Can AI simulate the Fed? - with Tara Sinclair

Paul Diggle

Hello and welcome to Macro Bytes the economics and politics podcast from Aberdeen. My name is Paul Diggle

Luke Bartholomew

And I'm Luke Bartholomew.

Paul Diggle

Today's guest is Tara Sinclair, Professor of Economics and International Affairs at George Washington University and a former Treasury official and Chief Economist at the job website Indeed. And we're going to be talking to Tara primarily about a fascinating new working paper she has co-authored called ‘FOMC in silico: A multi-agent system for monetary policy decision making. It's a really innovative use of AI personas, or AI avatars, to model the interactions and debate at the Fed about the setting of much policy. And Tara runs these counterfactuals, these sort of, these ‘what if’ experiments, these in silico computer simulations driven by a large language model of shocks to that decision making process, such as political pressure on the Fed or the effect on monetary policy of a rapidly weakening labour market. So it's a really timely piece of work, I think, both in its methodology as well as the questions it's grappling with. We're really excited for this conversation. Professor Sinclair, Tara Sinclair, welcome to Macro Bytes.

Tara Sinclair

Great. Thank you so much. I'm delighted to be here.

Paul Diggle

Well let's start, Tara, if we could, with you straightforwardly, to the extent that's possible, laying out the research question that you're trying to answer here and this modeling process that you use.

Tara Sinclair

Yeah. Well, so first of all, let me just acknowledge that, this is work done with a coauthor, Sophia Kazinnik at the Stanford Digital Economy Lab. And so she is really the brains behind the LLM personas and I am more coming at it from this research question approach and this need for a new way of getting at macroeconomic counterfactuals. And so, when she and I got together and started talking about her skills and my desire to do these kinds of research projects, we found an immediate match. And so, the way I look at our research project, is it's a bigger project than just this one paper where we're really trying to see if we can use these new large language models to help us to simulate Federal Open Market Committee decisions in different environments. I've always dreamed of being able to have something more like an experimental set up for monetary policy and being able to go out and give, you know, small changes in economic environments and see how we might see different monetary policy decisions unfold from that. And now we have that possibility with these large language models.

Luke Bartholomew

Yeah, I guess it's always been the dream of macro economists to be able to run natural experiments with the economy, which for various ethical and other reasons, has sometimes not been the easiest thing to do. But that is the wonderful thing about your methodology that it does give us that kind of power. So in terms of the first step, I guess what you do in the first instance is try and model each FOMC persona, each of the individual members of the FOMC and you call them things SimPowell, SimWaller, SimCook going through all the different members. So can you talk a little bit about sort of what's making up those simulated AI personas? What things about the different policy makers do you think it's important to capture in their simulated version?

Tara Sinclair

Absolutely. And this is really critical for where economists have to come in and very carefully train the personas, because if we are wrong in this step, then we're going to have a wildly different decision-making group than the actual FOMC members. And this is something where Sophia has past experience at the Richmond Fed. I have a past experience working with various Feds and also just studying the Fed member materials. And what's really great about the FOMC is that they do put out a lot of public information about how they’re approaching their decision making, what news they're focusing on, how they're parsing that. And so for each member we have member specific materials that we feed in and we create a separate persona for each one. And we give them recent speeches, biographical context to their entire CVs, district conditions for those that are president of the regional Feds, and we give them a role biography and for gleaning district context, all of that information. So we're really trying to carefully train these personas to look as much as possible like the actual human that they are simming, if you will.

Luke Bartholomew

And then having set up all of those different personas, I guess you then send them off into the board meeting room itself, where they can interact and deliberate. So how is it that you’re sort of modeling that process of the agents coming together and a decision emerging from that interaction?

Tara Sinclair

So this is where things get really cool because we actually, consider two different approaches here. So, we use the sim personas in both approaches - so the same sim personas. But in, one case we do another large language model approach of simulation whereas in the other we use a more standard in game theory, a generalized Bayesian voting model. So we just take the personas and have them vote in rounds of voting, without having any interaction that would lead to deliberation. And we think that this is a real critical addition to be able to have that deliberation, because that's much more similar to how committees make decisions in general, and how the FOMC has documented that they approach decision making more generally. And so we take our personas. They take in the information that we give them. So we give everybody a data sheet just what would be given to FOMC members generally. Now I'll note here that in this paper we're simulating the July 2025 meeting and we were simulating it just before the July 2025 meeting happened. So we only had the publicly available information. Ideally, we'd have the actual Tealbook materials that are provided, but those only come out at a five-year lag. And so here we had to kind of make up a data sheet based on the economic data that is publicly available that we have, but without having the staff forecasts that are only, provided in internal materials. But so each person goes off, they determine their own beliefs and their own, you know, initial perspective on what they want to vote, but then they come together if we're on that LLM track and we have, you know, an entire discussion amongst the personas that I frankly just think is the coolest thing, because that really is a new innovation of this generative AI type technology.

Paul Diggle

So you have this setup, and as you say, you're modeling the July FOMC meeting ahead of time. And we remind listeners that that was a Fed meeting in which the Fed funds rate was held at 4.25 to 4.5%. But interestingly, there were these two unusual dissents, these dovish dissents from Waller and Bowman two of the Trump appointed governors. And you also are sort of running these two hypothetical meetings at which there are in one there's increased political pressure on the Fed in a dovish direction and in the other, there's a negative labour market shock, downward revisions to the labour market data. So very relevant, timely sort of counterfactual experiments. What are you feeding into the model then to simulate those shocks? And what decision from the simulated committee comes of those shocks?

Tara Sinclair

Yes. So let me just set up first the baseline. So the baseline - we gave the personas all of their information including their most recent speeches, which particularly for SimWaller really matters because he had made quite a bit more dovish speech than his earlier speeches had been. And so that was important. And so they had that in the baseline. And what we found is that even there we still came in at kind of the upper end of the range of what the FOMC actually came out with. We had them doing very very precise numbers instead of a 25 basis point range which is what the FOMC actually comes out with. But once we added in the political pressure and we added that in both by weakening the amount of agenda setting power that the Chair had for leading the meeting, as well as by creating kind of a bigger set of incentives for a few members, most heavily weighted on Waller for them to have the sense that, well, if you were to be a little bit more dovish you might have a higher probability of becoming Chair in the next Chair decision from the President. And, so, sure enough, with that political pressure scenario, what we ended up seeing is more dissent and dispersion than what we saw on the baseline. And so in the in the baseline, we didn't see much dissent, which diverged from what we actually saw in the July 2025 meeting, where we did see, as you highlighted, those two dissents. Now, I'm not going to say that this research then confirms that those dissents came from political pressure. Let me let clear that there are still a lot of assumptions that are going into this. And these are still simulated agents not actual human beings who are explaining their decision making process or anything like that. But it is interesting to see that generally that political pressure scenario did result in more dissent. And in particular, it didn't really seem to move the range that much. Rather, it created more dissent. And then on the second scenario, and this is actually the one that I had always been dreaming of doing this is where I think things get really interesting, because FOMC meetings typically happen on a regularly calendared schedule, and this idea that we might get some dramatic new data just shortly after an FOMC meeting really makes you wonder, oh, would they have changed our minds and made a different decision if they had that employment data that came out just a couple of days later? And so we did a very mild scenario of that in the sense that we didn't give them the employment data, the new employment data that came out. We just gave them the revisions. But they were particularly dramatic revisions, which we didn't know about when we were first running the simulation, then we reran it, with this notably large revision in the employment numbers and in particular change in total non-farm payroll employment for May was revised down by 125,000. And the change for June was revised down by 133,000. So they both went from strong jobs growth numbers of around 144-147,000, down to in the teens for jobs growth. So that was a pretty big change. And people had said that change might have justified Waller and Bowman's dissent. And sure enough, when we provide that new information, we did get, once again, not too much, we got more movement in the range. But we also got, more and more dissents there as well.

Luke Bartholomew

I think Chair Powell rather dismissively talked about not wanting to set policy through the rearview mirror when it came to revisions right? But I think this is actually quite a nice formal way of actually getting at truly the question of if they had different data to hand how different the decision would be, because you do rather hope that if the data were different, their minds might be different.

Tara Sinclair

Yeah. Just to come in on that briefly, because on the one hand, yes, if the data were different, we would want them to respond differently. That's what data dependent means for the Fed. At the same time, there's this big difference between data dependent and data point dependent. And there's a really big question here of, okay, two months of revisions back a couple months ago. What does that really mean? Is that a data point? Is that really a reframing of the overall understanding of the economy? And I think that's a debate we're continuing to have.

Luke Bartholomew

Yep. No, that's a that's a very important distinction I think. But just sort of going back to the sort of the twin track way of how you model the decision making so the AI LLM version that allows for interaction and then the game theoretic Bayesian version, you then are able to compare the predicted policy under the different scenarios between those different decision-making tracks. And you call that difference the ‘behavioral wedge’. So I wondered if you could talk a little bit about sort of the directionality of that wedge, whether it's hawkish, dovish, is it meaningful in its size and kind of what you think the intuition behind that wedge might be.

Tara Sinclair

Yeah. Well so I'll be honest that in all three of our scenarios that we're looking at in this paper, the wedge is really quite small. And so people might think, okay, well then that doesn't really tell us much at all. Maybe we could just continue to use these Bayesian methods and not have that additional stage of the LLM interaction simulation. But we still think that it's really important to continue to track this because this is just one FOMC meeting and one set of simulations for three different potential scenarios. There's a whole bunch of different scenarios where this might make a much bigger difference. You know, we've just begun to think about all of the different scenarios and applications where we could do this. And I'll note one particular one that I think, could be quite interesting, which is right now we're focusing on one meeting, but this is a repeater game. These people are interacting with each other on a regular basis. And so we're really excited once we have, you know, the computing bandwidth to do this for many meetings and see how that repeated interaction affects this wedge, because we think that's really one place where we might see, a much, bigger, wedge. And what we think that'll show is ideally this idea of the value of deliberations and norms and incentives beyond that rational baseline. You know, a lot of people complain about economists – oh economists believe that people are rational. And I would very much argue against that. We don't believe that people are rational. We model that as a benchmark case and then look for places of deviation. And those places of deviation are really interesting. And I think this is a way of us being able to have that rational baseline, and then also to be able to explore more that deviation in those deliberations.

Paul Diggle

You said there were possibly a lot of other cases of this FOMC in silico environment that you could test using that environment. I mean could you test things like a different Chair, an alternative Chair, or an expansion of the mandate to include lower long-term interest rates? I mean, obviously two other key debates around the Fed right now.

Tara Sinclair

Yeah. So, absolutely. You know, one thing that I am very keen to explore, although all of the news is shifting so quickly on this we can't quite keep up is precisely on, you know, a different Chair and bringing in different personas. Now, of course, it's much easier for us to consider personas that are already part of the Fed system because they kind of have consistent sets of information that we can then build out a persona that we think is comparable to our other FOMC personas. I think one of the challenging things is that there are names now being mentioned that don't have Fed’s features for us to train on, and so they don't have that same structure. We'd have to go and get the other, pieces of information. And, you know, in certain cases a lot of information about those people might be harder to find and to train a consistent persona on, but that's definitely something that I think would be really interesting and something, you know, obviously that's a very hot topic right now. So that would be that would be really cool. Another thing that I think is probably a lot simpler, and back in normal times, when we were first talking about this project, I thought people would find interesting but now obviously there's more drama happening. But there's been a lot of discussion about the rotation rules of the different members, because we have at the regional Fed Presidents only vote some of the time and there's blocks in which they rotate. And so sometimes I have wondered, well, what if we changed up those blocks and had different parts of the country represented at different parts of the economic business cycle, and would that make a difference in the kind of monetary policy that we're getting? And I think that could be really interesting for maybe considering changing those rotation rules to robustify monetary policy, and in particular, because people often talk about how our regional Feds were determined, their locations were determined based on the population in 1914 or so. And so thinking about how different the country's population distribution is today, people have talked about having different representation, and we could explore some parts of that through changing up just who gets to vote when.

Paul Diggle

So your paper is situated in this sort of this new but growing literature from the past couple of years that uses LLMs and AI personas specifically in this sort of paradigm of homo siliceous, you know, to model and understand and increasingly to help improve group decision making. So the FOMC process is really interestingly amenable to this modeling device. You you've seen increasingly court cases, court decision making, done like this. And one can very easily imagine a Supreme Court set of AI personas. A particular domain of use cases that I think might be very relevant to our financial market investing audience would be their application to improving group decision making, so not just understanding, but improving monetary policy outcomes, but also investment committee outcomes or corporate board debates. I mean, do you have thoughts, Tara, on that, this framework being used not just to model but also improve group decision making?

Tara Sinclair

Absolutely. And I think that's exactly what we're trying to get at when we're looking at that wedge between the LLM simulation versus, you know, a rationality-based voting model. And so one alternative interpretation of that small wedge is that, all in all, the approach that we're simulating for the FOMC following, assuming that that's actually how the FOMC is following is getting us something pretty close to a rational outcome, which we do tend to think is going to be, kind of, the optimal outcome for decision making. And so being able to apply these sorts of models and compare to a rational outcome and see where they differ, and then being able to do lots and lots of simulations to see if we can tweak different aspects of the deliberation to move it closer to that rational outcome, could be a way of exploring how exactly do we structure the rules of deliberation? How much agenda setting power do we give to the Chair? How do we, you know, bring together a group that's going to bring us to a strong decision thatd suitable for this application versus where they might go off the rails.

Paul Diggle

Yeah. For our own part, at Aberdeen as an investment house, we've been starting to think about and explore use cases bringing AI agents with particular specialisms or trained on certain data or on policies or on research sources into meetings or chats. And you can do that now, with, Microsoft Teams, you can bring an agent in, you can make that a devil's advocate agent to challenge assumptions and reasoning. You could insert behavioral bias identifying agents or avatars into trading systems to tell you things are you sure about this trade? It might be falling into x, y, z behavioral bias. And I think in time you can imagine the method you use being applied to building AI personas or digital twins of senior leaders or key decision makers, which you then could perhaps use to road test or refine proposals. It’s a sort of digital red team before you then present it to the human. And you know, that sounds amazing and so interesting, but also you could build sort of near-future sci-fi dystopian sort of stories there as well. So are there drawbacks, so problems to the AI persona approach, either as applied directly to modeling the FOMC or more broadly as a way to model and improve group decision making?

Tara Sinclair

Well, I'm sure there's lots of potential areas of risk, but let me highlight two that we are very well aware of. The first is the immediate reaction to our paper from some corners of the internet was yay! We can replace the FOMC with AI agents and I think that is a very scary direction. I don't think that that's useful. I think we are very far from wanting to give over any of this important decision making power. You know, people have said well we could just have the Chair and maybe a Vice Chair and then have everybody else be AI. I don't think we want to go there, not just because I care about the employment of the other FOMC members, but because I think that they do bring real value. And the only way we're able to train the AI is on these people. And I don't necessarily want these same people being our FOMC members for life or anything. I think there's a lot of various scary things there. And so I just want to rule that one out.

Paul Diggle

Yeah. It's sort of rather than a Taylor rule , well why not just do it as an LLM you know rules-based model approach?

Tara Sinclair

Yeah. Right. And yeah and I think we're very far from that. I think discretion, human discretion is still very very important and I don't see a near-term world with any of the tools that are being developed currently that would shift my view on that. The other thing which Sophia and I are really aware of and concerned about, is how much our simulated AI is actually the actual people. Did we give them enough training and information? Are there still things about LLMs that will just go in a different direction from a person generally? I think that's definitely true, but in addition, I think that there's still a lot of responsibility on us to make sure that we've given them the appropriate information. And you could imagine a case where some really critical way of approaching a decision is something that the actual human FOMC member has not ever reveal to the public, but is actually a core part of their decision-making process. And if that is something that is not available to the public, then it's not in our training for our persona. And we can possibly, with enough data on their past decisions infer some of it. But it would be great to be able to get more information about their decision-making process than just what is revealed in carefully curated speeches that have gone through legal review and everything that's a lot of the content we have to train on. And I think that that's something where I would like that information as an economist and as a researcher, but as someone who also cares about privacy and the value of human contribution, I understand why there might be incentives more than ever for humans to keep some of that to themselves and not let it be available for training AI.

Luke Bartholomew

Well, that question of how much the AI persona are truly like the people you're trying to model kind of gets at the somewhat philosophical question I had, if you don't mind indulging me Tara, because you’ve talked a little bit about how you know economists don't really always assume rationality, that's sometimes used as just a benchmark that we compare against. But I do think it's true to say that, you know, the mode of modeling the economists tend to use is what's known typically, as you know, talked about as ‘instrumental reasoning’ - that we care a little bit less about the assumptions and what matters is the predictive quality of that model. And, you know, the archetypal example that's often given to demonstrate this kind of thinking I think we owe it to Milton Friedman is, you know,  imagine a billiards player, snooker player, pool player, whatever they are, and you might model them as if they were getting a protractor out before every shot, doing some geometry, and then taking the shot. And if you modeled the player like that you'd have a pretty good prediction of how they behaved. But obviously that's not really what they're doing. And then the analogy would be in economics, perhaps we assume things like profit maximization or a certain kind of maximization under constraints or a certain way in which institutions make decisions and the criticism that's often leveled against that kind of reasoning modeling is that it fails to give us true understanding of why decisions made or why outcomes come about. We haven't really cracked the institution, or inside the head, or inside the firm, to get what's going on. So I wonder, do you think of this sort of ‘in silico’ AI process that you’ve developed is as within that ‘as if’ kind of modeling tradition, or does it kind of give us a glimpse of potentially moving to something that can give us what some people would call a true understanding of how certain outcomes are brought about?

Tara Sinclair

Yeah, I think that's a fantastic and deep question that I'm going to find myself mulling about the rest of the day Luke

Luke Bartholomew

For better or for worse…

Tara Sinclair

Yeah but it's a deep philosophical question that kind of gets at the criticism of a lot of generative AI work where it's pretty black box and, is that black box a good thing because humans are also black box in how they're decisions are made, in fact, often to themselves. Right? I don't think the snooker player can actually explain to you exactly how they're making some of these decisions. There's a lot of instinct and gut and repeated training without actually calculating the math. And I think we see some of that in the LMs. But are they doing it like humans do it? I don't think so. I think they're mimicking in certain ways, and I think that they're mimicking in an ‘as if’ way. So can I have a both/and on that one?

Luke Bartholomew

A both/and is a perfectly good logical answer, especially when we're talking about AI encoding I'm sure. So perhaps, maybe then I'll ground us a bit more back into the hard economics and to pivot us to talk a little bit about the labour market and how AI might interact with that, because as Paul said at the top, of course, outside of your academic work, or in addition, I should say, you've also been at Treasury and you're the Chief Economist at Indeed, the job site, founded their hiring lab. So I guess a question from a microeconomic perspective, to start with, you know, increasingly one hears stories of, well, I think it's pretty clear, as we see CVs as well, that they are being generated by AI. And then in turn, there's lots of screening of CVs done by AI as well. So you had this whole matching process at least in the first instance, where maybe humans haven't had too much of a role. So I'm wondering and indeed, you could imagine to Paul's point about improving decision-making, maybe hiring panels, increasingly have AIs as part of them, maybe they take a bigger role in the interviewing. So, you know, do you think this is something that could potentially improve labour market matching, gets rid of some inefficiencies and some biases, or does it in turn bring its own inefficiencies, frictions, keyword mining or whatever it might be? And indeed, you know, AI has its own implicit biases that perhaps we're even less sure what they are, and so that brings concerns about labour market matching?

Tara Sinclair

So I am incredibly bullish about AI used in the labour market space. And it's going to take me a couple of steps to explain why. Because the first thing I think we need to agree upon is that resumes are terribly dystopian things, distilling people down to a one page piece of paper and then having someone spend less than a full second reading that and inferring whether that person is going to be a good fit for their particular job is, I think we're going to look back at that time period at some point in the future and just really be horrified by that classification of humans in that very flat way. And we know that people have felt this for a while, because I don't know if y'all remember it, but there was this brief phase of trying to make video resumes that fell very flat. They were they were terrifying in their own way, but it was an attempt to get some 3D ness, or at least some 2D ness even to, a job seeker and who they are and what they could bring to an employer. And so I am very excited about this idea that I can now, instead of going and writing a flat resume, I can go to an AI and I can answer some questions about myself and about what I'm looking for in a job. And I can be quite detailed, and it can then take that information and it can go and communicate with an eye on the other side, on the employer side, and determine if we're a good match. And I think that is going to expand opportunities and expand and improve matches in a way that all the technological advances to date have, not succeeded. And I actually think that part of the problem there has been this challenge in translating my skills and interests to a one page piece of paper, having that be read oftentimes by a recruiter who is not a specialist in what, the employer actually needs. They have to be told by a hiring manager, and everything kind of gets a little bit lost in translation. And having these AIs I really think, even though maybe people's first reaction is oh this is dystopian - having these AIs interact with each other, having humans interacting in inefficient ways, I think is also terribly dystopian. And I think this could actually improve things quite a bit.

Paul Diggle

Well then a more macro question to go with Luke's micro question on AI and the labour market. So there's obviously this enormous debate about AI as a net job destroyer or not, its impact on productivity growth, whether or not it's a general purpose technology. And right now, in the US, we are seeing this interesting combination in the macro data of a weakening labour market, weaker employment growth, but still pretty robust GDP numbers, especially given some of the shocks, the tariff shocks and so on. Is that AI having an impact on productivity therefore, this combination of a weakening, labour market but still pretty strong activity side?

Tara Sinclair

So I don't think it's AI yet. I, actually think that there's, a lot of complexity going on in the US economy. And I actually still see that the strong GDP growth as being driven on. Consumers still seem to be spending. Firms are investing in AI. So that's an area where there is a connection to AI. But I don't think that the improvement in productivity growth that we've been seeing in recent quarters really has a lot to do with AI. In fact, I think AI is still holding back that productivity because people are, you know, spending time trying to figure out how to use these tools rather than, effectively using those tools. So I actually think a lot of the productivity improvements we've seen so far are actually coming from the slower churn in the labour market. We've got people that have been staying at their jobs for a while and have gotten better at them after some pretty, rapid and remarkable amounts of transition coming after the pandemic. So I think there's room for AI to improve our productivity quite a lot. But as of yet, there's new research that just came out of the Yale Budget Lab last week that I really by Martha Gimbel and some coauthors, where they're really highlighting that as of yet, we're not really seeing any noticeable shift in our occupational mix that's any faster than what we've been seeing historically. And I think until we start to see those kinds of really noticeable shifts, most of the AI impact on the labour market that I'm seeing is more about relative demand shifts and more not on the occupation front, but more on the entry level. That's really the area that I'm most concerned about being impacted. But then again, that also shifts us towards keeping more of the higher productivity workers that have already settled into their jobs a bit. And so it's something we have some room for AI to expect to impact productivity and therefore GDP growth in the future, but I don't think it's explaining the data pattern today.

Luke Bartholomew

All right. Well I think that's all we have time for this week. So as ever, please do let me ask you to subscribe wherever it is that you get your podcasts. And then all that remains is for me to thank Tara, Professor Tara Sinclair, for joining us. And I recommend again, I commend very highly to you her paper ‘FOMC in silico’, a really fascinating methodology and just sort of way of thinking about policy and other trends in the economy at the moment. So thank you, Tara, and thank you all for listening. Speak again soon.

This podcast is provided for general information only and assumes a certain level of knowledge of financial markets. It is provided for information purposes only and should not be considered as an offer, investment, recommendation or solicitation to deal in any of the investments or products mentioned herein and does not constitute investment research. The views in this podcast are those of the contributors at the time of publication, and do not necessarily reflect those of Aberdeen. The value of investments and the income from them can go down as well as up, and investors get back less than the amount invested. Past performance is not a guide to future returns, return projections or estimates and provide no guarantee of future results.