How the analysis is made

Sources, evidence, scoring, how we check ourselves, and what the engine cannot do.

How a story is made

1. Ingest. The engine reads news feeds and official sources for the UK, US and India on a schedule, de-duplicates them by headline, and keeps the source link for every item.
2. Classify. Each article is classified once: event type, severity (1–10), direction and a confidence score.
3. Resolve entities. Companies and organisations named in the article are matched to a registry using identifiers first (company number, ticker, LEI), then name matching. Uncertain matches are queued for review rather than linked.
4. Trace the chain. The engine writes a causal chain from the reported trigger to the step that reaches a small business — for example a crude-oil move, then pump and utility prices, then the operating margins of fuel-heavy firms.
5. Score and publish. The story gets a 0–100 score and, where the evidence supports one, a cost direction for the businesses it reaches.

How sure we are: the evidence ladder

Every claim carries the strongest kind of source that actually backs it, and the page never shows a claim as more certain than that source. Government records and market data are strong enough to raise a warning. Past patterns and our own reading are shown fainter, and never raise a warning on their own. From strongest to weakest:

1. Filing — a regulatory or company filing: the Gazette, Companies House, IBBI, the SEC. Strong; can raise a warning.
2. Market data — a published price or official statistic: exchange-rate fixings, freight and commodity indices. Measured, not guessed; can raise a warning.
3. Historical analogue — a pattern seen in comparable past events. A pattern, not a fact; shown as context.
4. Inferred — reasoning by the model, with no stronger source behind it. Shown last; never raises a warning alone.

In a published chain, the first step is what the source reported. The steps after it are our traced reasoning: how the event could reach a business, not a forecast that it will.

Scoring

The impact score blends three signals with fixed weights: company financial health and leverage (50%), macro conditions (30%) and structural factors (20%), normalised to 0–100. Confidence combines the model's own sub-scores with the reliability of the source and a small bonus when independent sources corroborate the story. Analytical frameworks (for example discounted cash flow or Porter's Five Forces) are shown as explanations of a step; they are not inputs to the score.

How outcomes are measured

A story's cost call is the direction its final "reaches you" step gives small-business costs: rising or easing. It is read from the story's own words. If the headline or claim moves prices only the other way, no call is recorded. The call is dated at publication. After its horizon, it is checked against a published benchmark series for that region and input over the call's own window. Every result is published on the track record: hits, misses and calls still pending. Calls the benchmark could not check, and stories that made no clear cost call, are counted and not graded. The track record is measurement only: it never feeds back into how stories are scored.

Correction, 28 October 2026

Corrected on 28 October 2026. A decided price change reported by several outlets was counted once per report, and graded against a market price instead of the decision. Ofgem's October price cap rise of 4% was counted four times, and three of those were marked as misses because wholesale energy prices fell over the period, although the cap did rise on 1 October. The rule that one event counts once per cost compared the wording of the reports, not the event, so differently worded reports of the same decision survived it. Now a decided price change (an Ofgem price cap, a central-bank rate) counts once per cost, and is never graded against a market price it does not set. A call on the Reserve Bank of India's repo rate is graded on our recorded repo rate: a hit when the recorded rate moved the way the call said. We hold no recorded series of Ofgem's price cap, so a call on the cap is now marked unverifiable instead of being graded against wholesale prices. Calls on the Bank of England's and the US Federal Reserve's rates are counted once per decision and keep their usual grade. Previously shown on 28 October 2026: 58 hits, 43 misses and 36 pending. Earlier correction, 24 October 2026: the totals at the top of this page were labelled "corrected totals" but still counted calls the 19 October audit had found off-method, so they disagreed with the audit notice beside them. Every count on this page, and the downloadable dataset, now comes from one cost-call gate (method v5), applied when a call is filed and again whenever the record is counted: a call is graded only when its own claim states a realized move in a named cost (a price that rose or fell, a decided price change, or supply cut or restored), it is filed under that cost, and its direction is the one the claim states. A forecast, a proposal, a warning, a price level ("prices remain high") or a story about something else that mentions the cost in passing (central-bank commentary, company earnings, share moves) is not graded. Pending calls are held to the same gate. Previously shown on 24 October 2026: 94 hits, 72 misses and 133 pending. Earlier correction, 16 October 2026: the record graded some stories that make no cost call at all (company earnings, share moves, takeovers, features on a single business, forecasts and opinion), counted one event several times when several outlets reported it, and its headline rate included calls the table hides because the story's own words move the other way. A call now counts only when the story states a move, decision or disruption on the cost it is filed under, in the direction the story states; one event counts once per cost; and the rate counts exactly the calls shown. Previously published on 16 October 2026: 191 hits and 154 misses (55.4%). Earlier correction, 11 October 2026: until then this record graded the engine's call on the company in each story, counted flags that made no cost call as hits when the cost benchmark stayed flat, and included calls checked against a share price. It now grades only calls with an explicit direction for small-business costs (rising or easing), read from the story's final "reaches you" step, against the input's public cost benchmark over the call's own window. Flags with no clear cost direction are shown as a count and never graded. Previously published: 1,627 hits and 1,392 misses (53.9%). A first re-grade shown on the same day (200 hits, 788 misses, 20.2%) was itself wrong and was withdrawn after an audit: it counted calls checked against benchmarks that cannot move as misses (India's administered pump diesel price was unchanged on every print since June), graded some calls against a different cost than the one they named, and measured some monthly prints that came before the call. Calls against a benchmark that cannot register a move (one that stayed inside its no-change band for 90 days, or an administered price that changes only in occasional steps and did not step during the call) are now counted as unverifiable, a call must name the cost it is checked against, and at least half of the measured period must fall after the call. The figures on this page use the corrected method.

Which pages search engines index

Only a small set of stories are offered to search engines: UK or US stories in English with a cost direction, a chain of at least three steps, at least one named organisation, a link to a small-business input cost, a last step that names a cost for businesses in the story's own region, and either a high score or a checked outcome. Exchange notices, daily market wraps, deals whose last step reaches no business cost, duplicates of the same event and anything with investment-recommendation wording are left out. Everything else stays readable, but is marked not to be indexed.

Known limits

• The analysis is written by AI models and can be wrong. Every story links its source so you can check the trigger yourself.
• Coverage depends on what our sources publish and when; some sectors and regions are thinner than others.
• A chain describes one plausible route from an event to a business. Your own exposure depends on your contracts, suppliers and hedges.
• For India, direction and magnitude are withheld and pages are informational only, in line with SEBI rules.
• Nothing here is investment advice.

About · Editorial policy · Contact · Track record · Disclaimer