AI Identifies Early Signs of Ratings Downgrades
Machine learning captures additional risk signals from institutional investor behavior.
- By
- June 02, 2025
- CBR - Artificial Intelligence
The financial world typically uses rules and formulas to evaluate risk. Credit-rating agencies assign ratings on the basis of specific metrics such as debt-to-EBITDA ratios and interest coverage. Investment banks build risk models around carefully chosen factors. Industry risk managers create scorecards with weighted criteria. Regulators require specific capital ratios.
But investors’ actual decisions could provide a better gauge of risk, suggests research by Harvard’s Xavier Gabaix, Chicago Booth’s Ralph S. J. Koijen, New York University’s Robert J. Richmond, and Princeton’s Motohiro Yogo. They used machine learning to uncover the risk signals embedded in investors’ portfolio holdings. The approach “is akin to crowd sourcing credit analysis to a large group of institutional investors,” the researchers write.
As artificial intelligence has advanced, technology has been moving from rules-based systems to more sophisticated models that mimic human performance. Take self-driving cars, for example. The technology has involved explicit commands for certain driving scenarios—such as to stop when there’s an object ahead. With the addition of ML, however, a car can now figure out the outlines of roads, among other things.
Using similarly complex AI models could provide a more nuanced view of credit, the study suggests. Gabaix, Koijen, Richmond, and Yogo used ML to analyze millions of decisions made by mutual funds and insurance companies. The data—including trading data from regulators and bond characteristics and holdings information from commercial providers—covered 41 percent of the US corporate bond market from September 2002 to December 2022.
A new way to measure credit risk
The researchers’ embeddings-informed model captured risk signals that machine learning extracted from investors’ bond holdings. Compared with models based on traditional risk measures, it did a better job of explaining credit spreads across the full spectrum of bonds, from safer to riskier.
The researchers extracted 64 risk signals, or what they call firm embeddings, that they say capture most of the signals that institutional investors use to evaluate companies. Traditional credit ratings rely on their formulas to rank companies from AAA to C, but embeddings can be used to position companies in a more complex risk landscape, the researchers explain—much like a 3D coordinate system requires height, width, and depth, but with more dimensions at play.
The researchers incorporated the firm embeddings into a model and tested whether it could identify bonds at a higher risk of being downgraded. In their analysis, the embeddings-informed model provided more accurate signals of credit risk than did traditional models that included both credit ratings and an oft-used measure of distance to default.
For bonds rated in the BBB category, which amounts to about 42 percent of the investment-grade corporate bond market, the embeddings explained 67 percent of the variation in credit spreads, compared with 60 percent for the combination of credit ratings and conventional risk metrics, according to the study.
The researchers’ model captured risk signals linked to future downgrades that traditional methods missed. It flagged investment-grade companies that were likely to fall to junk status within a year, an indication that sophisticated investors tend to detect trouble before traditional models do.
Moreover, the researchers find in a related study that embeddings could be useful for analyzing the valuations and riskiness of equities too.
The credit-focused research suggests that firm embeddings could be used to refine risk-based capital requirements for insurers. The amount of money that insurance companies must keep on hand depends on the riskiness of their investments, as determined by regulators, who currently rely on traditional ratings. But if the amount were instead determined by both the ratings system and the embeddings model, about half of insurance companies would need to hold more capital, and half would hold less—with an average adjustment of 16 percent in their regulatory capital ratio. The researchers project that these new ratings wouldn’t cause big, unexpected changes each year, addressing a key concern of regulators who prefer steady capital requirements.
Investor behavior offers risk insights beyond traditional models, the study explains. By systematically extracting and using this information, the financial world might build systems that better reflect real-world risks and opportunities.
- Xavier Gabaix, Ralph S. J. Koijen, Robert J. Richmond, and Motohiro Yogo, “Asset Embeddings,” Working paper, September 2024.
- ——— “Upgrading Credit Pricing and Risk Assessment Through Embeddings,” Working paper, February 2025.
More from Chicago Booth Review
More from Chicago Booth
Your Privacy
We want to demonstrate our commitment to your privacy. Please review Chicago Booth's privacy notice, which provides information explaining how and why we collect particular information when you visit our website.