In the age of GenAI, why do we still need good models?

Jean-Philippe Bouchaud says models can guide artificial intelligence through regime shifts and away from overfitting

Key points Summary Q&A

• Jean-Philippe Bouchaud argues models should improve our understanding, not merely mimic reality.
• Financial data is scarce, non‑stationary and heterogeneous, so generative models often memorise and overfit.
• Models should guide GenAI by identifying what matters in markets, such as scale‑invariant feedback, self‑excitation and path dependence.
• A hybrid method – moment guided diffusion – combining maximum entropy, wavelets and stochastic interpolants, was proposed by CFM ML Lab with Stéphane Mallat’s group; its synthetic time series will be released open access.

These summaries are generated by an LLM, but edited by humans.

My last column was a plea for good models – models that improve our understanding of what is really going on behind the scenes and help us build a faithful intuition of the underlying phenomena. Their purpose is not merely to mimic reality, but to make it intelligible. 

But proponents of generative artificial intelligence (GenAI) challenge this view. Since models are – or at least should be – based on empirical observations, why not dispense with modelling altogether, the argument goes. Why not let machines learn from examples and generalise – as they do, impressively, with text and images? We have all seen stunningly realistic – yet fake – images of celebrities, generated after training on real examples. 

So why not create huge collections of fictitious but realistic price time series with the same technology and use them for what quants are supposed to do – generate risk scenarios, stress-test portfolios, price exotic options or train systematic trading strategies – all while bypassing the demanding struggle of traditional modelling?

Well, there are several reasons not to throw the baby out with the bathwater.

First and foremost, financial data is unwieldy. It is scarce, except at very high frequency. It’s non-stationary – can one blindly use data from 50 years ago to understand today’s markets? And it is strongly heterogeneous across stocks, sectors, countries and regimes. This means the dataset available for training is, in practice, very small. And in such settings, generative models are often just parrots: they memorise the data, but fail to generalise. In other words, they overfit. 

There is a danger of pulling black swans out of black boxes: without a deep understanding of what the algorithm is doing, the risk of blow-up only increases

The result is that synthetic future scenarios may amount to little more than replays of the past, which is not especially useful. To be able to generalise, generative models must be complex enough and exposed to enough data to capture the underlying structure of the system. Only then can the machine stop regurgitating what it has learned and start innovating.

This is why the right question is not whether GenAI can replace models, but how models can guide GenAI. One way to alleviate this predicament is precisely to model, in my preferred sense of the word: to try to understand root structures and mechanisms. 

In other words, the model is not the rival of the generator. It is what tells the generator what actually matters. 

Establish what is important to capture in market behaviour – scale-invariant feedback loops between past returns and future volatility, self-excitation and path dependence – then use this information to guide GenAI models.

This is not a new idea. It amounts to generating random paths that are as random as possible, while satisfying a set of constraints. These so-called maximum entropy methods have recently been revived by combining two powerful tools: wavelets, which capture the multiscale nature of volatility correlations; and stochastic interpolants, which generate new members of an ensemble while often capturing details invisible to the naked eye. This hybrid generative model – moment guided diffusion – was recently proposed by two members of the newly created CFM ML Lab, in collaboration with mathematics professor Stéphane Mallat’s group at Collège de France and Ecole Normale Supérieure. The corresponding synthetic financial time series will soon be available in open access for quant researchers. Stay tuned.

Related

Beyond epicycles: models must describe markets, not just fit them 

Modelling needs to embrace complexity in volatility patterns, says Jean-Philippe Bouchaud

A second reason not to give up on modelling is just as important. 

Learning from past data says little about regime shifts and interventions – what happens when the system is exogenously perturbed, accidentally or by design? There is a danger of pulling black swans out of black boxes: without a deep understanding of what the algorithm is doing, the risk of blow-up only increases. Think of the October 1987 market crash, which resulted from the runaway feedback loop of Black-Scholes delta hedging onto itself. Without a clear understanding of market impact, no history-based learning model can anticipate such systemic risks.

GenAI is clearly a fantastic new tool in the quant toolbox. But it should be seen as a tool for extending and enriching models, not as a substitute for them. The temptation of quick fixes and plug-in solutions should not make us forget the essential purpose of modelling – not merely to produce realistic outputs, but to build intuition about the world we live in. 

Good science is about getting the assumptions right. In the end, the machine has no skin in the game. When things go wrong, it will be the responsibility of the modeller to explain what happened. GenAI may generate possibilities, but only models tell us which ones are worth believing. 

And I would go even further. 

The best models are those that generate more open questions than precise answers, challenge received wisdom and encourage us to dig deeper. 

Editing by Louise Marshall

Further reading

• Comment

Rethinking model validation for GenAI governance

• 03 Feb 2026

• Comment

Responsible AI is about payoffs as much as principles

• 17 Dec 2025

• Comment

When AI models malfunction, address the problem not the math

• 16 Sep 2025

• Comment

The AI bot that left the garage

• 05 Aug 2025