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

FedSight AI: Multi-Agent System Architecture for Federal Funds Target Rate Prediction Yuhan Hou∗1, Tianji Rao∗1,2, Jeremy Matthew Tan∗1, Adler Viton∗1,2, Xiyue Zhang∗1 David Ye1 , Abhishek Kodi2 , Sanjana Dulam2 , Aditya Paul2 , YiKai Feng2 1Duke University, Durham, NC, USA 2BNY AI Hub, New York, NY, USA Abstract The Federal Open Market Committee (FOMC) sets the federal funds rate, shaping monetary policy and the broader economy. We introduce FedSight AI, a multi- agent framework that uses large language models (LLMs) to simulate FOMC deliberations and predict policy outcomes. Member agents analyze structured indicators and unstructured inputs such as the Beige Book, debate options, and vote, replicating committee reasoning. A Chain-of-Draft (CoD) extension fur- ther improves efficiency and accuracy by enforcing concise multistage reasoning. Evaluated at 2023-2024 meetings, FedSight CoD achieved accuracy of 93.75% and stability of 93.33%, outperforming baselines including MiniFed and Ordinal Random Forest (RF), while offering transparent reasoning aligned with real FOMC communications. 1 Introduction The Federal Open Market Committee (FOMC) sets the federal funds rate, and its decisions reflect diverse philosophies and regional concerns. Traditional econometric models assume static relation- ships, while machine learning methods improve accuracy but remain opaque and cannot incorporate unstructured inputs such as speeches or Beige Books, which strongly influence policy makers’ rea- soning [1–3]. Leveraging LLMs, prior multi-agent work simulates FOMC-style deliberations [4]. Building on this, we present FedSight AI: agents jointly analyze structured indicators and unstructured narratives, deliberate, vote, and produce forecasts with interpretable reasoning; a CoD mechanism streamlines multi-stage reasoning [5]. Our contributions are threefold: (1) one of the first forecast- ing systems treating FOMC decisions as deliberative institutional outcomes rather than black-box mappings; (2) integration of structured indicators and unstructured narratives in agent deliberations; and (3) FedSight CoD: 93.75% accuracy and 93.33% stability on recent meetings—outperforming MiniFed [4] and an Ordinal Random Forest [6]—with transparent, FOMC-aligned reasoning. 2 Related Studies Interest Rate Prediction. Forecasting interest rates has been widely studied through diverse approaches. Classical econometric models, such as the expectations hypothesis [7] and the Taylor Rule [8], provided interpretable but simplified frameworks. Later, time-series models including VARs and DTSMs [9, 10] incorporated macroeconomic dynamics, while machine learning approaches—random forests, boosting, deep neural networks, and LSTMs [3, 11]—captured nonlinear relationships and sequential dependencies. More recently, an Ordinal Random Forest model [6] is proposed and achieved strong predictive performance, offering a benchmark for our research. Yet these methods remain limited by opacity and by their inability to reflect the deliberative nature of collective policymaking, motivating alternative frameworks. ∗Equal contribution. Authors listed alphabetically. 39th Conference on Neural Information Processing Systems (NeurIPS 2025) Workshop: Generative AI in Finance. arXiv:2512.15728v1 [q-fin.GN] 5 Dec 2025 Multi-Agent System Frameworks. Parallel to econometric and Machine Learning (ML) models, advances in multi-agent systems (MAS) with LLMs emphasize autonomy, communication, and collaboration [12]. In finance, FinCon demonstrated agent collaboration for portfolio management, but with hierarchical constraints. More relevant is MiniFed [4], which simulates FOMC meetings through five stages of agent discussion, persuasion, and voting, showing high predictive accuracy and behavioral realism. Our framework builds on these developments but diverges structurally: each agent represents an independent FOMC participant with unique interpretations, enabling a more faithful simulation of policy deliberations and providing interpretable interest rate forecasts. 3 Data and Methodology 3.1 Data Description Our target variable is the change in the Federal Funds Target Rate (FFTR) at each FOMC meeting, expressed in basis points. Our dataset covers 16 scheduled meetings from February 2023 to December 2024. Structured Inputs. We compile structured predictors from six domains: inflation, monetary indicators, economic activity and growth, political environment, past rate decisions, and market expectations. These variables capture the standard economic signals considered in prior literature and are aligned to values available two days before each meeting. The complete variable list and definitions are provided in Appendix A.1 Table 2. Unstructured Inputs. Beyond numerical indicators, we incorporate unstructured data that reflect qualitative aspects of FOMC deliberations. First, the Beige Book2 provides anecdotal evidence from all 12 Federal Reserve districts, covering lab

…(Full text truncated)…

This content is AI-processed based on ArXiv data.