Financial AI Research: New Solutions and New Risks

On-demand webinar · 60 min · Recorded September 22, 2026

How can AI help financial researchers understand the economy, model risk and work with complex data?

In this video, Professor Charles-Albert Lehalle explores three areas of financial AI research: connecting financial decisions with the real economy through alternative data, using scientific machine learning for risk management, and applying language models to financial documents.

Using examples from maritime traffic, supply chains and analyst reports, he explains how embeddings make unstructured information usable in quantitative models. The discussion also examines the limits of these methods, including changing market conditions, biased datasets and reports that partially supersede earlier information.

The session closes with the risks of synchronized decisions and dependence on shared AI models, software and infrastructure.

This webinar is for ML researchers, quantitative researchers, AI engineers and technology leaders working with financial data and models.

Watch

Charles-Albert Lehalle

Professor at Ecole Polytechnique

Artem Elmuratov

Director of AI Education, Science and Research Partnerships

Register to get the recording

Key takeaways

Alternative data includes sources such as satellite images, maritime traffic, transaction records and company supply chains. These sources can help researchers understand economic activity, but their different formats make them difficult to combine and compare.

Embeddings represent information as numerical vectors that quantitative models can use. In the webinar, Charles-Albert Lehalle explains how this helps connect text, images and relationships between companies with financial research. Linking the data to the correct economic entities remains an essential part of the work.

Nowcasting means estimating current conditions, while forecasting looks ahead. Financial researchers can use alternative data to build a more timely picture of economic activity before it appears in conventional reports.

Maritime traffic, geolocation and supply-chain information can help reveal what is happening across the economy. Lehalle emphasizes that understanding the present is already a demanding and valuable task. A clearer picture of current conditions can inform decisions without implying that a model can reliably predict future events.

Financial risk management involves adjusting positions as prices and risk factors change. Scientific machine learning offers methods for solving high-dimensional equations used to model these problems. Reinforcement learning can also help researchers study how trading decisions affect prices and subsequent actions.

Lehalle highlights a central difficulty: strong average performance across simulated scenarios does not guarantee good performance along the one market trajectory that actually occurs. Financial markets also cannot be reset and replayed like a controlled experiment.

Financial language changes as new events, technologies and risks emerge. Models must also handle documents that overlap without clearly stating which earlier claims they replace.

For example, a short sector update might invalidate part of a longer company report while leaving the rest relevant. Finding both documents is not enough: a system must determine which statements still apply. Lehalle describes this as an unresolved research challenge for applying language models to financial information.

Models can help analyze transitions between market regimes represented in historical data. Anticipating a regime with no historical precedent is much harder, because the data may not reveal either its characteristics or how the economy could enter it.

Lehalle also describes a difficulty in detecting early warning signals from news. Removing repeated stories can expose new information, but repetition helps establish which developments matter. Researchers must balance early detection with enough evidence to distinguish an emerging event from noise.

Financial datasets can give an uneven picture of the economy. Transaction data may overrepresent certain groups, while information about larger companies may be easier to obtain and arrive sooner than information about smaller firms.

Lehalle uses earnings-call transcripts to illustrate how reporting delays can distort a model’s view of current conditions. Researchers need to examine coverage, timing and the mapping of data to companies. Reweighting can help address some sampling biases, but it does not automatically resolve every gap in the data.

Institutions using similar data and algorithms may make synchronized decisions. Because trading changes prices, those decisions can reinforce one another through market feedback.

Dependence on a small number of model, software or infrastructure providers creates another risk: a shared failure could disrupt many institutions at once. Lehalle distinguishes this operational dependency from the accuracy of an individual model. Understanding common dependencies and preparing backup arrangements are therefore part of assessing AI risk across the financial system.