397612 05: New Series 2001 one dollar bill notes pass through a printing press November 21, 2001 at the Bureau of Engraving and Printing in Washington, DC. The new dollar bills contain the signatures of U.S. Treasury Secretary Paul O”Neill and U.S. Treasurer Rosario Marin. (Photo by Alex Wong/Getty Images)
Getty Images
Move over, model UN – researchers at George Washington University have made history by successfully building an AI model of a Federal Reserve bank committee meeting.
The project, called “FOMC in silico,” simulates a meeting of the central bank’s Federal Open Market Committee, complete with AI agents as board members. The team had information about each member’s attitudes toward fiscal policy and prior voting records, and they put all of that into the hopper.
“No existing approach lets us observe rational and behavioral decision rules on the same information set,” authors Sophia Kazinnik and Tara M. Sinclair explained in the paper covering this project. “We address this gap with a dual track simulation framework that integrates an LLM-driven simulation of the FOMC meeting with a game theory voting model. In the rational (game theory) track, policymakers observe macro data, update beliefs via Bayesian inference, and vote to an equilibrium. In the behavioral (LLM) track, the agents get the same data, reason in natural language, engage in committee debate, and generate outcomes shaped not only by the data but also by persona heterogeneity and institutional norms that formal models struggle to capture.”
So in one of these “tracks,” there’s no natural language dialogue, just “Bayesian inference,” and in the other, they “talk” and come to decisions that way. Using terms like “persona heterogeneity” (or: people think differently) they are saying a mouthful – what the writers seem to be trying to express is that you can’t really get a robust simulation without a rendering of people talking to articulate their thoughts like they would in a real live meeting.
Anyway, the major finding of this digital twinning of the FOMC is that, under political pressure, the board members will, as represented by agents, disagree and fracture consensus.
Building the People
MORE FOR YOU
One of the more interesting aspects of the project paper, I thought, was the authors explaining the data work that was done to create digital board members with context.
“Within the framework, we start by automatically ingesting real-time macroeconomic data,” they write, specifically citing speeches from the Governor, district-level reports, and financial news. “We then build detailed profiles for each committee member, combining their historical policy stances, biographies, recent speeches, district conditions, and the current macroeconomic context.”
In a way, this is a lot like those frequent attempts by marketing teams to build “personas” to figure out how to advertise to an audience. Here, though, the research team is trying to model people’s thoughts on financial policy, which is why the enumerated outcomes include initial rate recommendations, made up of a preferred policy rate, a confidence score, and a summary.
Then come the political pressure scenarios presented in “parallel Monte Carlo simulations,” where the model attempts to replicate public criticisms of Jerome Powell.
“(A) political pressure scenario (mirroring public attacks by the POTUS on Chair Powell) reduces the Chair’s agenda weight, applies a dovish bias to his proposal, and imposes career-driven dovish shifts for members with high appointment probabilities,” Kazinnik and Sinclair write.
Essentially, by splitting “hawkish” and “dovish” sentiment, they set the stage for a more granular investigation of what each board member will be, and what the resulting game theory looks like, for a board that is ultimately trying to figure out whether to raise interest rates, lower them, or keep them the same.
The Committee Deliberates
In fact, to get a better idea of how all of this works, it’s helpful to look at the minutes of past, real, human FOMC meetings. Here’s an excerpt from the minutes of such a meeting earlier this year:
“Participants viewed the statement as having played an important role in building a common understanding among policymakers of the congressionally mandated monetary policy goals and of their monetary policy strategy. The statement and other means of policy communication also figured importantly in conveying that understanding to the public and helping to anchor inflation expectations, thereby supporting the effective transmission of monetary policy. Participants indicated that they were looking forward to their discussions at upcoming meetings and to hearing a range of perspectives at various Fed Listens events as well as at a planned research conference, and that they were approaching the review with open minds.”
The first sentence has something like 31 words and takes a while to say, let alone read and digest. I asked ChatGPT to come up with a simple version that keeps the detail of the original:
“The statement was seen as very important because it helped government leaders agree on what the Federal Reserve’s money goals should be, and how to reach them. It also explained these ideas to the public so people would understand what the Fed was trying to do, which helped keep people’s expectations about inflation steady. The group said they were excited to talk more at future meetings, at Fed Listens events, and at a research conference, and that they wanted to keep an open mind.”
You see how this works.
However, the simulation wasn’t looking at whether the board members are excited for upcoming events. It was aimed at action, at mirroring the positions that board members will take in various scenarios.
Then there’s this from the paper, which illustrates the old adage that if you want to come at the King, you’d better not take half-measures.
“Embedding reputation driven career concerns captures how members may adjust their positions to align with current or anticipated leaders. This creates a novel channel for political pressure on the Chair to spread: members reposition in anticipation of a leadership change, shifting the committee’s stance before the change occurs.”
The way I read this is that if the model introduces pressure on Powell as the Chair, it sees the agents scattering in terms of positions supporting his own, and taking new paths to suck up to some other leader.
The bottom line is that now that a team has simulated the FOMC with AI agents, it shouldn’t be long until we have tea leaf readers taking the same approach with corporate earnings meetings, PTA meetings, major league sports locker room meetings, and, well, all kinds of other group events. Stay tuned.