Researchers Use Large Language Models to Build First Comprehensive Database of U.S. Bank Runs
Key Takeaways
- ▸LLMs enable extraction and structuring of historical data from massive unstructured text archives at scale, solving a longstanding data scarcity problem that has limited economic history research
- ▸This research creates the first comprehensive bank run database, revealing previously invisible patterns in banking crises across 170+ years of U.S. history
- ▸LLMs' contextual language understanding identifies historical phenomena even when not explicitly named, a capability far superior to traditional keyword-matching methods
Summary
Economists Sergio Correia, Stephan Luck, and Emil Verner have published groundbreaking research demonstrating how large language models can unlock historical economic data. Using LLMs to analyze 374 million digitized newspaper articles from the Library of Congress's Chronicling America project, the team created the most comprehensive database of U.S. bank runs in history—spanning over 170 years of American financial crises.
The research solves a critical gap in economic history: official regulatory records document when banks closed but don't systematically indicate whether runs occurred, making it impossible for historians to distinguish between bank failures caused by runs versus other causes. The researchers built a data extraction pipeline that identifies candidate articles mentioning financial distress using keywords like "bank run," "deposit withdrawals," and "suspended," then prompts LLMs to extract structured information: which bank was affected, what transpired, timing, and how the institution responded.
Crucially, LLMs' contextual understanding surpasses simple keyword matching. Nineteenth-century newspaper prose varied widely in describing the same phenomena; reporters might write of depositors "clamoring for their money" without using the term "run," while other articles might mention a bank "running for office." LLMs capture these contextual nuances, avoiding the false positives and false negatives that plague rigid keyword approaches. The resulting dataset is now publicly accessible through a companion website where users can browse individual bank run episodes and access original newspaper articles.
- Public access to this historical banking data enables future research on financial regulation, systemic risk, and lessons for modern banking policy
Editorial Opinion
This work demonstrates how large language models are becoming indispensable tools for historical research and evidence-based policymaking. By applying LLM technology to digitized archives at massive scale, researchers have made accessible insights that were effectively hidden across millions of newspaper pages. The creation of a public database democratizes this historical knowledge—benefiting not just academics but policymakers and the public. This exemplifies how AI can illuminate the past to serve better decisions in the present.
