Skip to main content
← Previsões

Recorded before kickoff

The forecast. The timestamp. The result.

Explore retained forecast versions and see how their probabilities performed after the final whistle.

Browse forecasts ↓Public calibration ↓

Original probabilities stay unchanged. Results are graded separately. The archive starts when retention was enabled; earlier forecasts are not reconstructed.

The evidence

Latest published forecasts

All timestamps in UTC

Public calibration — frozen forecasts

Latest 500 completed fixtures with retained forecasts; one latest retained pre-match version per fixture. Includes all model versions shown below. No reconstructed forecasts enter this report.

485graded matches
0.5972Multiclass Brier (0–2)
1.0649Log loss

500 considered · 0 awaiting results · 15 excluded: provenance or result could not be verified

Calibration across home, draw and away outcomes
Probability bandMean predictedObserved frequencyOutcome observations
0–10%5.2%10.0%30
10–20%16.2%10.4%106
20–30%26.5%23.3%468
30–40%34.4%33.1%528
40–50%44.1%51.2%203
50–60%54.5%62.3%77
60–70%64.0%80.8%26
70–80%74.4%81.3%16
80–90%80.9%100.0%1
90–100%——0

Each match contributes three outcome observations; these are correlated, not three independent matches. Bands include the lower bound and exclude the upper bound, except 100%. Sparse bins are unstable. Lower Brier and log loss are better.

Model versions: ML_composition_v8:Advanced_joint_form_v73 (425) · ML_composition_v8:Fallback_regulation_v3 (55) · ML_composition_v8:INTL_regulation_v3 (5)

Updated 19 Sep 2026 · 18:35 UTC · Forecasts stay unchanged. Grading uses current corrected regulation results; extra time and penalties are excluded. Rounded probability vectors are normalized for scoring; log loss uses natural logs with a 1e-15 floor.