Scoring News Relevance to a Ticker

A newswire mentions "Apple" in a story about a supplier's earnings, in a macro piece that name-drops Apple as an example of a large-cap winner, and in a story announcing Apple's own product recall. All three contain the ticker AAPL somewhere in the text, but only one is actually about Apple in a way that should move a trading signal. A pipeline that treats every mention as equally important floods a strategy with noise and drowns out the rare story that genuinely matters.

Relevance scoring is the step that sits between "this ticker's name appears in this document" and "this document should update my belief about this ticker." It assigns each article-ticker pair a score, typically between 0 and 1, representing how central the company is to the story, rather than treating detection as a binary switch.

What goes into the score

A relevance model typically combines several cheap, interpretable signals before ever reaching for anything fancy:

• Position and frequency — a company named in the headline and repeated throughout the body is almost certainly a subject, not a footnote. A single mention buried in paragraph six of an unrelated macro piece is not.
• Entity prominence — how many other companies share the story? An article about one firm's earnings has one prominent entity; an "S&P 500 winners and losers" roundup has fifty, each diluted.
• Source and section metadata — a company-specific press release wire is a strong prior; a general market-commentary feed is a weak one.
• Learned relevance — a supervised classifier (often a fine-tuned transformer, see Sentence Transformers and Contrastive Fine-Tuning) trained on human-labeled article-ticker pairs, predicting relevance directly from the text and its position within it.

These are usually combined into a single score, either by a hand-tuned weighted sum for a simple production system, or by feeding the individual signals as features into a supervised model that outputs the final probability.

Worked example

Suppose a rule-based scorer uses three binary features — headline mention (0.5), single-entity story (0.3), dedicated company-news source (0.2) — summed and capped at 1.0. A story headlined "XYZ Corp Cuts Guidance" from a company newswire, mentioning no other tickers, scores $0.5 + 0.3 + 0.2 = 1.0$: maximal relevance. A "Tech Sector Roundup" story that mentions XYZ Corp only in paragraph four alongside nine other names scores $0$ on all three features, despite the ticker technically appearing. A production pipeline sets a relevance threshold — say 0.4 — below which the story is discarded from the ticker-specific feed entirely.

What this means in practice

Relevance scoring is a filter that sits upstream of every text-based signal a desk builds: sentiment scores, event flags, or embedding features are only as good as the documents feeding them. Skipping this step is a common way a backtest looks great on curated data but degrades in production, where the raw feed is full of loosely related noise. It also interacts with Novelty and Staleness Detection in News — a highly relevant story that's just a rehash of yesterday's news needs a second filter on top of relevance alone.