To explain the stock market’s reaction to earnings results, investors and other market participants have a standard tool that has been around for decades: the earnings surprise.
This measure, which compares reported earnings per share against analysts’ consensus earnings-per-share forecasts, indicates how much a company beat or missed analyst predictions.
Its intuitive nature has made it the primary go-to for describing post-announcement stock movements, even though the metric alone explains only about 5% of same-day stock moves. Adding in the rest of what researchers have learned over the years pushes that figure only a little higher.
However, new research by Ralph S.J. Koijen and Bradford Levy of the University of Chicago Booth School of Business suggests that AI can both significantly lift this percentage and help us understand why stock prices move the way that they do. The researchers find that the best AI models explained about 17% of same-day stock moves.
This performance boost was in part attributable to the models “learning” how to predict the moves. AI then recorded its insights into digital notebooks that humans could read and learn from.
Koijen, a distinguished service professor, and Levy, an assistant professor, conducted a live test rather than a simulation using past data. This eliminated the possibility of lookahead bias, which can creep in when a model has access to information that wasn’t known or available for the time period being studied. Their testing involved three frontier models and covered nearly 2,000 earnings announcements made in late 2025 by companies across a range of industries.
The AI models’ approach produced written explanations that could be read and analyzed, suggesting the models seemed to pick up on details in the earnings calls that were not captured solely by the numbers shared.
Some of those details were included in forward guidance given by executives and management commentary, while others were revealed through the way in which speakers discussed results in the call, such as the language used.
Many of the models’ insights were ones that seasoned financial professionals may find intuitive, such as that good news was often already reflected in a stock price. Instead of overreacting to strong results, the models treated those as an expected baseline. Record profits or 20% growth were not enough to move a stock.
Koijen and Levy expect these human-readable explanations can give researchers looking to understand what drives price moves more specific hypotheses to test. And as models advance and become more powerful, the explanations they generate could potentially be applied to smaller, cheaper models with similar results, suggesting a path to scaling these systems at lower cost.
These results are early, the researchers said, adding that most of the variation in stock moves following earnings releases remains unexplained.
So they have launched a competition, sponsored by the trading firm Optiver, and are inviting anyone to submit their own model that predicts how prices respond to earnings-call information. Models can be submitted through explainingmarkets.ai, which contains the contest’s rules.
Koijen said that forecasting these market movements is a “surprisingly challenging task” for people, even those armed with AI.
“Decades of research can explain just 8% of the variation,” he said. “We show that we can lift this to 20%, but there’s still a lot of room to go.”
—This article originally appeared on the Chicago Booth Review website.
