How NewsPips scores direction, conviction and impact

This page explains how NewsPips turns raw news flow into the per-instrument scores you see in the product: where the information comes from, how direction and conviction are produced, and — just as importantly — what those numbers do not mean. We publish this because a score you cannot interrogate is a score you should not trust, and the entire design of NewsPips assumes users will check its reasoning rather than take it on faith.

From hundreds of sources to one signal

NewsPips continuously monitors hundreds of news sources — wire services, financial press, official releases, and specialist outlets — alongside the scheduled economic calendar. Breadth is the point: market-moving news rarely breaks in one place, and no single publication catches everything early.

Breadth creates its own problem, though: one event generates dozens of near-identical headlines as outlets pick up the same story. Raw feeds bury the signal in repetition. So the pipeline's first substantive job is deduplication and clustering: articles are matched by URL, by title, and by semantic similarity of their content, and grouped so that one real-world event becomes one cluster of coverage rather than fifty separate items. A cluster carries useful structure of its own — how many independent sources corroborate the story, how coverage is building over time, and which account of the event is primary.

The result is an event stream, not a headline stream. Everything downstream — categorisation, impact assessment, instrument analysis — operates on events.

Scoring direction and impact per instrument

NewsPips does not score individual articles, and it does not produce one global "market sentiment" number. Analysis happens per instrument: each supported market (major FX pairs, gold, oil, and others) gets its own periodic assessment built from the clusters relevant to it, plus recent calendar outcomes.

Each analysis produces a directional bias — the direction the assessed evidence currently leans for that instrument — together with a conviction level and a written rationale. The rationale is not decoration; it is the product. Every material claim in it cites the underlying source articles, so you can trace any statement back to the coverage that produced it and judge the sourcing yourself. If an analysis says a hawkish surprise supports the dollar, the citation shows you which release, reported where.

The same event can, and often does, point different directions for different instruments — that asymmetry is precisely why the analysis is per-instrument rather than global.

What conviction means — and what it does not

Conviction is a strength-of-evidence readout. It reflects how the current cluster set stacks up: how significant the events are, how well-corroborated, how consistently they point the same way, and how directly they bear on the instrument. High conviction means the evidence is substantial and aligned. Low conviction means the picture is thin, mixed, or ambiguous.

What conviction is not is a win probability. A high-conviction reading does not mean the market is highly likely to move that way; it means the news case for that direction is currently strong. Markets can and do move against well-evidenced positions — because prices already reflected the news, because a larger force dominated, or because the situation changed an hour later. Treating conviction as a forecast confidence interval is a common misreading of the product, which is why we state it plainly here.

Where the numbers can be wrong

An honest methodology page lists its failure modes. Ours are these:

• Fast-moving stories. Analyses are periodic snapshots. In a rapidly developing situation, a reading can lag the latest headline, and early coverage of breaking events is often partly wrong in ways that correct over subsequent hours.
• Ambiguous events. Some events genuinely cut both ways — data that is strong for growth but hawkish for rates, for example. The analysis will take a position on which channel dominates, and that judgment can be wrong.
• Model limits. The analysis is produced by AI language models. They can misweigh evidence, over-read thin coverage, or occasionally misstate a detail. Citations exist partly as a check on exactly this: a claim you can verify is a claim you can catch.
• What isn't in the news. The pipeline reads public coverage. Flows, positioning, and anything that moves markets without generating articles are outside its field of view.
• Already-priced news. The analysis assesses what the news implies, not how much of that implication the market has already traded. A correct reading of stale news is not a useful one.

What the scores are NOT

To remove any ambiguity:

• The scores are not investment advice and are not personalised to anyone's situation, account, or risk tolerance.
• A directional bias is not a recommendation to buy or sell anything. It is a structured summary of what the current news flow implies, published with its evidence.
• No output of the system carries any assurance of accuracy or future performance. Trading involves substantial risk of loss, news-driven trading especially so, and no analytical layer removes that risk.

NewsPips is an evidence engine. The decisions — whether to act, in what size, with what protection — are not part of the product and cannot be outsourced to it.

Reading NewsPips well

The users who get the most from the product treat it as an evidence layer beside their own process, not a signal to follow. In practice that looks like: using the event stream and pre-event reminders to never be surprised by scheduled releases; reading the per-instrument analysis as a briefing — what happened, how well-sourced is it, which way does it lean — and then checking the citations on anything that would change your view; and weighing conviction as a measure of how seriously the news case deserves to be taken, not how certain the outcome is.

That workflow presumes some grounding in how news actually moves prices — why surprises matter more than headlines, and what happens in a release window. If you are building that foundation, start with how to trade news events and the economic calendar, explained. And if you want to see the methodology in action before reading another word, the homepage live demo shows the real event stream and current per-instrument analyses; plan options are on the pricing page.