Tara Sinclair on Building a Synthetic FOMC Through AI

Tara Sinclair is a professor and chair of the economics department at George Washington University. Tara returns to the show to discuss her ambitious paper simulating an FOMC meeting before it happens with LLM models, the process of building sim FOMC members, the importance of publicly funding economic data, the future of AI and macroeconomics, and much more. 

Subscribe to David's new Substack: Macroeconomic Policy Nexus

Read the full episode transcript:

This episode was recorded on October 27th, 2025

Note: While transcripts are lightly edited, they are not rigorously proofed for accuracy. If you notice an error, please reach out to [email protected]. 

David Beckworth: Welcome to Macro Musings, where each week we pull back the curtain and take a closer look at the most important macroeconomic issues of the past, present, and future. I am your host, David Beckworth, a senior research fellow with the Mercatus Center at George Mason University, and I’m glad you decided to join us.

Our guest today is Tara Sinclair. Tara is a professor and the chair of the economics department at George Washington University. From 2022 to ’24, Tara also served as a deputy assistant secretary for macroeconomics in the Office of Economic Policy at the US Department of Treasury. Tara and her co-author, Sophia Kazinnik, have a new paper out titled, “FOMC in Silico: A Multi-Agent System for Monetary Policy Decision Modeling.”

In this paper, Tara and Sophia use large language models, LLMs, to simulate the decision-making process of the FOMC right down to the personalities and debates inside the Eccles Building. It’s one of the most ambitious efforts yet to fuse AI with monetary policy analysis, with big implications for policy. Tara joins us to discuss this paper today. Tara, welcome back to the program.

Tara Sinclair: It’s great to be back. Thank you, David.

Beckworth: Well, it’s great to have you on. Now, this is your second appearance, Tara. You were previously on late 2024. We had a great time then, and we talked about your work on real-time data. In fact, you’d just been to a real-time data conference, so you’re real big on that. Your work at Treasury dealt a lot with that as well. You are also big on getting full funding for economic data, have the government supply all the data. 

Data and Policymaking

Here we are. We’re recording this near the end of October. GDP should be coming out, but it won’t. No employment reports. We’re blind on the real side of the economy. What are your thoughts on that? How do we move forward as policymakers when we don’t have the data?

Sinclair: Well, I’m really concerned about the missing data, particularly for October. At this point, it looks like we may never get a full jobs report for October. We may never get a full CPI report for October. The whole reason that we got the CPI data for September was because it was required by law for Social Security COLA adjustments. Now, it’s really unclear what’s going to happen in terms of data, now that we’re probably going a whole month without government involvement in data collection. That might be okay if we were in normal time, so we could just linearly interpolate or do a random walk to just move forward.

It may very well be the case that we’re driving without the headlights on, and then we’re going to turn them back on. All of a sudden, we’re going to see that the road has turned in another direction. That really has me quite concerned about missing data at this critical time, when there are so many changes, both in the economy in terms of thinking about a potential economic turning point, but also all of the changes in terms of policies that we’d really like to be able to see what’s going on, both in real time—as well, if we don’t ever produce this October data, then we’ll be missing a data point for us to be able to analyze it as we’re going in and doing that deep-dive research to analyze the impacts of various policies.

Beckworth: That is so interesting, the way you frame that. We are at a potential turning point. Is the labor market softening or not? Is it immigration or not? We don’t know. We may hit the wall, and we don’t have the data to tell us that we’re hitting the wall until it’s too late. Then also, the Fed’s making choices and policy based on ignorance, not on data. Likewise, the Trump administration has to make policy choices as well. We really don’t know what we’re doing without this data. It’s good to have people like you on top of this, beating the drums, get the data out there. Now, is there any hope we can turn to private sector, big data, or other real-time measures?

Sinclair: I have long argued that private sector data is complementary to government-provided data. I really don’t see that it can provide much of a substitute, particularly when we see economic structures changing, because we just don’t have the historical time series to be able to understand how we might see that change in private sector data the way that we can see it in government-provided data. We don’t have the representativeness.

If you just think of one example, let’s think about credit card data. There, that’s great for capturing trends in normal times. If we can assume that what’s happening in noncredit card transactions are following the regular pattern, we can just watch the credit card data. What about if people that are primarily relying on cash transactions are suddenly shifting in their behavior? We won’t see that. That could be a big problem for a turning point.

Beckworth: Okay, so we need the data. As someone who likes to think about macroeconomic policy from a nominal GDP perspective, this is not good. In fact, this is a great critique against nominal GDP targeting because you could have an administration who doesn’t get the data out. What do you do? Yes, this is not optimal at all. Now, Tara, you were both on the show not too long ago, but you were also at a recent conference that you invited me to on forecasting. The theme was related to AI, which is a nice segue into your paper. Tell us about this conference, and how can people participate in it if they’re interested?

Federal Forecasters Conference

Sinclair: Great. About every 18 months, we have this conference. It’s called the Federal Forecasters Conference. It’s organized by the Federal Forecasters Consortium. It’s primarily led by about 10 federal government agencies that all have their responsibilities around forecasting. Not even necessarily economic forecasting, but primarily economic forecasting. They get together, and we have a local conference. GW is the academic partner of that Federal Forecasters Consortium.

A couple of times, we’ve hosted it here at GW, and that’s where we hosted it this year. It’s just a great opportunity to bring together federal forecasters, practitioners from other locations, academics, and we can really have this two-way dialogue between academic researchers who may know the cutting edge of economic forecasting modeling, but they may not know what the real issues are that government and private sector forecasters are facing, and vice versa.

One of the things that I worry about, in particularly budget-constrained times for government forecasters, is they may not have the opportunity to get the in-service training and feedback on their modeling perspectives. This is an opportunity for them to both learn from academic forecasters that are working at the cutting edge about new types of models and techniques they could use, but  also to, during coffee conversations, talk about what they’re working on and get some potential feedback on model choices that might really improve the forecasts that they’re producing with taxpayer dollars.

Beckworth: It was a really interesting conversation. You had two presenters there. One of them really spoke to the potential consequences of AI, changing jobs, structural changes, including for forecasters. In fact, in the paper presented there, there was a discussion about people who collect data, who analyze data. This could be automated at some point with AI. Also, the art of forecasting.

The other presenter, he talked about, there’s these AI models. It’s kind of a black box. You just feed everything into it. They go through all versions of forecasts. They spit out the best one. There was interesting discussions. How can we use that? How can we explain that to our bosses? Just really fascinating conversation on the role AI is playing in the art of forecasting.

Sinclair: You did a fantastic job moderating that panel discussion.

Beckworth: Well, thank you.

Sinclair: I think everybody really appreciated that as well.

FOMC in Silico

Beckworth: Yes. Well, that’s a nice segue into your paper, which is also AI-based or themed. Again, that paper is titled, “FOMC in Silico: A Multi-Agent System for Monetary Policy Decision Modeling.” We’re going to jump right in. First, I must apologize to Tara and her co-author because I have taken advantage of this paper multiple times, both on my Substack, on Twitter, even on here a few times. I’ve used it to motivate a world that I imagine where the FOMC is largely run by AI, invoking Milton Friedman’s notion of a computer increasing the money supply gradually. That’s not the vision I have, but something along those lines.

That’s not the vision you have. That’s not the point of this paper, but I’ve had a lot of fun with it. I’ve created some art in ChatGPT, where there’s the FOMC sitting around a table and a bunch of robots, just a few. The chair and the two vice chairs are humans. That’s basically where it is. That’s not what this paper is about, despite my best efforts. You’ve corrected me several times, so I apologize. Let’s really talk about what the paper does get into. What was the motivation for this? What were you and Sophia thinking about as you got together to write this paper?

Sinclair: I think the big thing is, with macroeconomics writ large, one of the big challenges is that we really can’t run experiments. I see this as a first step in really creating a sandbox or perhaps a flight simulator for being able to test out different pieces of information, different processes for decision-making, all in silico, all within these LLMs to be able to learn about what shifts decision-making and policy, what might make a bigger difference, or a smaller difference.

I’ve always wondered if we had FOMC meetings more frequently or less frequently, or if they were more connected to timing of data releases. I think all of those sorts of questions might be able to be answered if we were able to, rather than having to somehow create multiple universes, where we’ve got FOMC members, where we can put them in very different situations. We could do that with computer simulations.

Beckworth: That is so interesting and so exciting to do basically experimental economics for macro because, typically, we think of experimental economics being done more applied micro settings.

Sinclair: Exactly.

Beckworth: This is really neat. This has always been the curse of macroeconomics, as you said. We can’t identify, and we can’t run natural experiments. We can’t put the world through multiple recessions to see if it’s sticky prices or sticky wages or sticky information that’s the main friction. This is so fascinating, and you guys are breaking ground here. Is anybody else doing this, or are you guys truly the ones leading the charge? You’re the avant-garde.

Sinclair: There are some computer scientists that have also created a simulated FOMC meeting. One of the frustrating things at looking at people who aren’t thinking about this from understanding macroeconomic policy is that it captures some aspects of the FOMC meeting, but it misses a lot of what we think of as being the key institutional knowledge that I think Sophia and I are really bringing to this paper.

Beckworth: Yes, so you guys bring in persuasion, institutional norms, career incentives, personalities, the real world. That’s what I want to get to. You talk, as you start the paper, about the shadow between the idea and reality. You have a T. S. Eliot line, “Between the idea and the reality falls the shadow.” Walk us through this shadow.

Sinclair: Sure. First of all, I will note that that is very much Sophia’s choice in terms of a T. S. Eliot quote. In terms of thinking about what we were trying to highlight there is that even if we use some of these rational voting models where we have the information about what the FOMC members views are, and then we have them vote, we’re still missing all of these dynamics in the middle about how people are interacting and how they might be able to sway each other’s opinions. The whole idea of having a committee, I think, is not just about representing different views, but allowing them to discuss those views and really weigh out the evidence and come to a better decision than what would happen with just the individual views and voting and not talking to each other. That’s what we think of as that shadow.

Beckworth: Okay. You use two different models. You use a clean Bayesian game, and then the LLM model. We’ll get to that and discuss that. You bring in humanity into this.

Sinclair: Exactly.

Beckworth: This is really fascinating. It really makes me think of all the applications going forward you could use this for. If you thought of any other ideas, I don’t want to maybe steal any paper ideas from you and put it into public, but any other applications you have in mind?

Sinclair: There’s two key ones that I’m really excited about, and that is as soon as we get some more time and resources. Despite the fact that this is wonderful use of AI, it does still require human time to make all of the planning and decisions, and setting it up. The first one is, obviously, we’re continuing to discuss who is going to be the next FOMC chair. Wouldn’t it be interesting to see how different candidate chairs would interact with the rest of the existing FOMC, and what we might expect votes to look like in those cases? That’s one that I’m very keen to work on.

Another one, and this was actually the initial plan of what we thought would be our starter project. Then we were like, “No, we need to go back one and write this paper.” Initially, we wanted to just immediately dive in and focus on the dynamics and, from meeting to meeting, have that dynamic aspect of things. Right now, we’re only simulating one FOMC meeting, the July 2025 meeting. It would be very cool to see how reputational effects and interactions over time might change what our results look like. In particular, I think one way to validate what we’re doing, I would really love to have them produce all of their summary of economic projections contributions for each of the sims and see if we can connect the dots.

Beckworth: That would be fascinating. I think there’s a lot of applications. Again, you guys are doing the cutting edge. Maybe tell us about your journey into this. LLMs, is this something you’ve been working on? AIs? This is not something anyone can just jump into. You’ve got to actually know how this works. You must have spent some time scaling up on it.

Sinclair: Yes, it was interesting. I was working at Treasury when we had that first burst of realization across the world that LLMs were really useful and interesting. I was immediately playing with it on my phone and things in my personal life because I couldn’t use it initially in my professional life because, as you probably remember, David, the security is quite tight on the technology that we used inside of Treasury.

It was something I was definitely interested in. Sophia has been at the Stanford Digital Economy Lab. This is what she’s been doing. Shortly after I returned to GW, we invited Sophia to come and give a seminar because she had sent me one of her papers she was working on, “Simulating the Survey of Professional Forecasters.” She sent it to me in part because I had previously contributed to the Survey of Professional Forecasters.

I am one of her simulated agents in that model. I was immediately just so excited. We spent the whole day just talking ideas and coming up with more ideas. I think this project came together pretty quickly. It’s very much Sophia’s the one with the deep knowledge of all the intricate details. She’s the real prompt engineer and all of that. I’m coming at it more from the enthusiastic macro researcher who’s like, “Can we do this? What about this? Well, what if we change this? How does this work?” I ask her a lot of questions.

Beckworth: At some level, this reminds me of the science-fiction books and now TV show Foundation. You have these engineers who predict. You’re not quite doing that, but you’re beginning to do that. What’s interesting and what’s the challenge for anyone going into this area is modeling and forecasting human interaction, right? That’s exactly what you’re doing. The deliberations, the humanity.

I remember reading some forecasting books a few years back. They were telling us how awesome it is we can now forecast weather farther out. It’s hard to forecast the economy. The reason, and we all know this, is because people’s responses are endogenous to what’s happening. People adjust. They adapt. You’re actually tackling this head-on with AI. You’re doing that with two different models. You have a baseline model. I think you call it an LLM deliberation model. Walk us through these two models, and what do they each bring to the table?

Sinclair: Sure. The way that we’re setting this up is we first want to create these simulated agents. We call them Sim Powell, Sim Waller, Sim Bowman. We give them each their own personality and simulation. We train them on their past speeches and other publicly available information. That’s going to be a key point that we only have publicly available information about them. We’re able to create these personas based on that information, and then we use those personas in two different ways. In one track, we just use those to create their own individual views, and then we use a Bayesian voting model to have them vote. We consider that to be our rational baseline. That gives us a benchmark to compare to. This is something that people have been doing in the past to try and capture some aspects of committee decision-making. 

It misses, again, that whole shadow in between of the deliberation step. Once we’ve set up these individual personas, then instead of having just that rational voting step, we instead have them interact and follow as closely as possible an actual FOMC meeting structure and schedule. We think that’s really where we can get some more behavioral responses.

Beckworth: Tell us more about how you collect data on each FOMC member. Sim Powell, Sim Waller, Sim Bowman, Sim Cook. Do you have to literally go out and get all of their speeches and interviews from television and feed that in?

Sinclair: A couple of things. We’re starting with OpenAI’s GPT-4o. That already has a lot of information inside of it. We can already say, “Hey, you’re going to be acting as if you are Waller. Tell me what you know about yourself.” We can ask those sorts of things. We do specifically get it to focus on some core pieces of information that we think are important to ensure that the personas are accurate. Speeches is really a key one. One of the interesting things that we found out is that it is important to recency-weight those speeches, because if we just give equal weight to the speeches, then the persona doesn’t look very much like who they are today.

Beckworth: Interesting.

Sinclair: We really want to look at their more recent speeches. There’s a couple of FOMC members where that’s particularly important. Across the board, we initially were just like, “Here are the speeches.” Then, when we got back, we were like, “That doesn’t sound quite right.” Then we were like, “Oh, wait. What if we say, ‘Here’s your most recent speeches? Give that 100% weight,’” and then let’s go back in time. You can still have those past speeches, but give them less weight. We would ask them to give little statements about what they’re thinking about the economy today. It started to come out much more like what we’re hearing from the actual FOMC members.

Beckworth: Imagine you’re a big hedge fund. You’ve got unlimited resources, practically unlimited resources. You want to replicate this exercise that you did. Do you think it’d be worthwhile for someone to go and look at speeches as well as what are their hobbies? We know Chair Powell likes rock and roll. He likes golfing. He loves playing the guitar. We know Governor Waller likes to do powerlifting and shoot big guns. Do you think there’s useful information beyond speeches that would help better predict or mimic the actual people?

Sinclair: I don’t know. Maybe. That’s definitely something that we could explore and test. That’s really the point of setting up this sandbox is that, now, we can go in and see if these sorts of features matter or don’t, particularly if we have them vote before an FOMC meeting, and then we can validate and see how it looks. We can have simulations with all these different features turned on and off to test exactly that. We haven’t done that so far. We’re just building the bare bones setup.

Beckworth: Right, but you’ve put in motion an idea that I think is going to really take off. Now, again, I’m just going way out there like I do, so I apologize, Tara. I mentioned hedge fund. Imagine if you’re someone at the NSA who has access to all the information on these people’s lives, right? It’d be an interesting exercise to see what it does. Maybe it wouldn’t be. Maybe there’s a lot of information that’s not important to find out.

Sinclair: Some interesting way is when I think the FOMC members actively are trying through their transparency and communication to share what they think is important for making their decisions. We’re taking them at their word there that they are being transparent about that. Now, if we did have concerns about transparency, or perhaps we think that there are things that maybe they don’t even know are affecting their decision-making process, then that might be something we’d want to explore more specifically. I think as economists, we tend to take things in steps. Going from pure rational decision-making to assuming that they are being transparent about the information that’s influencing their decision-making seems like a reasonable next step.

Beckworth: That is, again, so interesting. So many applications. I look forward to your further work on this. Now, you also have Chair Powell playing a dual role in this framework. Tell us why that’s important.

Sinclair: Well, it’s important because Chair Powell has highlighted that it’s important that he is both an individual voting member, and so his vote counts just one vote just like everybody else’s when they tally up the votes. As the chair, he’s the agenda-setter. That’s the explicit terminology that’s used when talking about these sorts of voting models. He’s the one who initially sets up where we’re going to focus for talking about things. In other words, say, “Hey, so we’re thinking to stay within this range today.” I will note, our sims are very specific. They’re actually fed funds rates to two decimal points, which is very different.

Beckworth: Wow.

Sinclair: This is something that keeping a range is actually more challenging for us, so we’re letting them be very, very specific, and then comparing that to the range that actually comes out from the FOMC. He’d be proposing a very specific interest rate, and then they’d debate around that.

Beckworth: How do you get them to deliberate in your model? Are they interacting? Are you asking questions?

Sinclair: Yes, it’s set up exactly like they’re having an FOMC meeting. Obviously, we don’t give them chairs, and we don’t set them around the big table. I do picture it that way in my mind, to be honest. I do picture them all in the computer, but sitting around chairs and looking at each other and talking to each other. They go through each of the standard steps. In fact, we even have a water-cooler-chats-ahead-of-time option, where they could discuss amongst themselves before going into the main meeting. Yes, we’ve set it up so that it follows exactly the way that the minutes say that an FOMC meeting is structured.

Beckworth: Okay, so you have these distinct and separate LLM models that represent each of the members—Sim Powell, Sim Waller—they’re interacting with each other. It’s a dynamic experience. They’re learning from each other. They’re growing off each other. That’s what captures this human element, which is so, so neat. Now, from all of this, you also produced something in your paper that was mind-blowing, a synthetic Beige Book. How did that come about?

Sinclair: This is something that, I admit, Sophia initially threw in there just as a little add-on. It now looks like it might become a paper of its own because people do really seem to like this synthetic Beige Book part of things. We use agent researchers, again, LLMs that are specifically trained to do research, to go out and produce. Each of the 12 regional feds contribute to the Beige Book, but this gives us an opportunity to collect, again, it’s going to be public data. It’s not like these agents are going and calling corporate contacts or small business contacts or anything like that. They’re getting it from publicly available information, newspaper reports, that sort of thing, and trying to build a synthetic Beige Book that is not limited to being released only every six weeks. We can update it any time. That way, if new information comes in, we could create a new Beige Book right away.

Beckworth: You’re giving the synthetic Beige Books to each of the sims, like, “Hey, use this information in here. Deliberate at the FOMC”?

Sinclair: Yes, so that’s another piece of information that we give. The information that the full FOMC, all the voting members get is they get a macro snapshot. We did choose which macro data we’re providing to them. I would love to give them the full Teal Book instead. Of course, that’s held inside the Fed, much guarded. This is something like, if the Fed were to take this model, they could give it the entire Teal Book. We’re giving it, basically, a one-pager of key data snapshots to make sure that they have the most recent data.

Then we give them the synthetic Beige Book, and then we remind the regional Fed presidents that are voting what region they’re representing so that they make sure that they keep that district context in mind for their deliberating. Those are the key things that we really focus on, is giving them the macro snapshot, the outcome from the previous meeting. They need to know that information so they know where they’re starting from, and then the Beige Book.

Beckworth: When you model the FOMC meeting, they’re interacting with each other, they’re discussing things, do you find things like debate, disagreements, like, “No, I think the rate should be this number,” “No, I think it should be higher”?

Sinclair: Yes, so there’s definitely that classic, it does look like FOMC minutes type of—

Beckworth: Amazing.

Sinclair: Then it’s, “The majority went this way, but there were some that felt—” all that classic FOMC language.

Beckworth: That, again, is so remarkable, so mind-blowing that you can do this and have that dynamic learning and interaction and so forth. In this process, you also bring in political pressure and institutional dynamics. You simulate a scenario where the chair’s authority is undermined by presidential attacks and career incentives, turn dovish. Tell us about that part of the story and the model.

Sinclair: Sure. Let me just set the stage just a little bit. Our baseline is we were first just trying to simulate the July 2025 FOMC meeting. I’ll note we were doing this before the July 2025 meeting. I think that that’s one thing that’s important in thinking about this because that ensures that, despite our best attempts of trying to tell LLMs not to use any information past X date, they want to please us. They want to give us all that information. Oftentimes, they will ignore that particular prompt. 

Actively doing it before that information was even available for them to go and get anywhere is important. Our baseline, we just gave them the regular data setup and everything like that, without telling them anything about the political context in which they’re operating in. That result came in on the high end of the range, but still within the same range that the FOMC voted for.

Then, because there was so much discussion, when we were thinking about, “Okay, we need a couple of examples to show how we might be able to use this model,” because that’s where we’re at. These are not our conclusions. These are just examples to try and test out how this model works. That first one, we’re like, “Well, we have to talk about political pressure.” We can’t not talk about political pressure if we’re going to talk about the July 2025 meeting.

What we did is we did it in two ways. Of course, other people could model political pressure in other ways. First of all, we gave the chair less agenda-setting power, meaning that there was less control over the consensus of the meeting from the chair. The second thing is that we basically set up that there could be career incentives for the personas if they were to consider being more dovish.

Sure enough, what we found is it didn’t affect very much the actual outcome of the meeting, except that it did create dissents. That’s really similar to what we actually saw in the July 2025 meeting. Just to be really clear, we’re not saying that those dissents that we actually observed in the real world in the July 2025 meeting were due to political pressure. We’re just saying that we did this simulated model, and we were able to politically pressure the model in a way that resulted in dissents.

Beckworth: Yes. Again, you’re recreating what actually happens in the real world. Not specific case, but just in general, you get political pressure, it does change your thinking. It changes your incentives, but it’s also reassuring. Your outcome still was within the range, which says, “Hey, this is a committee. You can’t stack the deck and change things.” That’s always been my comfort in the past few months when we’ve had some turmoil with who’s on the Board of Governors, who’s not.

This is a committee. You can’t just go in there. Even if there’s more people from one team, so to speak, they’re not all going to vote together. There’s going to be dissents. Your models just confirms that intuition I had. This is not like a slam dunk for any president. They can’t just take over the Fed. It’d have to be something much more radical. That, to me, was reassuring.

Sinclair: I think this is another reason why I wanted to push back against the idea of eventually having a mostly AI with just a couple of decision-makers. Having that representation across the country, across a range of different views, I think, is really important for decision-making. It may not matter most of the time, but it might matter at critical points when there is a lot of disagreement about the economic context we find ourselves in. I think July was one of those times.

Future Applications

Beckworth: Yes. I don’t foresee a world where AI does everything, particularly at the top. There’s going to have to be somebody who still bears the buck, that has the responsibility. The decision has to rest with a human, even if some of the things that you do get automated. At the end of the day, there’s art and there’s science. The art has to be done by a human, and they have to be held accountable for it. Still, reading this, it’s like looking back at Nixon-Burns era. There’s political pressure, although, again, the outcome’s very different, very targeted meeting, which is July ’25, not the 1970s. However, though, it does make me think. Have you thought about going into history and replicating Nixon-Burns or some other group?

Sinclair: Yes, we’d love to. We will probably, at some point, do so, but there will be challenges and questions about how well we are isolating the models from information that they do already know. That’s probably going to require building a completely bespoke LLM that doesn’t have access to historical information because, otherwise, we won’t be able to guarantee that they won’t be trying to give us the result that matches what actually happened rather than having that happen truly ex-ante.

Beckworth: Yes. If there are any big funders out there listening to the show, hit Tara up, and she’ll be glad to take the money to build this bespoke model. You can take a peek inside. She’ll let you participate. I think it would be fun to do this model. Here’s my idea. You come up with your all-star FOMC team. I come up with my all-star FOMC team from different periods, different eras. Let’s put them to the test. Let’s see what they do.

Bring in Volcker. Bring in Greenspan. Bring in Yellen. Bring in people that you think did a good job, whoever your people are. Set them in. Again, however you make them up. I don’t know. That’s the job you would do. Let them run with the data. I think it would be really cool, for example, it’s the day of an FOMC meeting, or it’s the week leading up to it, on CNBC, they had their, “Okay, let’s bring on our AI all-star FOMC panel.” You got Volcker. You got Greenspan. They say, “This is what we think should happen,” and then we can compare that to what actually did happen.

Sinclair: Yes, I love that. I may have to implement that immediately after this call. I think that that’s really great. We’ve definitely thought about using it for what now seems like more mundane ideas, but might be really relevant for the structure of the Federal Reserve system, such as, do we need all the representation of the different regional feds, or could there be a few that really cover the full national perspective? Does it matter the way that we rotate through the regional Fed presidents? If we grouped them differently, would that make a difference?

Beckworth: Oh, that’s interesting.

Sinclair: There’s a lot of those sorts of questions. I do want to highlight the other application that we used in this paper because I am personally really excited about it. I think it ties back to what we were talking about earlier regarding thinking about data and data availability. One of the things that I’ve often thought about with FOMC meetings is it’s like, “Oh, I wish this meeting were timed a little bit differently,” so that they could have had the employment report, or a little bit closer to CPI, or a little further away from CPI.

When exactly in the data release schedule the FOMC meets seems potentially pretty important. I’m sure that they think about all of that when they’re calendaring. This experiment that we did—this was just a surprise to us that we had the July 2025 meeting. Then two days later, we got the jobs report. It wasn’t just about the data that came in. It was about the revisions from the previous two months that we got. We got the July data, but we also got June and May revised. They were revised quite substantially downward.

Immediately, I messaged Sophia. I was like, “What would happen if we just gave our sims the different May and June data, and then had them rerun the meeting?” Now, we had to run it after the meeting actually happened, but we were still in the same sandbox, so I’m hopeful that they didn’t go out and get information about what happened at the meeting, and instead gave them this new data. Sure enough, they did become—not dramatically so, and that’s reassuring because one data point or two revised data points shouldn’t massively change things—but it did bring them both more dovish as well as more dissents.

Beckworth: Yes, so interesting. Major data revisions, would they have done things differently? Yes, they would have, but not too dramatic. Enough to notice and to matter. Again, it underscores the importance of getting good data in a timely fashion. This goes back to the point we made at the beginning of the show. It really is unfortunate that we’re going through this period where we don’t have good data for our policymakers to use in making their decisions. That’s really neat.

Broader Implications

What about some of the broader implications do you think for our profession? I’ve run VARs all the time. That’s one of the tools I learned in grad school. I put some structural identification on it. That’s aggregate data. I know there’s some stuff you can do with it, or even DSGE models. How could you take some of the lessons, the insights learned from what you’re doing, and bring it into the macroeconomics profession?

Sinclair: I do think that this is one way of bringing in behavioral micro foundations to some of our macroeconomic thinking. I don’t know exactly how to plug this into a DSGE model, but I could imagine that it could connect into heterogeneous agent like HANK-type models, and possibly bring in this sort of additional behavioral information in some way. I think this is a key thing that we, as macro researchers, need to stay on top of is that as this technology is advancing, if we aren’t using it to address economic questions, other people will be.

If we want to make sure that the economic way of thinking is still being used, we need to get in there and use this technology. This is one of the things when I saw that computer scientists were simulating FOMC meetings, I was like, “Wait a second. We have things to say here. We need to get in and make sure that the analysis that’s being used does fit into that longer economic understanding, that longer literature connects into economic theory, which I still think is fundamentally important for helping us think about these things, rather than just residing only in simulation space.”

Beckworth: Well, how fortunate that you are the chair of the economics department at George Washington University. I am genuinely curious when you think about PhD programs and going forward in the future. In fact, this came up at that conference, where we both participated. What do you teach a grad student? In the case of a forecasting model, this one presenter was talking about, “Hey, there’s a black box. Just throw everything in it, and let it spit out data.”

I’m not sure that’s the best answer. I can understand that maybe as an expedient answer, and maybe it’s something you can cross-check. I think he argued it’s another data input you could use, not to be the final answer. As you think about someone who’s going to start a PhD in the next few years, what should they be focusing on? Should they be learning DSGE? Should they be learning regular time series? Should they be doing LLMs?

Sinclair: Well, I think, honestly, our students are currently learning a little bit of each of those, and then deciding where they want to specialize. What I actually think is really important for students at every level, but perhaps particularly for PhD students, is to get that economic foundation and fundamentals and intuition. It’s all the more challenging when you’re working with black box outputs to really be able to put that in context if you haven’t developed that economic intuition and understanding. How do we know if this number seems really out of whack?

Well, we need to know some framework to think about what other variables should it be interacting with. Does it fit into a broader understanding? Maybe not doing the super big-scale models, but at least understanding the inner workings of a DSGE model for just developing that intuition, for that understanding of that interaction. Then, of course, macro, for a long time, has been moving more toward these micro foundations. I think there’s so much research that can still be done in terms of that connection between those micro foundations and these macro patterns that we see. Maybe LLMs can help us get more in that direction.

Beckworth: Okay, so we get richer models informed in part by what we learn from these LLM models that give us dynamic interactions, humanity, people, debate. Again, that’s so amazing: Sim Waller, Sim Powell, Sim Bowman having a debate, and you’re there watching it from the outside. Again, a remarkable time we live in.

Central Bank Governance and AI

Let’s go back to central bank and maybe central bank governance, and what they themselves can get out of this. Could we use these models to stress-test central bank governance? Maybe we should have more of the regional presidents voting all at once, or maybe fewer, or maybe rotate at a different pace. I don’t know. Any insights we can glean from what you’ve learned?

Sinclair: Yes. I don’t think we’ve got a clear takeaway on any of those questions yet, but I think we’ve got a framework now to start experimenting and thinking about those things. One thing that I hope is that perhaps the Federal Reserve might take a version of this model and be able to use information that they have inside, such as the Teal Book, to be able to do exactly those sorts of experiments to consider, like, “Oh, hey, maybe if we changed the order of rotation, or if we changed even the way that things are structured around the table.”

We’ve always heard that story about how Greenspan used to speak first. Bernanke switched it to speaking last. Maybe we could look into those sorts of things as well and how the meeting is structured, and change that around. Again, it would be a low-cost way of testing things out, and then they can test it in the real world. They can come up with maybe a couple of different options based on what they’ve seen. The FOMC members themselves could observe what is being proposed and maybe provide additional insight on whether that would work in the real world.

Beckworth: Yes. Lots of potential for governance reforms or improvements, and what works best. Just to play off of that, this whole notion of transparency and openness at the Fed. It used to be the case, as you mentioned, Greenspan spoke, and that was it. In fact, in 1994, I believe we start reporting the actual decisions. We have minutes coming out, and we start having governors and presidents giving speeches. Now, as some have said, we’ve gone too far. There’s too much noise. Do you think these models could be useful in helping us maybe find what’s the proper balance?

Sinclair: Absolutely. I think that could be very useful. Actually, one thing that I think would be fun, again, this would require inside the Fed information. I’m very much trying to sell this model to the Fed. I’m like, “Hey, Fed, maybe you want to market this.”

Beckworth: Please, Fed, listen and take note.

Sinclair: Because they could look at what you get with publicly available information, and then they could look at what you get with the information that they control inside the Fed. If there’s a big gap there, that suggests that the public may be surprised by what’s going to happen. That seems pretty useful information to have in advance.

Beckworth: Yes, so that could improve market functioning, less shocks to the system. That’s such a great point. Now, you used this to look at one meeting, right?

Sinclair: That’s right.

Beckworth: Could you foresee scenarios where you, looking forward—look, if I’m in the marketplace, we’re academics, I’m in a think tank, but we’re similar spirits, right? Put someone in the marketplace who’s trying to make money, right? They listen to the show like, “Wow, this is really a powerful tool. Why don’t I use this to forecast multiple meetings going forward, a dynamic version of this?” Any thoughts on that?

Sinclair: Yes, so that’s something we definitely want to do. That adds an amazing amount of complexity, but therefore interesting possible things we could learn from it because, now, we have the dynamic interaction amongst the FOMC members. We can also potentially have that real forecasting aspect to it that I think would both be potentially validating for our approach, because if we have them do something and then out of sample can check it, that’s, I think, quite valuable.

In addition, I do think it could be, at this point, a pretty expensive way to forecast FOMC meetings and the outcomes thereof. In as much as, in particular, not just the rate decision, but also the other aspects of the summary of economic projections, for example, might be market movers and might matter, then that’s something that being able to forecast each individual’s decision within this broader framework of consensus building could be quite informative, I think.

Beckworth: This is a really fascinating paper. We’ll provide a link to it. Listeners, please check it out, “FOMC in Silico.” Also, Tara’s made a really strong case to all of you listeners from the Fed to bring this model into the Fed and use it for your own benefit and for our benefit as well. The world will be a better place. 

I want to go from your paper to an AI conversation more generally in monetary policy. Again, we were at this conference that you helped organize. It was on the future of AI and forecasting. We mentioned earlier, there were some potentially ominous stories you could tell. If you’re not someone who can learn and ride the AI wave and benefit and be flexible, you might lose your job. If you’re someone who can go with it, you’re better off. What about monetary policy? Do you see any roles for the Federal Reserve in making more use of AIs, and what would it be? If I’m someone working at the Fed, what does that mean for me?

Sinclair: Well, I think people have made a similar argument for crypto that I’ll make for AI, which is that it’s going to be affecting our economy. Therefore, it is very important that our key policymakers and their advisers are very familiar with its capabilities. I think just, fundamentally, whether or not it would actually advance monetary policymaking in and of itself, I think being very familiar with how it works and what its constraints are will help them better understand how AI itself is going to impact the economy.

That’s something that, obviously, is a very hot topic. We saw it a lot at the Federal Forecasters Conference that we were both at. I think we’ve seen it a lot in other places where people really want to know, in particular, what it’s going to do for the labor market. That’s going to be very important for monetary policymaking in general.

Beckworth: Do you see a role for real-time data, one of your favorites, and AI coming together? For example, if I’m the actual chair, I’m actually Chair Powell, and I go into work in the morning, and I say, “Hey, AI, give me the most updated, real-time version of the Taylor rule based on inferences you can draw from spending, credit cards, other real-time data,” that seems like great possibility there.

Sinclair: It does. It does. I think we’re still at a point where we have to check it carefully. That’s really the key question is where in the workflow would AI fit, because we don’t want it too close to the top, because it does need to be reviewed and have someone make sure that it’s not off hallucinating in some way.

Beckworth: You would need somebody in monetary affairs division to build a model for the governors to use to check when they come in the morning. If you’re a governor, don’t go to ChatGPT and have it run the model for you because it might hallucinate.

Sinclair: I think that’s right.

Beckworth: Tara, this has been a fun conversation. I’m very hopeful for the future. Of course, I’ve got to ride the AI wave and not get in front of it and get swept over. What can we learn as researchers? How should we apply this? Now, maybe where have we misapplied its emphasis?

Sinclair: I think one of the key things is that economists have really focused a lot on AI impacts on the labor market. That’s a very important thing, and we still want to know more about that. We do want to know more about AI impacts on the economy more broadly. I want to make sure as well that economists are adopting AI technology and using it for our own research to advance our own research agendas as well. That’s one of the things I’m really excited about this paper is I think it sets an example of how we cannot just talk about AI as impacting the economy, but actually have it improve and advance our research agendas as well.

Beckworth: Okay. With that, our time is up. Our guest today has been Tara Sinclair. Tara, thank you so much for coming on and informing us about AI and the “FOMC in Silico.”

Sinclair: Great. Thank you.

Beckworth: Macro Musings is produced by the Mercatus Center at George Mason University. Dive deeper into our research at mercatus.org/monetarypolicy. You can subscribe to the show on Apple Podcasts, Spotify, or your favorite podcast app. If you like this podcast, please consider giving us a rating and leaving a review. This helps other thoughtful people like you find the show. Find me on Twitter @DavidBeckworth, and follow the show @Macro_Musings.