NBA Model Accuracy
Out-of-sample backtests of the per-game projection engine — projections vs. what actually happened, compared to the Marcel baseline
Per-game projections
Beat the Marcel baseline by 4% on per-game accuracy · p10–p90 coverage 66% (2025 backtest, 2,148 preds)
Player Accuracy
per-game rate vs naive · 2,148 predsskill 4%p10–p90 cov 66%
| Stat | N | Model/g | Base/g | Skill | Cov |
|---|---|---|---|---|---|
| Scoring | |||||
| Points | 358 | 2.53 | 2.68 | +6% | 64% |
| 3PM | 358 | 0.35 | 0.36 | +5% | 66% |
| Playmaking | |||||
| Assists | 358 | 0.7 | 0.73 | +5% | 63% |
| Defense & Boards | |||||
| Rebounds | 358 | 0.86 | 0.89 | +4% | 68% |
| Steals | 358 | 0.2 | 0.2 | -0% | 65% |
| Blocks | 358 | 0.14 | 0.15 | +5% | 67% |
How this is graded
methodology
Every completed season is projected walk-forward — the model trains only on seasons beforethe one it predicts, so nothing leaks. Projections are per-game averages, compared directly to each player's realized per-game line (min 25 games, so the rate is stable).
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.