Go to ACL ARR 2026 January homepageFedAgent: An Institution-Constrained Multi-Agent System Architecture for FOMC Meeting Simulation
Keywords: Multi-Agent System, Large Language Model, Real World Simulation, Federal Open Market Committee
Abstract: Decisions of the Federal Open Market Committee (FOMC) regarding the federal funds rate play a central role in shaping financial market conditions and macroeconomic dynamics. However, forecasting the decisions is challenging due to the complicated committee-based nature. Most existing approaches map economic indicators or textual signals directly to policy actions, abstracting away from institutional deliberation. We propose FedAgent, an institution-constrained multi-agent framework that models FOMC interest rate decisions as a structured process of collective reasoning. The framework decomposes policymaking into evidence construction, discrete policy option formation, role-conditioned deliberation, and rule-based aggregation, while explicitly modeling role heterogeneity and asymmetric informational salience. Evaluated on scheduled FOMC meetings from 2020 to mid-2025, FedAgent achieves higher accuracy and greater prediction stability than strong machine learning and LLM-based baselines, while providing interpretable diagnostics of committee disagreement and consensus.
Paper Type: Long
Research Area: AI/LLM Agents
Research Area Keywords: Language Modeling, NLP Applications
Contribution Types: Model analysis & interpretability, NLP engineering experiment, Data analysis
Languages Studied: English
Submission Number: 1674
Loading