Entity-level sentiment analysis for financial news

A news article rarely says one thing about one company. A supplier's shortage is bad news for the carmaker and may be good news for its competitor, and one article can name five companies. Entity-level sentiment analysis, also called targeted sentiment, asks about each company separately: is this article good or bad news for this company?

This guide shows how to run it with spinf, how to turn the scores into a daily sentiment index, and what it measured on our benchmark and in production.

One read, a question per company

spinf asks questions of the content with templates. A placeholder such as {company} asks the same question for every name you list, all over one read of the article:

{
  "id": "sentiment",
  "template": "\n\nIs this news positive, neutral or negative for {company}?\nAnswer:{?}",
  "options": [" positive", " neutral", " negative"],
  "combinations": [ { "company": "Northwind Motors" }, { "company": "Voltra" } ]
}

For each company you get the probability of each answer, and they sum to 1. A template takes up to 1,000 values, so one call can follow a whole watchlist. The same article can carry other questions too:

• Events: "Does it report a supply problem?", "A guidance change?", "A lawsuit?", as yes/no questions.
• Impact: "How much could this news move the share price of {company} (none, small, large)?"
• Topic: earnings, supply chain, regulation, deals, management.

The article is read once per call, and each extra question costs only its own tokens. That is what makes hundreds of questions per article practical.

From scores to a sentiment index

A single article score is noisy. The value for trading research and alternative data comes from aggregating many of them:

1. Remove the question's bias. Each answer comes with its floor, the same question asked with empty content. The lift over the floor is what the article itself says.
2. Place each score on its own history. Rank or standardise the lift against the same question over a baseline of articles from the same sources, so different questions and entities are comparable.
3. Aggregate per day, per entity and per source into indicators, then compare them with prices, flows or outcomes.

The Calibration guide in the docs covers these steps. The result is a measured read of what the news says. It is not a forecast.

In production: the Thinking Text indexes

Thinking Text built daily oil and equity indexes on these reads. It asks about five hundred fixed questions of every oil article and aggregates the answers into a daily read of the press:

• 15M news articles over five years, 2022–2026
• ~1T tokens processed across research iterations
• 0.85 correlation with crude in the best year, 0.80 pooled on WTI
• 92% of big 20-day moves read in the right direction

These figures are from the Thinking Text oil index, which publishes its score card daily. The Thinking Text story describes how the index was built.

Measured on the benchmark

On 108 synthetic news articles about fictional companies (255 company mentions), labelled by a language model and checked blind by a second model, spinf-12b (Gemma 4 12B, optimized) scored:

• 95% correct sentiment per company with 8 labelled example articles in the prompt, and 76% with no examples.
• 98% correct on "does it report a supply problem?" with 8 examples (94% with none).
• 99% correct on "does it report a guidance change?" with 8 examples (98% with none).
• 94% correct main topic out of 5 with 8 examples (88% with none, calibrated). The larger spinf-31b (Gemma 4 31B, on request) reached 95% with no examples.

Synthetic articles are cleaner than real news. On 300 public financial-news posts, which are short and informal, sentiment accuracy was 69% with spinf-12b and 82% with spinf-31b, with the same 8 examples in the prompt. Test on a sample of your own sources before relying on the numbers. The test sets, questions and code are public: spinf-benchmarks on GitHub.

Cost at archive scale

Only input is billed, at $0.09 per million tokens on spinf-12b. An 800-token article with 20 questions of about 25 tokens each is about 1,300 billed tokens, so a million articles cost about $117. The first 50M tokens are free.

Practical notes

• Use the names as they appear. Put the company name in the placeholder the way the articles write it, and keep the list per article to the names it mentions or your watchlist.
• Prefer simple questions for events. Yes/no questions about what an article reports give the clearest reads.
• Keep the questions fixed. Scores compare over time only under the same template, answers and model build.
• Send batches. Many articles per call share the per-call minimum and the floors of the shared questions.

Try it

• The news analysis page has a live example with two companies, and the cost at scale.
• Related guides: Probabilities from a language model and Zero-shot and few-shot classification.