AI Could Spot Financial Crises-If Regulators Can Solve a Fundamental Problem
Artificial intelligence can predict where financial trouble will emerge with remarkable accuracy. But new research shows it cannot explain why trouble is happening or how to fix it-a gap that could undermine the entire approach.
Since 2008, central banks have tried to monitor the financial system for warning signs of instability. The strategy, called macroprudential regulation, requires vast amounts of data on institutional portfolios and economic conditions. Until recently, regulators lacked access to that information.
That constraint has vanished. Regulators now collect massive datasets showing balance sheets across the financial system in real time. Antonio Coppola, an assistant professor of finance at Stanford Graduate School of Business, said this creates an opening for AI. "These models are potentially very capable," he said. "They could give you granular signals of where financial vulnerabilities are so that you can target your policies."
AI-driven models could be especially useful for monitoring shadow banking-hedge funds, ETFs, pensions, and other non-bank entities where systemic risk has shifted since traditional banks faced stricter post-2008 rules.
The Prediction-Explanation Trade-Off
Here's where the problem emerges. Economists have long worried about using historical data to forecast the future. As economist Robert Lucas argued in the late 1970s, patterns in old data can break down when policy changes. An AI model might accurately pinpoint where financial stress is building without understanding the underlying causes or whether a specific regulatory action would actually prevent a crisis.
There's a second problem: moral hazard. Banks might buy vulnerable assets, betting that regulators will intervene if losses mount. Alternatively, investors could shift risk away from assets the model is watching toward corners of the financial system that remain invisible to regulators.
Coppola and Christopher Clayton of Yale School of Management studied this dilemma in a new paper. They built a graph transformer-a deep learning tool designed to analyze financial holdings data. After training on 14 years of information, the model reconstructed investor positions with high accuracy. Even though training ended in 2019, it predicted trading behavior during the 2020 market crash.
A Hybrid Approach
The researchers concluded that predictive models work best when paired with traditional economic theory about how policy actually affects markets. "Our perspective is that those two things don't necessarily have to be in conflict with each other," Coppola said.
The model could also assess the risk of new investors and assets in real time without retraining-a practical advantage for regulators managing constantly shifting markets.
Still, Coppola cautioned that AI-driven macroprudential regulation remains years away from deployment. "This needs a lot more R&D," he said. "Central banks will want to think very carefully about this going forward."
For finance professionals, the takeaway is straightforward: AI's predictive power is real, but regulators will need to combine it with economic reasoning to avoid creating new risks while trying to prevent old ones.
Learn more about AI for Finance and AI Data Analysis applications in your field.
Your membership also unlocks:
