Failing Banks: Correia, Luck & Verner (2026)

What this is. This is a machine-extracted skeleton of the paper. Read the original at QJE (or the arXiv preprint) to replicate or extend the results.

TL;DR

Correia, Luck, and Verner build a new panel of US banks from 1863 to 2024, covering historical national banks and modern commercial banks, 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 complements aggregate evidence that banking crises are predictable from business failures (Gorton 1988) and declining bank equity (Baron, Verner, and Xiong 2021).

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 for banks above the 95th percentile in both insolvency risk and noncore funding (both samples)	Figure IV, p. 168	27% (historical and modern); unconditional = 2.5% (historical), 1% (modern); 10-20x 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 outflow-observed sample (1880-1934)	Table I, Panel A, col. 5, p. 172; text p. 180	In-sample AUC = 0.855; OOS AUC = 0.839; nearly identical to all-failures AUC (0.864 in the broader 1863-1934 sample)
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%)
R10	Insolvency-risk gradient in historical three-year failure probability	Figure II, Panel A, p. 164; text p. 163	Moving from below the 50th to above the 95th percentile of insolvency risk raises failure probability by 4 percentage points (1863-1934)
R11	Insolvency-risk gradient in modern three-year failure probability	Figure II, Panel B, p. 164; text p. 163	Moving from below the 50th to above the 95th percentile of insolvency risk raises failure probability by 10 percentage points (1959-2024)
R12	Noncore-funding gradient in historical three-year failure probability	Figure II, Panel C, p. 164; text p. 165	Moving from below the 50th to above the 95th percentile of noncore funding raises failure probability by 14 percentage points (1863-1934)
R13	Noncore-funding gradient in modern three-year failure probability	Figure II, Panel D, p. 164; text p. 165	Moving from below the 50th to above the 95th percentile of noncore funding raises failure probability by 5.5 percentage points (1959-2024)
R14	Credit-loss proxy dynamics before historical failures	Figure III, Panel A, p. 166; text p. 165	Nonperforming-loan proxy rises by 5 percentage points before failure (OREO proxy, available 1889-1904)
R15	Credit-loss and capital dynamics before modern failures	Figure III, Panel B, p. 166; text p. 165	In the five years before failure, NPLs rise by 10 percentage points and equity/assets falls by 8 percentage points (1959-2024)
R16	Noncore-funding dynamics before modern failures	Figure III, Panel D, p. 166; text p. 167	Noncore funding rises by 18% of assets over the decade before failure (1959-2024)
R17	Real asset boom and contraction before failure	Figure III, Panel E, p. 166; text p. 168	Real assets expand by 34% from 10 to 3 years before failure, then contract over the last 2 years; pattern occurs in both historical and modern samples and is larger in the modern sample
R18	OCC asset-quality assessment at suspension	Table V, p. 187; text p. 188	Average shares assessed as good = 0.36, doubtful = 0.47, and worthless = 0.18 (1863-1934; N = 2,917)
R19	Asset-quality measures predict receivership recovery	Table VI, col. 1, p. 189; text p. 188	Recovery-rate coefficients: good assets = 0.89 (SE 0.01), doubtful = 0.54 (SE 0.01), worthless = 0.08 (SE 0.02); R² = 0.936; N = 2,682
R20	Precision-recall prediction performance relative to failure base rate	Text p. 176	Out-of-sample PR-AUC is about 13 times the failure rate historically and more than 70 times the rate in the modern sample
R21	Share of run-associated failures plausibly attributable to runs on solvent banks	Figure VIII, p. 194; text p. 193-194	Among failures with deposit outflows over 7.5%, fewer than 8% could plausibly be runs on solvent banks at rho = v = 0; this rises to 13% at rho = v = 0.05 and 22% at rho = v = 0.1
R22	Deposit-outflow heterogeneity across institutional eras	Table II, Panel B, p. 178; text p. 179	Mean deposit growth: -14.40% (1880-1913), -15.98% (1914-1918), -21.73% (1929-1933 pre-holiday), +1.13% (1933-1934 post-holiday), -5.64% (1993-2006), and -2.14% (2007-2024)

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 bank assets from t - 3 to t, 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 $t$. The accompanying text also specifies bank-age controls, although the printed equation omits an age term. 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 $j = -10$. The PDF’s printed summation bounds read 0 to 10, but the following text and Figure III index event time negatively and identify -10 as the omitted period; the equation here follows that event-time indexing. 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. Table I reports AUC statistics rather than regression standard errors, and its notes do not state the standard-error treatment. 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 1880-1934, col. 5 vs. 0.864 for all failures in the broader 1863-1934 col. 4 sample, Table I), consistent with the paper’s description that fundamentals predict run-failures approximately as 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 funds collected relative to assets at suspension and additional assets received after 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.

Failure-event dynamics (Section IV.B, pp. 163-166, equation 1). The event-time regression in the Method section includes bank fixed effects, uses failing banks observed within ten years of failure, and omits year -10. Figure III reports historical annual coefficients for 1863-1934 and modern quarterly coefficients for 1959-2024, plotting only end-of-calendar-year coefficients in the modern panels. The figure caption does not state the standard-error estimator or clustering procedure. Equation (1) is the paper’s numbered main-text specification and is shown in the Method section with its locator.

Aggregate failure-rate regression (Section VII, pp. 183-184). The paper estimates the following time-series specification (Table III, p. 184):

$$\text{FailureRate}_{t} = \alpha + \beta\,\bar{p}_{t|t-1} + u_t$$

Here $\bar{p}_{t|t-1}$ is the equally weighted mean of bank-level predicted failure probabilities produced from an expanding window using information through $t-1$, with the first ten years as the initial training sample. It is estimated by era, with no fixed effects; Newey-West standard errors use truncation parameter $S = 1.3T^{1/2}$. The table reports annual sample sizes of 117 for the full sample, 61 for 1874-1934, 55 for 1970-2024, and 111 excluding 1929-1934.

Receivership recovery regressions (Section VIII.A, pp. 188-189). Table VI estimates cross-sectional regressions of each failed bank’s recovery rate on the listed assessment or balance-sheet predictors (Table VI, p. 189):

$$R_b = \alpha + \beta X_b + \epsilon_b$$

The sample is receiverships from 1863-1934. Columns 2-5 include a constant; column 1, which regresses recovery on good, doubtful, and worthless asset shares, omits the constant. The table uses 2,682 observations in column 1 and 2,727, 2,725, 418, and 1,998 in columns 2-5. It has no bank or time fixed effects. Standard errors are in parentheses, but the table note does not identify their estimator or clustering.

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.