Calibration over 871 decided games across 2023, 2024, 2025, 2026. Each row groups games by the home win probability that was forecast; a well-calibrated forecaster's "actual" column matches its "forecast" column. Brier score is mean squared error, lower is better.
scripts/tune.ts.| Forecast bin | Elo games | Forecast | Actual | Δ | Spread games | Forecast | Actual | Δ |
|---|---|---|---|---|---|---|---|---|
| 0%–10% | 0 | 0 | ||||||
| 10%–20% | 11 | 17% | 9% | · | 9 | 16% | 11% | · |
| 20%–30% | 38 | 26% | 34% | +8 | 36 | 26% | 22% | -4 |
| 30%–40% | 96 | 36% | 29% | -7 | 132 | 36% | 29% | -7 |
| 40%–50% | 159 | 45% | 39% | -6 | 170 | 43% | 41% | -2 |
| 50%–60% | 192 | 55% | 52% | -3 | 177 | 57% | 60% | +3 |
| 60%–70% | 178 | 65% | 69% | +5 | 214 | 65% | 71% | +6 |
| 70%–80% | 143 | 75% | 71% | -3 | 86 | 74% | 71% | -3 |
| 80%–90% | 51 | 84% | 94% | +10 | 45 | 84% | 93% | +9 |
| 90%–100% | 3 | 91% | 100% | · | 2 | 91% | 100% | · |
Δ is actual minus forecast in points of probability; a negative Δ in the high bins means overconfidence in favorites. Bins under 20 games show a dot.