Eight leagues evaluated
Domestic club football
A LightGBM club model estimates 1X2 and total goals. The current 1X2 artifact uses isotonic calibration fitted on a held-out sample. Historical evaluation covers eight European domestic leagues.
Cette page n’existe qu’en anglais.
The method
A useful probability needs good inputs, a fair test and an honest account of uncertainty. Here is how those pieces fit together.
From data to perspective
Historical results, team strength and available pre-match context form the input. Data must belong to the correct competition, season and team.
The domestic club model estimates match-result and total-goals probabilities. A competition simulator then applies the relevant schedule, league-phase and knockout rules.
Training uses earlier matches; evaluation uses the period that follows. We inspect calibration, log loss and Brier score, together with the sample and baseline.
Available results are shown with their competition and update information. Missing data or an incomplete simulation keeps projections pending.
The engines
Eight leagues evaluated
A LightGBM club model estimates 1X2 and total goals. The current 1X2 artifact uses isotonic calibration fitted on a held-out sample. Historical evaluation covers eight European domestic leagues.
Calendar and forecasts are separate
Champions League, Europa League and Conference League use a hybrid club-strength engine with their own league-phase and knockout rules. Available market probabilities are adjusted for margin; otherwise the calibrated club model supplies the match inputs.
Separate national-team model
The World Cup workspace retains a national-team model based on Elo and score distributions. Its ratings and evaluation belong to international football and are kept separate from UEFA club projections.
Verified UEFA fixtures and team directories can be published before projections are ready. Projections remain pending until a complete verified calendar and a fresh validated simulation are published. A visible calendar does not establish forecast availability or predictive performance; each workspace shows its current data status.
Testing the forecast
The September 2026 domestic evaluation uses five chronological walk-forward folds and 12,944 out-of-sample matches. Each fold preserves the order of matches so that its test period follows its training period.
Expected calibration error compares confidence with observed outcomes. Log loss and Brier score assess the full probability forecast. No single number describes every aspect of model quality.
The latest bookmaker baseline scores better than our model on log loss and Brier score. Historical calibration alone does not establish a forecasting advantage or profitable strategy.
Read the complete comparisonReading with care
Late lineups, injuries, team changes and incomplete provider coverage can affect a forecast. Always check the available update information.
Tournament projections depend on the schedule, model inputs and completed results available to that run. They are not a fixed prediction of the final table.
An out-of-sample historical test is not a prospective live record. New seasons and competitions can behave differently from the evaluation sample.