Sentiment analysis converts text — news articles, filings, social media posts — into numerical scores that can be studied alongside prices. The idea is intuitive: markets react to information, and much information arrives as text. The practice is more subtle.

How sentiment scores are built

Early approaches counted positive and negative words using finance-specific dictionaries. Modern systems use machine-learning models, including fine-tuned language models, that classify sentences in context. Better models handle negation and domain language — “beat expectations” versus “beat down” — but all of them reduce nuanced text to a small number of values.

Where signal tends to appear

  • Entity-level news about specific companies, especially unexpected events.
  • Changes in tone relative to a company's own history, rather than absolute levels.
  • Aggregates across many sources, which reduce the noise of individual items.

Where noise dominates

Social media sentiment is vulnerable to bots, coordinated campaigns and a small number of very active accounts. Sentiment that merely describes price moves after they happen — “shares plunge” — adds little predictive information. And widely available scores are, by definition, widely available: any edge tends to be arbitraged away.

Sentiment that merely describes yesterday's price move is a mirror, not a signal.

Evaluating a sentiment dataset

  • Are timestamps the moment of publication, or when the vendor processed the text?
  • How are duplicate and syndicated articles handled?
  • Is the model's methodology documented, and has it changed over the history provided?
  • Has the vendor's back-history been reconstructed with today's model? If so, it may contain look-ahead.

Event-focused tools such as LevelFields AI and scoring platforms like Danelfin incorporate sentiment as one input among several. As with any signal, it is most useful as context for research rather than as a standalone trigger.

Testing a sentiment signal honestly

Sentiment research is particularly prone to data-snooping. There are many sources, many scoring methods, many aggregation windows and many possible trading rules — and with enough combinations, something will appear to work historically. A disciplined test fixes the hypothesis first: for example, that a sharp drop in a company's news sentiment relative to its own 90-day average precedes underperformance over the following week. Only then is the data examined.

Results should be examined by sub-period, by sector and by market capitalisation. A signal that works only in one year, or only in the smallest and least liquid stocks, is unlikely to survive costs or the passage of time.

Combining sentiment with other data

In practice, sentiment is most often used as a filter or a conditioning variable rather than a standalone signal — for instance, avoiding new positions in companies experiencing unusually negative news flow, or prioritising research on companies where sentiment and fundamentals diverge. These uses demand less precision from the score and are more robust to its imperfections.

Language models and sentiment

Large language models can classify sentiment with more context than earlier methods, but they introduce new questions: results can change when the model is updated, and historical scores computed with today's model may embed knowledge the market did not have at the time.

Quorient Editorial Desk

Articles credited to the Editorial Desk are researched and written for Quorient and reviewed by the editor, Callum Reid, before publication, in line with our Editorial Policy. Spotted an error? Request a correction.