Neural networks cracked stock market logic through textual reports

Researchers from the University of Chicago Booth School of Business conducted a live test of three advanced AI models using data from nearly 2,000 real corporate financial reports from various industries at the end of 2025. The models were able to explain about 17% of stock price fluctuations on earnings report days, while the traditional market metric—earnings surprise—historically explains only a negligible fraction. Even integrating decades of academic knowledge only raises that figure to 8%. The neural networks outperformed human analytical models by comprehensively analyzing unstructured text data, including management forecasts, top executive comments, and linguistic features of their speech during investor conference calls. The AI systems found that strong financial results are often already priced in, so investors react more to the accompanying rhetoric. The results were presented in human-readable digital journals with textual explanations of market anomalies, which could help economists formulate new hypotheses about asset pricing.