FOMC in Silico: A Multi-Agent System for Monetary Policy Decision Modeling

We build a multi-agent LLM framework that simulates the Federal Open Market Committee (FOMC) decision process. Each agent in the simulation represents an FOMC member and receives real-time macroeconomic data, district level conditions, and other relevant information. In a backtest of 218 FOMC meetings from 2000 through June 2026, the framework predicts the policy rate with a mean absolute error of 8.3 basis points, comes within 25 basis points in 94 percent of meetings, and correctly classifies hikes, cuts, and holds in 91 percent. Accuracy is similar before and after the model's training cutoff, suggesting the results are not driven by memorization. We then use the validated framework for counterfactual experiments. Committee rules can block a politically captured chair, but not public persuasion; committee decisions differ from those produced by mechanical decision rules; composition matters far more than the chair or formal mandate; and deliberation builds consensus rather than improving forecasts. The framework offers a new laboratory for studying monetary policymaking and institutional design.