Model Performance
A probability model is judged by calibration over samples, not by single-match hits. Every engine change must move these numbers on out-of-sample matches — or it gets cut.
2026-27 Season · Cumulative (pregame predictions, no hindsight)
Lower log loss = better. Baseline = constant league base rates (home 48%, draw 18%, away 33%). Sample is still tiny (9 matches): treat every number here as noise until at least a half-season has accumulated.
Over 2.5 Calibration
Pregame Over 2.5 probability (same number as on each match analysis page) vs whether the final score actually had 3+ goals. Sample is still tiny (9 matches): treat band rates as noise until n ≥ 30.
| Model said | n | Model avg | Actual Over rate | Δ |
|---|---|---|---|---|
| 60–70% | 5 | 66.9% | 60.0% (3/5) | -6.9 pp |
| 70–80% | 2 | 75.5% | 100.0% (2/2) | +24.5 pp |
| 80–90% | 2 | 83.4% | 100.0% (2/2) | +16.6 pp |
Well-calibrated bands sit near Δ ≈ 0 (actual rate ≈ model average). Positive Δ = model under-called Over; negative = over-called.
Per Gameweek
| GW | Matches | Model log loss | Baseline | Δ | Picks |
|---|---|---|---|---|---|
| GW 1 | 3 | 0.5276 | 0.8532 | +0.3256 | 3/3 |
| GW 2 | 3 | 1.0470 | 0.9748 | -0.0723 | 2/3 |
| GW 3 | 3 | 1.0963 | 1.0504 | -0.0458 | 1/3 |