TLDR: This research proposes a multi-agent LLM framework that simulates the Federal Open Market Committee’s (FOMC) deliberative process to predict monetary policy decisions (Raise, Hold, Lower). LLM agents, each with distinct hawkish or dovish beliefs, iteratively debate and revise predictions based on qualitative texts (Beige Book) and macroeconomic data. The model, which includes a latent belief variable for interpretability, significantly outperforms baselines, demonstrating the importance of debate and diverse beliefs in accurate forecasting.
Forecasting the decisions of central banks, particularly the Federal Open Market Committee (FOMC) in the U.S., is a critical challenge in today’s uncertain economic climate. Traditional methods often fall short because they don’t account for the complex, deliberative nature of how these policy decisions are actually made.
A new research paper introduces an innovative framework that mimics the FOMC’s collective decision-making process using multiple large language models (LLMs) as interacting agents. This approach aims to provide more accurate predictions of monetary policy changes, such as whether interest rates will be raised, held, or lowered.
The core idea is to simulate a debate among FOMC members. Each LLM agent starts with a unique “belief” – for example, being “hawkish” (favoring tighter policy) or “dovish” (favoring more accommodative policy). These agents then make initial predictions based on both qualitative information, like the Federal Reserve’s “Beige Book,” and quantitative data, such as inflation and unemployment rates.
What makes this framework unique is the iterative debate process. Over several rounds, agents observe each other’s predictions and justifications, then revise their own. This simulates the real-world discussions and consensus-building that happen within the FOMC. To make this process more transparent, the researchers introduced a “latent variable” to represent each agent’s underlying hawkish or dovish belief. This variable helps explain how individual perspectives and social influence shape the collective forecast.
Empirical results show that this debate-based multi-agent LLM approach significantly outperforms standard LLM-based models in predicting FOMC policy decisions. The study evaluated the model on 60 FOMC meetings between 2000 and 2025, achieving an F1 score of 0.48. The research highlighted the crucial role of the Beige Book and macroeconomic indicators in achieving accuracy. Furthermore, the debate rounds were found to be essential in mitigating a common bias towards predicting a “Hold” decision, leading to more diverse and accurate outcomes.
Also Read:
- Urban-MAS: A New Approach to Understanding Cities with AI
- Enhancing LLM Collaboration: Overcoming Lazy Agents in Multi-Agent Reasoning
This novel framework not only improves prediction accuracy but also offers valuable insights into the dynamics of collective policy forecasting by explicitly modeling how individual beliefs and interactions influence the final decision. For more in-depth information, you can read the full research paper: Modeling Hawkish-Dovish Latent Beliefs in Multi-Agent Debate-Based LLMs for Monetary Policy Decision Classification.
