Jean-Philippe Bouchaud: Using Models to Prevent AI Overfitting and Navigate Regime Shifts

Jean-Philippe Bouchaud, Head of Research at Capital Fund Management (CFM), argues that traditional modeling remains essential in the era of generative artificial intelligence (GenAI) to prevent overfitting and navigate regime shifts in financial markets.

In a column published on April 13, 2026, Bouchaud asserts that the primary purpose of good models is to make reality intelligible rather than simply mimicking it. He suggests that these models are necessary to build a faithful intuition of underlying phenomena and to improve the overall understanding of the mechanisms operating behind the scenes.

The Role of Models in AI Guidance

Bouchaud posits that models serve as a critical guide for artificial intelligence. Specifically, he states that models can help steer AI away from overfitting and provide a framework for managing regime shifts.

This perspective contrasts with the views of some proponents of generative AI who challenge the continued necessity of traditional modeling. Bouchaud maintains that while GenAI possesses powerful capabilities, the structural understanding provided by good models is what allows for a deeper comprehension of the phenomena being analyzed.

Model Risk and Institutional Implementation

The integration of Large Language Models (LLMs) into the banking sector has created new pressures regarding model risk management. As banks deploy hundreds or thousands of new applications, the industry is debating the adequacy of existing supervisory guidance, such as the SR 11-7 framework for model risk.

A survey of 13 banks conducted by Risk.net reveals a divergence in how these institutions handle the validation of LLMs:

• Bank of America and Goldman Sachs subject LLMs to model validation by default.
• Other institutions, including Citi and JP Morgan, utilize a flexible approach that varies based on the perceived level of risk.

Academic and Professional Context

Bouchaud’s advocacy for structured modeling is supported by a professional background in statistical mechanics, disordered systems, random matrices, and quantitative finance. His research expertise includes agent-based models and the study of how data-driven models may lag during regime changes.

The tension between generative AI and traditional modeling centers on the distinction between pattern recognition and intelligible understanding. While GenAI can identify complex correlations within vast datasets, Bouchaud argues that the goal of a good model is to provide an intelligible explanation of the underlying reality.

Their purpose is not merely to mimic reality, but to make it intelligible.

Jean-Philippe Bouchaud

As financial institutions continue to scale their use of AI, the debate persists over whether the flexibility of generative systems can replace the rigor of traditional model validation or if the two must coexist to ensure stability during market transitions.