FedSight AI: Multi-Agent System Architecture for Federal Funds Target Rate Prediction

Battle Count: 67

Highly relevant for macro-driven quantitative strategies. Accurate FOMC rate predictions (93.75% accuracy, 100% directional accuracy) directly inform interest rate futures positioning, yield curve trading, FX strategies, and fixed-income portfolio management. The interpretable reasoning chains provide alpha signal context. However, the 16-meeting test set and LLM stochasticity limit production deployment reliability. The framework is more suited for strategic asset allocation and macro overlay than high-frequency trading.

Moderate complexity. Requires CrewAI framework setup, GPT-4o API access, structured data pipelines from multiple sources (FRED, Fed website, MacroMicro), unstructured text preprocessing (Beige Book, Dot Plots), clustering procedure for agent archetypes, and careful prompt engineering for multi-stage deliberation. The CoD mechanism adds prompt design complexity. No specialized hardware needed (CPU-only LLM inference), but orchestration of 5 agents with sequential/parallel task execution requires careful workflow design. ~26M tokens consumed across experiments.

Detailed methodology, agent architecture, data sources, and feature descriptions are provided. However, no code is released due to IP restrictions from the Duke-BNY capstone collaboration. The paper uses standard public data sources (Fed website, MacroMicro) and a well-known LLM (GPT-4o). The clustering procedure for agent archetypes and specific prompt engineering details are described but not fully reproducible without code. Approximately 26 million tokens were used across experiments.