How the forecasts work

Probabilities, not tips

PredictMargin estimates home-win, draw and away-win probabilities before kickoff. The values always add to 100%; a likely outcome is never a certainty.

Data starts free and stays replaceable

The launch version uses fixtures, delayed results and public historical data. Providers sit behind one interface so paid coverage can be added without changing public pages or the prediction record.

What is deliberately excluded

• Club crests, league marks and player photographs without separate permission.
• Odds, bookmaker feeds and affiliate data.
• Cancelled, abandoned or unresolved matches in training and evaluation.

A score model people can audit

A Poisson framework estimates team attack, defence and home advantage. A Dixon–Coles correction improves low-score estimates, and a normalized score grid produces outcome probabilities and expected goals.

How performance is scored

Log Loss, Brier Score and Ranked Probability Score reward useful probabilities and penalize confident mistakes. Calibration checks whether outcomes forecast near a percentage occur at about that rate.

The record cannot be rewritten after kickoff

Each forecast stores its model version, training cutoff, source timestamp and publication time, then locks at kickoff. Evaluation uses that locked forecast after the result is known.

Known limitations

Free data can be delayed and the baseline omits several match-day signals. Football remains noisy, so a calibrated probability is never a promise about one match.

• The free source may delay fixtures and final scores, so this is not a live-score product.
• The baseline does not yet include injuries, confirmed lineups, player quality or live expected goals.
• Recently promoted and data-poor teams receive conservative competition-level fallback parameters.
• Football remains noisy; a well-calibrated probability is never a promise about one match.