Champions League: how our match model actually scores
753 Champions League matches, 2021-22 to 2025-26, graded on home, draw and away.
The headline numbers
Season by season
The two most recent seasons ran under the 36-club league phase, which is why they carry more matches.
| Season | Matches | Correct | Accuracy | Log-loss | RPS | We said home | Home actually won |
|---|---|---|---|---|---|---|---|
| 2021-22 | 125 | 73 | 58.4% | 0.936 | 0.193 | 47.9% | 44.8% |
| 2022-23 | 125 | 73 | 58.4% | 0.896 | 0.183 | 47.5% | 48.8% |
| 2023-24 | 125 | 71 | 56.8% | 0.911 | 0.184 | 47.5% | 48.0% |
| 2024-25 | 189 | 108 | 57.1% | 0.951 | 0.213 | 47.8% | 50.8% |
| 2025-26 | 189 | 113 | 59.8% | 0.951 | 0.205 | 47.5% | 49.7% |
| All 5 seasons | 753 | - | 58.2% | 0.933 | 0.198 | - | - |
The last two columns are the average home-win probability we quoted, against how often the home side actually won.
Are the probabilities honest?
The gap between what we said and what happened. Smaller is better.
| Outcome | Calibration error | Brier score | How often it happened |
|---|---|---|---|
| Home Win | 4.8pp | 0.211 | 48.7% |
| Draw | 3.8pp | 0.151 | 18.6% |
| Away Win | 3.2pp | 0.185 | 32.7% |
The full reliability diagram, bin by bin, is on the model transparency page, and the same payload is available through our API.
The season engine is graded separately, season by season, in the 2025-26, 2024-25 retrospectives.
Model: C_hybrid_2015 + goals attack/defence (form_att/form_def), z_shrink=0.90, draw_coupling=0.03, frozen-protocol refit on seasons 2015-2021, holdout 2022-2026 evaluated once
Reuse this data — CC BY 4.0
The Champions League 1X2 model accuracy (backtest) dataset is free to reuse, including commercially, under CC BY 4.0 — attribution with a link is the only requirement. Copy-paste credit:
Data: <a href="https://www.predixsport.com/champions-league-model-accuracy">Champions League 1X2 model accuracy (backtest) by Predixsport</a> (CC BY 4.0)