Highlight Research

AI, Finance, and the Context Gap

As the financial world turns increasingly to alternative data and automation, a new question has emerged: how much context can we trust machines to understand?

At the Hi! PARIS Summer School 2025, Charles-Albert Lehalle, Professor at École polytechnique, revisited the fundamental shifts brought about by statistical learning. What once seemed impossible, turning satellite images, graphs, or unstructured text into actionable financial signals, has now become routine. But behind the technological progress lies a deeper transformation: how we define insight, risk, and value.

Key Takeaways
  • AI has unlocked access to unstructured data, turning language, images, and graphs into usable financial vectors.
  • Large Language Models (LLMs) may play a key role in financial education for retail investors, enhancing their ability to demand transparency from institutions.
  • Nonlinear models offer context-awareness, which is crucial for understanding economic signals in turbulent times.
  • The most overlooked bias in alternative data is projection bias, favoring clean, liquid, well-known assets and discarding less obvious signals.
  • The next wave of innovation lies in aligning technical mastery with long-term questions that matter to users, institutions, and systems.
Beyond optimization: Understanding context in a volatile system

In finance, the edge often lies in nuance. And nuance requires context.

Nonlinear models, once seen as overfit or opaque, are now central to understanding volatility. Inflation after a pandemic, for instance, does not behave like inflation during a trade war. AI, like language models distinguishing between a “blackboard” and a “board of directors,” must learn to detect these shifts. But unlike text, financial time series are sparse and ambiguous.

The challenge, as the speaker noted, is less about data availability and more about structure. Injecting macroeconomic knowledge or using dependency graphs could help models “read” signals more meaningfully, turning isolated numbers into informed estimates.

Financial literacy as a lever of change

Robo-advisors and AI-driven insights are often cast as tools for automation. But a more subtle transformation may come through education.

For retail investors? and those managing pension assets, the speaker pointed to a growing opportunity: personalized financial literacy. LLMs could help demystify portfolios, contextualize macroeconomic trends, and equip individuals to ask sharper questions. “More demanding users,” he argued, “are a healthy pressure on the system.”

This vision moves beyond product design and into governance. It positions AI not just as a tool for institutional advantage but as a mechanism to rebalance information asymmetry between experts and everyday investors.

Charles-Albert Lehalle

Charles-Albert Lehalle at the Hi! PARIS Summer School 2025

Hidden biases in alternative data

Alternative data, satellite imagery, credit card transactions, geolocation patterns, has added new texture to financial analysis. But it is far from neutral.

While selection bias is often acknowledged and corrected for, projection bias remains under-discussed. Analysts tend to map clean datasets onto the most liquid stocks or sectors. But these assets are also the most monitored, offering little informational surprise.

“Exhaustive data processing,” Charles-Albert argued, “is the only way to see what others miss.” And that often means resisting the urge to discard messy, hard-to-map inputs.

Reframing the research-practice gap

Having worked within both academic and institutional circles, including ADIA and CFM, the speaker underscored a paradox. The strongest systematic investment groups often rely on deeply internal knowledge systems. While this drives focus and cohesion, it can also create blind spots.

Some ideas are dismissed too quickly; others are adopted too casually. “Seeing nails everywhere once you have a hammer” is a risk familiar to both researchers and practitioners. The key is pragmatic humility, a mindset that treats academic literature as a resource to be tested, not a script to be followed

A new frontier for students and researchers

Asked what advice he’d give to students entering the field, the speaker answered with three themes:

  1. Helping users of financial markets better express and manage their risk exposure.
  2. Designing more adaptive, complex models for risk and optimization.
  3. Extracting insight from unstructured, nontraditional datasets.

But rather than prescribing a technical path, he urged a return to the fundamentals: find a question you care about. Once that’s clear, the methods will follow.