Season to date · 2026/27
Calibration
Raw accuracy is close to meaningless: it depends on which fixtures happened to be on the coupon, and it rewards always backing the favourite. The number that matters is Ranked Probability Score, tracked over time as a self-measure of forecast quality.
41/65Correct63%
0.2209Mean RPS65 scored
Reliability
Do the percentages mean anything?
Every forecast contributes three points, one per sign. If we are calibrated, predicted and observed should track each other down the table.
| Band | n | Predicted | Observed |
|---|---|---|---|
| 0–20% | 18 | 15.1% | 27.8% |
| 20–40% | 126 | 29.0% | 29.4% |
| 40–60% | 43 | 47.2% | 46.5% |
| 60–80% | 7 | 67.6% | 28.6% |
| 80–100% | 1 | 80.0% | 100.0% |
Draw watch
Predicted draws vs actual
Tracked separately because it is the failure this project has already walked into: a coupon that excluded X on four soft favourites in a round that produced five draws. A z-score inside ±2 is noise, not a bias worth correcting.
16.9Expectedover 65 matches
23Actual
1.72zwithin noise
Base rates
Outcome mix
How the forecast matches actually resolved this season — the base rate any single call is competing against.
35%Home wins
35%Draws
29%Away wins