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—using heterogeneous agents representing clustered policymaker archetypes—to forecast federal funds rate decisions. A Chain-of-Draft (CoD) prompting extension improves both accuracy and token efficiency. On 16 meetings from 2023–2024, FedSight CoD achieves 93.75% accuracy and 93.33% voting stability, outperforming prior MAS and ML baselines.

Problem

Traditional econometric and ML models treat FOMC rate decisions as a black-box mapping from indicators to outcomes, ignoring the deliberative, multi-stakeholder nature of committee policymaking. They also cannot incorporate unstructured inputs (Beige Books, speeches, dot plots) that materially influence policymakers' reasoning.

Method

FedSight AI is built on CrewAI and instantiates five heterogeneous agents in a fixed pipeline: an Analyst (interprets Fed Funds Futures), an Economist (proposes three candidate policy options—dovish, neutral, hawkish), and three Member agents derived via clustering of historical FOMC participants into archetypes (Regional Pragmatists, Academic Balancers, Central Policymakers). Each Member ingests both structured inputs (PCE, CPI, Treasury yields, GDP, unemployment, VIX, political indicators) and unstructured inputs (Beige Book verbatim text, verbalized dot-plot distributions, FedWatch probabilities). Members deliberate, exchange perspectives, and cast votes; the system emits a simulated FOMC statement. The CoD extension enforces minimal-step reasoning drafts (\(\leq 30\) words per step) followed by revision, reducing token consumption while preserving reasoning fidelity. A separate ICL variant stores agent reflections on historical meetings as long-term memory for in-context fine-tuning. All agents use GPT-4o.

Key Contributions

  • Multi-agent deliberative architecture that treats FOMC decisions as emergent institutional outcomes rather than statistical mappings.
  • Fusion of structured macroeconomic indicators with unstructured qualitative sources (Beige Book, dot plots, FedWatch) inside agent deliberation.
  • Clustering-based agent archetypes grounding heterogeneity in historical FOMC member characteristics without inflating system size.
  • Chain-of-Draft (CoD) adaptation for multi-agent settings, achieving best accuracy and lowest token cost simultaneously.
  • Six-dimensional evaluation framework (accuracy, agent accuracy, voting stability, semantic similarity, average tokens, MAE).

Results

  • FedSight CoD: 93.75% total accuracy, 90.22% agent accuracy, 93.33% voting stability, MAE = 0.0156, ~60k tokens/meeting.
  • FedSight baseline: 87.50% accuracy, 86.67% stability, ~75.7k tokens/meeting.
  • FedSight ICL: 87.50% accuracy, 88.54% stability, ~81.3k tokens/meeting (higher token cost than baseline).
  • vs. MiniFed on 2018 test set: FedSight CoD 100% vs. MiniFed 75%.
  • vs. Ordinal Random Forest on 2023–2024: FedSight CoD 100% directional accuracy vs. RF 62.5%.
  • vs. Linear Regression (structured features only): LR achieves 31.25% directional accuracy.
  • Semantic similarity to actual FOMC statements: ~73–75% cosine similarity across variants (baseline marginally highest at 74.58%).

Limitations

  • Test set is only 16 meetings (2023–2024)—insufficient for robust statistical significance claims.
  • Results depend on GPT-4o behavior; model version changes or hallucinations could break reproducibility.
  • Potential LLM training data leakage: GPT-4o may have seen historical monetary policy documents, though the authors argue deliberation over contemporaneous data drives gains rather than memorized recall.
  • Results may not generalize under major political regime change or structural macroeconomic shifts.
  • Code is not released due to university-industry IP constraints.

Relevance to Harnesses / Meta-Harnesses

FedSight AI is a domain-specific multi-agent harness: it orchestrates a fixed pipeline of specialized agents (Analyst → Economist → Members → Vote → Statement) over heterogeneous input modalities, directly illustrating how harness-level design choices (agent sequencing, role specialization, memory injection, prompting strategy) determine end-to-end system performance. The CoD vs. ICL comparison is effectively an ablation of two harness-level reasoning modules, showing that prompt-level efficiency constraints (CoD) outperform memory-augmented replay (ICL) on both accuracy and cost—a transferable finding for any multi-agent harness designer. The CrewAI-based implementation and the six-dimensional evaluation suite (accuracy, stability, semantic alignment, token efficiency) provide a concrete template for benchmarking harness variants in time-sensitive, decision-making domains.