Skip to content

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

🕒 Published (v1): 2025-12-05 16:45 UTC · Source: Arxiv · Venue: NEURIPS 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

FedSight AI is a multi-agent LLM framework that simulates FOMC deliberations by having heterogeneous agents analyze both structured macroeconomic indicators and unstructured narratives (Beige Book, dot plots) to predict federal funds rate decisions. A Chain-of-Draft (CoD) extension enforces concise multi-stage reasoning, achieving 93.75% accuracy and 93.33% voting stability on 2023–2024 FOMC meetings. The system produces interpretable outputs semantically aligned with official FOMC communications.

Problem

Traditional interest rate forecasting models (econometric, ML) treat FOMC decisions as black-box statistical mappings, ignoring the deliberative, multi-stakeholder nature of committee policymaking and failing to incorporate unstructured qualitative inputs (e.g., Beige Books, dot plots) that materially influence policymakers.

Method

FedSight AI is implemented in CrewAI with five heterogeneous agents: an Analyst (interprets Fed Funds Futures for market-implied probabilities), an Economist (formulates three candidate policy options: dovish/neutral/hawkish), and three Member agents representing clustering-derived FOMC archetypes (Regional Pragmatists, Academic Balancers, Central Policymakers). Member agents consume both structured indicators (PCE, CPI, unemployment, VIX, Treasury yields, M2) and verbalized unstructured data (Beige Book passed verbatim, dot plots converted to natural-language distributions). Agents deliberate, exchange perspectives, and vote; the system outputs a simulated FOMC statement. The CoD extension constrains each reasoning step to ≤30 words, followed by revision passes, reducing token use while preserving reasoning quality. An ICL variant embeds historical meeting simulations as long-term agent memory. Backtesting runs each of 16 meetings five times to account for LLM stochasticity.

Key Contributions

  • Multi-agent deliberative framework (CrewAI) that treats FOMC rate decisions as emergent institutional outcomes rather than regression targets
  • Integration of structured macroeconomic indicators with unstructured Beige Book, dot plot, and FedWatch inputs within agent deliberation
  • Clustering-based agent archetypes grounded in historical FOMC member characteristics (hawkishness, regional affiliation, policy focus)
  • Chain-of-Draft adaptation for multi-agent settings: enforces minimal-step reasoning drafts, improving accuracy and reducing token cost simultaneously
  • Six-dimensional evaluation framework covering accuracy, agent-level accuracy, voting stability, semantic similarity to official statements, token efficiency, and MAE

Results

  • FedSight CoD: 93.75% total accuracy, 90.22% agent accuracy, 93.33% voting stability, 73.82% semantic similarity, 60,464 avg tokens, MAE 0.0156 on 16 FOMC meetings (2023–2024)
  • FedSight baseline: 87.50% accuracy, 86.67% stability, 75,724 avg tokens, MAE 0.0313
  • FedSight ICL: 87.50% accuracy, 88.54% stability, 81,303 avg tokens (highest token cost)
  • FedSight CoD vs. MiniFed on 2018 test set: 100% vs. 75% accuracy, MAE 0.00 vs. 0.0625
  • FedSight CoD vs. Ordinal Random Forest on 2023–2024: 100% directional accuracy vs. 62.5%
  • Linear Regression baseline on structured variables only: 31.25% directional accuracy

Limitations

  • Test set of only 16 FOMC meetings (2023–2024) is too small for robust statistical conclusions
  • Results depend on GPT-4o behavior; model version changes, behavioral drift, or hallucinations could alter reproducibility
  • Potential LLM training data contamination with historical monetary policy materials (acknowledged but not fully resolved despite mitigation claims)
  • Results may not hold under significant political regime change or structural economic breaks
  • Code and data not publicly released due to university-industry IP constraints

Relevance to Agentic AI / LLM Agents

FedSight AI exemplifies the institutional simulation paradigm in agentic AI: rather than using agents as function-calling wrappers, it encodes real-world organizational structure (committee roles, voting procedures, dissent dynamics) into the agent topology itself. The clustering-based persona design is a concrete technique for grounding heterogeneous agent identities in empirical data, directly relevant to multi-agent system design research. The CoD integration demonstrates that reasoning efficiency interventions developed for single-model settings can transfer to multi-agent deliberative pipelines with compounding benefits. This work connects to the broader thread of using LLM agents to simulate human collective decision-making processes in high-stakes domains.