Tara Sinclair on Building a Synthetic FOMC Through AI

Episode Summary

Executive Summary: The episode centers on Tara Sinclair’s paper, which uses LLM-based multi-agent simulations to model FOMC deliberations, personalities, and voting dynamics as a macroeconomic sandbox. The conversation also highlights the urgency of missing government data, the limits of private-sector substitutes, and broader implications for forecasting, Fed governance, and how economists should adopt AI without abandoning economic foundations.

Main Topics: Missing government data and policymaking blind spots (Priority: 5/5): Sinclair warns that absent October GDP, jobs, and CPI releases could leave policymakers and researchers effectively "driving without the headlights on," especially during a potential economic turning point. FOMC in Silico: AI as a macroeconomic sandbox (Priority: 5/5): The paper uses LLMs to simulate FOMC members and deliberation in order to create a flight simulator for testing how information, meeting structure, and institutional processes affect monetary policy choices. Deliberation, not just voting (Priority: 5/5): Sinclair and Beckworth emphasize that existing committee models miss the crucial middle layer: how discussion, persuasion, norms, and personalities shape final outcomes beyond individual votes. Synthetic Beige Book and data inputs (Priority: 4/5): The project generates a synthetic Beige Book from publicly available information and feeds simulated members macro snapshots, prior outcomes, and district context to approximate real FOMC meeting information. Political pressure and institutional incentives (Priority: 4/5): The simulation explores how reduced chair agenda-setting power and dovish career incentives can create dissents without radically changing the overall rate decision, echoing real committee dynamics. Future applications and Fed governance experiments (Priority: 4/5): Potential uses include simulating alternative chairs, meeting timing, rotation rules, transparency levels, and the effect of inside-Fed information such as the Teal Book on expectations and market functioning. AI, forecasting, and macro training (Priority: 4/5): Sinclair argues economists should use AI to advance research, but students still need core economic intuition, DSGE fundamentals, and an understanding of the limits of black-box tools.

Key Arguments: Government data remain essential because private-sector data lack long historical series, representativeness, and robustness during structural change. Missing October data could permanently weaken both real-time policy decisions and later empirical analysis if the data are never produced. LLM-based agent simulation can serve as a laboratory for macroeconomics, where actual experiments are impossible. A committee’s value lies not just in votes but in discussion and persuasion; ignoring deliberation omits a key part of policymaking. Recency-weighted speeches materially improve the realism of simulated FOMC personas. Political pressure in the model mainly increased dissents rather than overturning the overall policy outcome, suggesting committee structure is resilient. The model can be used to test institutional design questions, including chair selection, rotation, meeting structure, and transparency. Economists should adopt AI themselves so the field shapes how these tools are used in policy and research, rather than ceding that role to computer scientists alone.

Data Points: Frequency of Federal Forecasters Conference: About every 18 months - Sinclair describes the recurring conference organized by the Federal Forecasters Consortium. Number of federal agencies involved: About 10 agencies - The consortium is led by roughly ten federal agencies with forecasting responsibilities. Meeting modeled in the paper: July 2025 FOMC meeting - The simulation was built before the actual July 2025 meeting occurred. Interest-rate precision in the simulation: Fed funds rates to two decimal points - The model lets simulated members propose highly specific rate levels, more precise than the real FOMC range. FOMC structure in simulation: 12 regional Feds contribute to the Beige Book - The synthetic Beige Book is built from public sources for each of the 12 Fed districts. Typical data-release timing example: Jobs report released two days after the July 2025 meeting - Sinclair says revised May and June employment data became available shortly after the meeting. Conference cadence reference: Every 18 months - Reiterated as the schedule for the Federal Forecasters Conference.

Pivotal Quotes: ""we're driving without the headlights on"" — Tara Sinclair: Used to describe the danger of missing key October economic data for policymakers and analysts. ""a flight simulator for being able to test out different pieces of information, different processes for decision making, all in silico"" — Tara Sinclair: Explaining the motivation for using LLMs to simulate FOMC deliberations. ""The whole idea of having a committee... is not just about representing different views, but allowing them to discuss those views"" — Tara Sinclair: Her explanation of why deliberation matters beyond simple individual voting.

Implications: The discussion suggests AI can become a serious macro tool for experimentation, forecasting, and institutional design, but only if anchored in economic theory and validated against real-world behavior. It also reinforces that high-quality public data remain indispensable to credible policy.