Bank Run, Interrupted: Modeling Deposit Withdrawals with Generative AI

Abstract

I study depositor behavior in panic-driven bank runs using an LLM-based survey simulation. I build a representative population of synthetic agents by assigning demographic attributes, expose them to a viral panic post, and randomize bank communication interventions. I validate model responses against human benchmarks, then generate systematic message variants that vary tone, strength, and content within the validated design space. I then include the estimated withdrawal propensities into a contagion model, which maps network nodes to depositor personas and propagates withdrawals through a single-layer proximity network. Direct, personalized bank communications with strong reassurances and explicit survival clauses substantially reduce withdrawal intent. This framework offers an efficient, inexpensive method for testing crisis messages and projecting how runs may unfold under alternative communication strategies.

Keywords: Large Language Models, Behavioral Finance, Generative AI, Policy Communication, Bank Runs, Survey, Experiment