Before the match
Freeze the information window
Each training row may use only information available before that fixture kicked off. Tables, ratings and recent-form values are joined “as of” that publication time.
Transparent by design
How one prediction becomes a public result
Football Proof AI turns historical match information into calibrated home, draw and away probabilities. Every forecast is published before kick-off, then settled against the final result on the same URL.
Published by Football Proof AI · Published · UpdatedInputs
14 active signals + 1 reserved market feature
The launch model uses 14 active inputs across five football-data groups. A fifteenth market-data slot is explicitly reserved and disabled. The active list is locked before evaluation so a good-looking backtest cannot be manufactured by repeatedly changing the rules.
Recent form
Points, goals scored and goals conceded across each team’s previous five completed fixtures.
- Home team points, last five
- Away team points, last five
- Home team goals scored, last five
- Home team goals conceded, last five
- Away team goals scored, last five
- Away team goals conceded, last five
Home and away splits
Separate recent records for the home team at home and the away team on the road.
- Home points per home match
- Home goal difference at home
- Away points per away match
- Away goal difference away
Rating strength
Our own Elo ratings, including a controlled regression toward the league average at each new season.
- Season-regressed home Elo
- Season-regressed away Elo
Match-up context
Head-to-head form is included carefully, with older meetings carrying less useful information.
- Recency-weighted head-to-head points edge
Recovery time
Rest days for both teams capture schedule pressure without guessing at unconfirmed line-ups.
- Rest-days difference
Reserved market feature
One schema slot is reserved for a 24-hour market-consensus prior. It stays disabled until coverage, timing and leakage checks pass.
- 24h market consensus prior — reserved, not active
No future knowledge
As-of joins prevent data leakage
Never allowed
Reject accidental hindsight
Final scores, later league positions and corrected future records cannot flow backwards into a prediction. This is the model’s most important engineering rule.
Model and calibration
LightGBM finds patterns; isotonic calibration fixes confidence
Estimate
LightGBM combines the 14 active signals to estimate the relative case for a home win, draw or away win. The reserved market feature does not enter the current model.
Calibrate
Isotonic calibration adjusts overconfident raw scores. A set labelled 60% should land close to 60% over enough matches.
Publish
The final probabilities, model version and timestamp are recorded at least 24 hours before kick-off.
Acceptance test
Walk forward, never shuffle
The model trains on earlier fixtures and predicts the next untouched time window. That window then rolls forward. It mirrors how supporters in Lagos, Nairobi and elsewhere would actually have received each forecast.
- Top-pick hit rate
- >48%
- Multiclass Brier score
- <0.22
Both gates must be passed on the walk-forward holdout before a model version can publish production forecasts. Hit rate shows how often the highest probability won; Brier score also penalises bad confidence. Read the complete time-ordered validation protocol, then reproduce every reported accuracy metric. Audit a complete 1X2 CSV in the browser.
Public verification
Judge the probabilities, not the personality
No official performance record is active
This method page intentionally withholds sample hit-rate and calibration figures. Production metrics will appear only when they can be derived from locked, settled forecasts. Until then, use the independent audit checklist to inspect the standard.