Tara Sinclair on Building a Synthetic FOMC Through AI

Episode Summary

Executive Summary: The episode explores Tara Sinclair and Sophia Kazanik’s paper, “FOMC in Silico,” which uses LLM-based multi-agent simulations to model FOMC deliberations, personalities, and committee dynamics. The conversation also covers the importance of timely government data, the limits of private-sector substitutes, and how AI could improve forecasting, central bank research, and understanding of monetary-policy decision-making without replacing human accountability.

Main Topics: AI Simulation of the FOMC: Sinclair explains how the paper builds simulated Federal Reserve personalities and uses them to recreate an FOMC meeting structure in silico, aiming to study committee decision-making and deliberation. Why Macro Needs Simulation: Because macroeconomics cannot run controlled experiments, the paper is framed as a sandbox or flight simulator for testing how policy settings, meeting design, and information flows affect outcomes. Government Data Gaps and Policy Risk: The hosts discuss missing October data, the likelihood that jobs and CPI reports may not arrive on time, and how policymakers are forced to operate with blind spots during a potentially critical turning point. Deliberation, Personality, and Institutional Context: The project goes beyond voting models by trying to capture persuasion, agenda-setting, career incentives, and committee interaction—features that are central to how the FOMC actually works. Synthetic Beige Book and Data Inputs: The paper also generates a synthetic Beige Book from public sources and feeds simulated agents macro snapshots and district context, illustrating how AI can create timely research inputs. Implications for Forecasting and Fed Governance: The discussion considers future uses such as testing chair candidates, rotation rules, meeting timing, transparency, and whether AI tools can help the Fed evaluate institutional design choices. Training Economists for an AI World: Sinclair argues that students should learn core economic intuition, plus enough DSGE/time-series/LLM literacy to interpret black-box outputs and apply AI responsibly in macro research.

Key Arguments: Missing government data is especially dangerous during economic turning points because policymakers can be forced to make decisions without a reliable read on labor markets or inflation. Private-sector data can complement official statistics but cannot replace them because it often lacks representativeness and historical continuity, especially when behavior changes across payment types or subgroups. Macroeconomics needs simulation tools because it cannot run real experiments; LLM-based agent models provide a new way to test how committee structures and information affect policy. Pure voting models miss the “shadow” between ideas and outcomes: deliberation, persuasion, and interaction are essential to understanding committee decisions. The paper’s simulated FOMC members are built from publicly available information, especially speeches, with recency weighting proving important for realism. Chair Powell is modeled both as one vote and as an agenda setter, reflecting his dual institutional role in the FOMC process. The simulation can incorporate political pressure and incentives; doing so produced more dissents without materially changing the policy range, suggesting committees remain resilient. Data revisions matter: when the model was rerun with revised payroll data, it became modestly more dovish and more dissent-prone, showing that updated information can change policy reasoning. AI should not replace human accountability at the top of monetary policy, but it can improve research, stress tests, forecasting, and institutional design. Economists should engage with AI directly so economic theory and institutional knowledge shape how these tools are used, rather than leaving the field to computer scientists alone.

Data Points: Conference cadence: about every 18 months - Frequency of the Federal Forecasters Conference organized by the Federal Forecasters Consortium. Number of agencies: about 10 federal government agencies - Agencies leading the Federal Forecasters Consortium. Treasury role duration: 2022 to 2024 - Sinclair served as Deputy Assistant Secretary for Macroeconomics at Treasury during this period. LLM version: GPT-4.0 - Used as the starting point for simulating FOMC member personas. Model horizon: July 2025 meeting - The paper simulates a specific FOMC meeting ex ante, before it occurred. Fed funds representation: to two decimal points - Simulated policy choices are modeled as specific rates rather than just a range. Federal Reserve structure: 12 regional feds - The synthetic Beige Book is constructed from the 12 regional Federal Reserve districts. Meeting timing: 2 days later - The July 2025 FOMC simulation was followed by a jobs report two days later, motivating a rerun with revised data. Data revision scope: 2 months - The rerun incorporated revised May and June payroll data in addition to July data. Speaker count in model: multiple simulated agents - Examples include Sim Powell, Sim Waller, Sim Bowman, and Sim Cook representing FOMC participants.

Pivotal Quotes: "“we really can't run experiments”" — Tara Sinclair: Explaining why macroeconomics benefits from LLM-based simulation and sandbox-style modeling. "“it's like driving without the headlights on”" — Tara Sinclair: Describing the danger of making policy when key government data are missing. "“between the idea and the reality falls the shadow”" — David Beckworth citing T.S. Eliot / discussed by Sinclair: Framing the gap between simple voting models and the real-world deliberative process inside the FOMC.

Implications: The episode suggests AI can become a serious research tool for macroeconomics, especially for policy simulation, forecasting, and institutional design. But it also reinforces that central banking still needs human judgment, accountability, and high-quality official data.