news sentiment / stock returns / quantitative research
Does News Sentiment Predict Stock Returns? We Tested 2.27M Company Days
Nvidia rose 24% after earnings, but buying the next morning’s open lost 1.4% by the close. We tested whether company news sentiment predicted returns after entry across 2.27 million company days and five holding periods.
See the actual return gaps, uncertainty, a separate news-change ranking test, and the limits of scoring older headlines with a 2026 classifier. These retrospective comparisons did not establish a reliable trading edge. Read the full results and methods in the article.
Transcript
0:00Nvidia jumped twenty-four percent after
0:02earnings. But if you bought
0:03at the next morning's open,
0:05you finished that day down
0:06one point four percent.
0:08The stock had already moved
0:09before your entry.
0:11Can news predict what happens
0:12after you buy?
0:13We tested that across millions
0:15of company days.
0:16Our archive had nearly nine
0:18and a half million deduplicated
0:20story and company matches across
0:21trading sessions. Each morning, we
0:24compared positive and negative news
0:25companies, then subtracted the market's
0:27price return over the same
0:29window. Each dot asks: did
0:31positive-news stocks beat negative-news stocks?
0:34We asked that after a
0:35day, a week, a month,
0:36a quarter, and a year.
0:38Four answers were too uncertain
0:40to call. The quarter looked
0:41more convincing: positive-news stocks finished
0:44about one percentage point behind.
0:46But checking five periods gives
0:47us more chances to find
0:48a pattern by luck.
0:50With this much uncertainty, the
0:51quarterly gap needed to reach
0:53about one point one eight
0:54percentage points. We measured just
0:56over one percentage point.
0:58It fell short.
0:59Extreme price moves mattered, too.
1:01Removing windows with a one-day
1:03rise over four hundred percent,
1:05or a fall over eighty
1:06percent, brought the next-session gap
1:08close to zero.
1:09Those moves can be real,
1:10so the main test kept
1:12them. Then we tested sudden
1:13changes and bursts of news.
1:15Both models ranked the same
1:17news-selected companies using recent prices;
1:19one also saw the news
1:21features. Across five hundred fifty
1:23paired daily cohorts, the measured
1:25average gap was near zero,
1:26with uncertainty stretching in both
1:28directions. It failed all four
1:30checks we set in advance.
1:32We built this classifier in
1:33twenty twenty-six, then used it
1:35to score older news.
1:37For most of those records,
1:38we used news-archive timestamps, rather
1:41than logs of our system
1:42receiving and scoring them at
1:43the time. We still need
1:45to test whether this setup
1:46makes money when run live.
1:48Read the full results and
1:49charts in the article.
Join the conversation
Loading conversation…