Cars Hommes: Why markets – and AI – behave like humans

November 3, 2025

_{This interview was originally written in Dutch. This is an English translation.}

For over thirty years, economist Cars Hommes has been researching how markets really work: not as rational calculators, but as complex systems full of behavioural patterns. He tells Financial Investigator how simple heuristics and experimental research can help us better understand the economy, and why central banks need to break away from their traditional models.

By Harry Geels

You are ranked among the highest in the ESB rankings. What do you think explains this ranking?

'I started researching complex systems and behavioural economics more than thirty years ago, long before it became mainstream. My group at the University of Amsterdam, the Centre for Nonlinear Dynamics in Economics and Finance (CeNDEF), has succeeded in publishing groundbreaking articles in top journals with PhD students and postdocs. An important turning point was the publication of my review article in the Journal of Economic Literature in 2021. This has strengthened the international position of our research. The core idea of our paper is that the economy consists of heterogeneous, ‘boundedly rational agents’ who use simple heuristics to make decisions and form expectations, because the environment is too complex to fully understand. An important finding is that strategic complementarity – when optimistic or pessimistic expectations reinforce each other – often leads to coordination failures, resulting in instability, booms and busts, and (almost) self-fulfilling prophecies. In other words, the dynamics of financial markets and the economy are often driven by trend-following behaviour. Models with heterogeneous expectations and “agent-switching heuristics” better reflect the observed fluctuations at the micro and macro levels than models based on a representative, rational agent. This has important policy implications: traditional rational models designed to promote desirable macro behaviour can fail.'

In addition to empirical research, you also do a lot of experimental research. What does that yield?

'I have been strongly influenced by my collaboration with the influential American economist William Brock. We developed models with heterogeneous expectations: instead of assuming that all market participants are rational, we distinguish between two groups: fundamentalists, who believe that the market converges towards equilibrium, and trend followers, who extrapolate patterns in the data. Switching between these two groups on the basis of relative performance appears to have a dynamic influence on the market. It leads to bubbles and crashes. We have indeed substantiated and empirically tested these insights both theoretically and experimentally in the CREED laboratory at the University of Amsterdam, using data on house prices, financial and macroeconomic data, among other things. Laboratory research is very rewarding. In the lab, for example, we simulate markets and observe how people and algorithms respond to price changes. This shows that positive feedback – typical of financial markets – promotes instability. This is less evident in traditional goods markets, which are much more stable. AI systems such as ChatGPT also exhibit behaviour in such experiments that is very similar to that of humans. Insofar as markets were efficient at all, given the heuristics and behavioural biases, AI is not going to make them more efficient. The opposite could well be the case.'

So AI will reinforce irrationalities?

'Coincidentally, we have just finished a paper in which we investigate what happens when we replace human subjects (usually students in our experiment) with AI agents (ChatGPT). The aim of these experiments is to investigate how rational or irrational ChatGPT agents are compared to human subjects. These experiments involve predicting prices with positive (trend-confirming or trend-following) and negative (corrective or stabilising) feedback, respectively. Positive feedback occurs in financial markets, because higher expectations lead to more demand and thus higher prices. Negative feedback occurs in consumer goods, where higher expectations lead to more production and thus lower prices. In the lab, positive and negative feedback lead to very different market behaviour. With negative feedback, the market remains stable and moves towards rational equilibrium. With positive feedback, the market usually does not converge, but instead exhibits irrational trend-following behaviour that leads to price fluctuations around the equilibrium. We wondered whether the difference between positive and negative feedback also occurs in AI agents. This indeed appears to be the case: ChatGPT agents converge towards rational equilibrium when there is negative feedback, and irrational bubbles occur when there is positive feedback, just as with human subjects. AI agents therefore exhibit human traits, rational or irrational, when it comes to predicting prices or exchange rates. The current generation of AI agents does not seem to differ much from human subjects.'

AI agents exhibit human traits, rational or irrational, when it comes to predicting prices or exchange rates.

Have you drawn any other surprising lessons from your experiments for professional investors?

'Most economic and financial experiments are conducted with students as test subjects. Few experiments have been conducted with financial specialists. The laboratory experiments on bubbles by Nobel Prize winner Vernon Smith are famous. In many of these experimental financial markets with students, bubbles and crashes occur.

In our experiments in the UvA-CREED lab, we also observed many bubbles, even in markets with large groups of 100 people. An important cause of the bubbles is that in a group of test subjects in a financial market, coordination on trend-following behaviour easily arises. Once you are caught up in such a bubble, it is better to go along with it, because if you are the only one to opt out, this will lead to major losses. On the other hand, you should not wait too long, because if the market crashes before you get out, this will lead to even greater losses. We also conducted an experiment with financial specialists. In this experiment, bubbles occur slightly less often, but they cannot be ruled out entirely.'

Behavioural macroeconomics is still in its infancy compared to traditional Dynamic Stochastic General Equilibrium (DSGE) models. To what extent can behavioural macroeconomic models contribute to better monetary and fiscal policy?

'DSGE models could not predict the credit crisis because these models are based on rational expectations. In a DSGE model with rational behaviour, a crisis is a major stochastic shock that cannot really be explained by the model. Since then, much has changed in macroeconomic models. For example, DSGE models with adaptive learning behaviour have been developed. In these models, expectations are not rational, but bounded rational. That is a lot more realistic. Behavioural macroeconomics has developed rapidly in recent years, but too often a behavioural macro model is still only a minimal deviation from the rational benchmark, for example with limited information, while everything else is rational. Macro models with adaptive learning behaviour represent a bigger step away from traditional models. Even more radical are agent-based models (ABMs), which use microdata and micro behaviour of bounded rational agents with simple decision rules (heuristics). In ABMs, micro-level interaction through complex networks can lead to boom-bust cycles, with large fluctuations in macroeconomic behaviour. There is growing interest in ABMs among central banks. The macroeconomic forecasts of ABMs can now compete with those of standard DSGE models. This also makes them a serious candidate for policy analysis.'

To what extent do you currently see self-fulfilling dynamics in the markets, for example in inflation expectations or AI hype?

'Inflation has been very low for a long time, around 2% to 3% from the beginning of 2000 until around 2020. Then, triggered by the COVID pandemic and the war in Ukraine, inflation rose rapidly, reaching more than 10% in the Netherlands in 2022. According to behavioural models, this inflation peak was triggered by a major external shock (a pandemic or a war), but was then amplified by an overreaction in prices and by trend-following inflation expectations. As far as AI is concerned, there is, of course, primarily a positive shock (a new technology). I don't know to what extent there is already an AI hype. But it could be that we are dealing with or will have to deal with an AI bubble, just like the dotcom bubble at the time. That was also difficult to predict. In hindsight, it is always easy to say that there was a bubble that was followed by a crash.'

In ABMs, micro-level interaction through complex networks can lead to boom and bust cycles, with large fluctuations in macroeconomic behaviour.

Central banks are increasingly trying to steer markets with forward guidance. From a behavioural economics perspective, are central banks overestimating their ability to manage expectations?

'Forward guidance and communication are policy measures that work almost perfectly in rational models. But the world is not rational, which is why these policy measures do not work as well in practice as expected. That is why it is important for central banks to make more use of realistic behavioural models and base their policies on them. Central banks would also do well to test their policies in laboratory experiments. Within macroeconomic behavioural models, it is indeed possible to manage expectations. We concluded earlier that an inflationary boom can be reinforced by trend-following behaviour. By raising interest rates more quickly, if necessary, the central bank can try to manage trend-following behaviour. An interest rate hike is actually a policy measure that provides negative feedback to output, making a complex adaptive macro system more stable.'

In a complex world, can a simple strategy be superior to an advanced model?

'In a complex world, simple strategies (heuristics) often work better than complicated rules. One example is the naive heuristic: do the same as in the previous period. If the market follows a pure ‘random walk’, this heuristic is optimal. That is why such a simple heuristic can work well in a complex system. Somewhat more advanced is the heuristic switching model, in which you can choose from a number of different heuristics, switching between them based on their relative returns. In addition to simple heuristics, we also have complex adaptive systems consisting of heterogeneous, bounded rational agents that use many different strategies or heuristics. Prices do not converge towards their rational equilibrium, but exhibit temporary bubbles and crashes. High debt levels, geopolitical tensions, rapidly moving global capital flows, climate transitions, etc. all contribute to greater uncertainty, which usually leads to lower fundamental prices and increased volatility. This applies to the rational model, but to a greater extent to a complex system.'

What are your other motivations and ambitions?

'Perhaps a personal note as a start to answering this question. It took a lot of perseverance to get where I am today. Initially, biases and irrationalities were seen as an exotic niche. However, I have always believed that we can better understand markets by taking human behaviour seriously. That path ultimately led to fundamental insights and recognition. And I am quite proud of that. I want to build a bridge between theory, experiment and policy. Our models are now used by central banks, including DNB and the Bank of Canada. As a researcher, that is of course extremely valuable: a theory that works in the lab, proves empirically sound and is relevant to policy analysis. Enough to write another book or two: one specialist book in which I further elaborate on my work on behavioural macroeconomics, and perhaps also a book for the general public. There are already accessible books on behavioural economics and behavioural psychology, such as Kahneman's Thinking, Fast and Slow, but there are hardly any on behavioural macroeconomics, so there is still a lot to be done.'