Using AI to Let History Speak About Bank Runs (libertystreeteconomics.newyorkfed.org)

Researchers have unveiled a novel approach to understanding historical bank runs by leveraging large language models (LLMs) to extract data from millions of digitized newspaper articles. This project has resulted in the creation of the most comprehensive database of U.S. bank runs, accessible through a newly launched website. By analyzing over 374 million newspaper articles, the team identified roughly 3,000 distinct bank run episodes, categorizing events such as suspensions and failures–with notable findings indicating that a significant number of bank runs occurred without resulting in failures. This development is significant for the AI and ML community as it demonstrates an innovative application of LLMs in historical research, overcoming the limitations posed by sparse regulatory data. By enabling a nuanced understanding of the dynamics behind bank runs, including their causes and geographic patterns, this research not only enlightens historical economic trends but also provides valuable insights for contemporary financial policy discussions. The public availability of this dataset invites broader exploration into the complexities of financial stability and the historical context surrounding bank distress, marking a substantial advancement in the intersection of AI, economics, and historical analysis.