Cette page n’existe qu’en anglais.

Evidence, in context

The numbers behind the forecast.

A transparent view of our domestic model evaluation, including the baselines and the limits of what these results show.

Historical evaluation · 2 September 2026

Out-of-sample matches
12,944
Walk-forward folds
5
Domestic leagues
8
Prediction targets
2

What the evaluation says

A useful test includes a strong baseline.

The bookmaker baseline has lower log loss and Brier score than our model on both targets in this evaluation. These results do not demonstrate an advantage over bookmakers.

Calibration measures one aspect of forecast quality. A low calibration error does not mean that individual matches are predictable, or that a strategy will be profitable.

Match result · 1X2

Model calibration error 0.0315

Five-fold historical evaluation. Lower log loss and Brier score are better.
PredictorLog loss ↓Brier ↓Accuracy
PREMATCH IQ · LightGBM1.06820.595950.96%
Bookmaker baseline0.97230.578053.28%
Elo baseline1.04380.623849.03%

Evaluation reference: validate_1x2_8leagues_20260902

Total goals · Over/Under 2.5

Model calibration error 0.0184

Five-fold historical evaluation. Lower log loss and Brier score are better.
PredictorLog loss ↓Brier ↓Accuracy
PREMATCH IQ · LightGBM0.67920.243156.48%
Bookmaker baseline0.67230.239858.45%
Poisson baseline0.70910.256253.55%

Evaluation reference: validate_ou25_8leagues_20260902

The match-result Brier score sums across three outcomes; the total-goals score is binary. Compare models within the same target, rather than comparing the two Brier scales directly.

Publication status

Historical tests are not a live record.

This page summarises saved evaluation reports. A public, match-by-match ledger of timestamped predictions and settled outcomes is still pending publication.

UEFA performance results are also pending. Domestic evaluation numbers do not validate Champions League, Europa League or Conference League tournament projections.

Read the measures

Calibration error
How closely grouped probabilities match observed outcome frequencies. The sample and grouping method affect this measure.
Log loss & Brier score
Scores that evaluate the predicted probability distribution against the outcome. Lower values are better on the same evaluation set.
Walk-forward evaluation
Successive tests on later matches, using earlier matches for training. It preserves time order when assessing generalisation.
Explore the methodology