AI Could Predict the Next Financial Crisis — But There’s a Catch

What if AI could predict the next financial meltdown? Sounds like a promising idea, yet as new research finds, the devil is in the details.

Since the 2008 financial crisis, the Federal Reserve and other central banks have focused on monitoring the financial system for warning signs of instability. This approach, known as macroprudential regulation, requires lots of data to provide both a comprehensive overview of the economy and details about institutional investors’ portfolios.

Until recently, regulators could only dream of having so much information. That was then, explains Antonio Coppola, an assistant professor of finance at Stanford Graduate School of Business. “Now we’re in an environment where data is not scarce anymore. Regulators have access to this massive data. They can see exactly what balance sheets look like throughout the financial system.”

That’s where AI comes in. Armed with big data and big compute, regulators could deploy “real-time, high-dimensional predictive models” to detect fire sales and other signs of distress before they ripple through the economy. “These models are potentially very capable,” Coppola says. “They could give you granular signals of where financial vulnerabilities are so that you can target your policies.” In particular, they could augment oversight of shadow banking— non-bank entities like hedge funds, ETFs, and pensions — where systemic risk has migrated since traditional banks faced tighter post-2008 regulations.

So what’s the problem? For one, economists are wary of predictive models, even when they’re highly accurate. As economist Robert Lucas noted in the late 1970s, using historical data to make forecasts can overlook underlying structural forces that are unaffected by policy interventions. In other words, AI-driven predictive models might pinpoint exactly where financial trouble is brewing, but they can’t explain why it’s happening or whether a specific policy would fix it.

Likewise, predictive models present a moral hazard — that is, they may inadvertently prompt investors to take on new risks. Banks and financial institutions could buy up vulnerable assets, assuming that the model will prompt regulators to step in if things go badly. Conversely, investors may back away from investments they know the model is scrutinizing. Instead, Coppola explains, they will shift their risk “to a corner of the financial system that remains more invisible to the regulator.”

Forecasting Fragility

This could present financial regulators with what Coppola describes as a “Faustian bargain,” in which they must choose between predictive precision and causal clarity. In a new paper, Coppola and Christopher Clayton of Yale School of Management delve into this dilemma and propose a way forward. “We’re clarifying which side of the problem you should be using the applied AI model for and where you should supplement it with more traditional economic models,” Coppola says. “Our perspective is that those two things don’t necessarily have to be in conflict with each other.”

To test how a predictive AI model might be most useful for regulators, Coppola and Clayton made their own. Specifically, they built a graph transformer, a deep learning tool designed to crunch data on financial holdings. After being trained on 14 years of data, the model was able to reconstruct investor positions with remarkable accuracy. Even though its training data ended in 2019, it accurately predicted trading behavior during the 2020 market crash at the start of the COVID pandemic.The model could also assess the risk of new investors or assets in real time without retraining.

Coppola and Clayton conclude that their model demonstrates the potential for “model-informed” regulation and provides “a blueprint for real-time regulatory approaches.” However, they also find that predictive models perform best when used in conjunction with existing economic theory about the causal impacts of interventions.

Despite his model’s success, Coppola stresses that AI-driven macroprudential regulation is not here just yet. He notes that this approach “needs a lot more R&D” and that “central banks will want to think very carefully about this going forward.” He hopes that his research offers useful insight into how regulators could tap into the predictive power of AI without injecting new kinds of uncertainty and risk into the system.