Market Sentiment from News: Signal or Noise?

Markets react to news within seconds, so it is tempting to believe that measuring the mood of headlines must predict prices. Sometimes it does. Much more often, sentiment is noise that only looks like a signal in hindsight.
What “sentiment” really measures
Sentiment analysis scores text as positive, negative or neutral. Applied to financial news, it tries to capture how the market might feel about a company, a sector or the economy. Modern language models are far better at this than older word-list approaches: they understand context, negation and the difference between “profit fell less than feared” and “profit fell”.
But a better reader does not automatically make a better signal. The key question is not “is this headline positive?” — it is “does this information change what the market already expects?”
Why most sentiment is noise
- It is already priced in. By the time a story is widely published, prices have often moved.
- Volume is not importance. A company can attract thousands of articles about a story that has little effect on its earnings.
- Duplication. One event is rewritten by dozens of outlets, making a single piece of news look like a flood of opinion.
- Context matters. Negative news about a competitor can be positive for the company you are studying.
Sentiment is useful only when it tells you something the price has not yet reflected.
Turning sentiment into something testable
At Foxiton we treat sentiment as one input among many, never as a signal on its own. That means removing duplicate stories, tying each item to the right company and time, measuring how surprising it is relative to recent coverage, and combining it with price and volume data.
Most importantly, every sentiment feature is tested historically — with strict care to use only information that was available at the time. A feature that cannot show a consistent relationship with subsequent price moves is dropped, however intuitive it seems.
So, signal or noise?
Both. Raw sentiment is mostly noise. Carefully filtered, de-duplicated and surprise-adjusted sentiment can carry real information — usually small, often short-lived, and only valuable when combined with other evidence. The research is in knowing which is which.
Key takeaways
Language models read news far better than older methods, but better reading is not the same as better prediction.
Most headline sentiment is already priced in, duplicated or irrelevant.
Sentiment becomes useful when it measures surprise and is tested strictly against historical outcomes.
THIS ARTICLE IS FOR INFORMATION AND EDUCATION ONLY. IT IS NOT INVESTMENT ADVICE OR A RECOMMENDATION TO BUY OR SELL ANY SECURITY.
NEXT ARTICLEAI Stocks: Separating Real Earnings from Hype

