share with $$A_0 = 0$$ (no other four-year option). Since $$A_0$$ is a latent type (endogenous to cutoff-crossing), the paper imposes a rank assumption: A-always-takers have weakly better treated potential outcomes than A-compliers, who have weakly better outcomes than A-never-takers. This assumption is supported by observed covariate patterns (Figure XIV top right panel, p. 486) and yields sharp bounds on the separate intensive- and extensive-margin LATEs (Figure XV, p. 489).

## Empirical specifications

All specifications follow Section II.D (pp. 449-450) and use the `regression-discontinuity-design` approach. The running variable for each applicant is her concorded test score minus the admission cutoff of her target application cell, with SAT scores divided by 40 to convert to ACT units for pooled estimation.

**Main RD specification (all results, R1-R8).** For outcome Y measured at a given year:

$$
Y_{iac} = \alpha_{ac} + \gamma_{ac}^{-} R_{iac} \cdot \mathbf{1}[R_{iac} < 0] + \beta_{\text{RF}} \cdot \mathbf{1}[R_{iac} \geq 0] + \gamma_{ac}^{+} R_{iac} \cdot \mathbf{1}[R_{iac} \geq 0] + \varepsilon_{iac}
$$

where $$\alpha_{ac}$$ are application-cell fixed effects, $$\gamma_{ac}^{-}$$ and $$\gamma_{ac}^{+}$$ are cell-specific running-variable slopes on each side of the cutoff, and $$\beta_{\text{RF}}$$ is the reduced-form discontinuity. The bandwidth is 3 concorded ACT points (approximately 120 SAT points) with a triangular kernel. Standard errors are clustered at the applicant level (Kolesár and Rothe (2018)). The first-stage regression substitutes D (enrollment at the target university) for Y:

$$
D_{iac} = \alpha_{ac}^D + \gamma_{ac}^{D-} R_{iac} \cdot \mathbf{1}[R_{iac} < 0] + \pi \cdot \mathbf{1}[R_{iac} \geq 0] + \gamma_{ac}^{D+} R_{iac} \cdot \mathbf{1}[R_{iac} \geq 0] + u_{iac}
$$

with $$\hat\pi = 0.150$$ (se = 0.003; F = 2,024, Figure III, p. 446). The LATE is $$\widehat{\text{LATE}} = \hat\beta_{\text{RF}} / \hat\pi$$. No additional control variables are required; the estimates are very similar when adding the full suite of pre-college covariates.

**Pooling across application cells.** The running variable is normalized to zero at each cutoff, and application-cell fixed effects $$\alpha_{ac}$$ absorb all between-cell variation. The pooled estimand is a precision-weighted average of cell-specific LATEs (Cattaneo et al. (2016)); cells with a more numerous and more complier-dense pool at the cutoff receive more weight.

**Earnings stacked panel (R3-R5, R7-R8).** For earnings outcomes over years 8-12, a long panel is constructed with one observation per individual per year. Year indicators are interacted with cell fixed effects and slopes; the coefficient on $$\mathbf{1}[R_{iac} \geq 0]$$ in this pooled specification delivers a precision-weighted average of the year-specific LATEs (Section III.C, p. 461, footnote 17).

**Intensive/extensive margin specification (R9).** Two separate fuzzy RD regressions replace Y and D with interactions involving the stratifier $$A$$ (whether the applicant has any other Texas public university offer). A series of RD regressions on $$Y$$, $$D$$, and $$A$$ interactions identifies the observable mean potential outcomes; the rank assumption then bounds the unobservable components (Online Appendix C; Figure XIV, p. 486).

**Covariate balance and density tests.** Figure IV (p. 447) verifies: (i) no discontinuity in log application density (McCrary (2008) test: RD estimate = 0.005, se = 0.011); (ii) smooth covariate-predicted BA completion through the cutoff (estimate = -0.0013, se = 0.0012); (iii) smooth covariate-predicted earnings through the cutoff (estimate = 19, se = 63). Specification checks across 12 alternative bandwidth and polynomial choices confirm robustness (Online Appendix Figure A.11).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Texas administrative education records via UT-Dallas Education Research Center (UT-Dallas ERC): Texas Education Agency (TEA) high school records and Texas Higher Education Coordinating Board (THECB) college application, admission, and enrollment records | Running variable (SAT/ACT scores), admission decisions, enrollment trajectories, credit accumulation, degree completion, financial aid - 35 TX public universities, cohorts 2004-2014 | No page yet |
| Texas Workforce Commission (TWC) quarterly earnings records | Labor market earnings outcomes, 8-12+ years post-application; quarterly "sandwich" earnings measure | No page yet |
| IPEDS (Integrated Postsecondary Education Data System, NCES) | Institution-level per-student expenditures, tuition sticker prices, for cost-benefit analysis | No page yet |
| National Student Clearinghouse (NSC) | College enrollment records outside Texas for 2008-2014 cohorts (two-thirds of sample), to verify no college enrollment elsewhere | No page yet |

Sample: Texas public high school graduates 2004-2014; RD bandwidth yields 234,271 applicants for education outcomes and 200,456 applicants for earnings outcomes. Earnings outcomes are available for a 12-year post-application window; cost outcomes lag by one year and are observed for 9-year balanced panels.

## When to read the full paper

Read [the original](https://doi.org/10.1093/qje/qjaf055) if you need: the full battery of specification checks across bandwidths, polynomial forms, and cutoff definitions (Online Appendix Figures A.10-A.12); the institution-by-institution and demographic-subgroup heterogeneity results with confidence intervals (Figures XI-XIII); the detailed cost-benefit accounting under alternative assumptions about room and board, earnings growth, and life-cycle extrapolation (Online Appendix Figures A.15-A.18); the formal exposition of the intensive/extensive margin bounding method and its proofs (Online Appendix C); or the comparison with prior estimates from Hoekstra (2009), Zimmerman (2014), Kozakowski (2023), Goodman, Hurwitz, and Smith (2017), Bleemer (2024), and Smith, Goodman, and Hurwitz (2025) reconciling this paper's smaller effect sizes with their larger estimates via differences in treatment size and institution type (Section III.D, pp. 463-465). The locators in the Core results table above point to the exact tables and figures.

## Attribution and rights

Source: peer-reviewed, *The Quarterly Journal of Economics* 141(1), 2026. This distillation was extracted by an LLM on 2026-06-28 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Mountjoy, Jack. "Marginal Returns to Public Universities."
> *The Quarterly Journal of Economics* 141, no. 1 (2026): 429-497.
> DOI: 10.1093/qje/qjaf055. Copyright The Author(s) 2025. Published by Oxford University
> Press on behalf of President and Fellows of Harvard College.
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Who's Afraid of the Minimum Wage?: Rao & Risch (2026)
# https://instituteforautomatedresearch.org/wiki/papers/qje/2026/rao-afraid-minimum-wage-measuring-2026/
# Distilled: Using matched IRS administrative tax records for roughly 271,000 independent U.S. businesses over 2010-2019 and a stacked difference-in-differences design on 19 state minimum wage changes, Rao and Risch find that firms in highly exposed industries do not lay off workers but modestly reduce part-time hiring, fully finance higher wage costs through revenue growth, and leave owner profits unchanged; firm entry falls roughly 2% and individual low earners gain earnings with stable employment rates. QJE 2026, CC BY 4.0. Eight core results with source locators, datasets, and the estimating equations.
# Tags: paper-summary, minimum-wage, labor-markets, small-business, firm-dynamics
==============================================================================

**What this is.** The paper's core results, the identification strategy, and the estimating
equations: enough to know what it found and how, without reading all 55 pages. To
replicate or extend it, read the full source at the
[original](https://doi.org/10.1093/qje/qjaf053).

## TL;DR

Rao and Risch construct the first matched firm-worker-owner panel from the
universe of U.S. pass-through tax returns, covering roughly 271,000 independent
businesses in highly minimum-wage-exposed industries over 2010-2019. Using 19
state minimum wage increases between 2013 and 2016 as quasi-natural experiments
and a stacked difference-in-differences design with 22 control states as clean
comparators, they estimate how independent businesses accommodate higher wage
floors. Firms do not lay off existing workers but modestly reduce part-time
hiring, ending up with about 1.5 fewer employment relationships per year. They
fully finance the higher wage bills through revenue growth: four years out,
revenues rise 3.3% of baseline while profits are statistically indistinguishable
from zero change. Firm entry falls roughly 2%, with surviving entrants
positively selected for efficiency. At the individual level, low earners and
young workers gain substantially in earnings (+18.9% and +21.8%) with
essentially unchanged employment rates. Minimum wages redistribute from
consumers to workers; owners escape the burden.

## Core results

Magnitudes are as reported from the source PDF. Locators are in the form
Table/Figure, page number. All coefficients are from the stacked
difference-in-differences specification at event-year s+4 (four years after
the initial minimum wage increase).

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Firms reduce part-time hiring; no layoffs; employment falls ~2% | Figure II Panel A, p. 390 | 1.5 fewer employment relationships per firm per year; own-wage employment elasticity -0.245 (s.e. 0.134) |
| R2 | Wage bills rise sharply after the minimum wage increase | Figure III Panel A, p. 392; Table II, p. 397 | +0.0143 of baseline revenue (s.e. 0.00167); 1.43 cents per dollar baseline revenue |
| R3 | Revenues rise and fully cover the added labor costs | Figure IV, p. 395; Table II, p. 397 | +0.0331 of baseline revenue; 3.31 cents per dollar; consumers bear 100% of incidence |
| R4 | Owner profits are unchanged: incidence falls on consumers, not owners | Figure V, p. 396; Table II, p. 397 | 0.0002 (s.e. 0.0029); null result; rules out losses larger than 0.37% of baseline revenues at 95% confidence |
| R5 | Firm entry falls; total active firms decline ~2% | Figure VI, p. 401 | -1.97% total active firms; entry rate -5.5%; exit flat (increase >0.53% ruled out at 95%) |
| R6 | Value added per worker rises; stronger among entrants | Table III, p. 402 | +4.6% across all firms (coeff. 0.0462, s.e. 0.0102); entrants +15.5% (0.1550, s.e. 0.0473); incumbents +3.1% (0.0313, s.e. 0.0149) |
| R7 | Low-earning workers gain earnings with near-zero employment effect | Figure VII Panel A, p. 409; Table VI Panel A, p. 411 | +$1,473 (+18.9%) at s+4; own-wage employment elasticity -0.013 (s.e. 0.036) |
| R8 | Young workers gain earnings with stable employment; retention rises | Figure VII Panel B, p. 409; Figure VIII, p. 413; Figure IX, p. 415 | +$1,995 (+21.8%) for ages 15-26 at s+4; employment elasticity 0.064 (s.e. 0.032); retention +1.30 pp for low earners |

**Overall (paper's conclusion).** Independent businesses are more adaptable than
the conventional narrative about small-firm vulnerability to minimum wage hikes
suggests. They absorb the cost shock by passing it through to consumers via
revenue growth and by shedding the least productive firms from the industry,
leaving owners whole and workers better off. The worker-reallocation channel
(from independent businesses to larger C-corporations), analogous to what
Dustmann et al. (2022) document for Germany, explains why firm-level employment
reductions do not translate into individual-level unemployment.

## Theory / model

The paper presents no formal model in the main text. A Cournot competition
model with heterogeneous production technologies appears in Online Appendix O,
following Besley (1989). The conceptual framework generates five empirical
predictions that structure the analysis.

The paper connects to the long-standing empirical debate on minimum wage
employment effects, complementing Card and Krueger (1995) and extending the
who-pays analysis of Harasztosi and Lindner (2019) to independent U.S.
businesses using matched tax data.

Under imperfect product market competition with fixed costs, a minimum wage acts
as a differential labor cost shock: firms relying more heavily on low-wage labor
face larger per-unit cost increases. The key predictions are:

1. **Employment.** Incumbent firms facing modest cost shocks need not reduce
   employment if they can pass costs forward. Employment reductions should be
   concentrated in part-time and low-earning positions.

2. **Revenue pass-through.** Under Cournot, market shares are proportional to
   margins. An industry-wide cost shock is easier to pass through than a
   unilateral price increase because the elasticity facing firms is the industry
   demand elasticity, not the firm demand elasticity. Revenue should rise to
   offset wage bill increases.

3. **Profits.** If revenue pass-through is complete, owner profits should be
   unchanged. Consumers bear the entire burden.

4. **Entry deterrence.** Higher fixed operating costs (relative to benefits)
   deter entry of firms that cannot cover the wage premium. The minimum wage
   raises the viability threshold, reducing entrant counts.

5. **Positive selection.** Firms that enter despite higher costs are more
   productive and efficient than the marginal entrants under the pre-reform
   wage floor, generating a positive shift in the productivity distribution
   of the industry.

**Identification.** The paper relies on state-level variation in minimum wage
policy. Treatment states are the 17 states, Washington DC, and the city of
Chicago that raised their minimum wages between 2013 and 2016. Control states
are 22 states that enacted no minimum wage increase between 2011 and 2019,
providing a set of clean controls (p. 385). The stacked design compares treated
firms in each reform cohort to all control firms over event time, avoiding the
negative-weight problem in staggered difference-in-differences that arises when
previously treated units serve as controls (Callaway and Sant'Anna (2021);
Goodman-Bacon (2021)). Pre-trend validation at s = -4 to s = -2 confirms
parallel trends across all primary outcomes.

## Method

The paper applies a panel stacked difference-in-differences design introduced
in Cengiz et al. (2019) and extended here to a firm-level setting. It builds
on `difference-in-differences`, `panel-regression`, and `event-study` estimators.

**Firm-level estimating equation (equation 1, p. 385).** For firm j in year t
belonging to reform cohort c:

$$
y_{jct} = \alpha + \sum_{s=-4,\, s \neq -1}^{4} \left(\beta_s \,\text{treat}_{jc} + \Gamma_s X_{jc}\right) \times \text{year}_{s=t} + \delta_{ct} + \psi_{jc} + \nu_{jct} \tag{1}
$$

where $$\text{treat}_{jc}$$ is an indicator for firm j being in a treatment state
in cohort c; $$X_{jc}$$ is a vector of baseline firm and market controls (size
categories, value-added deciles, two-digit industry, county density quintiles,
county employment-rate quintiles); $$\delta_{ct}$$ is a cohort-by-year fixed
effect; $$\psi_{jc}$$ is a firm-cohort fixed effect; and $$s = -1$$ (year before
the minimum wage increase) is the omitted base year. The DD estimator
$$\beta_s$$ represents the differential average outcome between firms in treated
and untreated states relative to the pre-reform base year. Standard errors are
clustered at the state-by-cohort level.

For outcomes scaled by baseline revenue the dependent variable is
$$y_{jct} = z_{jct} / \text{revenue}_{j,s-1}$$. For percent-change outcomes the
dependent variable is $$y_{jct} = z_{jct} / z_{j,s-1}$$.
Regressions are weighted by log baseline revenues.

**Individual-level estimating equation (equation 2, p. 387).** For individual i
in year t belonging to cohort c:

$$
y_{ict} = \alpha + \sum_{s=-4,\, s \neq -1}^{4} \left(\beta_s \,\text{treat}_{ic} + \Gamma_s V_{ic}\right) \times \text{year}_{s=t} + \delta_{ct} + \rho_{ic} + \nu_{ict} \tag{2}
$$

where $$V_{ic}$$ are individual controls (age, age squared, county density
quintiles, county employment-rate quintiles); $$\rho_{ic}$$ is an
individual-cohort fixed effect (replacing the firm fixed effect). For binary
employment outcomes the specification is a linear probability model (LPM),
with coefficients interpreted as percentage-point changes for the treatment
group relative to the control group.

## Empirical specifications

All headline estimates are at event-year s+4 (four years after the initial
minimum wage increase). The full event-study path (s = -4 to s = 4, omitting
s = -1) is shown in the figures.

**Firm-level analyses (Section IV).** The primary sample is a balanced panel of
134,974 independent businesses in highly exposed industries in treatment and
control states, measured in the base year. "Highly exposed" industries are
four-digit NAICS industries where at least 1% of workers are paid less than the
prevailing minimum wage, identified using CPS Monthly Outgoing Rotation Group
data (CPS MORGs) for the pre-reform period. Restaurants alone account for 42%
of minimum wage workers; together the selected industries cover more than
two-thirds of minimum wage workers (pp. 382-383).

The main firm-level outcomes are:
- Wage bill / baseline revenue (Figure III; Table II)
- Revenues / baseline revenue (Figure IV; Table II)
- Owner profits / baseline revenue (Figure V; Table II)
- Number of employment relationships (Figure II; Section IV.A)
- Value added per worker (Table III)

The COGS and other-deductions items complete the income statement decomposition
in Table II, which traces the incidence of each cost dollar:

| Component | All exposed | Restaurants | Other/retail |
|---|---|---|---|
| Wage bill | 0.0143 | 0.0201 | 0.0076 |
| Revenue (financing) | 0.0331 | 0.0294 | 0.0359 |
| COGS (nonlabor costs) | 0.0131 | 0.0053 | 0.0210 |
| Other deductions | 0.0035 | 0.0029 | 0.0039 |
| Owner profits | 0.0002 | -0.0001 | 0.0003 |

Table II coefficients are scaled by baseline revenue at s-1; the three asterisk
significance levels (p < .01, .05, .10) from the paper are omitted here; revenue
and wage bill are significant at 1% for all-exposed and restaurants.

**Extensive-margin analysis.** The collapsed dataset aggregating firm counts by
cohort, year, treatment status, and industry is used for entry, exit, and total
active firm regressions (Section IV.E). Regressions are weighted by base-year
firm counts and scaled by pre-reform firm counts.

**Individual-level analyses (Section V).** Two panels constructed from IRS
administrative data:
- Low-earning workers: 2% random sample of individuals earning less than $20k
  in any industry in the year before the minimum wage increase (s-1) who were
  also earning less than $25k or not working in s-2.
- Young workers: 2% random sample of individuals ages 15-26 in the year before
  the minimum wage increase, regardless of employment status.

Individual employment outcomes use an LPM; earnings outcomes use the log of
annual individual income estimated via equation (2) with individual-cohort fixed
effects. Own-wage employment elasticities are estimated as the percent change in
employment divided by the percent change in average annual earnings, calculated
using the delta method.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| IRS administrative tax records (linked firm-worker-owner panel) | Business income tax returns (revenues, COGS, deductions, profits) + Form W-2 wage data linked to owners and workers; universe of U.S. pass-through firms 2010-2019 | No page yet |
| CPS Monthly Outgoing Rotation Groups (CPS MORGs) | Identifies highly exposed industries by share of hourly workers paid below the prevailing minimum wage, pre-reform | No page yet |

Sample: approximately 134,974 firms in highly exposed industries (balanced panel)
or 271,000 firms per year (full unbalanced panel); 2% random samples of
low-earning and young workers. Period: 2010-2019. Frequency: annual.

## When to read the full paper

Read the [original](https://doi.org/10.1093/qje/qjaf053) if you are:
designing or evaluating minimum wage policy for a setting that covers
independent businesses; extending the incidence decomposition to other
cost shocks (payroll taxes, mandated benefits); studying how the IRS
linked firm-worker-owner panel (detailed in Online Appendix L) was
constructed; or exploring the productivity selection mechanism in the
Cournot model of Online Appendix O. The locators above point to exact
tables and figures.

## Attribution and rights

Source: peer-reviewed, *The Quarterly Journal of Economics* 141(1), 2026.
This distillation was extracted by an LLM on 2026-06-28 and is
**not human-verified or independently reproduced**. The CC BY 4.0 licence
permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Rao, Nirupama L., and Max Risch.
> "Who's Afraid of the Minimum Wage? Measuring the Impacts on Independent
> Businesses Using Matched U.S. Tax Returns."
> *The Quarterly Journal of Economics* 141, no. 1 (2026): 373-427.
> DOI: 10.1093/qje/qjaf053. (c) The Author(s) 2025.
> Published by Oxford University Press on behalf of President and Fellows
> of Harvard College.
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Political Foundations of Racial Violence: Testa & Williams (2026)
# https://instituteforautomatedresearch.org/wiki/papers/qje/2026/testa-political-foundations-racial-violence-2026/
# Distilled: Using a regression discontinuity design on close presidential elections in the post-Reconstruction South (1880-1900), Testa and Williams show that a narrow Democratic county loss raised Black lynching probability by roughly 10 percentage points, while Democratic-aligned newspapers amplified anti-Black crime narratives after those losses, foreshadowing the vote-suppression machinery of Jim Crow. The Quarterly Journal of Economics 2026, paywalled. Eight core results with source locators, datasets used, the identification strategy, and estimating equations.
# Tags: paper-summary, political-economy, economic-history, elections
==============================================================================

**What this is.** The paper's core results, the conceptual framework it tests, the identification strategy, and the estimating equations: enough to know what it found and how, without reading all 62 pages. To replicate or extend, read the original at [doi.org/10.1093/qje/qjaf045](https://doi.org/10.1093/qje/qjaf045).

## TL;DR

Testa and Williams use a regression discontinuity design on county-level popular-vote shares in presidential elections across the post-Reconstruction South (1880-1900) to show that a narrow Democratic Party loss in a county raised the probability of a Black lynching in the following four years by about 10 percentage points, equivalent to an 80% increase over the control mean. No comparable effect exists for white lynchings, ruling out a general violence response. Southern newspapers aligned with the Democratic Party amplified anti-Black crime narratives in the aftermath of Democratic losses, providing the coordination mechanism through which elite-fomented racial antagonism translated into mob violence. These effects are concentrated in counties with all-white, all-Democratic local elites facing a large Black electorate, consistent with Blalock (1967)'s power threat hypothesis. The paper further shows that Black lynchings had a positive mediating effect on Democratic electoral success in the early twentieth century, echoing Jones, Troesken, and Walsh (2017) on lynching and Black political participation. Racial violence helped consolidate the Solid South by suppressing Black political participation and foreshadowing the de jure vote-suppression of Jim Crow.

## Core results

Magnitudes and significance are as reported; \*/\*\*/\*\*\* = 10%/5%/1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Democratic county loss raises Black lynching probability by ~10 pp** (80% over control mean); robust across polynomials, bandwidths, and spatial controls | Table II Panel A col 3, p. 755 | beta = 0.104\*\* (SE 0.041); control mean = 0.13; optimal bandwidth ~15 pp |
| R2 | **No comparable effect on white lynchings** (placebo); coefficients small and insignificant across all specifications | Table II Panel A cols 5-8, p. 755 | beta = -0.009 (SE 0.013) in col 7; indistinguishable from zero |
| R3 | **Effect nearly doubles in previously uncompetitive counties** (the election constitutes new information about local political strengths): beta = 18.8 pp | Table II Panel B col 3, p. 756 | beta = 0.188\*\*\* (SE 0.072); 703 obs |
| R4 | **Anti-Black crime accusations in Democratic newspapers rose 29-88%** after Democratic county losses; effect driven by Democrat-affiliated outlets | Table V cols 1-5 and 7-8, p. 772 | beta = 0.126\*-0.168\*\* (SE 0.073-0.089); control mean 0.19-0.20 pages per 100 |
| R5 | **Democratic newspapers actively reported county presidential losses** (39% increase); Democratic papers show 99% increase vs. near-zero for non-Democratic papers | Table VI cols 1, 3-4, p. 776 | beta = 0.043\*\*\* (SE 0.016) for any coverage; beta = 0.119\*\*\* (SE 0.020) for Democratic papers |
| R6 | **Effects concentrated in counties with all-white/Democrat elite and large Black constituency** -- consistent with the Blalock power-threat profile; near zero elsewhere | Table VII Panel A col 1 and col 5, p. 779 | beta = 0.127\*\*\* (SE 0.044) in Democrat/white-elite sample; beta = 0.117\*\*\* (SE 0.065) for large-Black-population subsample |
| R7 | **Black lynchings in 1880-1900 predict Democratic electoral victory in 1904-1912**, even conditioning on prior Democratic performance; correlation is suggestive of electoral reversal | Table IX cols 1-2, p. 784 | beta = 0.021\*\*\* (SE 0.006) on any Democratic win 1904-1912 |
| R8 | **Causal mediation confirms Black lynchings as a positive channel** for the Democratic electoral reversal; indirect effect is positive and significant; ~25% of total effect mediated via lynching | Table IX col 5, p. 784; Online Appendix Table E.2 | Indirect effect = 0.004\* (SE 0.002); direct effect = -0.013 (SE 0.018) |

**Overall (paper's conclusion).** The post-Reconstruction Democratic Party used racial violence, amplified through partisan newspapers, as a strategic tool to suppress Black political participation when legal disenfranchisement was unavailable. Close presidential election losses served as focal signals that credibly threatened Democratic hegemony; the ensuing lynching surge helped reverse Democratic electoral fortunes and prefigured the formal vote-suppression mechanisms of Jim Crow. These findings qualify the prevailing economic explanation of lynching, which emphasizes Black-white labor competition, by showing that political factors were primary.

## Theory / model

The paper proposes no formal structural model. Its conceptual framework (Section II.B, p. 743) draws on Blalock (1967)'s power threat hypothesis, which posits that a dominant group increases its use of social control measures against a minority as the minority's political power grows. In the post-Reconstruction South, lynching of Black people was plausibly an instrument for maintaining white Democratic hegemony after emancipation removed formal slavery and before Jim Crow provided legal disfranchisement tools.

The framework posits two specific mechanisms through which a Democratic electoral loss could galvanize racial violence:

**Informational channel (Section II.B, p. 745).** Local political actors use recent vote shares to assess the relative strengths of competing groups. When actors lack complete information, even a close Democratic loss can serve as a focal point for coordination among members of the pro-Black opposition (following Anagol and Fujiwara (2016) and Granzier, Pons, and Tricaud (2023)). Anticipating such mobilization, local Democratic elites have an incentive to mount a violent preemptive backlash. This mechanism predicts that effects should be stronger where the loss constitutes more novel information, that is, in counties where Democrats had previously won by comfortable margins.

**Elite strategy channel (Section II.B, p. 746 and Section IV.B, p. 768).** Democratic newspapers, which dominated the Southern press and reported on county-level presidential results, could operationalize racial hatred by publishing anti-Black crime accusations (rape, murder, robbery). Such accusations served as narrative pretexts that lowered the coordination costs of lynch mobs. Glaeser (2005) models the supply side of this process: elites supply hatred to mobilize voters, dividing poor-white and Black coalitions. Ottinger and Posch (2024) document a related dynamic in which Southern elites used newspapers to mobilize white voters against populist political threats. In this paper, newspapers are the supply chain for manufactured racial antagonism directed specifically at suppressing Black political participation.

The framework is tested empirically; no equilibrium condition or Euler equation is derived. The two channels generate testable predictions: larger effects in previously uncompetitive counties (informational channel), and positive newspaper-violence correlation driven by Democratic-affiliated press (elite strategy channel). Both are confirmed in the data.

## Method

The primary identification strategy is a sharp regression discontinuity design (RDD) exploiting the county-level popular-vote threshold for a Democratic loss (or win) in presidential elections. The method builds on `regression-discontinuity-design` for causal identification and `panel-regression` for the newspaper analysis.

**Main RD estimator.** Equation (1), p. 748:

$$
\text{Any Lynching}_{c(s)\tau} = \beta \cdot \text{Democratic Loss}_{c\tau} + f(\text{Loss Margin}_{c\tau}) + \phi_\tau + \theta_s + \mathbf{X}'_{c\tau} \boldsymbol{\Gamma} + \varepsilon_{c\tau}
$$

where $$\text{Any Lynching}_{c(s)\tau}$$ is a binary indicator for at least one Black (or white) lynching in county $$c$$ of state $$s$$ in the four-year window after presidential election $$\tau \in \{1880, 1884, 1888, 1892, 1896, 1900\}$$. $$\text{Democratic Loss}_{c\tau}$$ is a binary indicator for whether the Democratic presidential candidate lost the county popular vote. $$f(\text{Loss Margin}_{c\tau})$$ is a flexible running polynomial (linear in the main specification) in the Democratic vote-share loss margin. $$\phi_\tau$$ is an election-period fixed effect, $$\theta_s$$ is a state fixed effect, and $$\mathbf{X}_{c\tau}$$ is a vector of spatial controls including quadratic polynomials in county longitude and latitude.

The local average treatment effect (LATE) is identified under the assumption that counties where Democrats barely lost are comparable in all pretreatment characteristics to those where they barely won -- a condition supported by balance tests (Table I, pp. 752-753) and a McCrary (2008) density test (p-value 0.4; p. 750).

**Bandwidth selection.** MSE-optimal bandwidths are computed following Calonico, Cattaneo, and Titiunik (2014), restricting estimation to county-elections close to the Loss Margin = 0 threshold. The core Black-lynching result (Table II col 3) uses an optimal bandwidth of approximately 15 percentage points, yielding about 1,481 observations.

**Newspaper RD estimator.** Equation (2), p. 769:

$$
\% \text{Accusations}_{n(c)t(\tau)} = \beta \cdot \text{Democratic Loss}_{c\tau} + f(\text{Loss Margin}_{c\tau}) + \phi_\tau + \Upsilon_{t(\tau)} + \alpha_{\sigma(c)} + \varepsilon_{nt}
$$

where $$\% \text{Accusations}_{nt}$$ is the share of newspaper pages (per 100) in newspaper $$n$$ in year $$t$$ (within the four-year period following election $$\tau$$) that contain anti-Black crime accusation phrases ("negro rape," "negro murder," "negro robbery" and variants). $$\Upsilon_{t(\tau)}$$ is a year-within-election-cycle fixed effect, and $$\alpha_{\sigma(c)}$$ is a newspaper-city fixed effect.

**Causal mediation.** Table IX column 5 combines the RD variation with a structural mediation analysis to decompose the effect of Democratic losses on downstream Democratic electoral success (1904-1912) into a direct effect and an indirect effect through the Black lynching channel, following a local average structural equation approach.

## Empirical specifications

All specifications focus on the 11 former Confederate states (the "Solid South") over the 1880-1900 presidential election cycle period, covering elections in November of each election year.

**Main RD (R1, R2).** Equation (1) with linear running polynomial, election-period and state fixed effects, and spatial covariates. Outcome: indicator for any Black (or white) lynching in the four years after election $$\tau$$. Estimated at MSE-optimal bandwidth. Standard errors clustered at the county level (counties reclassified if boundaries changed between elections; see p. 751). Reported in Table II Panel A, p. 755.

**Uncompetitive-counties subsample (R3).** Equation (1) restricted to county-elections where $$|\text{Loss Margin}_{c,\tau-1}| > 16.2$$ (the median vote margin among sample Democratic losses), so the loss in $$\tau$$ is relatively novel information. Reported in Table II Panel B, p. 756. Split-sample p-value tests the null of equal coefficients across subsamples.

**Robustness suite (Table III, pp. 760-761):** alternative clusterings (county, county-decade, state-election-period); specifications omitting covariates, spatial controls, or lat/lon polynomials; county fixed effects; county-pair fixed effects based on geographic proximity; quadratic controls for 1880 Black population shares; bandwidth multipliers of 0.5x and 1.5x; quadratic, cubic, and quartic running polynomials. All yield Black-lynching estimates of 0.073-0.160, all significant at $$p < 0.10$$.

**Newspaper RD (R4, R5).** Equation (2) estimated on a balanced panel of Southern city newspapers from newspapers.com (as of June 2023), linked to their contemporaneous county. Outcomes: % pages with anti-Black crime accusations (Table V), and probability of any county election reporting (Table VI). Election-cycle year fixed effects and newspaper-city fixed effects absorb within-newspaper trends. Subsampled by newspaper partisan affiliation from Gentzkow et al. (2014) and Gentzkow, Shapiro, and Sinkinson (2014).

**Elite composition and power-threat subsamples (R6).** Equation (1) interacted with whether a county had a Democrat-only elite, a white-only elite, and an above-median Black population share (Table VII). The split-sample p-values test whether the Black-lynching effect differs significantly between the Blalock profile counties and the remainder.

**Jim Crow moderation (Table VIII).** Equation (1) and Equation (2) estimated separately for county-elections before and after a state enacted any Jim Crow ballot requirement or poll tax (timing from Jones, Troesken, and Walsh (2012)). The lynching effect is significant pre-Jim Crow (beta = 0.160\*\*\*) and near zero post-Jim Crow (beta = 0.010), consistent with racial violence substituting for legal disenfranchisement.

**Downstream electoral analysis (R7, R8).** OLS regressions of Democratic presidential victory in 1904, 1908, or 1912 on a binary indicator for any Black lynching in the county during 1880-1900 election periods (Table IX cols 1-2). Causal mediation in col 5 adapts the baseline RD by fixing the Democratic loss margin at zero, then separately estimating the direct and indirect (through lynching) paths to later Democratic victory.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Historic American Lynching (HAL) Project (Hines and Steelwater 2023) | County-level Black and white lynching indicator outcomes for all states except Texas and Virginia | [Project HAL](/wiki/datasets/project-hal/) |
| Seguin and Rigby (2019) lynching data | County-level lynching outcomes for Texas and Virginia to supplement HAL | no page yet |
| U.S. Decennial Census (1870, 1880, 1900, 1910) | County demographics (population density, Black population share), slaveholder shares, Confederate veteran shares | [Census](/wiki/datasets/census/) |
| Clubb, Flanigan, and Zingale (2006) ICPSR presidential and congressional election returns | County-level vote tabulations for presidential elections 1880-1900 and congressional elections; main explanatory variable | no page yet |
| newspapers.com full-text archive (as of June 10, 2023) | Anti-Black crime accusation rates (% pages) and county election-reporting rates in Southern city newspapers | no page yet |
| Gentzkow et al. (2014) and Gentzkow, Shapiro, and Sinkinson (2014) newspaper political affiliations | Partisan affiliation coding for Southern newspapers during the sample period | no page yet |
| Kestenbaum (2023) Political Graveyard | Partisan composition of local officeholders (mayors, postmasters) matched to counties | no page yet |
| Logan (2020) racial composition data | County racial composition of elected officials | no page yet |

Sample: 11 former Confederate states; presidential elections 1880-1900 (6 elections); lynching window through approximately 1904; newspaper panel annually 1880-1900; downstream electoral analysis through 1912.

## When to read the full paper

Read the [original](https://doi.org/10.1093/qje/qjaf045) if you are:
studying the political economy of racial violence, ethnic conflict, or elite-fomented social antagonism; working on the causal effects of electoral outcomes on social behavior beyond officeholding; extending the close-elections RD design to new social outcomes; or examining the historical origins of Jim Crow and Black disenfranchisement. The online appendix contains the McCrary density test, robustness tables (B1-E5), causal mediation details, and congressional-election extensions. Replication data are at [doi.org/10.7910/DVN/08YUBP](https://doi.org/10.7910/DVN/08YUBP).

## Attribution and rights

Source: peer-reviewed, *The Quarterly Journal of Economics* (2026), 733-794. This distillation was extracted by an LLM on 2026-06-28 and is **not human-verified or independently reproduced**. The paper is paywalled; the verbatim PDF is not hosted here.

> Testa, Patrick A., and Jhacova Williams. "Political Foundations of Racial Violence in the Post-Reconstruction South." *The Quarterly Journal of Economics* (2026), 733-794. DOI: 10.1093/qje/qjaf045. © The Author(s) 2025. Published by Oxford University Press on behalf of President and Fellows of Harvard College. All rights reserved. Reproduced here as a brief extract for research purposes only.


==============================================================================
# Financing Infrastructure in the Shadow of Expropriation: Acharya, Parlatore & Sundaresan (2025)
# https://instituteforautomatedresearch.org/wiki/papers/rfs/2025/acharya-financing-infrastructure-shadow-expropriation-2025/
# Distilled: A theory of optimal infrastructure financing under double moral hazard (private-sector operator shirking and government expropriation of project returns). The second-best contract features government guarantees to financiers, government coinvestment, development rights, and tax subsidies, matching observed practice in public-private partnerships. Review of Financial Studies 2025, paywalled. Seven core results with source locators, the model equations, and the method.
# Tags: paper-summary, infrastructure, public-private-partnerships, contract-theory
==============================================================================

**What this is.** The core results, model equations, and propositions from this theory paper
on optimal infrastructure financing under double moral hazard: enough to understand what the
contract looks like and why, without reading all 50 pages. To replicate or extend it, read
the full source at the [original](https://doi.org/10.1093/rfs/hhaf007).

## TL;DR

The paper develops a principal-agent model of public-private partnerships (PPPs) in which
two moral hazard problems interact: a private-sector operator shirks if not given sufficient
incentives, and a government expropriates project cash flows if its incentive to do so
outweighs the cost. This double moral hazard simultaneously limits the feasibility (extensive
margin) and scale (intensive margin) of infrastructure projects, explaining persistent
infrastructure gaps globally. The second-best optimal financing contract combines
(a) government guarantees to financiers against project failure to discipline the government's
expropriation incentive, (b) direct government coinvestment when the project return is
sufficiently high, (c) development rights to private parties, and (d) tax subsidies.
These features match institutional arrangements common in practice (TIFIA, UK Infrastructure
Bank, Hong Kong MTR, etc.). India evidence shows 72% of stressed coal-power-plant failures
are attributable to public moral hazard.

The paper relates to three literature strands. On government expropriation and growth, Myers
(1997) establishes the debt overhang idea that motivated sovereign debt dynamics; here it applies
to private infrastructure. On infrastructure financing specifically, Perotti (1995) shows
credible privatization as a commitment device, and Martimort and Sand-Zantman (2006) characterize
optimal delegated management contracts; neither features double moral hazard. On double moral
hazard contracting under agency problems, Repullo and Suarez (1998) study venture capital, and
Bolton and Dewatripont (2004) is a standard text reference. India evidence is consistent with
Lewis and Bajari (2014) (contractor moral hazard in Minnesota highway procurement) and Gardner
and Henry (2023) (government moral hazard limiting infrastructure investment globally).

## Core results

Magnitudes and statements are as reported in propositions and tables; locators point to the
source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Double moral hazard limits feasibility and scale jointly**; neither moral hazard alone causes both problems | Props. 1-3, pp. 1382-1385; Table 1, p. 1385 | Three feasibility thresholds: (i) if $$(R-A_o) < \underline{\Gamma}$$, project unfunded absent government MH; (ii) if $$\underline{\Gamma} \leq (R-A_o) < \overline{\Gamma}$$, project unfunded due to government MH; (iii) otherwise funded but at limited scale when $$(R-A_o) < \Gamma^*$$ |
| R2 | **Optimal financing contract requires government guarantees** $$K_g > 0$$ when $$(R-A_o) \leq \Gamma^*$$; guarantees decrease in project return net of operator agency rent | Prop. 4, pp. 1386-1387; Figure 4, p. 1388 | $$K_g^* = A_g + \underline{R} - (R-A_o)$$ for $$\overline{\Gamma} \leq (R-A_o) < \Gamma_R$$; decreasing in $$(R-A_o)$$, positive throughout the region |
| R3 | **Pecking order: guarantees precede coinvestment**; government coinvestment $$I_g > 0$$ only when project return is sufficiently high relative to moral hazard severity | Prop. 4, pp. 1386-1387; Figure 4, p. 1388 | $$I_g^* = 0$$ for $$(R-A_o) < \Gamma_I$$; $$I_g^* > 0$$ and rising for $$\Gamma_I \leq (R-A_o) < \Gamma^*$$; maximal $$I_g^* = \bar{K}_0$$ at $$\Gamma_R$$ |
| R4 | **Higher government benefit from expropriating** ($$C$$) requires larger guarantees and a lower coupon; expropriation risk is worse in high-$$C$$ (developing-economy) settings | Prop. 5, p. 1389 | $$\partial K_g^*/\partial C > 0$$ and $$\partial R_f^*/\partial C \leq 0$$; developing economies need higher guarantees but their fiscal limits make these harder to provide |
| R5 | **Government fiscal resources increase scale but not feasibility**; resources available at the investment stage (date 0) weakly dominate those at the cash flow stage (date 1) | Props. 6-7, pp. 1389-1390 | $$\partial I^*/\partial \bar{K}_0 \geq \partial I^*/\partial \bar{K}_1 \geq 0$$; feasibility thresholds $$\overline{\Gamma}$$ and $$\underline{\Gamma}$$ independent of $$\bar{K}_0, \bar{K}_1$$ (Prop. 6a: $$\partial\overline{\Gamma}/\partial\bar{K}_0 = \partial\overline{\Gamma}/\partial\bar{K}_1 = 0$$) |
| R6 | **Development rights and tax subsidies** improve feasibility and scale; it is always optimal to set the tax-sharing rate $$\tau = 1$$ (fully share tax revenue with private sector) | Prop. 10-11, pp. 1395-1397 | Development rights $$D$$ and externalities $$X$$ raise $$\overline{\Gamma}$$ and $$I^*$$; distribution of rights to operator vs. financiers depends on which constraint is binding |
| R7 | **India evidence confirms double moral hazard**: 72% of stressed thermal power plant failures 2007-2011 due to public (government) moral hazard, 22% private, 6% unclassified | Table 5, p. 1400; Tables 2-3, p. 1398 | 34 stressed plants; 1,165 NHAI cases (40% of 2,912 highway disputes); government as petitioner in 123 of 139 arbitration-related cases |

**Overall (paper's conclusion).** The double moral hazard problem jointly limits the feasibility
and scale of privately financed infrastructure, explaining global infrastructure gaps. The
second-best optimal contract features government guarantees, coinvestment, development rights,
and tax subsidies in a pecking order determined by project return relative to agency rents.
These features are prevalent in practice. Third-party multilateral guarantees do not substitute
for government guarantees because they do not discipline the government's expropriation
incentive; building fiscal capacity and governance is more effective for closing infrastructure
gaps in developing economies.

## Theory / model

The model has three players: (a) the **government**, (b) a private **operator** who builds and
manages the project, and (c) private **financiers**. An infrastructure project has constant
returns to scale up to a maximum scale $$\bar{I}$$; total investment $$I = I_f + I_g$$ where
$$I_f$$ is from financiers and $$I_g$$ from the government. The per-unit payoff is $$R > 1$$ if
successful and zero otherwise. The model runs over four stages (Section 1, p. 1375, Figure 2
p. 1376): an **investment** stage (financial contract set), a **gestation** stage (operator
appointed, operational contract set), an **operating** stage (operator exerts effort), and a
**cash flow** stage (payoffs distributed).

**Operator's moral hazard.** Following Holmstrom and Tirole (1998), the operator exerts
high or low effort (p. 1376). High effort yields success probability $$p_h \in (0,1)$$, low effort
$$p_l < p_h$$. Let $$\Delta p \equiv p_h - p_l$$. Low effort gives the operator a private nonpecuniary
benefit $$BI$$ as in Jensen and Meckling (1976). If the project succeeds the operator receives $$R_o I$$;
if it fails the operator receives zero. The operator's incentive compatibility constraint (ICO)
is (p. 1379):

$$
p_h R_o I \geq p_l R_o I + BI \tag{ICO}
$$

Rearranging gives the familiar Holmstrom-Tirole (1998) condition: the operator requires an
agency rent of at least

$$
R_o \geq A_o \equiv \frac{B}{\Delta p}
$$

**Government's moral hazard.** The government can expropriate the operator's cash flows
by setting user fees below the contractual level (p. 1376). If the government expropriates, it
receives a net benefit $$CI$$ (with $$C > 0$$ in economies with weak institutions, $$C < 0$$
otherwise). The government will agree not to expropriate $$A_o$$ if its incentive compatibility
constraint (ICG) holds (p. 1380):

$$
p_h R_g - (1-p_h)K_g \geq p_l (R - R_f + C) - (1-p_l)K_g \tag{ICG}
$$

Rewriting:

$$
K_g + R_g \geq A_g \equiv \frac{p_l(A_o + C)}{\Delta p}
$$

where $$A_g$$ is the government's agency rent. Since $$C > -A_o$$ and $$p_l > 0$$, Assumption 1c
($$C > -B/\Delta p$$) ensures $$A_g > 0$$: double moral hazard is present whenever $$B > 0$$.

**Key structural insight (Section 2.2.3, p. 1383).** The government agency rent $$A_g$$ is
increasing in the operator agency rent $$A_o$$: the two moral hazards are intertwined. A higher
$$A_o$$ means the operator needs more cash flows in the success state, leaving less for the
government and sharpening the government's expropriation incentive.

**Payoff structure (Figure 3, p. 1378).** If the project succeeds (probability $$p$$):
financiers receive $$R_f I$$, operator $$R_o I$$, government $$R_g I \equiv (R - R_f - R_o) I$$.
If it fails (probability $$1-p$$): financiers receive guarantee $$K_g I$$, operator zero,
government $$-K_g I$$. The government has fiscal resources $$\bar{K}_0$$ at date 0 and $$\bar{K}_1$$
at date 1.

## Method

This is a pure theory paper; the method is constrained optimization (mechanism design).
The planner maximizes the net present value of the infrastructure project subject to incentive
compatibility constraints for the operator (ICO) and the government (ICG), individual
rationality constraints for the financiers (IRF) and the government (IRG), and no-default
(fiscal) constraints on government guarantees (NDK) and promised returns (NDR). The full
program is (Section 2.3, p. 1384):

$$
\max_{I_g \in [0,\bar{K}_0],\, I_f \geq 0,\, K_g \geq 0,\, R_f \geq \underline{R}} (p_h R - r)(I_g + I_f)
$$

subject to:

$$
(1-p_h)K_g + p_h R_f \geq r \frac{I_f}{I_f + I_g} \tag{IRF}
$$

$$
(1-p_h)K_g + p_h R_f + r \frac{I_g}{I_f + I_g} \leq p_h(R - A_o) \tag{IRG}
$$

$$
K_g + (R - A_o - R_f) \geq A_g \tag{ICG}
$$

$$
K_g(I_f + I_g) \leq \bar{K}_1 + \bar{K}_0 - I_g \tag{NDK}
$$

$$
R_f(I_f + I_g) \leq (R - A_o)(I_f + I_g) + \bar{K}_1 + \bar{K}_0 - I_g \tag{NDR}
$$

$$
I_g + I_f \leq \bar{I} \tag{MS}
$$

The total project scale is determined by the government guarantee via the NDK constraint:

$$
I = \min\!\left\{ \frac{\bar{K}_1 + \bar{K}_0 - I_g}{K_g},\; \bar{I} \right\}
$$

The paper proves eleven propositions by characterizing the optimal contract in three cases
depending on which constraints bind (Appendix A.2, pp. 1406-1408): (Case 1) ICG binds and
IRF is slack; (Case 2) ICG and IRF bind simultaneously; (Case 3) project is at maximal scale.
The full optimal contract formulas are given in equations (A1)-(A4) (pp. 1408-1409).

Extensions in Section 3 solve the program with: (i) costly state verification microfounding
$$\underline{R}$$ (Section 3.1); (ii) government's limited commitment to financiers with
default penalty $$\Phi$$ (Section 3.2, Prop. 8); (iii) random contract enforcement
(probability $$\delta$$ of enforcement, Section 3.3); (iv) private-party coinvestment by the
operator (Section 3.4, Prop. 9); and (v) development rights $$D$$ and government externalities
$$X$$ (Section 3.5, Props. 10-11). The extension with limited commitment and externalities
solves the program (Appendix B, p. 1410):

$$
\max_{I_g \geq 0,\, I_f \geq 0,\, K_g \geq 0,\, R_f \geq \underline{R},\, D_f \in [0,D]} (p_h(R+X+D) - r)(I_g + I_f)
$$

subject to modified IRF, IRG, ICG, NDK-LC, NDR-LC, and MS constraints, yielding a maximum
feasible project scale:

$$
\bar{I} = \frac{r\bar{K}_0 + (1-p_h)\min\!\{\Phi, \bar{K}_1\} + p_h\Phi}{(r - p_h D)} \tag{Eq. 1}
$$

## Empirical specifications

This paper has no econometric specifications. The empirical content is a structured case
study of two Indian infrastructure sectors serving as illustration of double moral hazard.

**Highway contracting in India (Section 4.1.1, pp. 1397-1399).** Data from Mehta and
Thomas (2022) on 2,912 highway contractual disputes 2007-2020 involving the National Highway
Authority of India (NHAI). Tables 2-3 (p. 1398) classify disputes by project lifecycle phase
and litigation driver. Key finding: 66% of disputes arise in the postaward/construction phase
(consistent with operating-stage moral hazard); NHAI is the petitioner in 123 of 139
arbitration-related cases and the defendant in 75 of 90 payment-related cases, consistent with
government moral hazard in payments.

**Stressed thermal power plants in India (Section 4.1.2, pp. 1398-1400).** Hand-collected
data on 34 stressed coal-fueled power plants initiated 2007-2011, sourced from monthly Broad
Status Reports of the Central Electricity Authority. Table 4 (p. 1399) shows 94% of failures
occur in the postaward/construction phase. Table 5 (p. 1400) classifies each plant's cause of
failure into private moral hazard (22%) and public moral hazard (72%), with 6% unclassified.
No regression or econometric test is performed; the classification is based on qualitative
case-by-case analysis (Appendix B1).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| NHAI litigation data (Mehta and Thomas 2022) | Illustrative evidence of double moral hazard in highway contracting | No page yet |
| Central Electricity Authority Broad Status Reports (hand-collected) | 34 stressed coal-fueled power plants; cause-of-failure classification | No page yet |

The paper introduces the 34-plant hand-collected dataset for the Indian power sector stress
episode (2007-2015). No quantitative empirical estimation is performed on either dataset.

## When to read the full paper

Read the [original](https://doi.org/10.1093/rfs/hhaf007) if you are: designing or evaluating
government guarantee programs for infrastructure; studying PPP contract design in settings
with weak institutions; extending the model to stochastic government expropriation or
multilateral guarantors; or seeking the formal proofs of Propositions 1-11 and the full
characterization of the optimal contract (Appendix A-B, pp. 1404-1411).

## Attribution and rights

Source: peer-reviewed, *The Review of Financial Studies* 38(5), 2025. Published by Oxford
University Press on behalf of the Society for Financial Studies. All rights reserved; paywalled.
This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or
independently reproduced**. Extract-only; no PDF hosted.

> Acharya, Viral V., Cecilia Parlatore, and Suresh Sundaresan.
> "Financing Infrastructure in the Shadow of Expropriation."
> *The Review of Financial Studies* 38, no. 5 (2025): 1368-1418.
> DOI: 10.1093/rfs/hhaf007. © The Author(s) 2025. Published by Oxford University Press.


==============================================================================
# Build or Buy? Human Capital and Corporate Diversification: Beaumont, Hebert & Lyonnet (2025)
# https://instituteforautomatedresearch.org/wiki/papers/rfs/2025/beaumont-build-buy-human-capital-2025/
# Distilled: Using French administrative data, this paper shows that firms enter distant sectors by acquisition (buy) rather than organically (build) because building human capital in distant sectors requires costly organizational integration of new workers. Review of Financial Studies 2025, CC BY 4.0. Seven core results with source locators, datasets used, the identification design (shift-share IV), and the estimating equations.
# Tags: paper-summary, corporate-diversification, human-capital, mergers-acquisitions
==============================================================================

**What this is.** The paper's core results, the human capital distance measure it introduces, and the
causal identification design (shift-share IV): enough to know what it found and how, without reading
all 35 pages. To replicate or extend it, read the full source at the
[original](https://doi.org/10.1093/rfs/hhaf004).

## TL;DR

Using exhaustive French administrative data on all firms entering new sectors from 2003 to 2014,
the paper shows that firms acquire incumbents (buy) rather than develop resources organically
(build) when their existing human capital is far from the target sector. The key friction is
integration cost: building human capital requires not just hiring workers with the right skills
but also investing organizational capacity to train, manage, and assign them to tasks. Firms with
limited organizational skills face particularly high build costs and are the primary drivers of the
result. Post-entry, firms that build in distant sectors underperform for at least three years;
firms that buy are unaffected by the human capital distance to the entry sector, consistent with
acquiring already-operational human capital that avoids integration costs.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*`/`\*\*\*` = 10%/5%/1%. Locators point
into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | HC distance between a firm and the entry sector is positively associated with the probability of buying (OLS) | Table 3, col. 1, p. 1345 | +0.39 pp per 1-SD increase in HC distance; ~20% of the unconditional buy probability; sig. at 5% |
| R2 | HC Bartik instrument confirms causal effect of HC distance on buy probability | Table 4, col. 1, p. 1350 | +0.26 pp per 1-SD HC Bartik; ~13% of unconditional buy probability; sig. at 5% |
| R3 | Firms that build hire more workers in entry-sector occupations; firms that buy downsize non-key occupations | Table 5, Panel A col. 2 and Panel B col. 2, p. 1352 | Build: +0.532\*\*\* in top-5 occupations per 1-SD HC distance; Buy: 0.045 (insignificant); difference sig. at 1% |
| R4 | Build firms hire from external labor markets; buy firms do not assimilate target workforce | Table 6, col. 3, p. 1354 | Build: +0.5 pp external hires per 1-SD HC distance (sig. at 5%); Buy: -4.7 pp; difference sig. at 1% |
| R5 | The build-or-buy HC distance effect is driven by firms with low organizational skills | Table 8, p. 1357 | Low top-layers: coef = 0.013\*\*\*; high top-layers: coef = 0.018 (insig.); similarly for SG&A and HR presence |
| R6 | Firms that build in distant sectors invest more in organizational skills; buy firms do not | Table 9, col. 1 and col. 3, p. 1358 | Build: +6.6 pp SG&A wage share (sig. at 1%), +0.369 pp top-layers share (sig. at 1%) per 1-SD HC distance; Buy: insig. |
| R7 | Build firms in distant sectors underperform for at least 3 years post-entry; buy firms are unaffected by HC distance | Table 11, Panel A col. 1, p. 1362 | Build: coef -1.160\*\*\* on HC distance in log(Sales) (~23% lower per 1-SD); Buy: +0.037 (insig.); survival rate build -0.166\*\*\*; difference in sales sig. at 1% |

**Overall (paper's conclusion).** Firms buy primarily when they cannot build: 98% of diversifying
entries are organic (build), and buy entries are concentrated in distant sectors. Building human
capital is costly because it requires organizational skills to integrate new workers, not just the
right hires. Firms that buy avoid these costs and reach profitability faster in new sectors.
Organizational skills are therefore non-rival: they allow firms to profitably enter multiple
sectors.

## Theory / model

The paper has no formal theoretical model but builds on two testable hypotheses derived from the
organizational economics literature (Prescott and Visscher (1980)). The framing of human capital
as a combination of occupation-specific skills follows Lazear (2009). The diversification
direction literature finds firms enter sectors requiring similar resources (Neffke and Henning (2013);
Hoberg and Phillips (2018); Boehm, Dhingra, and Morrow (2022); Tate and Yang (2024)).
The M&A motive of acquiring human capital is developed by Ouimet and Zarutskie (2020).

**Hypothesis 1 (build vs buy):** If building human capital requires integrating new workers
(training, managing, assigning to tasks), the cost of building increases with the level of
required human capital investment, proxied by HC distance to the entry sector. Therefore, firms
should be more likely to buy in distant sectors and more likely to build in close sectors.

**Hypothesis 2 (organizational skills):** Firms with limited organizational skills face higher
integration costs, so the HC distance effect on buy probability should be stronger for
low-organizational-skill firms. Moreover, firms that build in distant sectors should invest more
in organizational skills to reduce integration costs.

The paper defines the HC distance between firm $$g$$ and sector $$n$$ as one minus the cosine
similarity between the firm's occupation wage-share vector and the incumbent representative firm's
occupation wage-share vector (pp. 1340-1341):

$$
\text{HC distance}_{g,n} = 1 - \frac{\sum_i s_{g,i} \cdot s_{n,i}}{\sqrt{\sum_i s_{g,i}^2} \sqrt{\sum_i s_{n,i}^2}}, \tag{--}
$$

where $$s_{g,i}$$ is the share of firm $$g$$'s wage bill going to workers in occupation $$i$$, and
$$s_{n,i}$$ is the share of the sector $$n$$ wage bill in occupation $$i$$, computed by consolidating
all single-sector firms in the sector. HC distance ranges from 0 (identical workforce composition)
to 1 (no overlapping occupations). The paper uses 414 four-digit PCS-ESE occupation codes from
the French matched employer-employee data set (DADS, p. 1340).

## Method

The primary empirical tool is an OLS linear probability model of the build-or-buy decision, with
the shift-share IV (HC Bartik) used to establish causality. The paper exploits the richness of
the French employer-employee data to construct firm-level occupation vectors unavailable in
standard M&A databases.

**OLS baseline** (eq. 1, p. 1344). The main specification compares firms in the same sector of
origin that enter the same new sector in the same year:

$$
\mathbf{1}(\text{Buy})_{g,n,t} = \lambda_{n,o,t} + \delta \cdot \text{HC distance}_{g,n,t-1} + \beta X_{g,n,t-1} + \varepsilon_{g,n,t}, \tag{1}
$$

where $$\mathbf{1}(\text{Buy})_{g,n,t}$$ is one if firm $$g$$ acquires an incumbent in sector $$n$$ at
year $$t$$ and zero if it builds. The fixed effects $$\lambda_{n,o,t}$$ are sector of origin $$o$$
times entry sector $$n$$ times year $$t$$ fixed effects, absorbing all sector-pair-year factors
(synergies, barriers to entry, demand shocks). Controls $$X_{g,n,t-1}$$ include firm size, value
added, number of occupations, cash holdings, tangible assets, and total wages, all scaled by
workers. Standard errors are double-clustered at the sector of origin and sector of entry.

**Shift-share instrument (HC Bartik)** (eq. 2-3, p. 1348). To address omitted time-varying
firm factors, the paper constructs an instrument that captures variation in HC distance due to
changes in incumbent firms' human capital, holding firms' workforce composition fixed at their
2003 values:

$$
\text{HC Bartik}_{g,n} = \sum_i \hat{s}_{g,i,03} \cdot \Delta \hat{s}_{n,i,03,11}, \tag{2}
$$

where $$\Delta \hat{s}_{n,i,03,11} = \hat{s}_{n,i,11} - \hat{s}_{n,i,03}$$ is the change in the
normalized occupation share of sector $$n$$ from 2003 to 2011. This is equivalently:

$$
\text{HC Bartik}_{g,n} = -(\text{HC distance}^{03}_{g,n,11} - \text{HC distance}^{03}_{g,n,03}), \tag{3}
$$

so a high HC Bartik value means sector $$n$$ has become less distant from firm $$g$$ over 2003-2011.
Identification follows Borusyak, Hull, and Jaravel (2022): the shocks (occupation-level workforce
changes $$\Delta \hat{s}_{n,i,03,11}$$) are the source of variation and the shares ($$\hat{s}_{g,i,03}$$)
are the weights. The identification differs from the exposure-share approach of
Goldsmith-Pinkham, Sorkin, and Swift (2020). The identification assumption is that long-term
changes in incumbents' workforce composition are uncorrelated with firm-level unobservables.
The paper validates this assumption by regressing shocks on potential observable confounders
following Xu (2022).

**Event study around entry** (eq. 5-6, pp. 1351, 1361). For employment and performance outcomes,
the paper estimates dynamic regressions from $$t-4$$ to $$t+3$$ around entry, separately for firms
above and below the HC distance median, plotting coefficient paths to show the timing and
persistence of effects. Specifications include sector-of-origin, sector-of-entry, and year fixed
effects; standard errors are double-clustered.

## Empirical specifications

**Main build-or-buy regression (R1, R2).** The OLS specification (eq. 1) is estimated on 61,228
firm-sector-year observations from 2005 to 2014 (firms with at least 20 workers and at least 1%
of entry-year sales in the new sector). Table 3 reports four specifications: baseline sector-triplet
FE (col. 1), adding size-quartile FE (col. 2), adding main-sector FE to account for synergies
(col. 3), and adding firm FE to exploit within-firm variation across sectors (col. 4). The HC
distance coefficient is positive and significant at 5% across all four. The IV specification
(Table 4) is estimated on the 2011-2014 subsample (after the 2003-2011 shock period); Column 1
uses sector-of-origin by entry FE; column 2 adds sector-of-origin by entry by year FE.

**Employment decomposition (R3, R4).** Table 5 regresses 3-year employment growth ($$t-1$$ to $$t+3$$)
and its decomposition into top-5-occupation growth and other-occupation growth on HC distance,
separately for build (Panel A) and buy (Panel B) entries surviving 3 years post-entry. Table 6
further decomposes total employment growth into internal flows (from other subsidiaries) and
external flows (from outside the firm), revealing that builds rely exclusively on external hiring.
All specifications include sector-of-origin, sector-of-entry, and year FE; errors double-clustered.

**Labor market tightness and contract type (channels for R1).** Table 7 splits the main
regression by local labor market (LLM) tightness tercile (Panel A, 2010-2014 subsample) and by
fraction of permanent contracts tercile (Panel B). The HC distance effect on buy probability is
significant only in the tightest LLM tercile (coef = 0.023\*\*) and in the highest permanent-contract
tercile (coef = 0.024\*), consistent with higher build costs when hiring is harder.

**Organizational skills heterogeneity (R5, R6).** Table 8 splits the sample by three proxies for
organizational skills: (i) fraction of top-layer managers in the wage bill (Caliendo, Monte, and
Rossi-Hansberg (2015)), (ii) SG&A-related occupations wage share, and (iii) presence of any HR
worker. The HC distance effect is significant only for the lowest-skill tercile in all three
measures. Table 9 regresses the change in SG&A wage share, top-layers share, and HR adoption
between $$t-1$$ and $$t$$ on HC distance, separately for builds (Panel A) and buys (Panel B), finding
investment only among build entries.

**Post-entry performance (R7).** Table 11 regresses log(Sales), sales growth rate, survival
indicator, change in operating income per worker, and return on assets at $$t+3$$ on HC distance
for build entries (Panel A) and buy entries (Panel B). Table 12 adds an interaction between HC
distance and organizational skill indicators for build entries, finding that high-organizational-
skill firms suffer 20% lower sales vs 30% for low-skill firms in distant sectors (coefficient
on interaction HC distance x High SG&A = 0.245\*\*, Table 12 col. 1, p. 1363).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| French matched employer-employee data (DADS - Declarations Annuelles des Donnees Sociales) | Firm-level workforce composition by 4-digit occupation code (414 codes); used to compute HC distance | No page yet |
| French ownership links data (LIFI - Enquete sur les Liaisons financieres entre societes) | Identifies business group structure and subsidiaries; links firms to M&A targets | No page yet |
| French tax files (BIC - Benefices Industriels et Commerciaux) | Balance sheets and income statements at subsidiary level; firm-level controls | No page yet |
| French sales breakdown data (EAE/VAC - Enquete Annuelle des Entreprises / Ventilation des Ventes par Activite) | Identifies firm entries into new sectors (sales by sector, 3-digit SIC); sample period 2003-2014 | No page yet |
| SDC Platinum (M&A deals) | M&A transaction data (acquirer, target, deal date, ownership stakes) for buy entries 2003-2014 | [SDC Platinum](/wiki/commercial/sdc-platinum/) (licensed) |
| Bureau van Dijk Zephyr (M&A deals) | Supplementary M&A data matched with French administrative records; 7,165 M&A deals, 4,139 acquirers | No page yet |
| Pole emploi (French unemployment agency) occupational tightness | Local labor market (LLM) tightness measure by occupation; 348 LLMs, used for Table 7 robustness | No page yet |

Sample: 61,228 firm-sector-year entries from 2005 to 2014 (127,185 build; 2,415 buy). Administrative
data covers all French firms with at least 20 workers from 2003 to 2014.

## When to read the full paper

Use the [original](https://doi.org/10.1093/rfs/hhaf004) if you are:
studying the determinants of the build-versus-buy choice in diversification; measuring human
capital distance between firms and sectors; replicating the shift-share IV (HC Bartik) design
for workforce composition shocks; understanding how organizational skills affect labor integration
costs in M&As; or extending the analysis to non-French settings. The locators above point to the
exact tables.

## Attribution and rights

Source: peer-reviewed, *The Review of Financial Studies* 38(5), 2025. This distillation was
extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**.
The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Beaumont, Paul, Camille Hebert, and Victor Lyonnet.
> "Build or Buy? Human Capital and Corporate Diversification."
> *The Review of Financial Studies* 38, no. 5 (2025): 1333-1367.
> DOI: 10.1093/rfs/hhaf004. (C) 2025 The Author(s).
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Bail-Ins, Optimal Regulation, and Crisis Resolution: Clayton & Schaab (2025)
# https://instituteforautomatedresearch.org/wiki/papers/rfs/2025/clayton-bail-ins-optimal-regulation-2025/
# Distilled: In a tractable three-period dynamic contracting model with fire-sale externalities, the privately optimal bank contract combines short-term standard debt and long-term bail-in debt; the social optimum calls for joint regulation of both the level and composition of debt, rationalizing a leverage cap plus a TLAC requirement that can be satisfied with bail-in debt. Bail-ins replace bailouts as a recapitalization tool even without planner commitment. Review of Financial Studies 2025, paywalled. Five core results with source locators, the model, and its key propositions with equations.
# Tags: paper-summary, banking, financial-regulation, capital-structure, bail-ins
==============================================================================

**What this is.** This is a distilled skeleton of the paper for search and navigation. Read the
[original at Oxford Academic](https://doi.org/10.1093/rfs/hhaf002) to replicate or extend.

## TL;DR

Clayton and Schaab build a tractable three-period dynamic contracting model of bank liability
structure in the tradition of Innes (1990). Banks face an initial and a continuation monitoring
incentive problem. In the presence of fire sales from liquidations, the privately optimal contract
combines short-term standard debt (which forces liquidation in bad states and provides strong
incentives) and long-term bail-in debt (which avoids resource costs of liquidation by writing
down to pledgeable income). A social planner that internalizes the fire-sale externality
intervenes in both the level and the composition of debt: it prefers less standard debt (a
leverage cap / maximum leverage requirement) and less total debt (a TLAC requirement satisfiable
with bail-in debt). The model shows that bail-ins replace bailouts as a recapitalization tool
and that statutory provisions increasing the cost of bailouts improve welfare, providing a unified
rationalization of postcrisis regulation. The framework connects to demand-based explanations of
standard debt (Bolton and Oehmke (2019); Walther and White (2020)) and extends the macroprudential
literature on pecuniary externalities from Davila and Korinek (2018) to the bank contracting
setting.

## Core results

| # | Result | Locator | Magnitude as reported |
|---|---|---|---|
| R1 | Privately optimal contract uses both standard and bail-in debt | Proposition 1, Corollary 1, pp. 2822-2825 | Liability structure has three regions: liquidation ($$R_1 \le R_\ell^p$$), bail-in write-down ($$R_\ell^p < R_1 \le R_u^p$$), and no write-down ($$R_1 > R_u^p$$); implemented with short-term standard debt face value $$(1-b)R_\ell^p Y_0$$ and long-term bail-in debt face value $$(1-b)(R_u^p - R_\ell^p)Y_0$$ |
| R2 | Social optimum has same structure as private optimum but with additional wedges on standard and total debt | Proposition 3, Equations (24)-(28), pp. 2827-2829 | Socially optimal thresholds $$R_\ell^s \le R_\ell^p$$ and $$R_u^s \le R_u^p$$: planner uses less standard debt and less total debt than private banks; wedges $$\tau_\ell^s \ge 0$$ and $$\tau_u^s \ge 0$$ reflect social cost of liquidations $$\lambda^s \sigma \gamma^s$$ |
| R3 | All three model ingredients (initial incentive problem, continuation incentive problem, costly liquidation) are needed for bail-in debt to be part of the optimal contract | Proposition 2, p. 2826 | If $$B_0(e_0) = 0$$: bail-in debt alone suffices. If $$B_1 = 0$$: long-term debt alone suffices. If $$\gamma = 1$$: standard debt alone suffices. All three ingredients are needed for the combined structure |
| R4 | With planner commitment over bailouts, bail-ins dominate bailouts: the socially optimal contract with no bailouts is Pareto efficient | Proposition 4, p. 2832 | No bailouts ($$T_0 = T_1 = 0$$) is Pareto efficient; bailouts are redundant recapitalization when bail-ins are available; both instruments can achieve same state contingencies in bank debt contracts |
| R5 | Without planner commitment, Pareto-efficient debt levels still eliminate bailouts entirely; welfare is increasing in the cost of bailouts F | Proposition 5, p. 2834 | For any $$F \ge 0$$, Pareto-efficient debt levels $$(R_\ell^s, R_u^s)$$ result in no banks being bailed out; welfare is strictly increasing in F because higher F relaxes the no-bailout constraint (Equation 29) |

**Overall (paper's conclusion).** The paper provides a single contracting framework that
rationalizes both a maximum leverage requirement and a TLAC requirement as jointly optimal
responses to fire sale externalities. Bail-in debt is the instrument of choice because it
combines the incentive properties of equity with the cash-flow transfer properties of standard
debt, making it superior to outside equity as a loss-absorbing instrument. Bail-ins replace
bailouts in the regulatory toolkit; statutory provisions that increase the cost of engaging
in bailouts complement bail-in regulation by allowing looser regulatory constraints while
still preventing bailouts in equilibrium.

## Theory / model

The model has three periods ($$t = 0, 1, 2$$), a unit continuum of banks, penniless investors,
and arbitrageurs. Banks invest in a firm of variable scale $$Y_0 = A_0 + I_0 > 0$$ using their
own equity $$A_0 > 0$$ and investor funds $$I_0 \ge 0$$ (p. 2815).

**Project quality shocks.** At each of dates 1 and 2, the bank experiences a stochastic
quality shock $$R_t \in [\underline{R}, \overline{R}]$$ that adjusts project scale to
$$Y_t = R_t Y_{t-1}$$, so final scale is $$Y_2 = R_1 R_2 Y_0$$. The project pays one unit
of the consumption good per unit of final scale at date 2 but yields no dividend on date 1.
Shocks $$R_t$$ are independent and idiosyncratic with densities (p. 2815):

$$f_t(R_t | e_{t-1}) = e_{t-1} f_{tH}(R_t) + (1 - e_{t-1}) f_{tL}(R_t).$$

**Effort and MLRP.** Date 0 effort is continuous, $$e_0 \in [0,1]$$; date 1 effort is binary,
$$e_1 \in \{0,1\}$$. Higher effort increases the quality of the return distribution.
The paper assumes the monotone likelihood ratio property (MLRP): the likelihood ratio
$$\Lambda_t(R_t) \equiv f_{tH}(R_t) / f_{tL}(R_t)$$ is increasing in $$R_t$$.

**Private benefits.** The banker's date 0 private benefit is $$B_0(e_0) Y_0$$, where
$$B_0$$ is decreasing and concave with $$B_0(1) = 0$$. The banker's date 1 private
benefit is $$(1 - e_1) B_1 Y_1$$ for $$0 < B_1 < 1$$ (p. 2816).

**Resource constraints.** Along a history $$(R_1, R_2)$$, investor repayment $$x_1(R_1)$$,
$$x_2(R_1, R_2)$$ and bank consumption $$c_1(R_1)$$, $$c_2(R_1, R_2)$$ satisfy (p. 2817):

$$c_1(R_1) + x_1(R_1) = \alpha(R_1) \gamma R_1 Y_0, \tag{1}$$

$$c_2(R_1, R_2) + x_2(R_1, R_2) = (1 - \alpha(R_1)) R_1 R_2 Y_0. \tag{2}$$

Limited liability requires $$c_1(R_1), c_2(R_1, R_2) \ge 0$$.

**Investor participation constraint.** Investors must at least break even (p. 2818):

$$Y_0 - A_0 \le \mathbb{E}\bigl[x(R_1) \mid e_0 = e_0^*\bigr]. \tag{6}$$

**Fire sale and liquidation price.** A representative arbitrageur with borrowing constraints
generates a fire sale externality. The equilibrium liquidation price $$\gamma$$ satisfies the
market-clearing condition (pp. 2820-2821):

$$\gamma(\Omega) = \frac{\partial \mathcal{F}(\Omega)}{\partial \Omega}, \quad \Omega = \int_{\underline{R}}^{\overline{R}} \alpha(R_1) R_1 f_1(R_1 | e_0^*) \, dR_1. \tag{16}$$

When $$\partial^2 \mathcal{F} / \partial \Omega^2 < 0$$, more liquidations reduce the price,
creating the fire sale. The liquidation price elasticity is
$$\sigma = -(\Omega / \gamma)(\partial \gamma / \partial \Omega)$$.

**Date 1 incentive compatibility.** High effort $$e_1^*(R_1) = 1$$ is incentive compatible
if (p. 2818):

$$\mathbb{E}[c_2(R_1, R_2)(\Lambda_2(R_2) - 1) \mid e_1 = 0] \ge B_1 R_1 Y_0. \tag{9}$$

**Date 0 optimal effort.** The bank's optimal date 0 effort $$e_0^*$$ satisfies (p. 2819):

$$-B_0'(e_0^*) Y_0 = \mathbb{E}_0[c(R_1)(\Lambda_1(R_1) - 1) \mid e_0 = 0]. \tag{12}$$

## Method

This is a pure theory paper. The model is solved by characterizing the set of feasible contracts
(satisfying limited liability, resource constraints, investor participation, repayment monotonicity,
and incentive compatibility at both dates), then finding the contract that maximizes bank expected
utility subject to those constraints.

**Pledgeability reduction (Lemma 1, p. 2822).** The binary date 1 effort problem reduces to
a Holmstrom and Tirole (1997) style pledgeability constraint. The optimal contract sets
$$x_1(R_1) = 0$$ and repays investors on date 2 at a threshold $$R_2^u(R_1)$$. Bank
high effort on date 1 is incentive compatible if and only if $$c(R_1) \ge b R_1 Y_0$$,
where $$b = \int_{\overline{R}_2^u}^{\overline{R}} [R_2 - \overline{R}_2^u] f_{2H}(R_2) dR_2$$
is a constant (p. 2822, Equation 18).

**Mapping to promised liabilities.** The paper maps actual-repayment contracts to promised
"face value" liabilities $$L(R_1)$$: if $$L(R_1) \le (1-b)R_1 Y_0$$ the bank avoids
liquidation; if $$L(R_1) > (1-b)R_1 Y_0$$ the bank is liquidated with actual repayment
$$\gamma R_1 Y_0$$ (p. 2822).

**First-order conditions for the private optimum (Proposition 1, pp. 2822-2823).** The
privately optimal liquidation threshold $$R_\ell^p$$ and total debt threshold $$R_u^p$$
satisfy:

$$\underbrace{\frac{1 - \Lambda_1(R_\ell^p)}{(1 - e_0^*) + e_0^* \Lambda_1(R_\ell^p)} \frac{1}{|B_0''(e_0^*)|}}_{\text{Incentive Provision}} b \lambda G = \underbrace{b + \lambda(1 - b - \gamma)}_{\text{Liquidation Costs}}, \tag{20}$$

$$\underbrace{\frac{F_{1L}(R_u^p) - F_{1H}(R_u^p)}{|B_0''(e_0^*)|}}_{\text{Incentive Provision}} \lambda G = \underbrace{(\lambda - 1)(1 - F_1(R_u^p | e_0^*))}_{\text{Investor Repayment}}, \tag{21}$$

where $$\lambda > 1$$ is the Lagrange multiplier on the investor participation constraint and
$$G = \int_{\underline{R}}^{R_\ell^p} \gamma R_1 (f_{1H}(R_1) - f_{1L}(R_1)) dR_1 + \int_{R_\ell^p}^{\overline{R}} (1-b) \min\{R_1, R_u^p\} (f_{1H}(R_1) - f_{1L}(R_1)) dR_1$$.

**Social planner's first-order conditions (Proposition 3, pp. 2827-2828).** The planner's
optimum satisfies the same structural equations as the private optimum but with wedge terms
$$+\tau_\ell^s$$ and $$-\tau_u^s$$ on the right-hand sides of Equations (24) and (25):

$$\tau_\ell^s = \left(1 - \frac{1 - \Lambda_1(R_\ell^s)}{(1 - e_0^s) + e_0^s \Lambda_1(R_\ell^s)} \frac{1}{|B_0''(e_0^s)|} b L^s\right) \lambda^s \sigma \gamma^s \ge 0, \tag{26}$$

$$\tau_u^s = \frac{F_{1L}(R_u^s) - F_{1H}(R_u^s)}{|B_0''(e_0^s)|} L^s \sigma \gamma^s \ge 0, \tag{27}$$

where $$\sigma$$ is the liquidation price elasticity and $$\gamma^s$$ is the equilibrium
liquidation price. The wedge $$\tau_\ell^s$$ reflects the social cost of additional
liquidations induced by standard debt; $$\tau_u^s$$ reflects the social cost of more
total debt through the indirect effort channel.

**No-bailout constraint (Proposition 5, p. 2833).** In the case without planner commitment,
the planner does not bail out banks if the total losses from liquidations are high enough:

$$(1 - \gamma(\Omega)) \Omega Y_0 \le F. \tag{29}$$

## Empirical specifications

This is a pure theory paper with no empirical component. There are no regressions,
no estimation equations, and no data. The paper characterizes optimal contracts and regulatory
instruments analytically. The special case of linear private benefits ($$B_0(e_0) = b_0(1-e_0)$$)
is worked out in Internet Appendix B.3 as a closed-form tractable illustration.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| None | Pure theory paper; no data used | N/A |

## When to read the full paper

Read Clayton and Schaab (2025) if you are designing or evaluating:

- Optimal bank capital structure rules combining leverage caps and TLAC requirements
- The comparison between bail-in debt and outside equity as loss-absorbing instruments (Proposition 2 and Section 2.2.1), extending Dewatripont and Tirole (1994)
- The rationale for statutory provisions that increase the cost of bailouts, such as Dodd-Frank Act restrictions (Section 4.2); connects to Farhi and Tirole (2012) and Chari and Kehoe (2016)
- The good bank / bad bank approach to resolving large banks with partial liquidations (Section 5.2)
- The connection between the incentive-based explanation and the result of Keister and Mitkov (2023) that anticipated bailouts suppress bail-in debt issuance
- The connection between the incentive-based and demand-based (safety premium) explanations for the coexistence of standard and bail-in debt (Section 5.3)

## Attribution and rights

Clayton, C., and A. Schaab. 2025. "Bail-Ins, Optimal Regulation, and Crisis Resolution."
*The Review of Financial Studies* 38(9): 2810-2843. https://doi.org/10.1093/rfs/hhaf002.

Published by Oxford University Press on behalf of The Society for Financial Studies. All rights
reserved. For commercial re-use contact reprints@oup.com. This page is an LLM-distilled
summary (extract-only); it is not human-verified and has not been reproduced. Read the original
at the DOI above.


==============================================================================
# Uncertainty, Contracting, and Beliefs in Organizations: Dicks & Fulghieri (2025)
# https://instituteforautomatedresearch.org/wiki/papers/rfs/2025/dicks-uncertainty-contracting-beliefs-organizations-2025/
# Distilled: In a multidivisional firm, uncertainty aversion by managers creates endogenous disagreement that raises incentive costs; HQ can hedge this by designing contracts with cross-divisional exposure (equity or relative-performance pay), improving effort and aligning beliefs. Review of Financial Studies 2025, paywalled. Five core results with source locators, the model with its key equations, and the method.
# Tags: paper-summary, contract-theory, executive-compensation, organizational-economics
==============================================================================

**What this is.** The paper's core results, the theoretical model with its key
equations, and the method: enough to know what it found and how, without reading
all 44 pages. To replicate or extend it, read the full source at
[https://doi.org/10.1093/rfs/hhaf005](https://doi.org/10.1093/rfs/hhaf005).

## TL;DR

Dicks and Fulghieri study optimal incentive contracts in a two-division firm where
headquarters (HQ) and division managers are uncertainty averse in the sense of
Gilboa and Schmeidler (1989): they hold a set of admissible priors and evaluate
random variables by their worst-case expected utility. The paper extends the
classical moral hazard framework of Holmstrom (1979) and the linearity results of
Holmstrom and Milgrom (1987) to settings where agents lack a single prior on
the probability distribution of cash flows. Uncertainty creates two novel costs.
First, an "incentive effect": conservative beliefs about own division productivity
suppress effort, requiring higher pay-performance sensitivity. Second, an
"uncertainty discount": HQ and division managers disagree on the value of
compensation contracts because their positions in the hierarchy give them
different exposures to uncertainty, making participation constraints more costly.
The key insight is that HQ can partly resolve both costs by designing contracts
with cross-divisional exposure. Linking pay to the other division's output hedges
division managers' uncertainty, improves beliefs, and lowers incentive costs.
This motive for cross-pay is present even when divisions are uncorrelated, in
contrast to the informativeness principle of Holmstrom (1982). When uncertainty
is large enough, optimal contracts are pure equity (equal beta and gamma, beta = gamma),
which dominates relative-performance pay irrespective of cash-flow correlation,
unlike the prediction of Miao and Rivera (2016) for uncertainty-neutral agents.

## Core results

Magnitudes and significance are as reported. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Cross-division pay is optimal even absent correlation**, violating the informativeness principle; uncertainty aversion motivates uncertainty hedging through equity or relative-performance contracts | Theorem 1, p. 2201; Corollary 1, p. 2203 | Optimal cross-division exposure $$\lvert\gamma_d\rvert = \xi_d \beta_d > 0$$ whenever division managers face uncertainty ($$\eta > 0$$); pay-performance sensitivity $$\beta_d = 1/(1 + 3(1 - \hat{q}^d_d/q_d)) < 1$$ and effort are both decreasing in $$\eta$$ (eq. 20) |
| R2 | **Equity-based pay is optimal over relative-performance pay when uncertainty is sufficiently large** ($$\eta > \tilde{\eta}$$), irrespective of cash-flow correlation | Corollary 1(ii), p. 2203; Theorem 3, p. 2207; Theorem 4, p. 2209 | Pure equity contract: $$\beta_d = \gamma_d$$ with $$\beta + \gamma < 1$$ for $$\eta > 2\ln(3/2)$$; holds even for positively correlated divisions where standard theory predicts $$\gamma < 0$$ (relative performance) |
| R3 | **HQ uncertainty aversion makes relative-performance contracts even more costly**; the HQ uncertainty discount adds disagreement that raises the cost of having HQ hold a "short" position in the relative-performance hedge | Theorem 3, p. 2207; Section 4.1, pp. 2206-2208 | Pay-performance sensitivity $$\beta_d = \gamma_d = 1/(1+3(1-\hat{q}^d_d/\hat{q}^{HQ}_d)) < 1$$, with pay-performance sensitivity increasing in HQ uncertainty $$\eta^{HQ}$$ (Figure 5); pure equity optimal whenever $$\eta > \eta^{HQ} + 2\ln(3/2)$$ |
| R4 | **Internal hedging (cross-division pay) dominates external benchmarks** for uncertainty hedging; with large HQ uncertainty, optimal contracts exclude external hedges entirely | Lemma 5, p. 2213; Theorem 5, p. 2213 | HQ weakly prefers contract $$(\beta, \lvert\psi\rvert, 0)$$ over $$(\beta, 0, \lvert\psi\rvert)$$ when choosing between internal (division B) and external (variable C) hedges; if $$\eta^{HQ} > \tilde{\eta}^{HQ}$$, optimal contracts set $$\psi = 0$$ (no external hedge) |
| R5 | **Synergies reinforce the optimality of equity-based pay**; for any uncertainty level there exists a synergy threshold above which pure equity is optimal | Theorem 7, p. 2215 | For any $$\eta \geq 0$$ there is a threshold $$\bar{\zeta}$$ such that for all $$\zeta > \bar{\zeta}$$, the optimal contract has $$\gamma = \beta$$; at $$\eta = 0$$, pure equity is optimal only when $$\zeta = 1$$ (perfect effort substitutes) |

**Overall (paper's conclusion).** Uncertainty aversion provides a unifying
explanation for three otherwise puzzling compensation practices: the prevalence of
equity-based pay for lower-level managers (even when risk-bearing arguments do not
support it), the rarity of relative-performance contracts especially in high-uncertainty
environments such as young and innovative firms, and the optimism gradient
whereby senior managers hold systematically more favorable beliefs about firm
prospects than rank-and-file employees.

## Theory / model

The paper's model is a one-period, two-division firm. There are two divisions
$$d \in \{A, B\}$$, each run by a division manager supervised by HQ.

**Cash flows and effort.** Each division's cash flow is:

$$
Y_d = \mu_d + \varepsilon_d, \quad \text{where} \quad \mu_d = a_d q_d \tag{p. 2189}
$$

Division cash flows $$(Y_A, Y_B)$$ have a joint normal distribution $$N(\mu, \Sigma)$$
with homoscedastic variance $$\sigma^2$$ and correlation $$\rho$$. Effort $$a_d \in \mathbb{R}_+$$
affects the mean; the cost is $$c_d(a_d) = \tfrac{1}{2\theta_d} a_d^2$$, where $$\theta_d$$
is efficiency of effort. Division managers have CARA utility $$U(w) = -e^{-r w}$$
with coefficient $$r$$; HQ is risk neutral (in the base model).

**Linear incentive contracts.** HQ offers linear contracts. Division manager $$d$$'s
compensation is:

$$
w_d(Y) = s_d + \beta_d Y_d + \gamma_d Y_{d'} \tag{p. 2192}
$$

where $$s_d$$ is a fixed base pay, $$\beta_d$$ is pay-performance sensitivity on own
division output, and $$\gamma_d$$ is cross-division ("cross-pay") exposure. Setting
$$\gamma_d > 0$$ gives an equity component; $$\gamma_d < 0$$ gives relative-performance pay.

**Uncertainty aversion (MEU, Gilboa and Schmeidler (1989)).** Both HQ and division
managers are uncertainty averse: they evaluate payoffs by minimizing expected utility
over a set of admissible priors $$\mathcal{P}$$:

$$
\mathcal{U} = \min_{p \in \mathcal{P}} E_p[U(w)] \tag{eq. 1, p. 2190}
$$

The core beliefs set is defined using the relative entropy (Kullback-Leibler
divergence) of candidate distribution $$\hat{P}(x)$$ relative to reference $$P(x)$$
(Hansen and Sargent (2001)):

$$
R\!\left(\hat{P}(x) \,\middle|\, P(x)\right) \equiv \int \hat{p}(x) \ln\!\left(\frac{\hat{p}(x)}{p(x)}\right) dx \tag{eq. 3, p. 2190}
$$

The admissible set is $$\mathcal{P}(P) \equiv \{\hat{P}: R(\hat{P}(x)|P(x)) \leq \eta^P\}$$,
where $$\eta^P$$ is the uncertainty parameter. A higher $$\eta^P$$ means greater uncertainty
aversion and a larger set of admissible beliefs.

**Parametric approximation.** For tractability (following Dicks and Fulghieri (2019,
2021)), the core beliefs set for agent $$i \in \{HQ, A, B\}$$ is approximated as:

$$
K^i(q) \equiv \left\{\hat{q}^i \,\middle|\, D(\chi^i_A) + D(\chi^i_B) \leq \eta^i\right\}, \quad \chi^i_d = \left|\frac{\hat{q}^i_d - q_d}{q_d}\right|, \quad D(\chi) = -\ln(1-\chi) \tag{eq. 11, p. 2195}
$$

where $$\hat{q}^i_d$$ is agent $$i$$'s belief about division $$d$$'s productivity and $$q_d$$
is the reference productivity. This set is strictly convex with smooth boundaries,
guaranteeing that beliefs respond to changes in compensation contracts.

**Division manager utility.** Given beliefs $$\hat{q}^d$$ and action $$a$$, division
manager $$d$$'s certainty-equivalent utility is:

$$
u_d(\hat{q}^d, a) \equiv E\!\left[w_d|\hat{q}^d, a\right] - \frac{r}{2} Var(w_d) - c_d(a_d) \tag{eq. 5, p. 2192}
$$

where $$Var(w_d) = \sigma^2(\beta_d^2 + 2\rho\beta_d\gamma_d + \gamma_d^2)$$. The key feature
is that the expected wage $$E[w_d|\hat{q}^d, a]$$ depends on division managers'
beliefs about productivity of both divisions (through own-pay $$\beta_d$$ and
cross-pay $$\gamma_d$$), while $$Var(w_d)$$ does not.

## Method

The paper derives analytical solutions to a minimax contracting problem. HQ
maximizes expected profits subject to division managers' incentive constraints
(IC) and participation constraints (PC), while both HQ and managers minimize
over their worst-case beliefs. The problem is:

$$
\max_{\{w_d, a_d\}} \min_{\hat{q}^{HQ} \in K^{HQ}} \pi(\hat{q}^{HQ}) \equiv \sum_{d \in \{A,B\}} E\!\left[Y_d(a_d) - w_d|\hat{q}^{HQ}\right] \tag{eq. 6, p. 2193}
$$

subject to the division managers' IC constraints:

$$
\max_{a_d} \min_{\hat{q}^d \in K^d} u_d(\hat{q}^d, a) \equiv E\!\left[w_d|\hat{q}^d, a_d, a_{d'}\right] - \frac{r}{2}Var(w_d) - c_d(a_d) \tag{eq. 7, p. 2193}
$$

and PC constraints:

$$
\min_{\hat{q}^d \in K^d} u_d(\hat{q}^d, a_d, a_{d'}) \geq u_0 = 0 \tag{eq. 8, p. 2193}
$$

The solution strategy is three-step: (1) characterize how contracts determine
beliefs via Lemma 2; (2) derive equilibrium effort from beliefs via Lemma 3;
(3) characterize optimal contracts by substituting the binding PC into the
objective, yielding the reduced-form HQ payoff (eq. 10, p. 2194):

$$
\pi = \sum_{d \in \{A,B\}} \left\{E(Y_d(a_d)|\hat{q}^{HQ}_d) - \frac{r}{2}Var(w_d) - c_d(a_d) - \left(E[w_d|\hat{q}^d,a] - E[w_d|\hat{q}^{HQ}_d,a]\right)\right\} \tag{eq. 10, p. 2194}
$$

The fourth term is the "uncertainty discount" arising from belief disagreement;
it is novel and central to the paper's results.

**Key analytical result (Theorem 1).** With uncertainty-neutral HQ and
uncertainty-averse risk-neutral division managers, optimal contracts set
$$H_d = 1$$ (uncertainty-hedging ratio equal to one), where
$$H_d \equiv |\gamma_d| a_{d'} q_{d'} / (\beta_d a_d q_d)$$.
Optimal pay-performance sensitivity is:

$$
\beta_d = \frac{1}{1 + 3\!\left(1 - \hat{q}^d_d/q_d\right)} < 1, \quad |\gamma_d| = \xi_d \beta_d \tag{eq. 20, p. 2201}
$$

with $$\xi_d \equiv \frac{a_{d'} q_{d'}}{a_d q_d}$$. Both $$\beta_d$$ and effort $$a_d$$ are
decreasing in uncertainty $$\eta$$. Under symmetry, pure equity is optimal:
$$\beta = \gamma < 1$$.

**Theorem 2** (risk-averse division managers) shows that the optimal contract must
satisfy (eq. 21, p. 2203):

$$
\beta_d a_d q_d + r\sigma^2 \beta_d^2 = |\gamma_d| a_{d'} q_{d'} + r\sigma^2 \gamma_d^2 \tag{eq. 21}
$$

equating the total expected cost to HQ of a division manager's exposure to each
division, regardless of the correlation $$\rho$$. Cross-pay is always non-zero,
$$\gamma_d \neq 0$$, even when divisions are uncorrelated.

**Theorem 3** (uncertainty-averse HQ) yields pure equity at sufficiently high
uncertainty:

$$
\beta_d = \gamma_d = \frac{1}{1 + 3(1 - \hat{q}^d_d/\hat{q}^{HQ}_d)} < 1 \tag{eq. 25, p. 2207}
$$

Relative-performance pay creates a short position for HQ in the other division,
amplifying the beliefs disagreement between HQ (long position) and division
managers (short position), raising the uncertainty discount and making equity
strictly preferred.

The proofs use the envelope theorem applied to the minimax problem, with
closed-form first-order conditions derived under the parametric beliefs
approximation (eq. 11). The proofs of Theorems 1 and 2 appear in the appendix
(pp. 2218-2221); Theorems 3, 4, 5, and 7 proofs are in the supplemental materials.

## Datasets used

This paper is purely theoretical. It introduces no dataset.

| Dataset | Role in paper | Wiki page |
|---|---|---|
| No empirical data used | Theory paper with numerical illustrations only | N/A |

The empirical illustrations use baseline parameter values $$q_A = q_B = 10$$,
$$\theta_A = \theta_B = 2$$, $$\sigma = 10$$, $$r = 1$$ (stated in Section 2 footnotes,
p. 2195 and Figures 1-7).

## When to read the full paper

Use the [original](https://doi.org/10.1093/rfs/hhaf005) if you are:
designing incentive contracts under Knightian uncertainty or ambiguity aversion;
trying to explain equity-based compensation for division managers or rank-and-file
employees; studying why relative-performance pay is rare in practice despite
its theoretical benefits; or extending the model to multitasking, labor-market
equilibrium, or organization design with uncertainty. Theorem 1 (p. 2201) and
Corollary 1 (p. 2203) are the key analytical results; Figure 3 and Figure 6
(pp. 2204, 2210) illustrate optimal contracts under uncorrelated and correlated
cash flows respectively.

## Attribution and rights

Source: peer-reviewed, *The Review of Financial Studies* 38(7), 2025.
This distillation was extracted by an LLM on 2026-06-06 and is **not
human-verified or independently reproduced**. The paper is published under
Oxford University Press standard reuse rights (paywalled). Extract-only.

> Dicks, David L., and Paolo Fulghieri. "Uncertainty, Contracting, and Beliefs
> in Organizations." *The Review of Financial Studies* 38, no. 7 (2025): 2182-2225.
> DOI: 10.1093/rfs/hhaf005.


==============================================================================
# Investor Memory: Godker, Jiao & Smeets (2025)
# https://instituteforautomatedresearch.org/wiki/papers/rfs/2025/godker-investor-memory-2025/
# Distilled: Three lab and online experiments document a positive memory bias in investment outcomes: subjects overremember gains and underremember losses, which translates into overly optimistic beliefs, excess reinvestment, and overconfidence about stock-picking ability. Review of Financial Studies 2025, paywalled. Eight core results with source locators, datasets used, the experimental model, and the estimating specifications.
# Tags: paper-summary, behavioral-finance, investor-beliefs, memory-bias
==============================================================================

**What this is.** The paper's core results, the experimental designs, and the
estimating specifications with exact magnitudes: enough to know what was found and
how, without reading all 46 pages. To replicate or extend, read the full source at
the [original](https://doi.org/10.1093/rfs/hhaf006) or download the replication
data from [Harvard Dataverse](https://doi.org/10.7910/DVN/7K6ZPK).

## TL;DR

Across three incentivized experiments (N = 229, 498, 487), the paper documents
a positive memory bias for investment outcomes. After 1 week, subjects
overremember the positive returns and underremember the negative returns of
stocks they chose to invest in. This positive memory bias: (a) distorts beliefs
about stock quality by 8.16 percentage points toward optimism relative to the
objective Bayesian posterior, (b) doubles the rate of suboptimal reinvestment
(41.9% vs 20.5%), and (c) raises overconfident betting on one's own stock
picks by 47%. The mechanism is motivated memory suppression (not genuine
forgetting), consistent with the framework of Benabou and Tirole (2002): raising
the financial incentive for accurate recall eliminates the bias for subjects prone
to self-deception, and the bias disappears when subjects did not actively choose
their investment (passive endowment condition).

## Core results

Magnitudes and significance are as reported; \* p < .1, \*\* p < .05, \*\*\* p < .01.
Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Subjects **overremember positive investment outcomes** after 1 week | Table 2, p. 1614; Figure 1, p. 1613 | Delay: +0.89 outcomes overremembered (p = .000); Immediate: +0.27 (p = .058); difference p = .018 |
| R2 | Subjects **underremember negative investment outcomes** after 1 week | Table 2, p. 1614; Figure 1, p. 1613 | Delay: -0.73 outcomes underremembered (p = .000); Immediate: -0.28 (p = .037); difference p = .038 |
| R3 | **Beliefs are 8.16 percentage points too optimistic** in the Delay condition relative to the objective Bayesian posterior | Figure 2, p. 1616; Internet App. D | t-test p = .004; effect is absent in Immediate condition |
| R4 | **Memory bias directly predicts belief distortion**: each recalled positive outcome raises subjective belief by 6.20pp; each underremembered loss reduces it by 7.96pp | Table 3 cols 1-2, p. 1617 | 6.20\*\*\* (0.92) and -7.958\*\*\* (1.02); R2 = 0.52-0.55 |
| R5 | **Suboptimal reinvestment doubles** in the Delay vs Immediate condition (41.9% vs 20.5%) | Figure 3, p. 1619; Table 4 col. 2, p. 1620 | Odds ratio 3.542\*\*\* (1.44) for Delay treatment dummy; significant at 1% |
| R6 | **Memory bias magnitude predicts suboptimal reinvestment**: each overremembered positive outcome raises probability of suboptimal reinvestment by 59.9% | Table 4 cols 3-4, p. 1620 | Odds ratio 1.599\*\*\* (0.22) for positive-outcome memory bias; 0.617\*\*\* (0.10) for negative-outcome bias |
| R7 | **Overconfident betting is 47% higher** in Delay vs Reminder condition; Delay subjects bet 46.3 vs 31.5 points suboptimally | Figure 4, p. 1621; Table 5 col. 2, p. 1622 | 14.788\*\*\* (4.18) additional suboptimal points in Delay; p < .01 |
| R8 | **Motivated memory suppression drives the bias**: high recall incentives eliminate the memory bias for high-SDE subjects; only actively choosing investors exhibit the bias | Table 7, p. 1625; Table 6, p. 1623 | HighStake coeff for high-SDE: -0.469\*\* (0.19); NoChoice treatment effect 0.18 (p = .398, insignificant) |

**Overall (paper's conclusion).** Investors systematically overremember their
gains and underremember their losses, and this positive memory bias operates as
a microfoundation for three stylized facts in behavioral finance: gains weigh
more than losses when investors learn from experience (Kaustia and Knupfer (2008));
investors are more likely to repurchase stocks previously sold at a gain
(Strahilevitz, Odean, and Barber (2011)); and overconfident investors trade
excessively (Barber and Odean (2001)). The paper also extends evidence from
Kuhnen (2015) on asymmetric belief updating from financial information by adding
a memory channel. The bias is motivated (driven by self-image concerns) rather
than a passive cognitive limitation.

## Theory / model

The paper has no formal structural model in the body. The hypothesis tested is:

**H (Positive memory bias):** When subjects have time to form memories
(Delay condition, 1 week between observation and elicitation), they recall
more positive investment outcomes and fewer negative outcomes than actually
occurred, relative to a control group that elicits memory immediately
(Immediate condition).

The identification logic is a between-subject random assignment to Delay vs
Immediate conditions, with the 1-week gap as the treatment that activates
memory processes. The Immediate condition holds constant all non-memory
factors (information acquisition, salience, attention) that could influence
recall.

The Bayesian benchmark for subjective beliefs is the objective posterior
(equation 1, p. 1604):

$$
\mu_t^G(h_t) = \frac{1}{1 + \frac{1-\mu_0^G}{\mu_0^G} \times \left(\frac{\theta}{1-\theta}\right)^{t - 2n_t^+}} \tag{1}
$$

where $$\mu_0^G = 50\%$$ is the prior probability that the stock is good,
$$\theta = 60\%$$ is the probability that a good stock generates a positive
outcome each period, $$t$$ is the total number of observed outcomes, $$h_t$$
is the history of outcomes, and $$n_t^+$$ is the number of positive outcomes
observed. The objective Bayesian posterior $$\mu_t^G(h_t)$$ is the benchmark
against which subjective beliefs are compared to measure belief distortion.

The motivated-memory mechanism is tested via two conditions in experiment 2.
The **HighStake** condition increases the financial incentive for accurate
memory reporting from 8 to 50 GBP/USD per correct memory answer (following
Zimmermann (2020)), so that if the true information is still in memory
(suppressed, not deleted), high incentives should induce subjects to report
it more accurately. The **NoChoice** condition randomly endows subjects with
a stock rather than letting them choose, so that ego-relevance of outcomes is
reduced (following Mather, Shafir, and Johnson (2000, 2003)).

The Internet Appendix formalizes a model in which memory of investment
outcomes is systematically biased in a motivated way: memory bias stems from
quasi-Bayesian belief updating with a probability of underremembering
specific previously observed signals, where this probability depends on
whether the signals are consistent with the decision-maker's motivation
(Section 3, p. 1617 reference to the Internet Appendix model).

The parimutuel betting payoff in experiment 3 follows Enke, Graeber, and
Oprea (2023), equation 2 (p. 1611):

$$
\text{payoff}_i = \frac{b_i}{\frac{\sum_{l=1}^{10} x_l b_l}{\sum_{l=1}^{10} b_l}} + (100 - b_i) \tag{2}
$$

where $$b_i$$ denotes points bet by subject $$i$$ and $$x_i$$ is an indicator equal
to 1 if the subject's investment choice was optimal. This design provides an
objective measure of overconfidence: subjects should bet 0 if they believe
their choice was suboptimal and bet all 100 if they believe it was optimal,
so suboptimal betting (betting > 0 when the choice was suboptimal) directly
reveals overconfidence.

## Method

The paper uses a randomized between-subject design with exogenous variation in
the time span between observing investment outcomes and eliciting memory (Delay
vs Immediate conditions). Three experiments share a common framework but vary
in outcomes, conditions, and elicited beliefs/choices.

**Common structure.** Subjects observe 12 sequential investment outcomes from a
risky stock over 12 periods (each shown for 2 seconds on screen). Outcomes are
drawn from known distributions (positive/negative with known probabilities
depending on stock quality). After the observation phase, memory, beliefs, and
investment choices are elicited. The Delay condition places the elicitation in
week t+1 (1 week later); the Immediate condition places it in week t
(immediately after).

**Memory elicitation.** Subjects report how many positive and negative outcomes
they observed (experiment 1: how often each specific outcome value occurred;
experiments 2-3: total counts summing to 12). The memory bias at the individual
level is the difference between recalled and actually observed counts of
positive (or negative) outcomes.

**Identification via timing variation.** The comparison of Delay vs Immediate
conditions isolates the effect of memory from attention, salience, and
information processing, which are held constant because both groups observe the
same outcomes with the same level of engagement. A NoRecall condition in
experiment 1 (no memory task) confirms that the memory elicitation task itself
does not affect subsequent beliefs or investment decisions.

The **randomized-survey-experiment** primitive underlies all three experiments:
random assignment to conditions, incentivized elicitation of beliefs and choices,
and exogenous variation in a single treatment variable (delay vs immediate, or
HighStake vs Baseline, or NoChoice vs Baseline).

## Empirical specifications

**Memory bias (R1, R2).** The memory bias is regressed on a constant and tested
with a t-test against zero for each condition (Table 2, p. 1614):

$$
\text{MemoryBias}_{i,s} = \alpha + \beta \cdot \mathbf{1}[\text{Delay}_i] + \gamma \cdot \text{Session}_{is} + \varepsilon_{is}
$$

where $$\text{MemoryBias}_{is}$$ is the subject's recalled minus actually observed
count of positive (or negative) outcomes, and $$\text{Session}_{is}$$ is a session
fixed effect. Column 1 of Table 2 reports the t-test of the Delay group mean
against zero; column 3 tests the Delay-Immediate difference.

**Memory-based beliefs (R3, R4).** OLS regressions with session fixed effects (Table 3, p. 1617):

$$
\text{SubjProb}_{is} = \alpha + \beta_1 \cdot \text{MemBias}_{\text{pos},is} + \beta_2 \cdot \text{MemBias}_{\text{neg},is} + \gamma \cdot \text{ObjProb}_{is} + \delta \cdot \text{Session}_{is} + \varepsilon_{is}
$$

Dependent variable: subjective probability that the stock is good (1-100).
Columns 3 and 4 use belief distortion (difference between posterior log-likelihood
ratios of subjective and objective probabilities) as the dependent variable.
N = 188, R2 = 0.31-0.55. Standard errors in parentheses.

**Reinvestment behavior (R5, R6).** Logit regressions with session fixed effects (Table 4, p. 1620):

$$
\Pr(\text{Invest}_{is}) = F\left(\alpha + \beta \cdot \text{Delay}_i + \gamma_1 \cdot \text{MemBias}_{\text{pos},is} + \gamma_2 \cdot \text{MemBias}_{\text{neg},is} + \delta \cdot \text{Session}_{is}\right)
$$

Dependent variable: a dummy equal to 1 if the subject reinvested in the stock
(Inv.) or a dummy equal to 1 if the subject reinvested suboptimally from a
Bayesian perspective (Inv. (Subopt.)). Odds ratios reported. N = 152, pseudo-R2 = 0.05-0.13.
Sample restricted to subjects who invested in the stock.

**Overconfidence / suboptimal betting (R7, R8).** OLS regressions (Table 5, p. 1622):

$$
\text{PointsBet}_{is} = \alpha + \beta \cdot \text{Delay}_i + \gamma_1 \cdot \text{MemBias}_{\text{pos},is} + \varepsilon_{is}
$$

Dependent variable: number of points bet suboptimally (Points bet (subopt.)
= points bet when the stock chosen had fewer than 6 positive outcomes, i.e.,
was suboptimally chosen from a Bayesian perspective). N = 191 (col. 2),
N = 83 (col. 3 for Delay subjects only). R2 = 0.02-0.06. No session fixed
effects in experiment 3 (collected in one online session).

**Mechanism tests (R8).** OLS regressions split by median self-deceptive
enhancement (SDE) score (Table 7, p. 1625); Treatment (HighStake) is the
dummy for the high-incentive condition. The memory-suppression prediction is
that HighStake reduces bias for high-SDE subjects (column 1: coeff -0.469,
p < .05) but not for low-SDE subjects (column 2: 0.115, p = .60).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Experiment 1 lab data (Hamburg University, N = 229) | Primary identification of memory bias, beliefs, and reinvestment (Results 1-6) | No page yet |
| Experiment 2 online lab data (CESS Oxford, Xlab UC Berkeley, N = 498) | Memory suppression and active-choice mechanism tests (Result 8, Table 6-7) | No page yet |
| Experiment 3 online data (Prolific, UK, N = 487) | Overconfidence measure via parimutuel betting (Results 6-8) | No page yet |

All three experiments use primary hand-collected data: incentivized laboratory
or online experiments with randomized treatment assignment. No external financial
databases are used; the investment outcomes are computer-generated draws from
known distributions. Experiment 1 programmed in z-Tree (Fischbacher 2007);
experiments 2 and 3 in oTree (Chen, Schonger, and Wickens 2016). Experiment 3
preregistered at AsPredicted under ID 153791.

## When to read the full paper

Read the [original](https://doi.org/10.1093/rfs/hhaf006) if you are: studying the
microfoundations of investor overconfidence or asymmetric learning from
experience; replicating the experimental paradigm (exact outcome distributions
in Appendix D, instructions in Appendices A-C); extending the motivated-memory
model to field settings or longer horizons; or connecting this to
reinforcement learning in finance (Kaustia and Knupfer (2008)) and stock
repurchase anomalies (Strahilevitz, Odean, and Barber (2011)). The exact
Table/Figure locators above point to the key results. Replication code and data
are publicly available at Harvard Dataverse
([https://doi.org/10.7910/DVN/7K6ZPK](https://doi.org/10.7910/DVN/7K6ZPK)).

## Attribution and rights

Source: peer-reviewed, *The Review of Financial Studies* 38(6), 2025.
Published by Oxford University Press on behalf of the Society for Financial
Studies. All rights reserved. Commercial re-use requires reprints permission
from OUP. This distillation was extracted by an LLM on 2026-06-06 and is
**not human-verified or independently reproduced**. Extract-only: the OUP
standard publication reuse rights do not permit mirroring the verbatim PDF.

> Gödker, Katrin, Peiran Jiao, and Paul Smeets. "Investor Memory."
> *The Review of Financial Studies* 38, no. 6 (2025): 1595-1640.
> DOI: 10.1093/rfs/hhaf006. © 2025 The Author(s).
> Published by Oxford University Press on behalf of the Society for Financial Studies.


==============================================================================
# Effects of Credit Expansions on Stock Market Booms and Busts: Hansman, Hong, Jiang, Liu & Meng (2025)
# https://instituteforautomatedresearch.org/wiki/papers/rfs/2025/hansman-effects-credit-expansions-stock-2025/
# Distilled: Using China's staggered margin-lending deregulation (2010-2015) as a natural experiment, the paper finds causal evidence that credit expansions substantially raise stock prices, with the effect largely anticipated and front-run by unconstrained institutional investors. Review of Financial Studies 2025, paywalled. Seven core results with source locators, datasets used, the dynamic information-revelation model, and the empirical specifications (event study, RDD, panel regression).
# Tags: paper-summary, asset-pricing, credit, equities, leverage, china, event-study
==============================================================================

**What this is.** The paper's core results, the information-revelation model it contributes, and the three empirical strategies (event study, regression discontinuity, panel model) with their equations: enough to know what it found and how, without reading the full 43 pages. To replicate or extend the paper, read the original at [doi.org/10.1093/rfs/hhaf008](https://doi.org/10.1093/rfs/hhaf008).

## TL;DR

Using China's staggered deregulation of margin lending (three vintages introduced between 2013 and 2014 under a formula-based inclusion rule), the paper establishes causal evidence that an expansion of margin credit substantially raises stock prices, with 60-day cumulative returns roughly 25% higher for newly marginable stocks than for just-below-threshold controls. Crucially, this effect was largely anticipated: stock prices began rising months before formal margin eligibility, consistent with deep-pocketed institutional investors front-running the liberalization. Chinese mutual funds systematically overweighted soon-to-be-marginable stocks as each vintage approached, then sold after eligibility, and earned significant risk-adjusted profits from this strategy. A dynamic information-revelation model rationalizes the gradual exponential run-up with two parameters: the total price impact and the monthly rate of anticipation (roughly 5-7%). The paper also documents that stocks with more margin debt at the 2015 market peak experienced larger crashes, consistent with margin lending amplifying the bust as well. The results extend to equities the causal credit-to-asset-price evidence of Mian and Sufi (2009) and Di Maggio and Kermani (2017) for mortgage markets. Shadow margin data from Bian, Da, He, Lou, K. Shue, and Zhou (2023) is used to measure informal lending in the boom-bust analysis.

## Core results

Magnitudes and significance as reported; `\*` = 10%, `\*\*` = 5%, `\*\*\*` = 1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Pre-marginability run-up**: DGTW-adjusted returns for soon-to-be-marginable vs. never-marginable stocks are large and significant in months leading to deregulation, but near zero after | Table 2 Panel A, p. 1514 | -1 to 0 months: 2.1%\*\*\*; -3 to 0 months: 7.3%\*\*\*; -12 to 0 months: 26.9%\*\*\*; following marginability: statistically zero |
| R2 | **RD reduced-form**: stocks just above the marginability threshold saw roughly 25% higher 60-day cumulative returns than just-below-threshold stocks | Table 3 Panel B, p. 1522 | 4.6%\*\*\* (5 days), 8.3%\*\* (20 days), 24.6%\*\*\* (60 days), all at optimal bandwidth; fixed bandwidth: 4.6%\*\*\*, 8.2%\*\*, 24.2%\*\*\* |
| R3 | **Fuzzy RD causal effect**: 2SLS estimate of becoming marginable on cumulative returns, instrumenting with above-threshold status; first stage is near-perfect (96% jump in marginability) | Table 3 Panel C, p. 1522 | 4.7%\*\*\* (5 days), 8.5%\*\* (20 days), 26.0%\*\*\* (60 days) at optimal bandwidth; 25.2%\*\*\* at 60 days fixed bandwidth |
| R4 | **Mutual fund anticipatory buying**: funds systematically increased weights in soon-to-be-marginable stocks as the deregulation approached, interacted with time since last vintage | Table 4, p. 1528 | Interaction Top Quintile x Months Since Last Vintage: 0.0034\*\*\* (t=3.2, col 1); Top 150 x Months Since Last Vintage: 0.0027\*\*\* (t=2.5, col 2); implies roughly 8.1% total overweighting of top 150 stocks across two exchanges |
| R5 | **Mutual fund alpha**: TBM (to-be-marginable) holdings earned 0.56%/month more than NTBM holdings on a long-short basis; CAPM alpha of long-short = 0.56%/month | Table 5, p. 1530 | TBM raw return 1.65%/month (t=2.79); NTBM 1.09%/month; long-short 0.56%\*\* (t=2.10); CAPM alpha 0.56%\*\* (t=2.01) |
| R6 | **Dynamic panel model**: monthly rate of anticipation theta is 1.05-1.10 (text), implying 5-10% monthly exponential discounting of future credit shock; more than 60% priced in within six months before eligibility | Table 6, p. 1538 | Nonlinear NLS: theta = 1.051-1.099, delta_1 = 0.010-0.025\*\*\*; linear 2SLS: theta = 1.069, delta_1 = 0.035\*\*\* (base), 0.044\*\*\* (full sample) |
| R7 | **Margin debt and the crash**: stocks with higher margin debt at the June 2015 market peak experienced significantly larger crash returns in the subsequent 24-week bust | Table 7 Panel B, p. 1541 | OLS coefficient on Peak Total Margin: -0.266\*\* (SE=0.116); top quintile of margin stocks dropped 6 percentage points more than others; crash window Jun 12 to Nov 27, 2015 |

**Overall (paper's conclusion).** Major expansions of credit for stock purchases increase equity prices, and this effect is anticipated and priced in gradually by deep-pocketed investors who behave as front-runners. The model and estimates suggest that margin debt contributes meaningfully to stock market fluctuations, including both the boom and the bust.

## Theory / model

The paper has no formal asset-pricing equilibrium in the traditional sense; instead it develops a dynamic competitive stock-pricing model with information revelation about a credit supply shock (Section 3, pp. 1531-1538). The model is in the vein of Summers (1986) and DeLong, Shleifer, Summers, and Waldmann (1990), with deep-pocketed unconstrained investors receiving signals about a coming credit supply shock and front-running the price impact. The setup is:

Consider a market for stock with shares outstanding $$Q$$. The stock pays dividend $$\pi \sim N(0, \sigma_\pi^2)$$ at terminal date $$T$$, with the risk-free rate set to zero. A unit mass of unconstrained risk-averse investors have CARA utility $$-e^{-\gamma W}$$ and are price takers.

At $$t = 0$$ there is a permanent price-inelastic demand shock of $$\Delta$$ shares from previously constrained investors receiving margin credit. If the shock were entirely unanticipated, the price would jump at $$t=0$$ by $$m = \Delta \gamma \sigma_\pi^2$$. To capture anticipation, unconstrained investors receive signals $$m_t$$ about the demand shock $$m = \sum_{j=-\infty}^0 m_j$$ in each period $$t \leq 0$$. The signals $$m_t$$ are independent normal with mean zero and variance $$\sigma_t^2$$. The equilibrium price for any $$t$$ between $$-n$$ and $$0$$ is (p. 1532, equation 7):

$$
p_t = p^* + \sum_{j=-n}^{t} m_j - \gamma \left(\sum_{k=t+1}^{0} \sigma_k^2\right) Q \tag{7}
$$

The first term captures revealed information about future prices; the second is a risk discount that falls as uncertainty resolves.

**Exponential decay information structure.** To bring the model to data, the authors impose that for some $$\theta > 0$$ the variance of each signal is $$\sigma_t^2 = \beta(\theta)^t$$ (p. 1533). Under this parameterization uncertainty reduces exponentially as $$t \to 0$$. In a cross-section of treated stocks, the expected price at time $$t$$ is (p. 1533, equation 8):

$$
E[p_t | m > 0] = \bar{p} + \beta(\lambda + \gamma) \sum_{j=-\infty}^{t} \mathbf{1}\{j \leq 0\} \theta^j \tag{8}
$$

where $$\lambda = \frac{\phi(0)}{\Phi(0)} \frac{1}{\sigma_m}$$ and $$\bar{p} = p^* - \frac{\gamma \bar{\mu}}{1 - \frac{1}{\theta}}$$. The period-to-period expected price increase is:

$$
E[p_t - p_{t-1} | m > 0] = \begin{cases} \beta(\lambda + \gamma)\theta^t & \text{if } t \leq 0 \\ 0 & \text{if } t > 0 \end{cases}
$$

The parameter $$\theta$$ captures the exponential rate at which prices rise. Higher $$\theta$$ means anticipation begins earlier. The net impact of deregulation is recovered from the two parameters as: $$\Delta p_{-\infty} = E[p_0 | m > 0] = \frac{\delta_1}{1 - \frac{1}{\theta}}$$ where $$\delta_1 = \beta(\lambda + \gamma)$$.

**Identification assumption.** The timing of deregulation across vintages is orthogonal to vintage-specific demand shocks for stocks. The deregulation formula used publicly observable real-time data on market cap and volume, providing the basis for the regression discontinuity design.

## Method

The paper combines three complementary empirical strategies, each designed to address a different concern about identification.

**Event study (Section 2.1).** Reduced-form estimation of cumulative DGTW-adjusted returns comparing marginable and never-marginable stocks around announcement/implementation dates (pp. 1513-1517, equations 2-3):

$$
Ret_i^k = \beta_0 \text{Marginable}_i^k + \theta_k + \varepsilon_i^k \tag{2}
$$

The stacked version plots the full dynamic path (equation 3):

$$
Ret_{ikt} = \sum_{\ell} \beta_\ell \times \mathbf{1}\{t - \tau(k) = \ell\} \times Treated_{ik} + \gamma_{ik} + \lambda_{tk} + \varepsilon_{ikt} \tag{3}
$$

**Regression discontinuity (Section 2.2).** Exploiting the formula-based threshold in the inclusion index. Stocks with index above the cutoff $$C_E^k$$ qualify; the density test shows no bunching (t-stat = 0.96, Figure 3). The RD estimating equation is (p. 1519, equation 4):

$$
Y_i^k = \alpha_{0l} + \alpha_{1l}(Index_i^k - C_E^k) + \tau_i^k[\alpha_{0r} + \alpha_{1r}(Index_i^k - C_E^k)] + \theta_k + \lambda_{ind(i)} + \varepsilon_i^k \tag{4}
$$

where $$\tau_i^k$$ is the indicator for being above the threshold. A triangular kernel local linear approach is used. The fuzzy RD instruments marginability with $$\tau_i^k$$ (2SLS, Panel C of Table 3, p. 1522). Standard errors are clustered at the stock level; CCT robust confidence intervals are applied.

**Dynamic panel model (Section 3.3-3.4).** Two estimating equations derived from the theoretical model.

Nonlinear approach (equation 9):

$$
p_{it}^{treated} - p_{it-1}^{treated} = \underbrace{\delta_1}_{\beta(\lambda+\gamma)} \theta^t + \Delta \varepsilon_{it} \tag{9}
$$

Estimated by nonlinear least squares from pre-event price changes for treated stocks.

Linear approach (equation 10):

$$
p_{it} = \delta_1 D_{it} + \frac{1}{\theta} p_{it+1} + \alpha_i + \eta_t + e_{it} \tag{10}
$$

where $$D_{it}$$ equals one for treated stocks when $$t \geq 0$$. Estimated by 2SLS using forward leads $$D_{it+2}$$ and $$D_{it+3}$$ as instruments to address the dynamic panel bias (Arellano and Bond (1991), Arellano and Bover (1995), Malani and Reif (2015)).

**Mutual fund analysis (Section 2.3).** Portfolio-weight panel regression at the semiannual fund-stock level (equation 5):

$$
w_{ijt} = \alpha + \beta \, \text{Top ranking}_{it} \times \text{Months since last vintage}_t + \text{Controls}_{it} + \sigma_i + \theta_j + \gamma_t + \epsilon_{ijt} \tag{5}
$$

where $$w_{ijt}$$ is fund $$j$$'s dollar-value weight in stock $$i$$ at time $$t$$, scaled by total net assets. Standard errors are clustered by fund.

## Empirical specifications

**Event study (R1, pp. 1513-1517).** Sample: all stocks that either become marginable in vintages 1, 2, 3 or are never marginable. Three announcement/implementation events are pooled. Returns are DGTW-adjusted benchmarks from independent quintile sorts on size, book-to-market, and past-12-month returns (125 bins). Cumulative returns measured from 12 months before to 12 months after the announcement/implementation month. Vintage fixed effects $$\theta_k$$ included. Standard errors clustered at the stock level. Placebo tests use 1,000 iterations randomly drawing three placebo deregulation months from January 1993 to December 2008.

**Regression discontinuity (R2, R3, pp. 1518-1523).** Running variable is $$Index_i^k$$, centered at the vintage-exchange threshold $$C_E^k$$. Window of size one around the threshold. CCT covariate-adjusted MSE-optimal bandwidth (Calonico et al. (2019)) and fixed bandwidth 0.1 both reported. Local linear with triangular kernel. Industry (CSRC 2001) and vintage fixed effects included. Outcome $$Y_{ik}$$ is raw cumulative return over 5, 20, or 60 trading days after implementation. Fuzzy RD instruments marginability with above-threshold indicator; first stage shows 96% probability jump at threshold (Table 3 Panel A, p. 1522). CCT robust p-values reported. N ranges from 146 to 229 depending on bandwidth and horizon.

**Mutual fund panel (R4, R5, pp. 1526-1530).** Sample: semiannual fund-stock level observations, December 2011 to June 2014 (columns 1-3) or December 2011 to June 2015 (column 4). N = 324,860 or 455,846. Controls: percentile rank of market cap, book-to-market, turnover, past month and past year return, and their interactions with Months Since Last Vintage. Fund, stock, and time fixed effects. Standard errors clustered by fund. Performance test (Table 5): calendar-time portfolios January 2012 to September 2014, rebalanced semiannually, value-weighted within fund, Newey-West SEs with 11-month lag.

**Dynamic panel model (R6, pp. 1535-1538).** Monthly data, March 2009 to October 2015. Stock prices normalized to March 2009 price. Base sample includes three main vintages plus never-marginable stocks; full sample adds pilot vintages. Nonlinear NLS estimated from pre-event price changes. Linear 2SLS uses leads $$D_{it+2}$$ and $$D_{it+3}$$ as instruments; instruments validated by the Kleibergen-Paap F-statistic (37-41 across specifications). Two-way stock x year and month fixed effects included.

**Boom-bust analysis (R7, pp. 1539-1541).** Cross-sectional OLS at the stock level. Crash Return = cumulative log return, June 12 to November 27, 2015. Peak Total Margin = stock-level ratio of formal plus shadow margin debt to floating market cap, June 12, 2015. Controls: log floating market cap, book-to-market, turnover, past month and past year return as of end of May 2015. N = 1,645 stocks. Robust standard errors. Quantitative causal interpretation is withheld given the endogeneity of margin debt during the crash.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CSMAR (China Stock Market and Accounting Research) | Stock prices, trading data, book-to-market, fund semiannual holdings (complete portfolios), top 10 shareholder ownership data at quarterly frequency | No page yet |
| Shanghai and Shenzhen exchange margin debt data | Daily stock-level formal margin debt outstanding, collected directly from exchanges | No page yet |
| Proprietary shadow margin data (single large provider) | Daily stock-level shadow margin balance, approximately 5% of market; used for bust analysis | No page yet |
| China Securities Finance Co. (CSF) / regulatory formula | Vintage inclusion index inputs (market cap and trading volume); three-step screening rule per exchange | No page yet |

Sample: primary analysis March 2009 to May 2015 (daily), with the boom-bust extension through October 2015. Mutual fund analysis: December 2011 to June 2014/2015 (semiannual holdings).

## When to read the full paper

Read the [original](https://doi.org/10.1093/rfs/hhaf008) if you are: (a) studying credit-to-equity-price transmission with a quasi-experimental design (the RD, event study, and dynamic panel strategies are all described in full, with Internet Appendix robustness); (b) modeling anticipatory price dynamics in event-study settings more generally (the exponential information-revelation model and its panel estimator have broad applicability); (c) studying institutional front-running around predictable regulatory changes; or (d) analyzing the role of leverage in boom-bust cycles, particularly in markets where short-selling is limited. The locators above point to the exact tables and figures.

## Attribution and rights

Source: peer-reviewed, *The Review of Financial Studies* 38(5), 2025. Published by Oxford University Press. Paywalled; no CC licence. This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. Extract-only under fair-use/fair-dealing norms.

> Hansman, Christopher, Harrison Hong, Wenxi Jiang, Yu-Jane Liu, and Juan-Juan Meng.
> "Effects of Credit Expansions on Stock Market Booms and Busts."
> *The Review of Financial Studies* 38, no. 5 (May 2025): 1502-1544.
> DOI: 10.1093/rfs/hhaf008.
> (c) 2025 The Author(s). Published by Oxford University Press on behalf of The Society for Financial Studies. All rights reserved.


==============================================================================
# Proof-of-Work versus Proof-of-Stake: John, Rivera & Saleh (2025)
# https://instituteforautomatedresearch.org/wiki/papers/rfs/2025/john-proof-work-versus-proof-2025/
# Distilled: John, Rivera, and Saleh develop an equilibrium model showing that Proof-of-Stake blockchains generate higher security than equivalent Proof-of-Work blockchains under real-world parameter values, and that this advantage is particularly salient at high scale. Review of Financial Studies 2025, paywalled. Eight core results with source locators, the model equations, and the method.
# Tags: paper-summary, blockchain, fintech, cryptocurrency, defi, theory
==============================================================================

**What this is.** The paper's core results, the model it builds, and the propositions it derives: enough to know what it found and how, without reading all 50 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1093/rfs/hhaf013).

**Related literature.** The paper builds on Pagnotta (2022) (blockchain security via a crisis-of-confidence attack model), Hinzen, John, and Saleh (2022) (limited adoption model for Bitcoin), and Huberman, Leshno, and Moallemi (2021) (cutoff equilibrium structure for blockchain users). It extends Saleh (2021) (PoS sustainability) by providing the first comparative security analysis across protocols, and adopts Nakamoto (2008)'s 51% attack definition as the baseline PoW attack model.

## TL;DR

John, Rivera, and Saleh compare equilibrium security across Proof-of-Work (PoW) and Proof-of-Stake (PoS) blockchains in a unified OLG model. They derive a general condition (Proposition 3, eq. 52) showing that PoS is more (less) secure than PoW when the relative cost to attack PoS exceeds (falls below) a threshold. Applying real-world parameter estimates for mining depreciation (46.5%-89% annually) and DeFi returns (3.73%-4.85% annually), they show that the sufficient condition for PoS security dominance is met for any attacker leverage up to 453%-568% above PoW, a level far beyond what current overcollateralized lending markets permit. They also show that PoS's security advantage is particularly salient at high scale: scaling generates full insecurity for PoW at lower scale than PoS, and achieves full security for PoS at lower scale than PoW.

## Core results

Magnitudes and significance are as reported. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | PoS generates higher validation investment than PoW for all model parameters because stakers avoid depreciation and energy costs | Proposition 3, eq. 52-53, p. 1979; Section 3, p. 1979 | PoS validation investment exceeds PoW whenever PoW and PoS generate identical validator revenues; ratio given by threshold zeta = [e^{-2rho} psi] / [e^{-2rho} psi + 1 - (1-delta)/(1+kappa_h)] <= 1 |
| R2 | Under real-world parameters (delta_1Y in [46.5%, 89%]; psi_1Y in [3.73%, 4.85%]) PoS generates higher equilibrium security than PoW | Corollary 1, p. 1984-1985; Tables 1-2, pp. 1983-1984 | PoS more secure than PoW whenever lambda <= 453% (most conservative) or lambda <= 568% (less conservative); lambda approx 1 in practice so condition is always satisfied |
| R3 | The PoS security advantage requires only mining hardware depreciation (delta_1Y) > DeFi return (psi_1Y) and a standard supermajority threshold (eta <= 2/3) | Proposition 4, eq. 54-55, p. 1980; Remark 1, p. 1981 | Sufficient condition: lambda <= (1/2)(1 + log(1+delta_T)/log(1+psi_T)); horizon-invariant per eq. 55 |
| R4 | Scaling generates full insecurity (pi^p = 0) for a larger set of attacker benefits under PoW than under PoS | Proposition 5, eq. 57, p. 1986; Section 4.1 | Sigma_0^{PoS} subset of Sigma_0^{PoW}: the set of attacker benefits generating full insecurity at scale is weakly larger for PoW |
| R5 | PoW becomes fully insecure at lower scale than PoS for any fixed maximal attacker benefit | Proposition 6, eq. 59, p. 1987 | Lambda_0^{PoS} >= Lambda_0^{PoW}: the scale at which PoW loses positive security is weakly lower than for PoS |
| R6 | Scaling generates full security (pi^p = 1) for a larger set of attacker benefits under PoS than under PoW | Proposition 7, eq. 64, p. 1989; Section 4.2 | Sigma_1^{PoW} subset of Sigma_1^{PoS}: scaling achieves full security for PoS in cases where it does not for PoW |
| R7 | PoS achieves full security at lower scale than PoW for any fixed maximal attacker benefit | Proposition 8, eq. 66, p. 1990 | Lambda_1^{PoS} <= Lambda_1^{PoW}: PoS reaches full security at a weakly lower scale than PoW |
| R8 | Equilibrium user fees decrease in blockchain scale and go to zero as scale grows, reducing validator revenues and undermining security; this negative effect is more acute for PoW than PoS | eq. 60, p. 1987; Section 4, pp. 1985-1991 | f^p_{(i,t)} = (c_{(i,t)} / Lambda) integral_0^{c_{(i,t)}} x dG(x); fee decreases in Lambda (1/Lambda factor); the two scaling effects on security (fee reduction vs. adoption increase) favor PoS more broadly |

**Overall (paper's conclusion).** Under current real-world conditions, PoS blockchains are more secure than equivalent PoW blockchains for any realistic scale. The security advantage of PoS relative to PoW is particularly salient at high scale: PoW is rendered fully insecure at lower scale, and PoS reaches full security at lower scale, than the respective counterpart. The results provide a formal foundation for understanding blockchain protocol design in the tokenomics and decentralized finance context.

## Theory / model

The paper models a single blockchain as either PoW or PoS. The economy is populated by overlapping generations of agents, each living three periods.

**Users.** Each generation-$$t$$ agent $$i \in [0,1]$$ possesses a unit endowment and has utility:

$$
\mathcal{U}^p_{(i,t)} = \max\{U^p_{(i,t)}, 1+\sigma\} \tag{1}
$$

where $$1+\sigma$$ is the outside option return (p. 1961) and $$U^p_{(i,t)}$$ is the expected utility from adopting the blockchain under protocol $$p \in \{PoW, PoS\}$$. An agent adopts if and only if her wait disutility cost $$c_{(i,t)}$$ falls below a cutoff $$c^p$$, determined in equilibrium. The agent allocates her endowment between DeFi investment (fraction $$1-\omega^p$$) and blockchain validation (fraction $$\omega^p$$) to maximize:

$$
\omega^p_{(i,t)} = \arg\max_{\omega \in [0,1]} \omega \cdot \mathbb{E}_t[R^p_t] + (1-\omega) \cdot \mathbb{E}_t[R^{DeFi}_t] \tag{2}
$$

where $$R^p_t$$ is the period $$t$$ to $$t+2$$ return from validation and $$R^{DeFi}_t = \frac{P_{t+2}}{P_t} \cdot (1+\psi)$$ is the DeFi return, with $$\psi$$ the gross DeFi return and $$P_t$$ the cryptoasset price (p. 1966, eq. 11).

**Wait disutility and fees.** The blockchain has transaction rate $$\Lambda > 0$$. User $$i$$ pays a fee $$f^p_{(i,t)} \geq 0$$ to prioritize processing. The expected wait time when paying fee $$f$$ given others' fees $$f_{-(i,t)}$$ is (p. 1963, eq. 6):

$$
W^p(f, f_{-(i,t)}) = \frac{1}{\Lambda} \times \int \mathbf{1}\{f^p_{(j,t)} \leq f\} \, dG(c_{(j,t)}) \tag{6}
$$

Each adopting agent optimally solves:

$$
f^p_{(i,t)} = \arg\max_{f \geq 0} \left[ \Pi^p_{(i,t)} - f - c_{(i,t)} \cdot W^p(f, f_{-(i,t)}) \right] \tag{4}
$$

and the equilibrium fee function satisfies $$f^p_{(i,t)} = \frac{c_{(i,t)}}{\Lambda} \int_0^{c_{(i,t)}} x \, dG(x)$$ for adopters (Lemma 1, eq. 33, p. 1975).

**Blockchain cryptoasset supply.** The total supply follows:

$$
M_{t+1} = M_t e^{\rho}, \qquad B_t = M_t(e^{\rho}-1) \tag{7-8}
$$

where $$\rho \geq 0$$ is the asset growth rate and $$B_t$$ is the block reward distributed in period $$t$$ (p. 1963).

**PoW validation return.** A miner born in period $$t$$ buys mining hardware at upfront cost per unit hash power $$1+\kappa_h$$ (energy plus equipment), mines in period $$t+1$$, and liquidates equipment at depreciated value $$1-\delta$$ (p. 1964). The lifetime return to mining is (eq. 9, p. 1965):

$$
R^{PoW}_t = \frac{H_{t+1}(1-\delta) + \left(B_{t+1} + \frac{\int f^{PoW}_{(i,t-1)} dG(c_{(i,t-1)})}{P^{PoW}_{t+1}}\right) P^{PoW}_{t+2}}{H_{t+1}(1+\kappa_h)} \tag{9}
$$

where $$H_{t+1}$$ is aggregate hash power, $$B_{t+1}$$ are block rewards, and $$\delta \in [0,1]$$ is hardware depreciation. The key feature is that miners must pay both energy ($$\kappa_h$$) and equipment depreciation ($$\delta$$) costs, making PoW validation more expensive than PoS staking.

**PoS validation return.** A staker born in period $$t$$ purchases coins at price $$P^{PoS}_t$$, stakes them in period $$t+1$$, and liquidates in $$t+2$$. Letting $$S_{t+1}$$ be total staked capital (p. 1966, eq. 10):

$$
R^{PoS}_t = \frac{S_{t+1} P^{PoS}_{t+2} + \left(B_{t+1} + \frac{\int f^{PoS}_{(i,t-1)} dG(c_{(i,t-1)})}{P^{PoS}_{t+1}}\right) P^{PoS}_{t+2}}{S_{t+1} P^{PoS}_{t+2}} \tag{10}
$$

Crucially, stakers recover their principal $$S_{t+1} P^{PoS}_{t+2}$$ in full (no depreciation, no energy cost), so only the opportunity cost of foregone DeFi investment matters.

**Attacker.** A malicious agent draws benefit $$\Gamma_t \sim U[0, \bar\Gamma]$$ from disrupting the blockchain each period, and solves (p. 1967, eq. 12):

$$
\max_{A_t \geq 0} \Gamma_t \cdot \nu^p_{t+1}(A_t) - A_t \tag{12}
$$

where $$\nu^p_{t+1}(A_t)$$ is the probability of a successful attack given resources $$A_t$$. For PoW, success requires $$A_t \geq H_{t+1}(1+\kappa_h)$$ (51% attack, eq. 13). For PoS, success requires acquiring and staking a fraction $$1-\eta$$ of total staked coins, i.e., $$A_t \geq \alpha \cdot S_{t+1}$$, where $$\alpha := \frac{1-\eta}{\lambda \eta} \in [0,1]$$ is the relative cost of attacking PoS (eqs. 15-17, p. 1969-1970).

## Method

The paper derives closed-form stationary cutoff equilibria for both protocols. The solution method is: (1) solve for optimal fees (Lemma 1), (2) derive the equilibrium return to validation for each protocol (Propositions 1 and 2), (3) characterize the equilibrium adoption cutoff $$c^p$$ as the supremum of cutoffs at which all agents with $$c < c^p$$ prefer the blockchain to the outside option, and (4) compare survival probabilities $$\pi^p$$ across protocols.

**PoW equilibrium** (Proposition 1, pp. 1975-1977). The equilibrium fraction of wealth invested in mining is (eq. 34):

$$
\omega^{PoW}(c^{PoW}) := \min\left\{\frac{(1-e^{-2\rho})G(c^{PoW}) + fees(c^{PoW})}{G(c^{PoW})\left(e^{-2\rho}\psi + 1 - \frac{1-\delta}{1+\kappa_h}\right)}, 1\right\} \tag{34}
$$

The equilibrium hash power and survival probability (eqs. 36, 38-39):

$$
H^\*(1+\kappa_h) = \min\left\{\frac{(1-e^{-2\rho})G(c^{PoW}) + fees(c^{PoW})}{e^{-2\rho}\psi + 1 - \frac{1-\delta}{1+\kappa_h}}, G(c^{PoW})\right\} \tag{36}
$$

$$
\pi^{PoW} = \min\left\{\frac{H^\*(1+\kappa_h)}{\bar\Gamma}, 1\right\} \tag{38}
$$

**PoS equilibrium** (Proposition 2, pp. 1977-1978). The equilibrium staked capital and survival probability (eqs. 45, 47-48):

$$
S^\*(c^{PoS}) = \min\left\{\frac{(1-e^{-2\rho})G(c^{PoS}) + fees(c^{PoS})}{e^{-2\rho}\psi}, G(c^{PoS})\right\} \tag{45}
$$

$$
\pi^{PoS} = \min\left\{\frac{\alpha \cdot S^\*}{\bar\Gamma}, 1\right\} \tag{47}
$$

**Key comparison threshold.** Proposition 3 (eq. 52-53) derives the threshold:

$$
\zeta := \frac{e^{-2\rho}\psi}{e^{-2\rho}\psi + 1 - \frac{1-\delta}{1+\kappa_h}} \leq 1 \tag{53}
$$

Then $$\alpha \leq \zeta \Rightarrow \pi^{PoW} \geq \pi^{PoS}$$ and $$\alpha \geq \zeta \Rightarrow \pi^{PoS} \geq \pi^{PoW}$$. Because $$\zeta \leq 1$$ always holds, PoS validation investment exceeds PoW investment whenever both generate the same revenues. Whether this advantage overcomes PoS's lower attack threshold ($$\alpha \leq 1$$) depends on which side of $$\zeta$$ the parameter $$\alpha$$ lies.

**Sufficient condition (Proposition 4).** If the PoS supermajority threshold $$\eta \leq 2/3$$ and annual depreciation $$\delta_{1Y} > \psi_{1Y}$$ (DeFi return), then the condition $$\alpha \geq \zeta$$ holds whenever (eq. 54):

$$
\lambda \leq \frac{1}{2}\left(1 + \frac{\log(1+\delta_T)}{\log(1+\psi_T)}\right) \tag{54}
$$

This is horizon-invariant because $$\frac{\log(1+\delta_T)}{\log(1+\psi_T)} = \frac{\log(1+\delta_{1Y})}{\log(1+\psi_{1Y})}$$ for all $$T$$.

## Empirical specifications

The paper is a theory paper with no regression specifications. The calibration in Section 3.3 provides real-world parameter estimates, comparing the model's sufficient condition to data.

**Mining equipment depreciation** (Table 1, p. 1983). Using theblock.co ASIC price index (three efficiency tiers, August 2021-August 2023), the paper documents average annual price declines of 60.4%-66.38% across all tiers. Academic sources (Prat and Walter 2021: 47.7%-61.3%; De Vries and Stoll 2021: 46.5%-89%; Stinner 2021: 61.45%) confirm $$\delta_{1Y} \in [46.5\%, 89\%]$$.

**DeFi return estimates** (Table 2, p. 1984; Table 3, p. 1985). Using Compass Financial Technologies daily ETH staking return data (October 2023 to June 2024): average $$\psi_{1Y} = 4.32\%$$ (min 3.73%, max 4.85%). AAVE lending data (Chaudhary, Kozhan, and Vishwanath-Natraj 2023) gives 0.13%-2.28% for ETH and up to 8.957% for USDC. All estimates satisfy $$\delta_{1Y} > \psi_{1Y}$$ over the entire range.

**Parameter bounds.** Applying the most conservative values ($$\delta_{1Y} = 46.5\%$$, $$\psi_{1Y} = 4.85\%$$) to eq. 54 gives $$\lambda \leq 453\%$$. Less conservative values ($$\delta_{1Y} = 55\%$$, $$\psi_{1Y} = 4.32\%$$) give $$\lambda \leq 568\%$$. Since current lending markets are overcollateralized and $$\lambda \approx 1$$ in practice, the sufficient condition is always satisfied (Corollary 1, pp. 1984-1985).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| theblock.co ASIC Price Index (three efficiency tiers, Aug 2021-Aug 2023) | Estimate annual mining hardware depreciation rate $$\delta_{1Y}$$ | No page yet |
| Compass Financial Technologies daily ETH staking return data (Oct 2023-Jun 2024) | Estimate annual DeFi return $$\psi_{1Y}$$ for PoS staking | No page yet |
| AAVE ETH/USDC lending rate data (Chaudhary, Kozhan, Vishwanath-Natraj 2023) | Robustness estimate of $$\psi_{1Y}$$ via DeFi lending rates | No page yet |

Note: the core results are theoretical propositions derived from the equilibrium model; the datasets above are used only to calibrate model parameters and verify that the sufficient condition (eq. 54) holds under realistic values. The paper is primarily a theory paper with no econometric identification.

## When to read the full paper

Use the [original](https://doi.org/10.1093/rfs/hhaf013) if you are: (1) building a model of blockchain security or protocol design; (2) evaluating whether a PoW or PoS blockchain is more secure for a specific parameter range (the threshold $$\zeta$$ in eq. 53 is the key object); (3) studying DeFi ecosystem interactions with blockchain security; or (4) calibrating mining depreciation and DeFi yield parameters for a blockchain model. The proofs of all eight propositions are in Appendix B (pp. 1994-2002); replication code is at https://doi.org/10.7910/DVN/5VTIMI.

## Attribution and rights

Source: peer-reviewed, *The Review of Financial Studies* 38(7), 2025. This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The paper is paywalled; reproduction rights follow OUP standard publication reuse rights. Extract-only.

> John, Kose, Thomas J. Rivera, and Fahad Saleh. "Proof-of-Work versus Proof-of-Stake: A Comparative Economic Analysis." *The Review of Financial Studies* 38, no. 7 (2025): 1955-2004. DOI: 10.1093/rfs/hhaf013.


==============================================================================
# Unmasking Mutual Fund Derivative Use: Kaniel & Wang (2025)
# https://instituteforautomatedresearch.org/wiki/papers/rfs/2025/kaniel-unmasking-mutual-fund-derivative-2025/
# Distilled: Using SEC Form N-PORT data, this paper shows that most mutual funds (59%) use derivatives to amplify, not hedge, equity returns, contrary to prior belief. Five derivative strategy clusters are identified via K-Means Clustering; long index users dominate and underperform nonusers despite attracting abnormally high institutional flows. Review of Financial Studies 2025, paywalled. Eight core results with source locators, datasets used, the method, and empirical specifications.
# Tags: paper-summary, mutual-funds, derivatives, fund-behavior, asset-pricing
==============================================================================

**What this is.** The paper's core results, the data and method it contributes (SEC Form N-PORT derivative positions with monthly PnL plus K-Means Clustering of derivative strategies), and the empirical specifications behind each result: enough to know what it found and how, without reading all 47 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1093/rfs/hhaf001).

## TL;DR

Using newly available SEC Form N-PORT data (July 2019 to December 2022), Kaniel and Wang (2025) are the first to directly measure how mutual fund derivative positions contribute to fund returns. Contrary to the common belief that funds use derivatives to hedge, the paper finds that 59% of derivative users employ derivatives to amplify equity returns (positive derivative-nonderivative return correlation). Using K-Means Clustering on the allocation of derivative underlying assets, the paper identifies five persistent derivative strategy clusters. Long index users (41% of derivative users) dominate: they use long equity index derivatives to gain market exposure and amplify fund returns, and they contribute the bulk of the measured amplification. Despite this strategy, long index users do not outperform nonusers in normal times or during crisis periods. During COVID-19, they doubled derivative use to short indices, suffered losses when the Fed announced emergency measures on March 23, 2020, and then lost again on newly opened short positions as markets rebounded.

## Core results

Magnitudes and significance are as reported; `\*` = 10%, `\*\*` = 5%, `\*\*\*` = 1%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Derivatives contribute substantially to fund returns: **average monthly DIR is -6.5 bps** but with a standard deviation of 78 bps; the *derivative relative contribution* exceeds 0.1 for 30% of fund-month observations | Table 1 Panel B, p. 1130; Figure 1, p. 1132 | mean DIR = -6.5 bps (std 78); mean non-DIR = 20.7 bps (std 531); 10% of obs have derivative relative contribution > 0.6 |
| R2 | **59% of derivative users have a positive correlation between DIR and non-DIR**, indicating amplification, not hedging; median correlation 0.17; equity derivative users 64% positive, median 0.43 | Figure 4, p. 1144; Table 4 Panel A, p. 1144 | median DIR-non-DIR correlation: long index = 0.67; long stock = 0.12; short index = -0.58; short stock = -0.25; nonequity = -0.06 |
| R3 | **Derivative strategies are highly persistent**: gross notional exposure auto-regresses at 0.83-0.96 across nontoken strategy groups; probability of staying in major user group 85-94% throughout sample | Table 3, p. 1139; Figure 2 Panel C, p. 1138 | AR coefficient: long index 0.832\*\*\* (t=21.82); long stock 0.915\*\*\* (t=12.16); short index 0.883\*\*\* (t=22.53); short stock 0.956\*\*\* (t=10.15); token users only 0.084 (t=3.95) |
| R4 | **Long index users receive 0.2% more monthly flows than nonusers** (2.4% annually), driven by institutional share classes; all other derivative users receive 0.1% more | Table 9 Panel B, p. 1161; Table 10 Panel A, p. 1163 | long index dummy coef = 0.201\*\* (t=2.21) in col 1; institutional-flow channels confirmed via share-class regressions; retail flows not significantly different |
| R5 | **Long index users underperform nonusers by 0.36-0.62 pp annually** on all five risk-adjusted performance measures over 2011-2022; all derivative users underperform slightly but the gap is largest for long index | Table 9 Panel A, p. 1161 | long index FF5 alpha: -1.45\*\*\* vs nonusers -0.83\*; long index minus nonusers difference = -0.62\*\* (t=-2.20) annually |
| R6 | **During COVID-19 outbreak, long index users underperformed other derivative users by 4.85% per month**; derivatives accounted for 0.85% (18%) of the gap; active equity trading drove the rest | Table 8 Panel A, p. 1153; Figure 5, p. 1154 | long index DIR contribution: -47.60 bps/month; all others DIR: +37.11 bps during outbreak; long index - all others DIR gap = -84.71\*\*\* bps |
| R7 | **During COVID-19 recovery, long index users also failed to outperform**, gaining only 6.3 bps from DIR vs all others DIR of -54.92 bps (total DIR gap = +61.22\*\*); active DIR gap = 5.42 bps (n.s.), so active derivative trading was negligible | Table 8 Panel B, p. 1153; Figure 5, p. 1154 | long index fund return vs all others difference: 260.29 bps\*\*\* monthly during recovery; but driven by equities (non-DIR = 199.08\*\*), not derivatives |
| R8 | **Institutional investors allocated extra flows to high-tracking-error long index funds pre-COVID** (consistent with a risk-shifting channel), which then shifted to short derivative positions during the crash but still failed to outperform | Table 10, p. 1163 | high-CTE long index dummy coef = 0.284\*\*\* (t=3.16) for institutional flows; high-CTE users increased short notional exposure by 17.5\*\*\* pp vs low-CTE users |

**Overall (paper's conclusion).** The majority of derivative-using mutual funds, especially long index users, employ derivatives to amplify equity returns rather than hedge. This amplification strategy does not yield superior performance in normal times or in crises, yet these funds attract abnormally high institutional flows. The paper tests the prediction of Glode (2011) that mutual funds underperform in normal times but outperform in crises: long index users fail the crisis-period outperformance prediction. Institutional investors appear to seek funds that deviate from benchmarks during crises (a risk-shifting rationale), but the strategy failed on the realized price path due to the unexpected Fed intervention during COVID-19.

## Theory / model

The paper has no formal theoretical model. It develops and tests the following testable hypotheses against the N-PORT data:

**H1 (Amplification vs. Hedging).** Prior work by Koski and Pontiff (1999) surveyed mutual fund managers and found most claimed to use derivatives for hedging, with only a small minority reporting amplification. Cao, Ghysels, and Hatheway (2011) use N-SAR data and find hedging evidence by comparing return distributions. Frino, Lepone, and Wong (2009) study derivative use and fund flows with options and futures. All three relied on coarse data that could not directly estimate the derivative PnL contribution. This paper tests the hedging hypothesis directly: if funds amplify, the derivative-induced return (DIR) and the nonderivative-induced return (non-DIR) will be positively correlated. If funds hedge, the correlation will be negative. The key metric is the *signed derivative relative contribution*, the ratio of DIR to non-DIR (p. 1133):

$$
\text{DIR}_t = \frac{\text{PnL}_t^{\text{Realized}} + \text{PnL}_t^{\text{Unrealized}} - \text{PnL}_{t-1}^{\text{Unrealized}}}{\text{TNA}_{t-1}}
$$

$$
\text{Derivative Relative Contribution}_t = \left| \frac{\text{DIR}_t}{\text{non-DIR}_t} \right|
$$

where $$\text{PnL}^{\text{Realized}}$$ and $$\text{PnL}^{\text{Unrealized}}$$ are monthly realized and unrealized profit-and-loss from all derivative positions as reported in N-PORT, scaled by total net assets in the previous month (p. 1133). Non-DIR is defined as fund return minus DIR.

**H2 (Strategy Clustering).** Funds with similar derivative strategies will cluster along the dimension of their underlying-asset allocation. K-Means Clustering on the 12-dimensional allocation vector should yield economically interpretable clusters corresponding to recognized trading motives (amplification, hedging, information trading, nonequity risk management).

**H3 (Performance).** Amplifying derivative users should not necessarily outperform in normal markets (the strategy adds risk without guaranteed alpha). In crises, performance depends on the realized path; a strategy that bets against the market could succeed or fail depending on the crisis trajectory.

**Identification.** There is no causal identification strategy. All results are descriptive. The paper exploits the granularity of N-PORT (monthly fund-level PnL by derivative position, including swaps not previously covered in CRSP or N-SAR) to document facts not measurable with prior data.

## Method

**K-Means Clustering (Section 3.1, pp. 1134-1136).** The key input for each fund-quarter is the 12-dimensional allocation vector $$x = (x_1, \ldots, x_{12})$$ of notional amounts across 12 categories (equity index long/short, individual stock long/short, interest rate long/short, FX long/short, commodity long/short, other long/short). K-Means minimizes intracluster Euclidean distances and maximizes intercluster distances:

$$
\min_{C_1,\ldots,C_k} \sum_{j=1}^{k} \sum_{x_i \in C_j} \| x_i - \mu_j \|^2
$$

where $$\mu_j$$ is the centroid of cluster $$j$$. The optimal number of clusters $$k$$ is chosen by the Silhouette Method, yielding $$k = 5$$. The five clusters are labeled: long index (41.4% of derivative users), long stock (12.6%), short stock (8.5%), short index (11.4%), and nonequity (26.1%) (Table 2, p. 1136). This builds on `panel-regression` for the persistence and performance analyses that follow.

**Derivative Performance Measurement.** N-PORT provides monthly realized and unrealized PnL for each derivative instrument. The paper hand-collects daily security-level returns for each derivative position by matching security names in N-PORT to Yahoo Finance and Bloomberg, allowing construction of *hypothetical DIR* (return assuming static quarterly holdings). The difference between actual DIR and hypothetical DIR isolates active within-quarter derivative trading (p. 1152).

**Extended Sample via CRSP (Section 5, pp. 1158-1161).** For performance and flow analysis, the paper extends to 2011-2022 using CRSP mutual fund holdings. Because CRSP does not provide gross notional exposure, the paper identifies long index users as funds where over 80% of derivative positions are long equity index contracts, using intensive manual matching of security names. Equal-weighted portfolio returns are then regressed on Fama-French factor models.

## Empirical specifications

All regressions use standard errors clustered at the fund level. The main specifications are:

**Derivative strategy persistence (Table 3, p. 1139).** For each derivative user type $$g$$:

$$
\text{GrossNotionalExposure}_{f,t} = \alpha + \beta \cdot \text{GrossNotionalExposure}_{f,t-1} + \varepsilon_{f,t} \tag{1}
$$

with fund fixed effects and Lipper-style fixed effects. Results (R3): $$\hat{\beta}$$ ranges from 0.832 (long index) to 0.956 (short stock), with R-squared 0.68-0.96 for nontoken users.

**Excess cash and equity holdings response to flows (Table 5, p. 1148).** Separate panel regressions by derivative strategy group:

$$
\Delta \text{ExcessCash}_{f,t} = \alpha + \beta \cdot \text{Flow}_{f,t} + \text{Controls}_{f,t} + \text{TimeFE}_t + \text{StyleFE}_f + \varepsilon_{f,t} \tag{2}
$$

where excess cash is fund cash minus 20% of gross notional exposure (excluding call/put purchases and short equity positions), following An and others (2021). Long index users show $$\hat{\beta} = -0.0383^{**}$$ (t=-2.21), the only group with a negative relation, while all others show positive relations consistent with standard cash management.

**Flow-performance sensitivity (Table 6, p. 1149).** Fund next-month flows on past-year performance, controlling for lagged flows, expense ratio, turnover ratio, log TNA, past-year return volatility, with Lipper-style and time fixed effects:

$$
\text{Flow}_{f,t+1} = \alpha + \gamma \cdot \text{Perf}_{f,t-12:t} + \text{Controls}_{f,t} + \text{TimeFE}_t + \text{StyleFE}_f + \varepsilon_{f,t} \tag{3}
$$

Short equity users show the highest flow-performance sensitivity ($$\hat{\gamma} = 10.46^{***}$$ on raw return), consistent with hedge-fund-like investor base.

**Long-run performance (Table 9, p. 1161).** Equal-weighted portfolios formed by derivative user type; excess returns regressed on Fama-French factor returns using a 2011-2022 CRSP sample. Long index users show FF5 alpha of -1.45\*\*\* (t=-2.91) vs nonusers -0.83\* (t=-1.90), a -0.62\*\* difference (t=-2.20) annually (R5).

**Fund flows by strategy (Table 9 Panel B, p. 1161).** Monthly fund-level flow regressions on derivative strategy dummies and their interaction with a retail-share-class indicator:

$$
\text{Flow}_{f,t} = \alpha + \delta_{\text{LI}} \cdot \mathbf{1}[\text{LongIndex}]_f + \delta_{\text{AO}} \cdot \mathbf{1}[\text{AllOthers}]_f + \text{Controls}_{f,t} + \text{TimeFE}_t + \text{StyleFE}_f + \varepsilon_{f,t} \tag{4}
$$

Two-way clustered standard errors at fund and time levels. Long index coefficient: 0.201\*\* (t=2.21) through 0.184\*\* (t=2.15) across performance-measure variants.

**COVID return decomposition (Table 8, p. 1153).** Monthly fund returns decomposed into DIR and non-DIR; each further split into hypothetical (passive) and active components using hand-collected daily security returns. Comparisons made separately for outbreak (Feb-Mar 2020) and recovery (Apr-Jun 2020) periods. Long index - all others DIR gap = -84.71\*\*\* bps during outbreak.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| SEC Form N-PORT (monthly, quarterly) | Primary data: derivative holdings, notional amounts, monthly realized and unrealized PnL by instrument; fund total net assets; portfolio weights | No page yet |
| CRSP Mutual Fund Database (2010-2022) | Extended sample for performance and flow analysis; fund characteristics and returns; derivative identification via holdings | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| Morningstar Direct | Fund reported benchmarks (Lipper investment styles) | No page yet |
| SEC EDGAR Form N-1A (prospectus) | Principal Investment Strategy section; textual analysis of derivative-related discussions and keywords | [SEC EDGAR](/wiki/datasets/edgar/) |
| Yahoo Finance and Bloomberg (hand-collected) | Daily security-level returns for individual derivative positions; matched to N-PORT security names manually | No page yet |
| County-level COVID-19 statistics (New York Times) | Pandemic severity measures for cross-sectional variation analysis (Section IA.2) | No page yet |

Sample: N-PORT primary sample July 2019 to December 2022 (3,106 active domestic equity funds, 1,079 derivative users). Extended CRSP sample 2011-2022.

## When to read the full paper

Read the [original](https://doi.org/10.1093/rfs/hhaf001) if you are studying: (i) mutual fund derivative regulation, since the paper documents that most amplification is unhedged and questions whether derivative access benefits investors; (ii) constructing fund classification schemes using N-PORT data (Section 3 with the K-Means approach); (iii) measuring how derivatives affect fund tracking error during crises (Figures 6 and 9 in the source); or (iv) analyzing the flow-performance puzzle for derivative-using funds, particularly the risk-shifting channel evidence (Section 5 and Table 10). The Internet Appendix contains additional cross-sectional variation tests (SAH orders, industry concentration) and robustness checks for the COVID analysis.

## Attribution and rights

Source: peer-reviewed, *The Review of Financial Studies* 38(4), 2025, pp. 1120-1166. Published by Oxford University Press on behalf of The Society for Financial Studies. All rights reserved. Standard OUP publication-reuse rights; not CC-licensed. This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**.

> Kaniel, Ron, and Pingle Wang. "Unmasking Mutual Fund Derivative Use."
> *The Review of Financial Studies* 38, no. 4 (2025): 1120-1166.
> DOI: 10.1093/rfs/hhaf001.
> Replication code: Harvard Dataverse, https://doi.org/10.7910/DVN/TQCGER.
> Extract-only; paywalled source.


==============================================================================
# How to Dominate the Historical Average: Li, Li, Lyu & Yu (2025)
# https://instituteforautomatedresearch.org/wiki/papers/rfs/2025/li-dominate-historical-average-2025/
# Distilled: Proposes a conservative-slope forecast for the equity premium that sets the predictive slope to a small positive constant (1/A), reducing bias relative to the historical average while matching its zero estimation variance, and proves ex ante that this forecast first-order stochastically dominates the historical average whenever the population predictive slope is nonzero. Review of Financial Studies 2025, CC BY-NC-ND 4.0. Seven core results with source locators, datasets used, the theoretical framework, and the empirical method.
# Tags: paper-summary, asset-pricing, return-forecasting, equity-premium, time-series
==============================================================================

**What this is.** The paper's core results, the theoretical framework (bias-variance trade-off in OOS forecasting, first-order stochastic dominance theorems), and the method (conservative-slope forecast with parameter A): enough to understand what it found and how, without reading all 31 pages. To replicate or extend it, read the full source at the [original](https://doi.org/10.1093/rfs/hhaf010).

## TL;DR

The paper proposes an OOS equity premium forecast: instead of setting the predictive slope to zero (historical average) or estimating it by OLS, use a small positive constant slope $$\delta = 1/A$$ (where A is a large positive number calibrated to the lower confidence bound of the estimated slope). The method has zero estimation variance, matching the historical average, but a lower bias when the population slope is nonzero. The paper proves theoretically (Theorems 1-4) that this forecast first-order stochastically dominates the historical average, and shows empirically on 23 predictors from Goyal and Welch (2008) that 15 of 23 generate significantly positive OOS $$R^2$$ at the 90% level. Goyal, Welch, and Zafirov (2024) confirmed the Goyal and Welch (2008) findings using a larger predictor set, providing the direct motivation for the paper. The dividend-to-price ratio achieves an OOS $$R^2$$ of 2.1% (p = .019) at A = 100, versus an insignificant 0.2% for OLS. Clark and West (2006) show finite-sample estimation noise makes OLS $$R^2$$ negative under the null of no predictability; the proposed method avoids this by using a constant slope with zero variance.

## Core results

Magnitudes and significance are as reported; `\*`/`\*\*` = 10%/5%. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Method OOS $$R^2$$ for dp predictor (A=50): statistically significant improvement over HM | Table 3, p. 3110 | $$R^2 = 3.4\%$$, p-value = .044; OLS $$R^2 = 0.2\%$$, p-value = .477 |
| R2 | Method OOS $$R^2$$ for dp predictor (A=100): gains statistical power as A increases | Table 3, p. 3110 | $$R^2 = 2.1\%$$, p-value = .019; OLS $$R^2 = 0.2\%$$, p = .477; CT++ $$R^2 = 1.8\%$$, p = .286 |
| R3 | Method OOS $$R^2$$ for dp predictor (A=500, A=1,000): very conservative slopes still beat HM at 99% significance | Table 3, p. 3110 | A=500: $$R^2 = 0.5\%$$, p=.009; A=1,000: $$R^2 = 0.2\%$$, p=.008 |
| R4 | Across 23 predictors, 15 (8) have positive OOS $$R^2$$ at 90% (95%) significance | §5 / Internet Appendix C, p. 3090-3091 | 15 of 23 at 90%; 8 of 23 at 95%; OLS and CT generate statistically insignificant $$R^2$$ for most (Goyal and Welch 2008) |
| R5 | Method first-order stochastically dominates HM for dp predictor: empirical CDF of MSE everywhere above HM CDF | Figure 6, p. 3112 | A=100 CDF (MSE) > HM CDF for all MSE thresholds in 40-year rolling windows; confirmed with kernel smoothing in Internet Appendix C.5 |
| R6 | Simulations confirm method's OOS $$R^2$$ distribution is entirely to the right of zero; OLS can be negative | Figure 3, p. 3105 | A=50 centered at ~4-5 (broadest), A=100 at ~3-4, A=200 at ~2, A=500 at ~1 (narrowest spike); OLS distribution spans $[-10, +10]$ with nontrivial probability of $$R^2 < 0$$ |
| R7 | Previously published confidence bounds (Campbell and Shiller 1988a) add value to OOS forecasts when used as the predictive slope | §5.4, p. 3112-3113 | A=209 (95% lower bound from Campbell and Shiller 1988a) produces positive OOS $$R^2$$ from 1987 onward; demonstrates prior study estimates are not data mining |

**Overall (paper's conclusion).** A conservative deterministic predictive slope, calibrated to a lower confidence bound near zero, provably dominates the historical average and empirically dominates OLS and Campbell-Thompson forecasts on most of the 23 standard predictors from Goyal and Welch (2008). The method is an ex ante validated benchmark for time-varying expected return models.

## Theory / model

The paper models the equity premium return as a linear predictive relationship (Equation 1, p. 3092):

$$
r = \mu + bx + e, \tag{1}
$$

where $$b$$ is the population predictive slope coefficient of the predictor $$x$$, $$e$$ is a residual with zero mean uncorrelated with $$x$$, $$x$$ has zero mean, and $$\mu$$ is the unconditional expected return. The historical average sets the slope on $$x$$ to zero. The unknown $$\mu$$ is estimated by the historical average $$\hat{\mu}$$:

$$
\mu = \hat{\mu} + \xi, \tag{2}
$$

where $$\xi$$ has zero mean. Substituting gives (Equation 3, p. 3093):

$$
r = \hat{\mu} + bx + \epsilon, \quad \epsilon \equiv e + \xi. \tag{3}
$$

The method's forecast for the next return is $$\hat{\mu} + \delta x$$ where $$\delta = 1/A$$ if $$b > 0$$ and $$\delta = -1/A$$ if $$b < 0$$ (Equation 5, p. 3093):

$$
\delta = \begin{cases} 1/A & \text{if } b > 0 \\ -1/A & \text{if } b < 0. \end{cases} \tag{5}
$$

**Theorem 1** (p. 3092): Under conditions that $$\delta$$ is a constant between 0 and $$b$$ and the pdf of the error vector $$\mathbf{e}$$ is strictly decreasing in $$\|\mathbf{e}\|$$, the forecast $$\mu + \delta x$$ first-order stochastically dominates the forecast $$\mu$$ (population mean) for predicting return $$r$$. That is, applying loss $$-\text{MSE}$$, the forecast $$\mu + \delta x$$ gives at least as high a probability of achieving any MSE threshold, and strictly higher probability for some.

**Theorem 2** (p. 3093): The forecast $$\hat{\mu} + \delta x$$ first-order stochastically dominates the historical average $$\hat{\mu}$$ under the same conditions on $$\delta$$ and the distribution of $$\boldsymbol{\epsilon} = e + \xi$$. Since $$\xi$$ can correlate with $$x$$ (Stambaugh 1999 bias), the monotonicity condition on the pdf of $$\boldsymbol{\epsilon}$$ is imposed (the $$t$$ and normal distributions satisfy this).

**Corollary 1** (p. 3093): Under Theorem 2's assumptions, the MSE using $$\hat{\mu} + \delta x$$ satisfies:

$$
\text{MSE}_{\hat{\mu}+\delta x} \overset{d}{=} \text{MSE}_{\hat{\mu}} - \eta, \quad \eta \geq 0. \tag{4}
$$

The MSE improvement $$\eta$$ is a nonnegative random variable, so the method has weakly lower MSE in expectation and first-order stochastically lower MSE.

**Theorem 3** (p. 3095) extends stochastic dominance to the case where the sign of $$b$$ is inferred with error (probability $$p$$ correct, $$q$$ wrong). For a given predictor realization $$\mathbf{x}$$, the forecast $$\mu + dx$$ (where $$d$$ is $$\delta$$, $$-\delta$$, or 0 according to the sign inference outcome) first-order stochastically dominates the historical average when $$p/q > \max(B_1, B_2)$$, where $$B_1 > (2b+\delta)/(2b-\delta)$$ (approximately 1 when $$\delta/b$$ is near zero, and 3 at the maximum $$\delta/b = 1$$). Statistical significance at the 95% level gives $$p/q \geq 0.95/0.05 = 19$$, well above the cutoff of 3.

**Theorem 4** (p. 3097) restates Theorem 3 for the estimated-mean setting where $$\mu$$ is replaced by $$\hat{\mu}$$, yielding Corollary 2 (p. 3098, Equation 9):

$$
\text{MSE}_{\hat{\mu}+dx} \overset{d}{=} \text{MSE}_{\hat{\mu}} - \eta, \quad \eta \geq 0. \tag{9}
$$

**OOS MSE decomposition.** The expected MSE difference between the historical average and the method is (Equation 10, p. 3098):

$$
E\!\left[\text{MSE}_{\hat{\mu}} - \text{MSE}_{\hat{\mu}+\hat{b}x}\right]
= \left(b^2 - \text{Bias}(\hat{b})^2 - \text{Variance}(\hat{b})\right)x^2 - 2E\!\left[(\hat{\mu}-\mu)\hat{b}\right]x. \tag{10}
$$

The historical average's slope is zero and therefore unbiased but uses no predictive information. A regression slope reduces the first two terms but can inflate the variance term to the point where it dominates, yielding a negative OOS $$R^2$$. The method uses a deterministic $$\delta$$, so the variance of $$\hat{b}$$ is zero and Equation (10) simplifies to $$(b^2 - \text{Bias}(\hat{b})^2)x^2 > 0$$ whenever $$\hat{b}$$ is between 0 and $$b$$. This is the core intuition: a constant nonzero slope beats both the historical average (zero slope, biased) and OLS (unbiased mean but high variance).

**Gradient descent interpretation.** The method is a one-step gradient descent update of the historical average toward greater predictability, using the sign (but not the magnitude) of $$b$$ as the gradient signal and step size $$1/A$$ (Equation 11, p. 3099):

$$
a_1 = a_0 - \gamma f'(a_0), \tag{11}
$$

where $$a_0 = 0$$ (the historical average slope), $$a_1 = \delta$$, and $$\gamma = 1/A$$.

## Method

The implementation has three steps.

**Step 1: Obtain the sign of $$b$$.** Sign can come from (a) economic theory, as in Campbell and Thompson (2008), who restrict the OLS slope to have the theoretically expected sign, or (b) statistical inference: use the confidence interval $$[\hat{b}_L, \hat{b}_U]$$ for population slope $$b$$. If $$0 < \hat{b}_L$$ the slope is significantly positive; if $$\hat{b}_U < 0$$ it is significantly negative. The method builds on `time-series-forecasting` (predictive regression) but replaces the OLS slope with a constant.

**Step 2: Choose A.** Setting $$1/A$$ to the lower confidence bound $$\hat{b}_L$$ ensures with near certainty that $$\delta$$ is between 0 and $$b$$ (Equation 7, p. 3093):

$$
0 < \hat{b}_L \leq b \quad \text{with 95\% probability.} \tag{7}
$$

For the dividend-to-price ratio, Campbell and Shiller (1988a) Table 4 give a predictive slope of 0.129 (SE = 0.057); the standardized predictor has a standard deviation of 0.277, so the 95% lower confidence bound for the standardized slope is $$(0.129 - 1.96 \times 0.057) \times 0.277 = 0.0048$$, implying $$A = 209$$ (p. 3094).

**Step 3: Standardize the predictor.** Predictors are standardized to zero mean and unit variance using a 20-year rolling backward-looking window, so $$\delta = 1/A$$ measures the effect of a one-standard-deviation change in the predictor on the forecast annual return (p. 3094, 3107).

**Simulation design.** A VAR(1) for log returns $$r$$, log dividend-to-price ratio $$dp$$, and log dividend growth $$\Delta d$$ following Cochrane (2008) provides the data-generating process (Equations 12-13, p. 3101). Parameters are calibrated to the sample (Table 1, p. 3102): population predictive slope $$b_r = 0.143$$, $$\rho = 0.8911$$. The historical average (HM) uses a rolling 20-year window. OOS $$R^2$$ is computed per Equation 14 (p. 3103):

$$
R^2 = \frac{\text{MSE}_{\hat{\mu}} - \text{MSE}_{\hat{\mu}+\delta x}}{\text{MSE}_{\hat{\mu}}}. \tag{14}
$$

## Empirical specifications

**Data.** Annual value-weighted CRSP market returns (post-1926) and S&P 500 Index returns (pre-1926), 23 predictors from Goyal and Welch (2008) extended through 2017 (Table 2, p. 3105): bm, cape Shiller, cay, corpr, csp, de, dfr, dfy, dp, dy, ep, eqis, ik, infl, ltr, lty, tbl, tms, ntis, svar, and four Robert Shiller series. Predictors are standardized using rolling 20-year windows. Forecasts start 20 years after the sample start year for each predictor.

**OOS $$R^2$$ evaluation (R1-R4, R7).** For each predictor, compute annual one-step-ahead OOS forecasts using the method with $$A \in \{50, 100, 200, 500, 1000\}$$, OLS, CT+, and CT++ (Campbell and Thompson 2008). Compute $$R^2$$ via Equation (14). Report one-sided p-values using Diebold (2015) heteroscedasticity-adjusted test, cross-checked by Harvey, Leybourne, and Newbold (1997). For dp, this corresponds to a sample of 146 annual observations.

**Stochastic dominance evaluation (R5).** Estimate the empirical CDF of OOS MSE over 40-year rolling windows of annual successive one-year-ahead forecasts using dp predictor (Figure 6, p. 3112). Compare the method CDF (A=100) to the historical mean CDF. First-order dominance requires the method CDF to lie everywhere above (to the left of) the HM CDF. Kernel smoothing (Internet Appendix C.5) confirms the finding.

**Bias-variance simulation (R6).** Generate 10,000 simulation samples of the VAR in Equations (12)-(13) with parameters from Table 1. In each sample compute the OOS $$R^2$$ using a rolling 20-year window for HM and a rolling 20-year window for $$\delta = 1/A$$ for the method. Figure 3 (p. 3105) reports the kernel density of the $$R^2$$ distribution across simulations for $$A \in \{50, 100, 200, 500\}$$ and OLS.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| CRSP value-weighted index return | Annual market return post-1926 | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| S&P 500 Index returns (Shiller website) | Annual market return pre-1926 | [Shiller data](/wiki/datasets/shiller-data/) |
| Goyal and Welch (2008) predictor data (Amit Goyal's website) | 19 predictor series for equity premium forecasting, extended to 2017 | No page yet |
| Robert Shiller data (http://www.econ.yale.edu/~shiller/data.htm) | 4 additional predictors: cape Shiller, infl Shiller, lty Shiller, Trcape Shiller | [Shiller data](/wiki/datasets/shiller-data/) |

Sample: 23 predictors with start years ranging from 1872 to 1947 (Table 2, p. 3105), all ending 2017. Forecasting starts 20 years after the predictor start year. Frequency: annual.

## When to read the full paper

Use the [original](https://doi.org/10.1093/rfs/hhaf010) if you are: (a) constructing a competing equity premium forecast and need the formal ex ante dominance proofs for a given distributon assumption; (b) selecting the parameter A for a specific predictor using the confidence-bound rule (Internet Appendix C.1 contains calculations for all 23 predictors); (c) investigating whether previously published coefficient estimates add OOS value (Section 5.4); or (d) comparing to shrinkage estimators such as Ridge and Lasso (Internet Appendix C.6). The key tables are Table 3 (OOS $$R^2$$ for dp) and Internet Appendix Table C.1 (all predictors); key figures are Figure 1 (cumulative OOS performance), Figure 3 (simulation $$R^2$$ density), and Figure 6 (stochastic dominance CDF).

## Attribution and rights

Source: peer-reviewed, *The Review of Financial Studies* 38(10). This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The CC BY-NC-ND 4.0 licence permits non-commercial reproduction with attribution and no derivatives; the verbatim PDF is not hosted here.

> Li, Kai, Yingying Li, Changlei Lyu, and Jialin Yu. "How to Dominate the Historical Average."
> *The Review of Financial Studies* 38, no. 10 (2025): 3086-3116.
> DOI: [10.1093/rfs/hhaf010](https://doi.org/10.1093/rfs/hhaf010).
> Replication code: Harvard Dataverse, https://doi.org/10.7910/DVN/9VNJUN.
> Licensed under [CC BY-NC-ND 4.0](https://creativecommons.org/licenses/by-nc-nd/4.0/).
> This page is a distilled extract by the Institute for Automated Research.


==============================================================================
# Real Effects of Centralized Markets: Martin (2025)
# https://instituteforautomatedresearch.org/wiki/papers/rfs/2025/martin-real-effects-centralized-markets-2025/
# Distilled: Using staggered NYMEX steel futures introductions (2008, 2012) as natural experiments in a difference-in-differences framework, this paper finds that centralizing derivative markets reduces price dispersion in the physical product market by 6 pp (CV), lowers product prices by 3-4%, increases producer hedging, shifts market share toward low-cost firms, and reduces producer operating profits by 1.6-1.9 pp. Review of Financial Studies 2025, CC BY 4.0. Seven core results with source locators, datasets used, hypotheses, and the empirical specifications.
# Tags: paper-summary, derivatives, commodity-markets, market-microstructure
==============================================================================

**What this is.** The core results, hypotheses, and empirical specifications of this paper, distilled from the PDF. To replicate or extend, read the full source at the [original](https://doi.org/10.1093/rfs/hhaf011) or access the replication code at the [Harvard Dataverse](https://doi.org/10.7910/DVN/FHSYBL).

## TL;DR

The paper asks whether centralizing derivative markets has real effects on the underlying product markets. It exploits two staggered introductions of NYMEX steel futures contracts in the United States: hot-rolled coil (HRC) futures in October 2008 and busheling scrap (BUS) futures in September 2012. Using a difference-in-differences strategy comparing treated steel products (HRC, BUS) to similar untreated products (cold-rolled coil, plates, heavy-melting scrap, shredded scrap), the paper finds that futures markets: (1) reduce physical product price dispersion by about 6 percentage points (CV), consistent with futures prices acting as public reference prices; (2) increase producer hedging of commodity price risk; (3) make market shares more sensitive to production costs, reallocating share toward low-cost producers; (4) reduce product prices by 3-4%; and (5) reduce producer operating profits by 1.6-1.9 percentage points and stock market valuations by 4-5%. The results are consistent with centralized futures markets fostering competition in the underlying product market through improved price transparency and risk management.

## Core results

Magnitudes and significance are as reported; `\*\*` = 5%, `\*\*\*` = 1% level. Locators point into the source PDF.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | **Futures introduction reduces price dispersion (CV)** by 6 pp for treated relative to control steel products | Table 2 Panel B col 1, p. 2158 | Coefficient on Post x Futures_product = -0.057\*\*\* (SE 0.011); stable across cols 2-5 with demand, supply, trade controls |
| R2 | **Futures introduction reduces price dispersion (SD)** by ~$39/ton for treated products | Table 2 Panel A col 1, p. 2158 | Coefficient = -38.636\*\*\* (SE 7.737); stable across specifications; parallel pre-trends confirmed (Figure 2) |
| R3 | **Treated producers significantly increase commodity hedging** after futures introduction | Table 3 col 1-3, p. 2161 | Coefficient on Post x Futures_firm = 0.258\*\*\* (0.089), 0.248\*\*\* (0.091), 0.215\*\* (0.096); parallel pre-trends (Figure 3) |
| R4 | **Market shares become more sensitive to costs** after futures: EAF producers gain 0.5-0.8 pp more share per 10% iron-ore/scrap price increase | Table 4 col 1-3, p. 2163 | Post x Futures_firm x EAF x Iron/Scrap = 0.054\*\*\* (0.016), 0.078\*\*\* (0.023), 0.073\*\*\* (0.020) |
| R5 | **Product prices fall 3-4%** for treated relative to control products after futures introduction | Table 5 col 1-5, p. 2164 | Post x Futures_product: -0.032\*\*\* (0.005) to -0.040\*\*\* (0.007); significant at 1% with full controls; parallel pre-trends (Figure 4) |
| R6 | **Producer operating profits fall 1.6-1.9 pp** for treated relative to control firms after futures introduction | Table 6 col 1-7, p. 2166 | Post x Futures_firm: -0.016\*\*\* (0.004) to -0.022\*\*\* (0.005); significant at 1% across all seven specifications; parallel pre-trends (Figure 5) |
| R7 | **Producer stock prices fall 4-5%** around news events increasing likelihood of a futures contract | Table 7 col 1-4, p. 2169 | Futures_firm coefficient on CAR_{-2,+2}: -0.049\*\*\* (0.010) to -0.037\*\*\* (0.012) across market-adj., CAPM, 3-factor, 4-factor models |

**Overall (paper's conclusion).** Centralized futures markets reduce price dispersion and producer markups in the physical steel product market, increase cost-sensitivity of market shares, lower prices, and compress producer profits and valuations. The results are consistent with two channels: futures prices as public reference prices (improving buyer search and competition, as in Grennan and Swanson 2020 on hospital-supplier pricing) and improved risk management (relaxing financial constraints and enabling aggressive market-share investment by low-cost producers, following Perez-Gonzalez and Yun 2013 on weather derivatives and firm investment). Both channels increase product market competition.

## Theory / model

The paper has no formal structural model. Instead it derives testable hypotheses from two theoretical channels in the literature.

**Reference price channel.** Following Janssen, Pichler, and Weidenholzer (2011) and Duffie, Dworczak, and Zhu (2017), in decentralized search markets buyers have limited information about prices offered by other sellers. Price dispersion is a manifestation of this ignorance (Stigler 1961). When a futures market publishes reference prices, buyers can compare a seller's offered price against the publicly observed futures benchmark, improving their bargaining position and enabling more effective search. The equilibrium predictions are:

- Price dispersion decreases.
- Market shares become more sensitive to production costs (buyers identify low-cost sellers more easily).
- Average prices fall (sellers reduce markups).

**Risk management channel.** Futures markets may improve or impair producers' hedging ability. The net effect is theoretically ambiguous: centralization lowers counterparty risk and increases liquidity (Telser and Higinbotham 1977; Telser 1981; Vuillemey 2020), helping hedging. But increased price transparency may reduce risk-sharing opportunities (Hirshleifer 1971; Goldstein and Yang 2022), harming hedging. Empirically, the paper finds net hedging increases. Following Froot, Scharfstein, and Stein (1993) and Chevalier and Scharfstein (1996), improved hedging stabilizes cash flows, enabling otherwise liquidity-constrained firms to invest in market share by lowering prices. The prediction: if hedging improves, prices fall further and market shares reallocate toward low-cost producers.

**Identification logic.** The key identifying assumption is a parallel trends condition: treated products (HRC, BUS) would have followed the same trends as control products (cold-rolled coil, plates, heavy-melting scrap, shredded scrap) absent the futures introductions. Three facts support this: (1) treated and control products exhibit comparable ex ante price volatility, a requirement for futures viability; (2) parallel pre-trends in outcomes for products and firms; (3) placebo tests show no differential evolution for non-U.S. producers of treated products following the U.S.-targeted introductions.

## Method

The headline empirical design is a stacked difference-in-differences (DiD) exploiting the two staggered NYMEX futures introductions (HRC in October 2008, BUS in September 2012). For each introduction, the paper constructs a symmetric event window of 40 publication dates (product-level) or 9 years (firm-level) around the futures start date, then stacks the two panels and estimates a single DiD coefficient.

**Product-level specification** (equation 3.1, p. 2155, price dispersion and price):

$$
\text{PriceDispersion}_{k,p,t} = \beta \cdot \text{Post} \cdot \text{Futures}_{product,k} + \alpha_{k,p} + \alpha_{k,t} + \varepsilon_{k,p,t} \tag{3.1}
$$

where $$k$$ indexes the futures introduction (HRC, BUS), $$p$$ indexes the steel product, $$t$$ indexes the publication date. $$\text{Futures}_{product}$$ equals one for hot-rolled coils (HRC introduction) and busheling scrap (BUS introduction). $$\text{Post}$$ equals one after trading begins. $$\alpha_{k,p}$$ and $$\alpha_{k,t}$$ are product and publication-date fixed effects specific to each introduction event. Standard errors are clustered by publication date. The coefficient $$\beta$$ measures the change in price dispersion for treated versus control products after futures introduction.

**Firm-level hedging specification** (equation 4.1, p. 2160):

$$
\text{Hedge}(1/0)_{k,i,y} = \beta \cdot \text{Post} \cdot \text{Futures}_{firm,k,i} + \sum_{\tau=-4}^{4} \left( \theta_\tau' X_{k,i} \right) \mathbf{1}\{y=\tau\} + \alpha_{k,i} + \alpha_{k,y} + \alpha_{k,j,y} + \varepsilon_{k,i,y} \tag{4.1}
$$

where $$i$$ indexes firms, $$y$$ indexes years, $$j$$ indexes 3-digit NAICS industries. $$\text{Futures}_{firm}$$ equals one for HRC producers (BUS introduction: ferrous scrap sellers). Baseline controls (log assets, firm age, sales growth) are measured at the last pre-introduction quarter and interacted with year fixed effects. Standard errors clustered by firm.

**Market share cost-sensitivity specification** (equation 4.2, p. 2162):

$$
\text{MarketShare}_{i,q} = \beta \cdot \text{Post} \cdot \text{Futures}_{firm,i} \cdot \text{EAF}_i \cdot \text{Iron/Scrap}_q + \alpha_i + \alpha_q + \alpha_{j,q} + \varepsilon_{i,j,q} \tag{4.2}
$$

where $$q$$ indexes year-quarters, $$\text{EAF}_i$$ indicates electric arc furnace producers who benefit from cheap scrap, and $$\text{Iron/Scrap}_q$$ is the quarterly price ratio of iron ore to scrap prices. $$\beta$$ measures the change in EAF producers' market-share sensitivity to input price variation after futures introduction.

**Event study specification** (equation 4.3, p. 2168):

$$
\text{CAR}_{i,e} = \beta \cdot \text{Futures}_{firm,i,e} + \alpha_e + \alpha_i + \varepsilon_{i,e} \tag{4.3}
$$

where $$e$$ indexes news events related to futures introductions, and CAR is measured over the 5-day window $$[-2, +2]$$ around the event using market-adjusted returns, CAPM, Fama and French (1993) three-factor, and Carhart (1997) four-factor models.

The builds on `difference-in-differences`, `panel-regression`, and `event-study` technique primitives. The identification rests on `natural-experiment`: the staggered NYMEX product selection decisions are treated as quasi-exogenous events (exchanges chose which products based on liquidity-versus-basis-risk trade-offs, not potential real-economy externalities).

## Empirical specifications

**Price dispersion (R1, R2).** Panel: 19,653 product-publication-date observations. Window: 40 publication dates (SteelBenchmarker biweekly releases) before and after each introduction, stacked across HRC and BUS. Outcomes: SD(Price) and CV(Price) across reporting firms per product-date. Robustness: columns 2-5 add domestic demand controls (GDP growth, key-sector output interacted with Futures_product), supply controls (U.S. production growth, capacity utilization), and trade controls (import growth rates), all interacted with the treatment indicator. Standard errors clustered by publication date. Dynamic coefficient plot (Figure 2) confirms no pre-trend and immediate post-treatment break (p. 2158-2159).

**Producer hedging (R3).** Firm-year panel: 2,993 observations. Window: firms in years $$y=-4$$ to $$y=4$$ relative to each futures start, stacked across HRC and BUS. Outcome: indicator for any mention of commodity derivatives in SEC annual report. Fixed effects: firm, year, and industry-year (3-digit NAICS) interacted with each introduction. Robustness: column 3 adds controls-times-year interactions. Parallel pre-trends confirmed (Figure 3). Profits-steel price correlation (Table A.7) provides additional evidence: treated producers' profits become less correlated with steel prices after futures introduction (pp. 2160-2161).

**Market share cost-sensitivity (R4).** Restricted to HRC introduction (EAF/BOF distinction applies only to raw steel). 1,419 firm-year-quarter observations from Q1 2007 to Q3 2010. Fixed effects: firm, year-quarter, and industry-year-quarter (3-digit NAICS). Controls: log assets, firm age, sales growth measured at last pre-introduction quarter interacted with year-quarter FE. All interactions of Post, Futures_firm, EAF, and Iron/Scrap included. Standard errors clustered by firm (p. 2162-2163).

**Product prices (R5).** Same panel as price dispersion (19,653 observations). Outcome: ln(average price per ton). Specification mirrors equation 3.1. Demand, supply, and trade controls added in columns 2-5. Dynamic plot (Figure 4) shows price decline emerges only post-futures and persists to end of sample (pp. 2163-2164).

**Producer profits (R6).** Firm-year-quarter panel: 5,095 observations. Window: $$q=-7$$ to $$q=7$$ relative to each introduction, stacked. Outcome: operating profit/beginning-of-quarter total assets (Compustat oibdpq/atq). Fixed effects: firm and industry-year-quarter. Seven robustness columns control for: business cycle, iron/scrap price exposure, import competition, industry segment controls, and M&A exclusions (Table 6 cols 2-7). Significant at 1% throughout. Dynamic plot (Figure 5) confirms parallel pre-trends and persistent post-introduction decline (pp. 2165-2167).

**Stock market valuations (R7).** 1,106 firm-event observations across five HRC events (2007-2008) and two BUS events (2012). CAR measured over 5-day window $$[-2, +2]$$ using CRSP daily returns benchmarked against market-adjusted, CAPM, FF3, and Carhart four-factor models. Event-study OLS includes firm and event-date fixed effects. Standard errors clustered by firm. Robustness: results hold for alternative windows $$[d-2, d+3]$$, $$[d-3, d+3]$$; removing any single event; excluding below-median market return days (pp. 2168-2170).

**Placebo tests.** Non-U.S. firms (from Compustat Global/Refinitiv) selling treated products in countries without steel futures show no significant changes in hedging, market share sensitivity, profitability, or CAR around the NYMEX introductions (Table 9, p. 2174-2175). U.S. firms selling products used as controls in the main tests also show no differential reaction (Table A.15).

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| SteelBenchmarker (proprietary price database) | Product-level bi-weekly reported transaction prices for 6 steel products; primary source for price dispersion and price-level tests (R1, R2, R5) | [SteelBenchmarker](/wiki/commercial/steelbenchmarker/) (licensed) |
| Compustat North America Fundamentals Quarterly | Firm accounting data (assets, sales, profitability), identification of treated firms by product description, construction of profit variable (R6) | [WRDS / Compustat](/wiki/commercial/wrds/) (licensed) |
| SEC EDGAR (annual report text) | Firm product descriptions and commodity derivative mentions for hedging analysis (R3); treatment status classification | [SEC EDGAR](/wiki/datasets/edgar/) |
| CRSP daily stock returns | Stock prices for CAR computation (R7) and treatment status confirmation | [WRDS / CRSP](/wiki/commercial/wrds/) (licensed) |
| U.S. Geological Survey (USGS) | Steel production quantities by product for production quantity tests and trade controls | No page yet |
| Bureau of Economic Analysis (BEA) | Quarterly GDP growth and key steel-consuming sector output for demand controls | No page yet |
| PPI for Iron Ore and Steel Scrap (BLS) | Iron ore-to-scrap price ratio for market-share cost-sensitivity test (R4) | No page yet |
| Compustat Global / Refinitiv | Non-U.S. firm data for placebo tests | No page yet |

Sample (firm-level): Compustat North America, 2003-2017. Sample (price-level): SteelBenchmarker, January 2007-December 2017. All firm-level variables winsorized at 1st and 99th percentiles.

## When to read the full paper

Use the [original](https://doi.org/10.1093/rfs/hhaf011) if you are:
studying real effects of financial market innovations on product markets (Tables 2-7 with full robustness); examining how price transparency and risk management interact as channels for competition; replicating the DiD or event-study design (replication code at the [Harvard Dataverse](https://doi.org/10.7910/DVN/FHSYBL)); generalizing findings to other industries with similar market structure (Section 5.2); or assessing the external validity of the parallel trends assumption (Figures 2-5, Table A.3-A.5).

## Attribution and rights

Source: peer-reviewed, *The Review of Financial Studies* 38(7), 2025. This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

> **Attribution (CC BY 4.0).** Martin, Thorsten. "Real Effects of Centralized Markets: Evidence from Steel Futures."
> *The Review of Financial Studies* 38, no. 7 (2025): 2140-2181.
> DOI: 10.1093/rfs/hhaf011. © 2025 The Author(s).
> Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
> This page is an **adaptation** by the Institute for Automated Research:
> core results extracted and re-expressed; **changes were made**.


==============================================================================
# Social Connectedness in Bank Lending: Rehbein & Rother (2025)
# https://instituteforautomatedresearch.org/wiki/papers/rfs/2025/rehbein-social-connectedness-bank-lending-2025/
# Distilled: Using Facebook's Social Connectedness Index, Rehbein and Rother show that bank lending volumes, borrower-friendly loan terms, and bank profitability all increase with social connectedness between bank and borrower counties, while fintech lending is unaffected. Review of Financial Studies 2025, paywalled. Eight core results with source locators, datasets used, the empirical model, and three formal specifications.
# Tags: paper-summary, banking, credit-supply, social-networks, geographic-lending
==============================================================================

**What this is.** The paper's core results, the economic hypotheses (information channel vs. favoritism), and the main empirical specifications with equations: enough to understand what was found and how, without reading all 51 pages. To replicate or extend it, see the full source at the [original](https://doi.org/10.1093/rfs/hhaf014) and the replication archive at the Harvard Dataverse.

## TL;DR

Rehbein and Rother exploit geographic variation in the Facebook Social Connectedness Index (SCI) to show that bank lending is shaped by the social ties between bank and borrower counties. A 10% increase in social connectedness between counties raises cross-county SME loan volumes by roughly 6-9% and mortgage loan volumes by a similar margin, after controlling for physical distance, cultural dissimilarity, and a broad set of geographic and economic factors. The relationship is stronger when screening incentives are high (e.g., no government guarantee, not securitized) and is absent for fintech lenders whose decisions are algorithm-based. Loans to high-connectedness borrowers carry lower interest rates, lower LTV ratios, and lower delinquency and default rates. Banks with more socially connected loan portfolios have higher ROA and ROE. At the aggregate level, borrower counties more socially proximate to bank regions receive more lending and experience higher GDP growth and employment, especially if they are small-firm intensive, an effect confirmed by a shale-boom natural experiment.

## Core results

Magnitudes and significance are as reported in the paper; `\*` = 10%, `\*\*` = 5%, `\*\*\*` = 1%.

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | County-to-county **SME loan volume increases strongly with social connectedness**; the relationship persists across a broad array of geographic and economic controls | Table 2, col. 1 (baseline); Table 3, col. 11 (full controls), pp. 2773-2775 | Baseline elasticity 0.91\*\*\* (SE 0.06); with all controls 0.64\*\*\* (SE 0.06); R2 rises from 0.70 to 0.91, leaving little room for further omitted-variable attenuation |
| R2 | **Mortgage loan volumes** also strongly increase with social connectedness | Table 2, col. 8, p. 2773 | Elasticity 1.10\*\*\* (SE 0.05); robust to OLS, alternative clustering, HQ-based bank location |
| R3 | High social connectedness is associated with **lower interest rates** on originated mortgages | Table 7, col. 2, p. 2787 | Coefficient on log(SCI) = -0.497\* (SE 0.259); 1-SD increase in log(SCI) lowers rate by ~1 bp |
| R4 | High social connectedness predicts **lower delinquency and default rates** | Table 7, cols. 3-4, p. 2787 | Delinquency coefficient -0.004\*\*\* (SE 0.000); default coefficient -0.002\*\* (SE 0.000); 1-SD change implies 0.8 pp lower delinquency and 0.4 pp lower default |
| R5 | **Bank ROA** increases with portfolio social connectedness | Table 10, Panel A, col. 1, p. 2793 | Coefficient 0.05\*\*\* (SE 0.02); 1-SD increase (1.4 log-units) raises ROA by 0.07 pp; effect larger for heavy SME lenders |
| R6 | **Bank ROE** increases significantly with portfolio social connectedness | Table 10, Panel A, col. 2, p. 2793 | Coefficient 0.74\*\*\* (SE 0.27); 1-SD increase raises ROE by 1.04 pp; nonperforming loans decrease (coefficient -0.35\*\*\*, SE 0.10) |
| R7 | Borrower counties more socially proximate to bank regions experience **higher real GDP growth**, particularly in small-firm counties | Table 12, cols. 2-3, p. 2797 | Baseline coefficient 1.719\*\*\* (SE 0.553); 10% increase in social proximity raises GDP growth by ~0.3 pp for counties at 95th percentile of small-firm share (= 10·(1.660+0.060·19)/100); confirmed with shale-boom IV (Table 13, col. 3) |
| R8 | **Employment** also rises with social proximity to bank regions, concentrated in small-firm counties | Table 12, cols. 4-5, p. 2797; Table 13, cols. 4-5, p. 2800 | 10% higher social proximity: coefficient 0.021\*\*\* (SE 0.006) on log(employment); twice as large at high small-firm-share counties; shale-boom IV confirms effect |

**Overall (paper's conclusion).** Social connectedness between bank and borrower regions is a distinct and economically important dimension of the geography of bank lending. It is not subsumed by physical or cultural distance and explains lending patterns consistent with both an information channel and favoritism, though the bank profitability result suggests that on average information reduction rather than favoritism is the dominant mechanism.

## Theory / model

The paper has no single formal theoretical model. It motivates the analysis with two competing hypotheses.

**Information hypothesis.** Social connections reduce information asymmetries between bank and borrower regions. Loan officers in socially connected counties obtain soft information about local economic conditions (through their own or their networks' acquaintances in the borrower region) and can make better lending decisions. Formally, if social connectedness $$\text{SCI}_{i,j}$$ increases, the precision of banks' private signal about borrowers in region $$j$$ rises. Better-screened borrowers receive lower rates and have lower delinquency; banks earn higher ROA. This prediction aligns with the classical framework of Diamond (1984) and Boot (2000) on delegated monitoring.

**Favoritism hypothesis.** Social connections may also lead to conscious or unconscious preferential treatment of borrowers in connected regions. Haselmann, Schoenherr, and Vig (2018) document rent-seeking in elite networks; the paper considers whether a similar mechanism operates at the population level. Unlike the information channel, favoritism predicts that banks' loan profitability might not improve (and could fall) if resources flow to less creditworthy but socially connected borrowers. The paper tests both channels by examining loan terms, loan performance, and bank profitability jointly (Section 2.6).

**Identification logic.** The social connectedness measure (Facebook SCI, equation 1, p. 2766) is cross-sectional (2016) and predetermined relative to the 2017 lending outcomes in the baseline. It is not randomly assigned, so the authors proceed in two ways. First, they add an extensive set of geographic and economic controls including the physical and cultural distance controls emphasized by Degryse and Ongena (2005) and Agarwal and Hauswald (2010), and use the Oster (2019) coefficient-stability argument to rule out omitted-variable bias. Second, for the real-effects results (Section 5.2), they exploit shale-boom liquidity shocks (Gilje, Loutskina, and Strahan (2016); Gilje (2019)) as a natural experiment: unanticipated increases in deposits at shale-exposed bank branches raise lending potential for banks with branches in boom counties, and counties socially connected to those banks receive more lending, without being directly affected by the boom.

## Method

**Social connectedness measure.** The SCI is introduced by Bailey et al. (2018b) and defined at the county-pair level as (equation 1, p. 2766):

$$
\text{social connectedness}_{i,j} = \frac{\text{number of friendship links}_{i,j}}{\text{population}_i \cdot \text{population}_j} \cdot \text{scaling factor}
$$

It quantifies the relative probability that a person in county $$i$$ is acquainted with a person in county $$j$$. The log of this variable is the main regressor throughout.

**Baseline loan-volume specification.** County-pair-level Poisson pseudo-maximum-likelihood (PPML) regression (equation 2, p. 2770):

$$
\text{volume of loans}_{i,j} = \exp\!\left[\beta \cdot \log(\text{social connectedness})_{i,j} + \gamma_1 \cdot \log(\text{physical distance})_{i,j} + \gamma_2 \cdot \text{cultural distance}_{i,j} + M_{i,j} + \alpha_i + \alpha_j\right] \cdot \epsilon_{i,j}
$$

where $$\alpha_i$$ and $$\alpha_j$$ are bank-county and borrower-county fixed effects, $$M_{i,j}$$ is a vector of county-pair controls (same-state FE, common-border FE, percentile FE for commuting, migration, trade, industry-share differentials, GDP and unemployment differentials, highway travel costs, and flight-and-drive time). Standard errors are clustered at the bank-county and borrower-county levels. PPML is used because loan volumes are non-negative and often zero; it accommodates heteroskedasticity and yields consistent elasticity estimates. The coefficient $$\beta$$ is the elasticity of loan volume with respect to social connectedness.

**Loan-type heterogeneity.** For the screening-incentive tests, the estimation adds county-pair-by-loan-type observations (equation 3, p. 2781):

$$
\text{volume of mortgage loans}_{i,j,k} = \exp\!\left[\beta \cdot \log(\text{social connectedness})_{i,j} \cdot \text{loan type}_k + \text{distance percentile FE}_{i,j} \cdot \text{loan type}_k + \alpha_{i,j}\right] \cdot \epsilon_{i,j,k}
$$

where $$k$$ indexes loan types (guaranteed, securitized, low-LTV, fintech), and county-pair fixed effects $$\alpha_{i,j}$$ absorb all time-invariant pair characteristics.

**Portfolio social connectedness (bank-profitability analysis).** Each bank's portfolio social connectedness is defined as the loan-weighted average of cross-county connectedness (equation 5, p. 2790):

$$
\text{portfolio social connectedness}_{b,y} = \sum_i \sum_j \text{social connectedness}_{i,j} \cdot \frac{\#\,\text{loans}_{b,i,j,y}}{\text{total}\,\#\,\text{loans}_{b,y}}
$$

Bank profitability is then regressed on this measure, extending the Loutskina and Strahan (2011) empirical model of bank performance (equation 6, p. 2791):

$$
\text{profitability}_{b,s,t} = \beta \cdot \log(\text{portfolio social connectedness})_{b,t} + \gamma_1 \cdot \log(\text{portfolio physical distance})_{b,t} + \gamma_2 \cdot \text{portfolio cultural distance}_{b,t} + \gamma_3 \cdot \text{loan concentration}_{b,t} + \gamma_4 \cdot \text{further bank controls}_{b,t} + \alpha_s + \alpha_t + \epsilon_{b,s,t}
$$

**Social proximity to banks (real-effects analysis).** Equation 7 (p. 2794) weights total bank assets in county $$i$$ by social connectedness:

$$
\text{social proximity to banks}_{j,t} = \sum_i \text{social connectedness}_{i,j} \cdot \text{total bank assets}_{i,t}
$$

Real outcomes in county $$j$$ are regressed on lagged log(social proximity to banks), controlling for physical and cultural proximity to banks, county and year fixed effects, and industry-share controls (equation 8, p. 2796).

**Shale-boom IV.** Bank exposure to shale-boom liquidity windfalls (equations 9-11, pp. 2799-2800) instruments lending potential. County $$i$$'s boom exposure is the share-weighted average of its banks' shale-boom exposures; county $$j$$'s social proximity to shocked banks replaces social proximity to all banks in the IV regressions (Table 13).

## Empirical specifications

**Section 2 (loan volumes, county pairs):** PPML on the full 9.1-million-pair cross-section (2017 lending data, 2016 SCI). Main estimates: Table 2 (baseline, columns 1-8), Table 3 Panel A (progressive controls, columns 1-11 for SME loans; columns 12 for mortgages). Heterogeneity by county type: Table 4 (five county-type interactions, PPML, N = 7,022,043). Heterogeneity by loan type: Table 5 (fintech vs. traditional banks, guaranteed, securitized, low-LTV; PPML, N = 11,870,270 for mortgage-by-type pairs). All columns include bank-county and borrower-county FE; columns with controls add same-state, common-border, and percentile FE.

**Section 3 (loan terms and performance, loan level):** OLS on 1,268,200 Fannie Mae/Freddie Mac 30-year fixed-rate mortgages originated 2000-2008, observed until 2018 (equation 4, p. 2785). Outcomes: LTV (col. 1), interest rate in bp (col. 2), delinquency indicator (col. 3), default indicator (col. 4). Controls include FICO score, DTI, first-time buyer indicator, log(loan amount), LTV (for performance columns), bank origination-year and borrower-state origination-year FE, and physical/cultural distance percentile FE. Standard errors clustered at bank-county and borrower-county.

Interest rate dispersion (Section 3.3): OLS at the county-pair level, dependent variable is the within-pair SD of interest rates (Table 8); controls include same-state, common-border, and percentile FE for physical and cultural distance. N = 6,999 county pairs with at least four loans.

**Section 4 (bank profitability, bank-quarter level):** OLS on 18,914 bank-quarter observations (844 banks, 2009-2017). Outcomes: ROA, ROE, % NPL (Table 10 Panels A and B). Controls: bank-state and year FE, log(portfolio physical distance), portfolio cultural distance, loan concentration, and further bank-level controls (log total assets, securities/assets, real estate loans/assets, C&I loans/assets, unused commitments/assets, letters of credit/assets, deposits/assets, interest expenses/deposits).

**Section 5 (real effects, county-year level):** OLS on 3,021 counties, 2009-2017 (N = 24,152-24,161). Outcomes: log(loan volume), real GDP growth, log(employment) (Table 12). Shale-boom IV regressions replace social proximity to all banks with social proximity to shocked banks (Table 13, N = 22,047-22,053). Controls: county and year FE, log(physical and cultural proximity to banks), industry-share controls, commuting/migration controls for employment. Standard errors clustered at the county level.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Facebook Social Connectedness Index (Bailey et al. 2018b) | Main explanatory variable; county-pair relative probability of Facebook friendship | [Facebook SCI](/wiki/datasets/facebook-sci/) |
| Community Reinvestment Act (CRA) data, FFIEC | County-to-county SME loan volumes for 2017 (and 2004-2018 for time series); bank-level loan counts | [CRA (FFIEC)](/wiki/datasets/cra-ffiec/) |
| Home Mortgage Disclosure Act (HMDA) data, FFIEC | County-to-county mortgage loan volumes for 2017; loan-type classification | [HMDA](/wiki/datasets/hmda/) |
| Fannie Mae and Freddie Mac Single Family Loan-Level Datasets | Loan-level mortgage data (2000-2008 originations, observed through 2018): LTV, interest rate, FICO, DTI, delinquency, default | [Fannie / Freddie loan-level](/wiki/datasets/fannie-freddie/) |
| FDIC Call Reports | Bank profitability (ROA, ROE, % NPL) and bank characteristics (2009-2017); branch-location data for assigning loans to bank counties | [Call Reports](/wiki/datasets/call-reports/) |
| NBER county distance database | Physical distance (as-the-crow-flies, miles) between county centroids | No page yet |
| Bureau of Economic Analysis | County-level real GDP growth and industry-share data | No page yet |
| Bureau of Labor Statistics | County-level employment; unemployment differentials | No page yet |
| U.S. Census Bureau | Commuting, migration, common-border, and county-level population data | No page yet |
| National Transportation Center / Oak Ridge | Highway travel costs and flight-and-drive time between county pairs | No page yet |
| Gilje, Loutskina & Strahan (2016) / Gilje (2019) | Shale-boom well counts and bank branch locations in boom counties for IV construction | No page yet |

Sample summary: cross-sectional county-pair analysis uses 2016 SCI and 2017 CRA/HMDA lending data, covering over 9 million county pairs. Loan-level analysis: 1,268,200 mortgages originated 2000-2008. Bank-profitability analysis: 844 banks, 2009-2017. Real-effects analysis: 3,021 counties, 2009-2017.

## When to read the full paper

Read the [original](https://doi.org/10.1093/rfs/hhaf014) if you are: studying the geographic determinants of bank lending beyond physical distance; researching social networks and credit markets; applying the Facebook SCI to a new financial context; building a social-proximity-to-institutions measure analogous to Kuchler et al. (2022); evaluating the information-vs.-favoritism debate in relationship banking; or designing a shale-boom IV for bank lending. The replication archive at Harvard Dataverse (https://doi.org/10.7910/DVN/T3G5MD) covers all tables and figures.

## Attribution and rights

Source: peer-reviewed, *The Review of Financial Studies* 38(9), September 2025. This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The paper is paywalled (Oxford University Press standard reuse rights; not CC); this page is extract-only.

> Rehbein, Oliver, and Simon Rother. "Social Connectedness in Bank Lending." *The Review of Financial Studies* 38, no. 9 (September 2025): 2759-2809. DOI: 10.1093/rfs/hhaf014.


==============================================================================
# Financial Consequences of Pretrial Detention: Slutzky & Xu (2025)
# https://instituteforautomatedresearch.org/wiki/papers/rfs/2025/slutzky-financial-consequences-pretrial-detention-2025/
# Distilled: Using quasi-random assignment of court commissioners in Maryland as an instrument, this paper finds that pretrial detention causally raises household insolvency rates, driven by chapter 7 bankruptcy, judgment liens, and foreclosures in areas of declining house prices, with effects spilling over to family members rather than defendants themselves. Review of Financial Studies 2025, paywalled. Eight core results with source locators, datasets used, the identification strategy, and the estimating equations.
# Tags: paper-summary, household-finance, criminal-justice, bankruptcy, foreclosure
==============================================================================

**What this is.** The paper's core results, identification design, and estimating equations: enough to understand what was found and how, without reading all 45 pages. To replicate or extend, read the original at [https://doi.org/10.1093/rfs/hhaf009](https://doi.org/10.1093/rfs/hhaf009).

## TL;DR

This paper asks whether pretrial detention, which holds individuals in jail before trial because they cannot afford bail, causes subsequent household financial distress. Using Maryland criminal court data (2000-2016) matched to bankruptcy filings from Gross, Notowidigdo, and Wang (2014), foreclosure, and judgment lien records, it exploits the quasi-random assignment of court commissioners to cases as an instrument for detention decisions. A more lenient commissioner reduces the probability of detention, and this variation is unrelated to defendant characteristics. The main finding is that pretrial detention causally increases household insolvency, raising chapter 7 bankruptcy rates by 0.79 percentage points (30% of the mean) and judgment lien rates by 0.56 percentage points (35% of the mean) within three years. Foreclosures increase significantly (2.9 pp, 23% of the mean) only in ZIP codes with declining house prices, consistent with home equity acting as a liquidity buffer. The financial burden falls primarily on family members, not defendants themselves, suggesting cohabiting relatives (parents, partners) posting or guaranteeing bail bear the cost. Commercial bail bonds play a partial role, but insolvency effects persist even in samples that largely eliminate income-loss and criminal-record channels.

## Core results

Magnitudes and significance are as reported; `\*\*` = 5%, `\*\*\*` = 1%. All regressions instrument detention with the residualized leave-out mean commissioner leniency measure (first-stage F-stat exceeds 6,000 unless noted).

| # | Result | Locator | Magnitude |
|---|---|---|---|
| R1 | Pretrial detention **raises chapter 7 bankruptcy** at the 1-3 year horizon; null for chapter 13 | Table 5, p. 3353 | +0.44 pp at 1 year\*\*, +0.76 pp at 2 years\*\*, +0.79 pp at 3 years\*\* (30% of 2.65% mean); chapter 13 insignificant at all horizons |
| R2 | Pretrial detention **raises judgment lien rates** at the 3-year horizon | Table 6, p. 3355 | +0.56 pp at 3 years\*\* (35% increase relative to 1.6% mean); insignificant at shorter horizons |
| R3 | **Foreclosure effects are null overall** but large in areas with declining house prices | Table 7, p. 3356 | Full sample: +0.92 pp at 3 years (insignificant); negative-HPI subsample: +2.9 pp\*\* at 3 years (23% of mean); null in positive-HPI subsample |
| R4 | Overall household insolvency (bankruptcy + lien + foreclosure) rises by **2.4 pp at 3 years** | Table 8, p. 3359 | +1.7 pp\*\* at 2 years, +2.4 pp\*\*\* at 3 years (16% of 14.8% mean); clean zero in backward-looking placebo tests (Figure 8) |
| R5 | **Insolvency burden falls on family members**, not defendants | Table 9, p. 3362 | InsolventDef: +0.0027 (insignificant, mean 1.6%); InsolventFamily: +0.0075\*\* (22% of 3.4% mean); InsolventUnrelated: +0.0123 (insignificant) |
| R6 | **Commercial bail bonds amplify but do not fully explain** the insolvency effect | Table 10, p. 3365 | ROR/commercial-bond sample: +4.2 pp\*\*\*; excluding commercial bonds: +1.9 pp\*; effect persists when income-loss and conviction channels are controlled |
| R7 | Insolvency effects are **stronger for younger defendants** and **mortgaged, shorter-maturity properties** | Table 11, p. 3367 | Young defendants (age < 30): +2.8 pp\*; mortgage-financed properties: +3.9 pp\*\*; short-maturity: +4.1 pp\*\*\*; consistent with older relative bearing bail costs |
| R8 | **Failure to appear magnifies** the insolvency effect for defendants and family members via bond forfeiture | Table 13, p. 3369 | FailToAppear x Detained interaction: +0.0308\*\*\* (InsolventDef, col 2); +0.0088\*\*\* (InsolventFamily, col 3); insignificant for InsolventUnrelated (col 5); OLS result, causal interpretation limited |

**Overall (paper's conclusion).** Pretrial detention imposes significant household financial costs that extend beyond the defendant to cohabiting family members, most plausibly through the direct cost of posting or guaranteeing bail. The 3-year bankruptcy effect (0.79 pp) is comparable in magnitude to Dobkin et al. (2018), who find that hospital admissions raise bankruptcy rates by 0.4-1.4 pp. Home equity cushions households from insolvency: the foreclosure effect is concentrated in areas with declining house prices, where households cannot tap equity to meet liquidity shocks. Commercial bail bonds play a partial but not exclusive role. The findings add to the literature on the collateral damage of the criminal justice system and are relevant to ongoing debates about bail reform.

## Theory / model

The paper has no formal economic model. The identification logic and tested hypotheses are as follows.

**Hypotheses tested.** Three potential channels link pretrial detention to household insolvency (pp. 3330-3332):

1. Liquidity shock from bail costs (cash bond, commercial surety bond fee, property bond pledge) borne immediately by the defendant or family.
2. Loss of current income from prolonged detention, reducing the defendant's ability to contribute to household obligations.
3. Long-run income reduction through higher conviction rates and reduced future formal employment, documented by Gupta, Hansman, and Frenchman (2016) and Dobbie, Goldin, and Yang (2018).

The empirical tests are designed to separate these. The commercial bail bond subsample test (Table 10) restricts to cases where defendants obtain immediate release via commercial bonds, which filters out the income-loss channel while preserving the liquidity-shock channel. The finding that insolvency effects persist in this sample (R6) is consistent with the bail-cost mechanism.

**Double-trigger hypothesis for foreclosure.** Foote, Gerardi, and Willen (2008) show that both a liquidity shock and negative home equity are needed for foreclosure. The heterogeneous foreclosure effects (R3) are explicitly framed as a test of this: the liquidity shock from pretrial detention triggers foreclosure only when home equity is insufficient to buffer it (pp. 3356-3358).

**Identification.** Assignment of commissioners to cases in Maryland is treated as quasi-random conditional on court-by-year, ZIP-code-by-year, month, day-of-week, sex, race, and charge fixed effects. The exclusion restriction requires that the commissioner's leniency affects household financial outcomes only through the pretrial detention decision. The paper provides two sets of evidence: (i) the residualized instrument is uncorrelated with all observable defendant and case characteristics (Table 2, column 3; joint F-test p-value = 0.499), and (ii) the first-stage coefficient on the instrument is large and symmetric across subsamples defined by race, sex, age, and geography (Table 3).

## Method

The paper applies an instrumental variables design built on two existing approaches: the leave-out commissioner leniency instrument of Dahl, Kostol, and Mogstad (2014) and Dobbie, Goldin, and Yang (2018), applied to a new outcome domain (household finance). The technique genealogy runs through `instrumental-variables` and `panel-regression`.

**First stage: residualized leave-out mean (p. 3345, Equation 3).** For each commissioner $$j$$ in year $$t$$, the leniency instrument is constructed as:

$$
\text{ReleasedRIV}_{ctj} = \left(\frac{1}{n_{tj} - n_{itj}}\right)\left(\sum_{k=0}^{n_{tj}} \text{Released}^*_{klt} - \sum_{c=0}^{n_{itj}} \text{Released}^*_{ict}\right) \tag{3}
$$

where $$n_{tj}$$ is the total number of cases commissioner $$j$$ sees in year $$t$$, $$n_{itj}$$ is the number involving defendant $$i$$, and $$\text{Released}^*_{ict}$$ is the residual release decision after partialling out defendant and case characteristics $$X_{ict}$$ (Equation 2):

$$
\text{Released}^*_{ict} = \text{Released}_{ic} - \gamma \, X_{ict} = \text{ReleasedRIV}_{ctj} + \epsilon_{ict} \tag{2}
$$

Leaving out cases involving the focal defendant avoids the mechanical correlation that would arise if the instrument included the defendant's own case. The instrument is then transformed into $$\text{DetainedRIV} = 1 - \text{ReleasedRIV}$$ so that results are reported as effects of detention rather than release.

**Second stage (p. 3344, Equation 1).** The primary estimating equation is:

$$
Y_{ict} = \beta_0 + \delta \, \text{Released}_{ic} + X_{ict}\beta + \epsilon_{ict} \tag{1}
$$

where $$Y_{ict}$$ is the cumulative insolvency indicator (bankruptcy, judgment lien, or foreclosure) for individual $$i$$ in case $$c$$ in year $$t$$, $$\text{Released}_{ic}$$ is the endogenous binary treatment variable, and $$X_{ict}$$ is a vector of case- and defendant-level controls. In the 2SLS version, $$\text{DetainedRIV}$$ instruments for $$\text{Detained}_{ic}$$.

**First-stage strength.** The first-stage F-statistic is 13,526 in the main sample (Table 3, col 1; Table 4B), well above the Stock and Yogo (2005) and Olea and Pflueger (2013) thresholds. The coefficient on $$\text{ReleasedRIV}$$ in the first stage is approximately -0.946 (Table 3, col 1), meaning a one-unit increase in the leniency instrument shifts the probability of release by 94.5 percentage points. The coefficient is stable across subsamples (range 0.898 to 1.082, Table 3, cols 2-10).

## Empirical specifications

**Outcome variables.** The insolvency indicators are defined as cumulative dummy variables: $$Y_{ict}^{\tau} = 1$$ if a bankruptcy, judgment lien, or foreclosure occurs within $$\tau$$ periods of the initial hearing date. Horizons are 3 months, 6 months, 1 year, 2 years, and 3 years. Bankruptcy data (PACER) cover 2000-2008; foreclosure and lien data (Maryland Judiciary + ZTRAX) cover 2000-2016.

**Fixed effects and standard errors.** All regressions include court-by-year, ZIP-code-by-year, month, day-of-week, sex, race, and charge fixed effects. Standard errors are clustered at the commissioner level.

**Bankruptcy specifications (Table 5).** The 2SLS regression instrumenting $$\text{Detained}_{ic}$$ with $$\text{DetainedRIV}_{ctj}$$ yields:

| Horizon | Coefficient on Detained | SE | Mean |
|---|---|---|---|
| 3 months | +0.0003 | (0.0013) | 0.0029 |
| 6 months | +0.0005 | (0.0015) | 0.0057 |
| 1 year | +0.0044\*\* | (0.0021) | 0.0110 |
| 2 years | +0.0076\*\* | (0.0031) | 0.0198 |
| 3 years | +0.0079\*\* | (0.0037) | 0.0265 |

First-stage F-stat: 10,004. N = 306,722.

**Judgment lien specifications (Table 6).** Same specification, first-stage F-stat = 13,526, N = 502,546. The coefficient becomes significant at the 3-year horizon (+0.0056\*\*, SE = 0.0023), consistent with liens being a last resort after other repayment channels are exhausted.

**Foreclosure specifications (Table 7).** Full sample (N = 275,325): coefficient at 3 years = +0.0092 (SE = 0.0074), insignificant. Negative-HPI ZIP codes (N = 107,357): +0.0291\*\* at 3 years (SE = 0.0123). Positive-HPI ZIP codes (N = 166,155): -0.0033 (SE = 0.0089), null.

**Overall insolvency (Table 8).** Combines bankruptcy, lien, and foreclosure for the ZTRAX-matched sample (N = 275,325). Coefficient at 3 years: +0.0242\*\*\* (SE = 0.0090), mean = 0.1479 (SD = 0.3550).

**Family spillover specifications (Table 9).** 2SLS (same instrument as benchmark, first-stage F = 6,305.53). Insolvency events are matched using both addresses and full names, separating defendant's name (InsolventDef), family members' names (InsolventFamily), same last name (InsolventName), and unrelated individuals (InsolventUnrelated). N = 275,325, all horizons 3 years.

**Placebo tests.** Figures 4, 5, and 8 plot coefficients at backward-looking horizons (-3 years, -2 years, -1 year, -6 months, -3 months). In all cases the coefficients are small and statistically indistinguishable from zero, confirming the identifying assumption that the instrument is uncorrelated with pre-existing insolvency trends.

## Datasets used

| Dataset | Role in paper | Wiki page |
|---|---|---|
| Maryland Judiciary public access database | Criminal case records: 1.08 million cases 2000-2016; commissioner ID, release decisions, bail types, charge categories | [Maryland Judiciary](/wiki/datasets/maryland-judiciary/) |
| PACER (Public Access to Court Electronic Records) | Consumer bankruptcy filings 2000-2011: 318,000 filings in Maryland; chapter 7 and chapter 13 type, filing date, address | [PACER](/wiki/datasets/pacer-bankruptcy/) |
| ZTRAX (Zillow Transaction and Assessment Database) | Real estate transactions 1993-2020: 9 million Maryland transactions; foreclosure events post-2007, property-level matching | [ZTRAX](/wiki/commercial/ztrax/) (licensed) |
| Maryland Judiciary civil court records | Judgment lien filings 2000-2020: 386,938 lien filings; plaintiff/defendant address, filing date | [Maryland Judiciary](/wiki/datasets/maryland-judiciary/) |
| Federal Housing Finance Agency HPI | ZIP-code-level annual house price index; used to split sample into negative/positive HPI growth subsamples | No page yet |

Sample: over 500,000 criminal cases in Baltimore City, Montgomery County, and Prince George's County District Courts. 78% from Baltimore City. 81% Black defendants, 83% male, median age 30.

## When to read the full paper

Read the original if you are: studying the economics of the bail system or pretrial detention reform; working on household insolvency and liquidity shocks more broadly; replicating or extending the leave-out commissioner leniency instrument to other outcomes or jurisdictions; or testing the double-trigger hypothesis for foreclosures with a new source of liquidity shocks. Tables 5-9 contain the headline regressions; the Internet Appendix (available on the RFS website) contains robustness tests including a lagged IV, non-residualized IV, apartment-inclusive sample, and absorbing-state tests.

## Attribution and rights

Source: peer-reviewed, *The Review of Financial Studies* 38(11), November 2025. This distillation was extracted by an LLM on 2026-06-06 and is **not human-verified or independently reproduced**. The paper is paywalled (OUP standard publication reuse rights); no verbatim reproduction. Access the original at [https://doi.org/10.1093/rfs/hhaf009](https://doi.org/10.1093/rfs/hhaf009).

Citation: Slutzky, Pablo, and Sheng-Jun Xu. "The Financial Consequences of Pretrial Detention." *The Review of Financial Studies* 38, no. 11 (2025): 3329-3373. DOI: 10.1093/rfs/hhaf009.