Transparency centre
AI methodology and accuracy measurement
Last updated:
Tahminator forecasts are decision-support summaries built from official bulletins and verifiable match data. The goal is a repeatable, measurable workflow, not a promise of certainty.
1. Input layer
Schedule, competition, teams or players, official market lines, odds distribution, goal/point/set markets and official live status are read together. Missing data is marked as missing and is not replaced with an invented score.
2. Scenario layer
Suitable markets such as match result, handicap, totals, both teams to score, sets or periods are compared by sport. A main selection is shown only when supporting signals align; high-risk fixtures can be skipped.
3. News and availability
Injury, suspension, lineup and team-news effects are recorded only when the source context is verified. A revision does not silently erase the original forecast; the reason and timing remain visible.
4. Confidence band
Confidence combines market separation, market agreement, data completeness, risk filters and verified news impact. It is not a guaranteed probability, and skipped fixtures are excluded from average confidence.
5. Result comparison
A forecast is evaluated only after an official result is final. Won, lost or pending is calculated by comparing the saved selection with the verified result. No score is written when the result is unavailable.
6. Performance metric
Only settled matches with a saved forecast are counted. Pending, postponed, cancelled and skipped fixtures are excluded. If the sample is small, the panel reports insufficient data instead of presenting a misleading rate.
7. Data sources, methodology and confidence score
The model confidence score is not a measured success rate or win probability. Normalised bulletin odds and verified historical match statistics are shown separately. Automated analysis; no human review is recorded. Schedules and results use iddaa and Statshub records. Form and head-to-head history appear only when matching dated local archive or OpenFootball data is available. OpenFootball datasets link to CC0 licences; these do not cover other providers or team logos. Complete coverage of every league is not claimed.
8. Weather and travel report
Experimental uncertainty deduction, capped at 3 points: temperature ≤0 °C or ≥32 °C, 1; wind ≥10 m/s, 2 (≥15, 3); snow, 2; rain averaging ≥2 mm/h, 1. This is not a win probability or measured performance effect and applies to both sides. A verified journey over 1,500 km with less than 36 hours of recovery adds 1 risk point to the away-win selection. The temperature difference compares forecasts at the same kick-off time, not acclimatisation. Straight-line distance is not an itinerary or travel duration. No away travel deduction applies without a verified journey. Values are forecasts for kick-off, not observations; intermediate hours may be interpolated. The precipitation period is stated separately. Started matches retain their recorded analysis. MET Norway · CC BY 4.0.
9. How verifiable form contributes to confidence
When both football teams have five recent, sourced results, goals scored and conceded inform the analysis, with home/away splits when sufficient records exist. Form weight is 15%; confidence changes by at most ±3 points. Match-day, future and incomplete records are excluded; a fresh official bulletin is required. Match result, goal totals, both teams to score and protective choices are compared before selection. Scores below 60 are skipped; choices scoring 60 or more are considered when data and risk checks pass; direct choices scoring 70 or more take priority. Odds only break near-equal confidence ties. The last pre-match record is preserved after kickoff. This experimental score is not a measured win probability.