Foundation Models and Generative AI for Financial Time Series

Overview

This axis of research focuses on the prediction of financial time series using foundation models. The wide variety of these time series—ranging from order book data to historical data on stocks, bonds, or derivatives—and the inclusion of external explanatory variables make this problem particularly complex in practice.

Transformers have been successfully used to build large language models (LLMs) by leveraging the attention mechanism to reinforce strong dependency relationships within language. However, the dependency structure of financial time series proves much more difficult to model due to the variability in statistical behavior across datasets and over time within a single dataset. We are therefore exploring several research avenues, including more theoretical aspects of foundational models, to better understand the low- and high-frequency behaviors of financial time series while ensuring the model’s scalability.

Research Axis Leads

Laurent Carlier

Head of AI Lab at BNP Paribas Global Market

Laurent Carlier holds a master’s degree in applied mathematics from Centrale Paris and a master’s degree in quantitative finance from Sorbonne University. He has more than twenty-five years of experience in the banking sector. He initially led several quantitative research teams in London, Tokyo, and New York. For the past ten years, he has been leading the Data and AI Lab in Paris, which supports BNP Paribas’ Corporate and Investment Banking – Global Markets division. His primary areas of focus include designing and implementing projects aimed at automating tasks, increasing productivity, and improving pricing and risk management for financial products using artificial intelligence, whether traditional or generative.

Mathieu Fontaine

Associate Professor at Télécom Paris

Mathieu Fontaine earned his Ph.D. on alpha-stable processes for signal processing. He then completed a postdoctoral research fellowship at RIKEN AIP (Japan), where he worked on applied mathematics for real-time audio signal processing. He is currently an associate professor at Télécom Paris. His research focuses on the theoretical aspects of structured time series such as audio signals, as well as on fundamental models for automatic listening applications. More broadly, he is interested in the intersection of stochastic modeling and deep learning for time-varying signals.

François Roueff

Professor at Télécom Paris

François Roueff earned a master’s degree from École Polytechnique in 1995, followed by a Ph.D. in signal processing from the École Nationale Supérieure des Télécommunications in 2000. He is currently a professor at Télécom Paris, Institut Polytechnique de Paris, within the S2A group at the LTCI laboratory. His research focuses on time series modeling, with specific topics related to the modeling of low- and high-frequency financial data: statistical inference for long-memory models, forecasting in a time-varying context, and modeling of point processes.