NHL Model Accuracy
Out-of-sample backtests of the skater projection engine — projections vs. what actually happened, graded on a full-season (per-82-GP) basis against the Marcel baseline
Skater projections
Beat the Marcel baseline by 6% on out-of-sample accuracy · p10–p90 coverage 69% (2025 backtest, 2,400 preds)
Player Accuracy
per-game rate vs naive · 2,400 predsskill 6%p10–p90 cov 69%
| Stat | N | Model/g | Base/g | Skill | Cov |
|---|---|---|---|---|---|
| Scoring | |||||
| Points | 600 | 8.7 | 9.14 | +5% | 70% |
| Goals | 600 | 4.56 | 4.73 | +4% | 66% |
| Assists | 600 | 6.01 | 6.44 | +7% | 72% |
| Volume | |||||
| Shots | 600 | 21.01 | 23.07 | +9% | 67% |
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. Each skater's realized total is scaled to a full 82-GP season and compared to the projection, so games-played luck doesn't distort the grade.
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.