Model ledger
Model scoreboard
Served calibration rows for soccer_prop_lgbm. This is what the props board and player reads actually use. 3,037 player profiles across 8,406 archived matches · mode: in-season.
Model ledger
Held-out calibration · soccer_prop_lgbm
Temporal holdout with isotonic calibration. Skill is log-loss improvement vs per-player base rates (73,406 prop rows). Lower log-loss and Brier are better. ECE targets stay under 8%.
| Market | Rows | Log loss | Brier | ECE | Skill |
|---|---|---|---|---|---|
| Shots 2.5+shots_o25 | 73,406 | 0.232 | 0.068 | 1.00% | +23.0 pp |
| SOT 0.5+sot_o05 | 73,406 | 0.413 | 0.133 | 0.66% | +21.1 pp |
| Anytime goalanytime_goal | 73,406 | 0.228 | 0.063 | 0.44% | +20.1 pp |
| Shots 1.5+shots_o15 | 73,406 | 0.405 | 0.130 | 1.12% | +20.0 pp |
| SOT 1.5+sot_o15 | 73,406 | 0.167 | 0.044 | 0.98% | +17.6 pp |
| assists o05assists_o05 | 73,406 | 0.198 | 0.051 | 0.91% | +16.1 pp |
| Tackles 1.5+tackles_o15 | 73,406 | 0.513 | 0.174 | 2.75% | +12.8 pp |
| Fouls 0.5+fouls_o05 | 73,406 | 0.605 | 0.210 | 2.89% | +12.6 pp |
These metrics describe out-of-sample fit, not forward betting ROI. A calibrated model can still lose to the closing line. See methodology for how numbers reach the site.
Hub coverage
| League | Players | Matches | Prop markets | Fixtures hub |
|---|---|---|---|---|
| Premier League | 636 | 1845 | 13 | 50 |
| La Liga | 636 | 1637 | 13 | 70 |
| Bundesliga | 535 | 1849 | 13 | 36 |
| Serie A | 640 | 1674 | 13 | 50 |
| Ligue 1 | 590 | 1401 | 13 | 45 |
Forward-graded receipts: report card. Full process and limitations: methodology.