Does News Sentiment Actually Predict Stock Moves? A 2026 Reality Check

Somewhere on your trading dashboard right now, there is probably a sentiment gauge. A dial, a smiley, a green "Somewhat-Bullish" chip next to a headline — the visual language of 2026's favorite data product. AI reads the news so you don't have to, scores it from bearish to bullish, and the score sits there beside the price chart, radiating significance. The one thing the dashboard never tells you is whether that number predicts anything at all.

Where the scores actually come from

The supply chain is straightforward. Language models — once crude bag-of-words classifiers, now genuinely capable readers — ingest headlines and articles, tag each story with tickers, and emit a score: this piece is +0.4 bullish on this stock, that one is −0.2 bearish. Data vendors license these feeds; retail platforms surface them as gauges, filters, and "AI news insights."

The reading part mostly works. Modern models are good at telling an upbeat earnings write-up from a lawsuit story. The leap of faith happens after the reading: the assumption, everywhere implied and nowhere tested, that today's tone tells you something about tomorrow's price.

Sentiment as decoration

Here is the uncomfortable structural fact about how sentiment gets sold: it is almost always displayed, and almost never tested. The gauge sits next to the chart, and your eye does the rest — you see a green chip and a rising stock and infer causation from adjacency. It's astrology-adjacent UX wrapped around legitimate NLP: the score is real, the implied predictive claim is unexamined.

And the arrow of causation is at best ambiguous. News reacts to prices at least as much as prices react to news; a stock that jumped gets bullish coverage about the jump. Concurrent correlation is easy to find and easy to mistake for foresight. Whether tone leads returns — with what lag, for how long, in which direction — is an empirical question that has to be computed per ticker, per window. Which is exactly what the gauges never do.

What testing it actually looks like

We can offer something better than a hand-wave here, because we publish this kind of test daily — and the results are usefully humbling.

One recent study on our research feed asked whether XOM's most euphoric news days — the top decile of sentiment over roughly three years, 49 days — marked a turning point, with weaker returns over the following week. Answer: no. The reversal story simply didn't show up in the forward five-day numbers. Another asked whether days with top-decile geopolitical news volume (Iran, Hormuz, sanctions) changed XOM's behavior — and there the effect was unmistakable, just not the one the gauges imply: the average intraday high-low range expanded to about 7% of the closing price on spike days versus 4.7% otherwise, a statistically solid result. The news moved the stock. It just didn't say which way.

That pair of findings is a fair summary of what honest sentiment research tends to produce: tone and coverage often tell you something about volatility — how violently a stock will trade — while the directional, get-in-before-the-move signal the dashboards gesture at is faint, unstable, or absent once you actually run the numbers.

The question to put to any sentiment feature

So when a platform shows you a sentiment score, the question is not "is the AI reading the news correctly?" It probably is. The question is: has anyone tested this score against subsequent returns for this ticker, and will they show me the result — the lag, the sample, the decay, the caveats? A vendor with a real edge would publish that study. A gauge with no study attached is telling you about the news, not the future.

Our bias, disclosed

This is the trades.run blog, and the reason we keep writing variations of "test it" is that testing it is literally the product. On trades.run, per-ticker news sentiment isn't a gauge — it's a dataset, sitting alongside prices, earnings, and insider activity. You can ask, in plain English, whether sentiment leads returns for the stock you care about, and a frontier AI model writes and runs the actual analysis code: real event study, stated window, sample sizes, test statistics, caveats. An automated reviewer grades the methodology before anything is published. Sometimes the answer flatters the sentiment feed; more often, as above, it deflates it. Both outcomes ship, because the platform's job is the answer, not the vibe.

News sentiment in 2026 is a real signal wrapped in a fake promise. The AI that reads the headlines has gotten remarkably good; the products that display its scores have gotten no more honest about what those scores can do. Treat every gauge as a hypothesis. Then run the test — here or anywhere that will actually show you the numbers.