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

We develop a multi-agent framework for modeling the Federal Open Market Committee (FOMC) decision making process. The framework combines two approaches: an LLM-based simulation and a Monte Carlo implementation of a generalized Bayesian voting model. Both begin from identical prior beliefs about the appropriate interest rate for each committee member, formed using real-time data and member profiles. In a simulation replicating the July 2025 FOMC meeting, both tracks deliver rates near the 4.25-4.50% range's upper end (4.42% LLM, 4.38% MC). Political pressure scenario increases dissent and dispersion: the LLM track averages 4.38% and shows dissent in 88% of meetings; the MC track averages 4.39% and shows dissent in 61% of meetings. A negative jobs revision scenario moves outcomes lower: LLM at 4.30% (dissent in 74% of meeting), and MC at 4.32% (dissent in 62% of meeting), with final decisions remaining inside the 4.25-4.50% range. The framework isolates small, scenario-dependent wedges between behavioral and rational baselines, offering an in silico environment for counterfactual evaluation in monetary policy.