💻 computer science

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

This paper introduces an agent-based model utilizing constrained large language models to simulate how social contagion and networked communication among heterogeneous depositors amplify bank run risks, revealing that within-bank connectivity and cross-bank spillovers interact nonlinearly to accelerate withdrawal cascades and accurately reproduce the failure patterns of recent banking crises.

Original authors: Chris Ruano, Shreshth Rajan

Published 2026-02-18

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Chris Ruano, Shreshth Rajan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: Why Bank Runs Are Different Now

Imagine a bank run like a fire in a crowded theater. In the old days (the "classic" economic models), the fire spreads because people see smoke, panic, and push their way to the exit. If you see others running, you run too. This is called a coordination problem.

But the authors of this paper argue that in the digital age, the "smoke" isn't just physical; it's a viral tweet, a Reddit thread, or a group chat. The fire spreads not just because people see a crowd, but because they see a digital mob screaming "Run!" on their phones.

This paper asks: How does that digital screaming match actually cause a bank to collapse? And can we build a computer simulation to predict it before it happens?

The Experiment: A Digital Sandbox

To answer this, the authors built a video game simulation (an Agent-Based Model) of a bank run. Instead of using boring math equations that assume everyone thinks exactly the same way, they created a virtual world with three main characters:

The Banks: They have cash in the vault and long-term investments (like bonds). If too many people ask for their money at once, the bank has to sell those investments quickly at a discount (a "fire sale"), losing money in the process.
The Depositors (The Players): These aren't just numbers; they are simulated people with different personalities. Some are brave, some are scared. Some care about the bank's actual health (the "fundamentals"), while others care mostly about what their friends are saying.
The Social Network (The Megaphone): This is the most important part. The players are connected on a fake social network that looks exactly like Twitter/X during the 2023 banking crisis. Some people have millions of followers (influencers), and most have just a few.

The Secret Weapon: The "AI Brain"
Instead of programming the players with rigid rules like "If the bank loses 5%, I run," the authors used a Large Language Model (LLM) (like a smart AI chatbot) to act as the players' brains.

You tell the AI: "Your bank has $100M, but 20% of people withdrew yesterday. Your friend just posted on the feed that the bank is shaky. Are you scared?"
The AI thinks, reasons, and decides: "I'm going to withdraw my money and post a warning."
This makes the simulation feel much more human and chaotic than old-school math models.

What They Discovered: The Three Big Lessons

After running the simulation thousands of times, they found three surprising things:

1. The "Echo Chamber" Effect

Analogy: Imagine a room where everyone is whispering. If the room is small and quiet, you can hear the truth. But if you put everyone in a room with a giant echo machine, a tiny whisper becomes a deafening roar.
The Finding: When depositors are highly connected on social media, panic spreads faster and harder. Even if the bank is actually doing okay, if the "echo chamber" says it's failing, everyone runs. The more connected the group is, the faster the bank collapses.

2. The "Tipping Point" (The Phase Transition)

Analogy: Think of a dam holding back water. You can add a little bit of water (bad news) and the dam holds. But once you cross a specific line, the dam doesn't just crack; it shatters instantly.
The Finding: There is a specific threshold of "information spillover" (how much bad news jumps from one bank to another).

Below 10% spillover: The panic stays contained.
Above 10% spillover: The panic jumps like a virus.
This means a tiny change in how fast news travels can turn a stable bank into a failed one overnight.

3. The "Double Whammy" of Overlap

Analogy: Imagine two houses next to each other. If they share a fence (overlapping depositors) AND they share a ventilation system (social media), a fire in one house will burn the other down much faster than if they just shared a fence.
The Finding: The danger isn't just that people have accounts at two banks (overlap) or that they talk on social media (network). The danger is the combination. When people who bank at "Bank A" also talk to people who bank at "Bank B," the fear travels instantly. The simulation showed that even a healthy bank (like First Republic) could be dragged down by a sick neighbor (SVB) purely because their customers were talking to each other.

The Real-World Test: SVB and First Republic

The authors tested their model against the real-life collapse of Silicon Valley Bank (SVB) and First Republic Bank in March 2023.

The Setup: They fed the model the real balance sheets of these banks and the real Twitter activity from that week.
The Result: The AI simulation perfectly predicted the order of events:
1. SVB collapsed first (because it had too many uninsured depositors and a tight-knit, panicked community).
2. First Republic collapsed next (even though it was healthier, the panic jumped to it via social media).
3. A generic "Regional Bank" survived (because its customers didn't talk to the SVB crowd).

The model even predicted that uninsured depositors (people with more than $250k in the bank) would run much faster than insured ones, which is exactly what happened in real life.

Why This Matters for You

The authors aren't trying to say "Social media causes bank runs." They are saying: Social media is a super-charger.

If a bank has weak fundamentals (it's already shaky), social media doesn't matter much. But if a bank is okay but has a lot of overlap with a failing bank, social media can turn a small problem into a total disaster.

The Takeaway for Regulators:
Banks and regulators used to stress-test banks by looking at their balance sheets (how much cash they have). This paper says they need to start stress-testing the social network too.

Who talks to whom?
Do the customers of Bank A hang out with the customers of Bank B?
How fast can a rumor travel between them?

If you can measure the "social connection" between banks, you can predict which ones are at risk of catching a "digital virus" before the money actually runs out.

In a Nutshell

This paper built a digital sandbox where AI agents act like real humans on Twitter. It proved that in the modern world, fear is contagious, and the speed of that contagion depends on how our social networks are wired. If we want to stop bank runs, we need to understand the network, not just the math.

1. Problem Statement

The paper addresses a critical gap in modern banking stability analysis: the inability of traditional equilibrium models to explain the speed, synchronization, and propagation of bank runs in the digital age.

Limitations of Existing Theory:
- Diamond-Dybvig (1983): Explains why runs exist (strategic complementarities) but relies on "sunspots" (extrinsic belief shifts) for equilibrium selection. Sunspots are unmeasurable and untargetable by regulators, making them poor tools for stress testing.
- Global Games (Morris-Shin): Offers unique equilibrium selection via noisy private signals but relies on parametric assumptions about information structures that are difficult to validate in real-world, endogenous communication environments (e.g., social media).
The Modern Reality: The March 2023 banking crisis (Silicon Valley Bank, First Republic) demonstrated that runs are faster and more networked than canonical queue-at-the-branch models suggest. Runs are driven by social correlation (overlapping depositor communities and communication networks) rather than just balance-sheet fundamentals.
Core Question: How do beliefs synchronize and propagate in real-time through digital networks, and how can this be modeled to create actionable supervisory tools?

2. Methodology

The authors develop a process-based Agent-Based Model (ABM) that explicitly simulates the causal pathway of information flow, decision-making, and balance-sheet feedback.

A. Model Architecture

The simulation operates in discrete time steps ( $t=1, \dots, T$ ) involving three core entities:

Banks:
- Balance Sheet Mechanics: Cash-first withdrawal processing, followed by fire-sale liquidation of illiquid assets at a discount ( $h_b$ ).
- Constraints: Per-round processing caps (10% of deposits) to simulate operational friction; endogenous Probability of Default (PD) updates based on withdrawal intensity.
- Failure Rules: A bank fails if it cannot meet withdrawals or if cumulative withdrawals exceed a "seizure threshold" (e.g., 25% for SVB, 35% for First Republic).
Depositors (Agents):
- Heterogeneity: Agents vary in risk tolerance and the weight they place on fundamentals ( $\alpha$ ) vs. social information ( $\beta$ ).
- Insurance Status: Agents are labeled "insured" or "uninsured" based on bank-specific ratios (e.g., SVB ~93% uninsured), driving different withdrawal incentives.
- Network Structure:
  - Within-bank: Heavy-tailed, core-periphery networks calibrated to Twitter/X retweet data (hub-dominated diffusion).
  - Cross-bank: Sparse "bridge" edges connecting overlapping depositor bases between banks.
Communication Layer:
- Agents observe lagged social posts from neighbors and withdrawal rates.
- LLM Integration: Instead of hard-coded decision rules, agents use a constrained Large Language Model (LLM) (GPT-4.1-mini) as a behavioral policy function. The LLM maps a structured state description (fundamentals, insurance status, social posts) into a discrete action (Withdraw/Stay) and an optional social post.

B. Validation and Calibration

Empirical Anchors: Parameters are tuned using March 2023 data (SVB, First Republic, and a Regional control bank) and high-frequency payment data.
LLM Validation: The LLM policy is validated against laboratory coordination experiments (Garratt and Keister, 2009). The zero-shot LLM successfully reproduces qualitative comparative statics (e.g., higher withdrawal rates in high-vulnerability regimes) without being fine-tuned on run data.
Theoretical Benchmarks: Results are compared against Diamond-Dybvig equilibria and Morris-Shin global game predictions to ensure the simulator interpolates between static equilibria and instantaneous coordination.

3. Key Contributions

Methodological Shift: Proposes a process-based ABM as a superior framework for analyzing modern runs compared to equilibrium selection models. It treats the "causal pathway" of information propagation as the primary object of study, moving beyond unobservable sunspots.
Formalizing Social Correlation: Defines social correlation (overlapping depositor communities and communication networks) as a distinct, measurable risk channel that amplifies balance-sheet stress, analogous to correlated asset exposures but operating through beliefs.
Behavioral Innovation: Demonstrates that off-the-shelf LLMs can serve as disciplined, auditable behavioral priors in economic simulations. The study shows that LLMs can capture complex coordination dynamics without reinforcement learning training, provided they are constrained by strict action interfaces and validated against experimental data.
Quantitative Findings: Delivers specific simulation results on the non-linear dynamics of bank runs, identifying phase transitions and interaction effects between network topology and information spillovers.

4. Key Results

The model was run across 4,900 parameter-seed configurations (245 combinations $\times$ 20 seeds).

Reproduction of March 2023 Ordering: The model successfully reproduces the observed failure sequence: SVB $\to$ First Republic $\to$ Regional Survival.
- SVB fails first (30% withdrawal).
- First Republic fails second (40% withdrawal) despite having better initial fundamentals, driven by contagion.
- The Regional bank survives (6% withdrawal) due to a "flight-to-safety" effect and low social spillover.
Phase Transition in Contagion:
- There is a sharp phase transition in failure risk for the second bank (First Republic) as the information spillover rate increases.
- Tipping Point: Around a spillover rate of 0.10. Below this, the second bank usually survives; above it, failure probability jumps to 80–100%.
Non-Linear Interaction of Channels:
- Isolation: Neither "bridge depositors" (shared customers) nor "network structure" alone generates significant contagion (5% failure rate in isolation).
- Synergy: When combined, they generate 65% failure rates. Adding information spillovers pushes this to 85%.
- Implication: Weak channels in isolation become powerful amplifiers when combined. Stress testing must account for channel combinations, not just isolated exposures.
Uninsured vs. Insured Dynamics: Uninsured depositors withdraw at 81.3% compared to 34.5% for insured depositors, confirming that strategic complementarity drives the run, consistent with coordination theory.

5. Significance and Policy Implications

Supervisory Tool: The paper argues that regulators should treat social correlation as a measurable risk amplifier. Banks with overlapping depositor communities (e.g., shared VC networks, industry clusters) are vulnerable to synchronized belief shifts.
Stress Testing: Traditional stress tests focusing only on balance sheets are insufficient. Scenarios must include "social stress tests" that model how narratives propagate through specific network topologies.
Resilience: Communication policy (speed and credibility of bank responses) acts as a state variable. In a digital run, the ability to break the feedback loop of fear is as critical as liquidity buffers.
Future Direction: The authors propose extending the framework to create operational social-correlation risk estimators using observable network data (retweets, co-activity) to predict run risk for specific institutions.

Conclusion

This paper provides a proof-of-concept that process-based agent-based modeling, powered by LLM agents and calibrated to real-world network data, can capture the complex, non-linear dynamics of modern bank runs. It moves the field from explaining why runs are possible (equilibrium theory) to modeling how they propagate (causal process), offering a new toolkit for supervisory monitoring and stress testing in the digital age.