New Study Finds US Markets Underreact to Hard Data, Overreact to Soft Stories via AI Classification

A recent study accepted at the ICAIF 2026 workshop, titled "Buy the Rumor, Sell the News: When Is News Priced In?", analyzed 4.57 million financial news articles covering approximately 3,000 US stocks from 2023 to 2026 using a large language model (LLM). The researchers distilled an LLM teacher into a compact classifier via active learning, assigning each article one of 17 event tags and five attributes, while clustering articles to separate first reports from follow-up coverage.

By measuring beta-adjusted abnormal returns around 1.68 million stock-day events, the study identified three key results. First, price moves associated with news concentrate before and at publication: the cumulative move in the news direction by the close of publication day is 2.8 times its value twenty days later. For rumor-flagged events, directional returns are concentrated around the rumor itself, with little incremental return at subsequent linked non-rumor events.

Second, compared to neutral-sentiment controls, markets underreact to numbers and overreact to stories. Quantified fundamental news (earnings, dividends, guidance, analyst actions) continues to drift in the direction of the news for weeks, whereas soft story-driven news (launches, macro commentary, leadership changes) gives back its move. Third, news carries width as well as direction: neutral news raises publication-day volatility.

The study produces a table of measured drift for each event tag, usable as a prior in news-conditioned forecasting models. This suggests that for readers using AI in investment research, distinguishing between the "quantitative" and "narrative" nature of news is critical for assessing market absorption speed and potential trading signals.