Financial research

The problem

Coverage volume is easy to measure and almost always the wrong measure. Syndication inflates it, editorial stance skews tone, and both move for reasons that have nothing to do with the company being covered.

How does a news API support financial research?

Event-level rather than article-level data, with a fixed outlet set so a series measures the subject rather than the sample. Precise time bounds support event studies around announcements.

What the workflow looks like

1. Fix the outlet set before computing anything, and hold it constant across the series.
2. Use event-level counts rather than article counts.
3. Bound windows to the minute for announcement studies.
4. Query the company language as well as English — domestic coverage is earlier and more specific.

What will a news API not do for financial research?

Coverage follows announcements rather than preceding them, so this is not an information edge and should not be modelled as one. Publication timestamps are also inconsistent between publishers, which limits how finely sub-hour sequencing can be trusted.

The mistake teams make here

The mistake is letting the outlet set drift and then interpreting the resulting movement. A series computed over whatever outlets the query happened to reach changes when the index adds publishers, when a publication's feed changes, or when a query's ranking shifts — and all three look exactly like a change in coverage. Fixing the set and holding it constant is unglamorous and is the difference between a measurement and an artefact. The second mistake is treating a news-derived series as forward-looking. Coverage follows announcements, so anything built on it is measuring the reporting of a fact the market already has, and a backtest will look encouraging precisely because coverage volume correlates with everything that was happening at the time.

How would you know financial research is working?

Backtest before you deploy. Take a period you already understand, run the query as it would have run then, and check whether the signal appears when you know it should. News-derived signals are unusually prone to looking predictive in-sample because coverage volume correlates with everything; a signal that survives an honest out-of-sample window is worth having and most do not.

What to build in the first week

Reproduce something you already know before computing anything new. Pick a company and a quarter, fix an outlet set, and rebuild a coverage series you can check against events you remember. What you are testing is not the analysis but the plumbing: whether the window boundaries are what you think, whether the outlet set is stable, whether event-level and article-level counts differ the way they should. Every one of those will surprise someone the first time, and finding out while checking a known period costs nothing, whereas finding out in a result that has already been circulated costs a great deal more.

What are the alternatives for financial research?

Licensed financial news terminals carry exchange-grade timestamps, entity identifiers tied to security masters and the redistribution rights a fund needs — if the output touches trading decisions, that is the compliant path. Filings and market data are authoritative where news is merely early. This is the wide, cheap, fast layer for research and idea generation, not a system of record.

The queries that implement it

• Follow earnings coverage for a listed company — /stories in keyword mode
• Query an exact window, down to the minute — /search in keyword mode
• Find who reported a story first — /search in hybrid mode

Each recipe makes the case for its parameters. Response fields, failure modes and scheduling cost are in the technical documentation.