CFM Talks To: Xavier Gabaix

Xavier Gabaix is Pershing Square Professor of Economics and Finance at Harvard’s economics department. He received his undergraduate degree in mathematics from the Ecole Normale Supérieure (Paris) and obtained his PhD in economics from Harvard University. His research focuses on finance, macroeconomics, and behavioral economics. He received the Fischer Black prize given every two years to the best financial economist under 40, the Bernacer prize given to the best European economist under 40 working in macroeconomics and finance, and the Lagrange and Allais Prizes. His research has been published in the American Economic Review, Econometrica, the Quarterly Journal of Economics, the Journal of Finance, and Nature. He is a Research Associate of the National Bureau of Economic Research and of the Center for Economic Policy Research.

He co-organised the CFM Summer 2026 Paris conference “Persistent Puzzles and Paradoxes in Economics: Are Radical Paradigm Changes on the Horizon?” André Breedt spoke with him following the conference.

CFM: The recent conference was explicitly interdisciplinary, drawing speakers from physics, economics, and complexity sciences, amongst others. Why this conference, why now, and do you think there is enough common ground and shared language to make it productive?

XG: The value of periodically bringing these communities together is that it forces each group to confront their own assumptions.

For instance, Economists will comfortably write down an equilibrium price; econophysicists will immediately ask how you reach that price? That is a genuinely good question — one we often cannot answer cleanly. Likewise, when an econophysicist says, “I simulate the economy with a set of rules,” the economist’s natural response is: “but those rules are arbitrary, and if the policy environment shifts, or volatility changes, how does the model adapt?” The econophysicist might reply we need to recalibrate the model, or we hope it will be fairly invariant. This mutual questioning of these very basic premises is, to me, a very healthy intellectual exercise.

And communication between the two communities works better than one might expect. Of course, when you get into a deep debate of what “equilibrium” really means, there are philosophical controversies, but at the level of pragmatics — policy functions, GDP, consumption — people agree on the language well enough to have a productive exchange.

As for timing, the 2008 financial crisis was a genuine wake-up call. It became much easier since to openly say that the standard near-perfect-markets framework has serious limits — limits we always knew were there but papered over. That created more genuine appetite on the economics side for the kind of modelling physicists and complexity scientists have been doing. Researchers like Jean-Philippe Bouchaud and colleagues at CFM; Doyne Farmer and his collaborators, and many others — have been making increasingly credible claims that their models captured dynamics the standard framework missed, notably during the financial crisis and again during COVID, where e.g., heterogeneity could be modelled much more flexibly. That accumulation of results made this a good moment to revisit many of the interdisciplinary research efforts.

CFM: Given some of these results, what tools or methods from e.g. statistical physics do you then think mainstream economists could or should be adopting? Is there low-hanging fruit?

XG: Not especially low hanging, because the methodological styles are genuinely very different. In economics, the accent is most often on ‘clean’ models, with fairly low dimensional parameterisations and causal identification. In econophysics, there is much more comfort with large numbers of free parameters and much less obsession with causal identification in the econometric sense. It is not obvious how either community simply imports the toolkits and techniques of the others.

That said, something I found genuinely thought-provoking at the conference — and Jean-Philippe has written several papers on this — is the difficulty of actually reaching ‘equilibrium’ in complex networked economies. If you simulate – and you need to simulate because it is too hard to solve analytically – a network of firms, adjusting prices to each other, sometimes say by being incentivized to switch suppliers when costs shift, you find that it takes an enormous time to reach any sort of stability. The economy seems to fluctuate without any obvious exogenous shock driving it. We see clear cases where a single event — Lehman, the oil spike following the closure of the Strait of Hormuz, take your pick — produces large ripple effects. But there are many other episodes where it is genuinely unclear what is moving the economy.

The suggestion that this persistent fluctuation might be intrinsic — not driven by exogenous shocks hitting an otherwise stable system — is very intriguing for economists.

CFM: The conference framing took direct aim at the equilibrium concept. Some would argue it should simply be abandoned. Is that the right question to ask?

XG: I don’t think abandoning equilibrium is the right frame. Let’s take the price of oil: there are constant disturbances, political shocks, demand swings, regulatory changes, but at any given moment there is a price at which oil clears the market. That is what we mean by an equilibrium price. It shifts daily, but it is always there. The same applies to any security. The concept is not wrong — it is simply describing that supply and demand meet at some price.

Where things become much less clear is when the price of an asset depends on expectations of future prices. There is still, at any instant, a price at which asset X trade. In the minimal sense, that is an equilibrium. But whether that price is “correct” in any deeper sense — whether it reflects rational expectations of future cashflows — is a very different question. The market price is not the ‘rational’ price; it is whatever buyers and sellers have agreed on.

So, I think the wisdom is to stop arguing about the word “equilibrium” and ask instead: for a given market, on a given day, what is the best model of how that price was determined? An econophysicist might say: there are complicated adaptive dynamics that might cycle, while an economist might say: supply met demand at that price. But the dispute is how much does the price depend on future oil demand, or future expectations, or is demand itself distorted by over-extrapolation of recent trends? The real debate is about the mechanism, not whether to keep the word “equilibrium.”

CFM: The conference placed significant emphasis on the data revolution — transaction records, firm-to-firm payment networks, real-time invoicing systems, etc. Do you believe the data we now have gives economists and policymakers tangible advantages that were simply unavailable a decade or two ago?

XG: We have vastly more data on certain things as you mentioned — but on the questions that matter most for systemic risk and crisis prevention, how much of the available data is useful data? Take a deceptively simple question: who holds US Treasury bonds? Or who carries the exchange rate exposure in a major currency pair? Not just on the cash side but including derivatives. And who might unwind those risks in a crisis? That full data is effectively unavailable.

There are studies that have assembled partial pictures from regulatory filings on specific instruments, but the overall map of who holds what in the global financial system is not easily apparent. So, are we better equipped for any future financial crisis in terms of policies? I really don’t think so.

For credit card spending, retail transactions, sectoral output in near real time — yes, there has been real progress, and at the conference there were compelling presentations on what can be extracted from granular payment and invoice data. But the honest caveats are that many of these datasets are good to describe, rather than necessarily to predict.

I’m very open-minded, and of course more data is better. But the case still must be made. Why is that? You can have all the micro data in the world, but many of these datasets tend to span five to ten years at most. That gives you, at best, one or two business cycles. You cannot train a reliable macro model — let alone a large language model intended to represent how the economy functions — on one business cycle. There simply are not enough distinct macro regimes in the data. Now, while there are not many business cycles in the aggregate, there are many ‘micro business cycles’ where within industries one could imagine doing some more refined analysis.

For trading purposes, particularly at higher frequencies, the calculus is different, and that is genuinely useful. But for medium-to-long-run macro forecasting — one to three years — I am skeptical that the current abundance of some micro data solves the fundamental problem.

CFM: You mentioned that granular micro data may not help much in extrapolating to genuinely new macro regimes. That connects to your own work on large firms as drivers of aggregate fluctuations. Can you expand on that tension?

XG: The issue is one of the state space. Suppose a large shock originates in a concentrated sector — AI, a financial sector crisis, a sustained oil shock. What will the impact be on the rest of the economy? We just cannot say with complete certainty. You might for the sake of the argument have data on all the biscuits and liters of gasoline bought by consumers, but it is unlikely to reveal how much they will buy if there is a large change in the system. The historical data tells us a lot about how consumers behave within the range of conditions we have already observed. It tells us very little about how they will behave when the structural shock is large enough to push the economy into a genuinely new region of its state space. Machine learning cannot solve this: it can interpolate beautifully within the observed distribution, but extrapolation beyond it requires a model.

What economists have done is use structural models precisely to bridge between what the data identifies in normal times and what might happen under large departures from normal. My argument is just that vast quantities of micro data on spending or transactions — while helpful at the margin — do not substitute for that structural reasoning. The crucial missing piece is the data we do not have: balance sheet exposures, credit exposures between firms, position data across major asset classes. That is where the leverage for understanding crises lives.

CFM: On the issue of AI and prior-free discovery — both for policymakers and researchers — there is a tension between pattern recognition without theoretical grounding and theory-constrained modelling. How do you think about that?

XG: For macro-frequency data, I think pure empirical pattern-finding without any theoretical story is simply not reliable. Even if you pool 100 countries, business cycles are correlated across them, so the effective number of independent observations is still small. You cannot responsibly run an LLM over ten years of macro data and trust what comes out without a prior.

But “prior” does not have to mean dogma. It means a coherent narrative about what might be going on — what sector or mechanism is generating the signal, what other observables should co-move if the story is right. You then check whether those auxiliary predictions hold. If you find something in the data without a prior — a strategy with a high Sharpe ratio that you cannot explain — the right response is to dig harder for the mechanism, not to abandon the search for one. The story disciplines the search, and it is also what allows you to decide when the regime has shifted and the pattern has likely broken down.

For very high-frequency trading, this changes: you have enough independent observations to do reliable out-of-sample testing purely empirically, and the cost of a false positive is bounded. The macro context is very different.

CFM: The conference pre-reading described rational expectations as a consensus target for replacement but noted there is yet no consensus replacement. Did anything at the conference change that picture?

XG: Not fundamentally, but there was healthy progress in several directions. My own work has focused on tractable models of limited attention — agents who do not pay attention to every price change, making decisions based not on ‘full rationality’. These models are quite tractable and fit a large range of empirical regularities well. They are particularly useful for describing 95% of situations where the economy is operating not too far from its ‘normal’ range.

What is much less clear — and this was openly discussed at the conference — is what bounded rationality models should say about large departures from ‘normality’: wars, pandemics, financial crises. In those situations, it is not even clear what ‘rational’ means, since the events themselves are largely unprecedented and their consequences deeply uncertain.

A separate strand of work presented at the conference, especially experimental macroeconomics on expectation elicitation in controlled macro environments — is promising to test competing models of expectation formation, rather than simply picking among them on theoretical grounds.

In sum: we have workable, empirically disciplined models of limited rationality for small to moderate deviations from the norm, which covers about 95% of cases. For large deviations, the field does not yet have a settled framework.

CFM: The conference also touched on normative questions — what people value: community, fairness, inequality. Do you think economists give sufficient weight to these dimensions?

XG: My view is that the economics profession as a whole pays more attention to values than is commonly credited. There is a substantial literature on perceptions of fairness, on which inequalities people consider just versus unjust, on the role of merit versus luck in economic outcomes. That body of work is not marginal to the profession.

Where I would push back a little is the implicit assumption that more emphasis on values would improve policy. A critical function economists serve is reminding people that inequality is not simply a pathology. Some degree of inequality is necessary to provide incentives — without differential rewards for effort, investment, and risk-taking, the productive base of an economy deteriorates. That is not an ideological position; it is a straightforward consequence of how incentive systems work. The real question is always which inequalities reflect genuine differences in contribution, and which are products of rent-seeking, inherited advantage, or luck. Economists have tools to make that distinction and refining those tools matter more than shifting the field’s normative compass.

CFM: Did anything at the conference genuinely challenge your own convictions — particularly given that your work is known for tractability and elegance?

XG: Yes, of course. Once again, the simulation-based findings on convergence to equilibrium are the most persistent source of intrigue. The claim — that in a network of interacting firms, the time to reach any stable price configuration is very long, and that the economy therefore fluctuates endogenously without ever really settling — is intellectually very challenging if true empirically.

My caveat is that many results which are genuine in simulations turn out not to be first order empirically. There is a long history in complexity theory of theorems proving that certain optimisation problems are computationally intractable, cursed by dimensionality — and yet in practice humans navigate these problems routinely. We get up in the morning, plan our day, drive to work. These are, formally, high-dimensional dynamic programming problems, that the simplest theories would deem untractable. The resolution is that the real world has structure — lower-dimensional regularities — that precisely allow agents to do well without solving the full formal problem. The same may apply to the convergence question: the path to equilibrium may be exponentially long, but the economy might be structured in a way that collapses that to something much faster.

So, I hold both things at once: the simulation-based research is worth taking seriously, and they do not automatically translate into empirical claims about the economy’s actual dynamics. Navigating that is genuinely difficult and genuinely interesting — which is part of why this is such an enjoyable field.

CFM: Final question: if you had unlimited resources and a single project to fund, what would it be?

XG: Data release. I would advocate for the systematic disclosure of financial position data — who holds what across the major asset classes: Treasury bonds, corporate bonds, equities, derivatives, currency exposures. Not in real time, where legitimate competitive and commercial concerns apply, but with a lag of say two to three years — first to regulators, then publicly available to academia. The analogy I would draw is to fundamental physics: if the key constraint on progress is that you cannot observe what you most need to observe, no amount of theoretical cleverness closes that gap.

For macroeconomics and financial stability research, the missing data is balance sheet and position data for major financial institutions, large corporations, and the hidden offshore structures through which a large share of global capital flows. We are not meaningfully better equipped to handle a future financial crisis than we were in 2008, in part because the structural opacity of the global financial system has not been reduced.

Whenever I speak at the Fed, the BIS, or the IMF, I make this point. The institutions that could mandate disclosure have the regulatory reach to do it. The resistance comes from the private financial sector, which does not want positions disclosed in real time — a concern I understand and largely accept. But there is no good argument against disclosure with a multi-year lag for academic and regulatory use. That is the first thing I would recommend to do to improve economics: make that happen, and a large share of the most important open questions in macroeconomics and finance become tractable.

DISCLAIMER
The text is a transcript of an interview with Xavier Gabaix in August 2026 and has been edited for length and clarity. The views and opinions expressed in this interview are those of Prof. Gabaix and may not necessarily reflect the official policy or position of either CFM or any of its affiliates. The information provided herein is general information only and does not constitute investment or other advice. Any statements regarding market events, future events or other similar statements constitute only subjective views, are based upon expectations or beliefs, involve inherent risks and uncertainties, and should therefore not be relied on. Future evidence and actual results could differ materially from those set forth, contemplated by or underlying these statements. In light of these risks and uncertainties, there can be no assurance that these statements are or will prove to be accurate or complete in any way. The content of this document does not constitute an offer or solicitation to subscribe for any security or interest.