How Predictly forecasts
Between your question and the percentage on screen is a fixed sequence of stages. Here they are in order, with the thresholds that are actually applied.
The stages, in order
- NormalisationThe sentence becomes an event: what is being talked about, which outcome is in question, by when. This step is independent of the language and the phrasing used.
- IdentificationThe event gets a computed key. Two phrasings of the same event — in English, in French, forwards or backwards — produce the same key.
- Already settled?If the resolution date has passed, or the question is explicitly about the past, processing stops here and Predictly says so. No forecast is produced about a settled fact.
- Ambiguous subject?If the name in the question has several readings, they are named and the question is handed back. No research is run on two subjects at once.
- ReuseIf a recent forecast of the same event is still inside its freshness window, that is what comes back. Two people asking the same thing at the same moment trigger one research run.
- Research planTargeted research questions are built for the kind of event: a match, a product launch and an election do not call for the same checks.
- ResearchThe plan runs in parallel across several providers: grounded web search, news, press feeds, an international event database, an encyclopedia, and sports data where the subject calls for it.
- Qualifying the sourcesEvery retrieved document is typed and weighted, and copies of one wire story are collapsed into a single story.
- Extraction and clusteringClaims are extracted and tied to the source carrying them and the outcome they support. Ten articles repeating one report count as one.
- Evidence gateIf what remains is too thin, no forecast is made: Predictly returns a “not enough reliable evidence” answer rather than a number.
- Calculation and presentationProbabilities are computed per possible outcome and sum to 100. They are rounded to the nearest 5 points, and the answer is presented in your language.
What counts as a source
Documents are not all worth the same, and Predictly does not pretend otherwise. Every source is classified, and its class decides how much it weighs.
- Official
- The party itself or the body that runs the event: a league, a company, a government.
- Journalism
- Specialist or recognised major news organisations.
- Reference
- Encyclopedias and reference works: background, not news.
- Other
- Everything else. Counted, never enough on its own.
Relayed statements
A secondary site reporting an attributed statement — “UEFA confirmed”, “the studio announced” — carries the speaker’s authority at one remove. Its class is lifted one step, and no further.
Freshness
A source loses weight as it ages, and how fast depends on the horizon. For Saturday’s match, last week’s article is already old; for a launch two years out, it is recent.
Copies count once
One piece of information repeated by fifteen sites is not fifteen pieces of information. Before anything is computed, documents telling the same story are grouped, and the group counts as a single item of evidence.
Without that, a heavily syndicated subject would look solid simply because it had been copied. The question Predictly is trying to answer is not “how many articles?” but “how many independent findings?”.
The evidence gate
Three conditions are enough to decline. If any one holds, no probability is shown.
- Fewer than two independent findings: nothing corroborates anything.
- Fewer than two distinct publishers: a single voice, however often repeated.
- No recognised source, and fewer than four distinct publishers: neither authority, nor agreement broad enough for the agreement itself to be the evidence.
That third condition exists because no list of recognised sites covers the whole web. A finding can be solid without a known name behind it — provided enough separate publishers report it, once copies have been collapsed.
Confidence is not probability
These are two different measurements, and confusing them is the commonest misreading.
- Probability
- How likely the event is to happen.
- Confidence
- How well documented Predictly’s answer is.
An unlikely event can be reported with high confidence: “15 %, and that is well established” is a coherent sentence.
How it is computed
Confidence is computed on the server from four things: the number of independent findings, the share of authoritative sources, the average reliability of the dossier, and the gap between the leading outcome and the next. The model never chooses it, and it is not derived from the percentage.
High confidence is granted only to a dossier that has all of it: a sufficient overall score, at least four independent findings, a clear gap to the second outcome, and at least two items that are reported findings. A fixture list and an encyclopedia entry can say what is scheduled, not how it will go.
How long a forecast stays current
A forecast is stored per event and reused while it is fresh, then the research runs again. How long depends on how fast the subject moves.
- Crypto
- 1 hour
- Sport
- 3 hours
- Business
- 4 hours
- Politics, world
- 6 hours
- Other subjects
- 12 hours
- Tech, culture, science
- 24 hours
Those windows shorten as the date approaches: at most 3 hours within the final fortnight, at most 1 hour within the last two days. Tonight’s match is reconsidered far more often than next year’s election.
Same question, same answer
Because the event is identified before any research is done, rephrasing changes nothing. “Will PSG beat OM?”, “PSG v OM, who wins?” and the same question in French land on the same key, and therefore the same forecast, as long as the news has not moved.
What changes the answer is information, not wording. When an injury is announced or a date is denied, the next forecast takes it into account.
Methodology version
The engine described here is forecast-v3. Forecasts are stored with the version that produced them, and a change of version invalidates what was computed under the previous one.
For reading the percentages themselves, see understanding probabilities.