Social Contagion and Bank Runs: An Agent-Based Model with LLM Depositors

TL;DR

This paper advances a process-based ABM to study modern bank runs by explicitly modeling information diffusion and social coordination among depositors. It combines cash-first withdrawal dynamics, fire-sale liquidation, and an endogenous stress index with an LLM-driven depositor policy to probe how social networks amplify withdrawals, both within banks and across banks via depositor overlap. The results show a clear phase transition in contagion driven by information spillovers and network structure, and they reproduce observed orderings of SVB/First Republic/regional bank distress, including higher withdrawal rates among uninsured depositors. The framework suggests that supervisory stress testing should incorporate social-correlation channels, not just balance-sheet fundamentals, and offers a path to operational supervisory metrics that quantify cross-bank contagion risk through information networks.

Abstract

Digital banking and online communication have made modern bank runs faster and more networked than the canonical queue-at-the-branch setting. While equilibrium models explain why strategic complementarities generate run risk, they offer limited guidance on how beliefs synchronize and propagate in real time. We develop a process-based agent-based model that makes the information and coordination layer explicit. Banks follow cash-first withdrawal processing with discounted fire-sale liquidation and an endogenous stress index. Depositors are heterogeneous in risk tolerance and in the weight placed on fundamentals versus social information, communicating on a heavy-tailed network calibrated to Twitter activity during March 2023. Depositor behavior is generated by a constrained large language model that maps each agent's information set into a discrete action and an optional post; we validate this policy against laboratory coordination evidence and theoretical benchmarks. Across 4,900 configurations and full LLM simulations, three findings emerge. Within-bank connectivity raises the likelihood and speed of withdrawal cascades holding fundamentals fixed. Cross-bank contagion exhibits a sharp phase transition near spillover rates of 0.10. Depositor overlap and network amplification interact nonlinearly, so channels weak in isolation become powerful in combination. In an SVB, First Republic, and regional bank scenario disciplined by crisis-era data, the model reproduces the observed ordering of failures and predicts substantially higher withdrawal rates among uninsured depositors. The results frame social correlation as a measurable amplifier of run risk alongside balance-sheet fundamentals.

Figures (6)

• Figure 1: Influence concentration in SVB-related retweet activity (Cookson et al. replication). A small fraction of accounts generates a disproportionate share of amplification, motivating a hub-dominated diffusion process and a core--periphery social graph within the SVB depositor network.
• Figure 2: Hourly Twitter/X activity during the crisis window: SVB-topic activity (aggregating $SIVB and $SVB where available) versus other tickers. FRC co-moves strongly with the SVB topic, while SFST remains comparatively insulated in both levels and log scale, motivating the use of $q_{\text{safe}}$ to discount SVB-origin spillovers to the safe-bank benchmark.
• Figure 3: Selecting the "safe" regional benchmark. Scatter of (fundamental distance to FRC) versus estimated SVB$\rightarrow$bank spillover strength (distributed-lag sum of positive coefficients). SFST is selected as the bank closest to FRC in fundamentals while exhibiting minimal SVB-topic spillover, and is used to define $q_{\text{safe}}$.
• Figure 4: LLM behavioral validation (first-round choices). The figure compares the LLM’s sampled first-round action frequencies to the experimental benchmark across low/medium/high vulnerability regimes for the Round (high return, fragile) vs. Square (lower return, safer) banks. The LLM reproduces the qualitative comparative statics: as vulnerability rises, mass shifts toward the safer bank and early withdrawal becomes more prevalent primarily in the high-vulnerability regime.
• Figure 5: Phase diagrams showing First Republic failure probability across bridge multiplier and network degree multiplier, at three spillover rates. Yellow star marks Twitter-calibrated parameters. Transition from stability to contagion occurs between spillover rates 0.0 and 0.30.