Correia, Luck, and Verner explain in Liberty Street Economics how AI helps map bank runs

Correia, Luck, and Verner explain in Liberty Street Economics how AI helps map bank runs

On July 7, 2026, Sergio Correia, Stephan Luck, and Emil Verner describe using large language models to identify bank runs in historical newspaper articles, where official records often show bank closures but not whether a run preceded them. Their approach searches for candidate articles, then uses models to extract events and distinguish runs, suspensions, and failures. The authors say this helps capture varied descriptions that keyword matching can miss; they compare outputs from multiple models and report high agreement with hand-coded samples. The resulting database contains more than 3,000 run episodes from 1863 through 1934. The authors’ central distinction is that most recorded runs did not involve failure, even though most distress episodes in the dataset involved a bank failure. They also describe runs spreading during major crises and local distress appearing beyond the national crisis narrative. The database makes newspaper accounts available for readers to examine, while giving researchers material for questions that were difficult to study with regulatory records alone. The authors say their companion post uses the data to investigate when and why runs led to failure and the implications for policy discussions. The post does not specify the size of the hand-coded samples used for validation.

Publication: 5dd3c258-e36e-7834-6083-63c1aca07806

attribution — pass: The reading correctly credits Sergio Correia, Stephan Luck and Emil Verner, the post's bylined authors, and Liberty Street Economics, the Federal Reserve Bank of New York blog. The methods and findings are framed as the authors' own descriptions and results. This is a single authored post, not a listing.

Sergio Correia, Stephan Luck, and Emil Verner

summary — pass: The title, summary and question faithfully convey the post. The authors used LLMs to mine newspapers for bank runs that official records miss. The reading explains the keyword-then-model pipeline, the validation, the more-than-3,000 episodes from 1863-1934, and the key run-versus-failure distinction. It also covers geographic and crisis patterns and the companion post. LLM is defined beside the reading, and the context about regulatory records explains why the work matters. The question follows from newspaper-coverage limits without leading.

But such records do not indicate whether a run occurred before failure.

date — pass: The July 7, 2026 date matches the post's dateline, the citation and the UTC day of originalPublished. The page was modified on 2026-08-07, but every cited passage is plainly part of the body of the original article. The reading does not use the later navigation or related-story links and contains no hindsight framing. Kind is publisher-page, so no archive disclosure is needed.

July 7, 2026

support — pass: Each claim is supported verbatim: the LLM extraction, the advantage over keyword matching, the validation against hand-coded samples, the episode counts, the run-versus-failure finding and the companion post. The noted gap is accurate, since the post gives no hand-coded sample size. The stated uses (public browsing, research questions) are supported. Body text appears original despite the later edit flag.

We find 1,906 runs without failure and 1,515 runs with failure.

media — pass: No media was selected and no verified media candidates were supplied, so nothing is misidentified or irrelevant.