The Federal Reserve Bank of New York has published research applying artificial intelligence and natural language processing to one of finance's oldest problems: understanding why bank runs happen and how they spread. The paper, titled "Using AI to Let History Speak About Bank Runs" and released July 12, 2026 on the Liberty Street Economics blog, represents a fresh approach to financial stability research by training models on digitized archives of banking panics dating back over a century.

Mining Financial History at Scale

The researchers describe using large language models to process historical newspaper reports, regulatory filings, and central bank correspondence from major panic events including the Panic of 1907, the Great Depression banking crisis, and more recent episodes. By applying NLP techniques to these unstructured historical documents, the team claims they can identify patterns in how rumors about bank solvency translated into actual withdrawals—patterns that traditional economic models often miss or oversimplify.

What the Data Reveals

According to the research summary, the AI-driven analysis suggests that depositor behavior during bank runs is more sensitive to social network effects than previous models assumed. Rather than purely rational responses to fundamental information about a bank's balance sheet, the historical record shows significant clustering of withdrawal behavior based on geographic proximity and social connections—findings with direct implications for modern digital banking where information spreads instantaneously through social media.

Implications for Modern Financial Stability

The timing of this research is notable given ongoing debates about deposit insurance reform and the appropriate regulatory response to potential instability in the regional banking sector. If AI can better predict which conditions trigger panic behavior, regulators might design more targeted interventions. The researchers caution, however, that historical patterns may not map directly onto today's financial system with its instant communications and highly interconnected institutions.

Key Takeaways

  • Researchers at the New York Fed applied NLP and LLM techniques to historical banking crisis data spanning over 100 years
  • The AI analysis suggests social network effects play a larger role in bank run dynamics than traditional economic models assumed
  • Findings have potential implications for deposit insurance policy and modern financial stability frameworks

The Bottom Line

This is exactly the kind of applied AI research that actually matters—using modern machine learning tools to extract actionable insights from historical data, rather than chasing the latest hype cycle. If the Fed's models can help regulators get ahead of the next Silicon Valley Bank-style panic, that's a win for everyone except the doomers.