The four data layers
Fixtures and final scores
The configured sports-data integration is API-Football. Provider fixture IDs anchor competition, team, kick-off, status and score updates. A configured integration does not imply that every competition or field is available at every moment.
Historical features
Completed matches are transformed into recent points and goals, home/away splits, season-regressed Elo ratings, recency-weighted head-to-head context and rest-day difference. These values are calculated by Football Proof AI rather than supplied as a provider prediction.
Model and calibration
A multiclass LightGBM model generates raw 1X2 scores. One-vs-rest isotonic calibration and renormalisation convert them into home, draw and away probabilities. Each accepted artifact has a model version and hashes recorded in its model card.
Settlement and performance
When the provider marks a fixture final, the score is mapped to home, draw or away and appended as a settlement. Hit rate and Brier score are derived from the locked probability and latest valid settlement, not manually typed headline numbers. Every settlement revision remains available in the match's public receipt with its result-event provider payload hash and supersession links.
Time integrity and leakage control
Training and inference use as-of joins: a match may use only information that existed before that match's kick-off. Matches sharing a time window are processed from the same pre-match snapshot so one completed result cannot leak into another simultaneous prediction. Walk-forward evaluation recreates that sequence across time.
Deliberately excluded or separated
- A 24-hour market-consensus prior has a reserved schema slot but is disabled until timing, licensing, coverage and leakage checks pass.
- BTTS and Over 2.5 probabilities remain unavailable until separately trained and validated models exist; a 1X2 model does not manufacture them.
- Kito explanations are created only after a record is locked. Deterministic evidence mode is always available; optional AI response planning is separately labelled, and the external model can select only a full or compact view of server-authored evidence. Neither is a model feature or can change the official forecast.
- Demo rows are excluded from official RSS items, calendar events and CSV performance exports.
Known limitations
Provider feeds can be late or corrected; team naming and competition coverage can change; unconfirmed line-ups and injuries may be absent; newly promoted teams have thinner history. The site should expose stale or unavailable states rather than hide those limits. A data source is necessary for a forecast, but it does not make the forecast certain.
Read How it works for the feature and evaluation protocol, then open Model cardsfor registered artifact evidence. The public data documentation explains how to reproduce a single prediction's canonical receipt hash without trusting the page renderer.