Social Media as a Bank Run Catalyst: Cookson et al. (2026)

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.

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 negative hourly estimate is larger and more precisely estimated in the above-median-retweet subsample, 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; \* = 10%, \*\* = 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 estimate is larger in magnitude for high-retweet than low-retweet activity	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; adding the full set of in-run tweet controls reduces the pre-exposure estimate to near zero	Table 4, Panel C, cols. 2-6, p. 15	Contagion tweets: 6.696*** (SE = 1.333); run tweets: 7.252*** (SE = 1.374); pre-exposure falls to -0.129 (SE = 0.618) with all controls
R9	Run-related discussion was concentrated at the distressed banks	Table 2, Panel B, p. 9	SIVB: 6,528 run tweets and 9,662 contagion tweets; FRC: 1,249 run tweets and 1,368 contagion tweets; 90th percentile: 3 run and 2 contagion tweets
R10	Pre-run Twitter exposure tracks bank size, not baseline balance-sheet run risk	Table 3, col. 1, p. 10	log(MCap) = 0.767*** (SE = 0.043); % Loss MTM = -0.027 (SE = 0.040); % Uninsured = -0.051 (SE = 0.050)
R11	Total deposit outflows do not show the uninsured-deposit amplification result	Table 5, col. 3, p. 16	% Uninsured = 1.37*** (SE = 0.502); % Uninsured x % Loss MTM x Pre-Exposure = 0.673 (SE = 0.536), insignificant
R12	The hourly attention result is similar with 1-, 2-, and 4-hour tweet windows	Table 7, Panel A, cols. 1-3, p. 20	Post-Mar-9 x high attention: -0.074* (SE = 0.045), -0.084* (SE = 0.046), -0.100** (SE = 0.042)
R13	Negative sentiment predicts immediate losses for run-relevant tweet types, not for high-run-risk banks generally	Table 8, cols. 2, 4, 6, 8, p. 21; Table 9, p. 22	VADER Neg x Tech Community = -11.14** (SE = 5.45); x Run Tweet = -12.91*** (SE = 2.81); x Contagion Tweet = -20.63* (SE = 11.19); similar with follower and retweet controls
R14	Classical balance-sheet run risk predicts stock losses even without Twitter pre-exposure	Table 4, Panel A, col. 1, p. 13	% Uninsured = 4.249*** (SE = 1.026); % Uninsured x % Loss MTM = 1.742** (SE = 0.824)

Overall (paper’s conclusion). Social media exposure amplifies classical bank run risks, with evidence pointing to Twitter attention and social propagation via retweets, not negative sentiment. A similar interaction appears for uninsured deposit outflows, though quarterly data make that evidence suggestive about timing; effects are evident at both cross-sectional and hourly frequencies. The paper interprets these patterns as consistent with social media serving as a coordination device for depositors, linking them to 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 associated primarily with bank size and investor fanaticism (Pedersen (2022)), and has near-zero correlation with traditional run risk factors (Fig. 6b, p. 11: correlation of pre-exposure with run risk is 0.02). This pattern is consistent with cross-sectional variation in pre-exposure capturing social media reach rather than ex ante run vulnerability, though the observational design does not establish that interpretation causally. The content of pre-period tweets is also unrelated to banking distress topics (Fig. A.5 in the Appendix). For the hourly tests, a narrow 5-minute window around individual tweets (adapting the approach of Bianchi et al. (2023)) limits the scope for confounding news and price dynamics. For the deposit outflow tests, quarterly FDIC Call Report data provide complementary, suggestive evidence; Drechsler et al. (2024) deposit beta measures are also used as controls.

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. This OLS cross-section covers 277 publicly traded bank holding companies. Table 3 reports robust standard errors in parentheses (Table 3, p. 10).

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). This OLS cross-section uses 277 banks, has no fixed effects, and reports robust standard errors in parentheses (Table 4, pp. 13-15).

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. This OLS cross-section has no fixed effects and reports robust standard errors. The baseline sample is 275 banks, falling to 262 in columns with additional controls (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). The return and news controls also enter interacted with the post-run indicator; the baseline Table 6 sample has 14,786 bank-hour observations in the high-run-risk columns. The model is estimated by OLS, with standard errors clustered by bank (Table 6, p. 18).

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. The interaction specification adds the two sentiment scores interacted with the relevant tweet-type or high-run-risk indicator:

$$\Delta p_{i,t} = a + b\,\text{VADERPos}_{i,t} + c\,\text{VADERNeg}_{i,t} + \theta_1\bigl(\text{VADERPos}_{i,t}\times D_{i,t}\bigr) + \theta_2\bigl(\text{VADERNeg}_{i,t}\times D_{i,t}\bigr) + \eta\mathbf{X}_{i,t} + \gamma_i + \varepsilon_{i,t}$$

Here $D_{i,t}$ is, in turn, the Tech Community Flag, Run Tweet, Contagion Tweet, or High Run Risk indicator. The Table 8 and 9 tweet-level samples contain 36,659 observations in controlled specifications; all include bank fixed effects, while even-numbered specifications add prior price and news controls. Standard errors are clustered by bank-day (Tables 8-9, pp. 21-22).

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
CRSP US Treasury and Inflation Indexes	Treasury bond price changes by maturity for % Loss MTM construction	WRDS (licensed)
Compustat bank fundamentals (via WRDS)	Market capitalization, market-to-book, liquid assets, CRE loan share	WRDS (licensed)
I/B/E/S analyst coverage (via WRDS)	Number of analysts as information-environment control	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 (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 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. © 2025 Elsevier B.V. All rights reserved.