Modeling Hawkish-Dovish Latent Beliefs in Multi-Agent Debate-Based LLMs for Monetary Policy Decision Classification

Battle Count: 70

The paper directly addresses FOMC policy rate decision prediction, which is highly relevant to quantitative trading strategies involving interest rate derivatives, FX markets, and equity sector rotation. The multi-agent debate framework provides interpretability into how different policy stances influence decisions. However, the moderate F1 score (0.476) and the 2-week prediction horizon limit immediate practical trading application. The framework could be integrated into macro-event-driven trading systems or used as a signal component in broader quantitative models.

The framework requires: (1) API access to GPT-4o-mini with structured outputs, (2) 7 agents with predefined belief profiles, (3) iterative debate loop (up to 10 rounds), (4) data pipeline for Beige Book text, FRED macroeconomic indicators, and Bloomberg policy rate data, (5) prompt engineering with specific templates, (6) majority voting and consensus detection logic. The Bayesian theoretical framework is for interpretability and not directly implemented in the experiments. Moderate complexity due to multi-agent orchestration and data integration requirements.

The paper provides detailed prompt templates, experimental settings (7 agents, 10 max rounds, temperature=1, GPT-4o-mini), dataset description (Beige Book corpus from prior work, FRED data, Bloomberg terminal), and ablation study configurations. However, no code repository is mentioned, and the Beige Book corpus reference is to a Japanese-language conference paper. The dataset spans 2000-2025 with 60 selected meetings. Reproducibility is moderate as it relies on API access to GPT-4o-mini and specific data sources.