NHL Model Accuracy
Out-of-sample backtests of the goalie projection engine — projections vs. what actually happened, graded per stat against the Marcel baseline — rate skills (SV%/GAA) directly, counting stats over a full starter workload
Goalie projections
Beat the Marcel baseline by 7% on out-of-sample accuracy · p10–p90 coverage 54% (2025 backtest, 248 preds)
Player Accuracy
per-game rate vs naive · 248 predsskill 7%p10–p90 cov 54%
| Stat | N | Model/g | Base/g | Skill | Cov |
|---|---|---|---|---|---|
| Rate | |||||
| Save % | 62 | 0.0122 | 0.0132 | +8% | 52% |
| GAA | 62 | 0.3031 | 0.3265 | +7% | 58% |
| Volume | |||||
| Saves | 62 | 111.3248 | 123.144 | +10% | 48% |
| Shutouts | 62 | 1.914 | 1.9582 | +2% | 56% |
How this is graded
methodology
Every completed season is projected walk-forward — the model trains only on seasons before the one it predicts, so nothing leaks. Rate skills (SV%, GAA) are graded directly; counting stats (saves, shutouts) scale the realized total to a full starter workload, so games-played luck doesn't distort them.
Skill= how much lower the model's error is than the Marcel baseline (3-year weighted rate + age curve, the standard projection floor). Coverage is how often the real result landed inside the p10–p90 band — ~80% is ideal calibration. Goalie bands run tighter than ideal — a known calibration lever.