Model Accuracy and Backtest Scorecard
Public 2026 out-of-sample check for Gridiron Analytics. The season projection is the stored predicted PPG held flat each week. Rolling xFP uses up to three prior games of expected points. The baseline is last week's actual PPR. Lower MAE is better.
Model MAE
5.10
Baseline MAE
5.58
Within 3 PPG
37.3%
Top-24 hit rate
41.7%
Positional MAE: model vs baselines
Completed weeks: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11.
QB, RB, WR, TE
| Pos | Model MAE | Model RMSE | Rolling xFP MAE | Last week MAE | Within 3 | Top 24 | n |
|---|---|---|---|---|---|---|---|
| QB | 7.07 | 8.70 | 7.03 | 7.75 | 24.8% | 82.2% | 278 |
| RB | 4.99 | 6.75 | 4.42 | 5.29 | 40.4% | 65.9% | 658 |
| WR | 5.14 | 6.61 | 4.79 | 5.71 | 33.6% | 54.0% | 989 |
| TE | 4.18 | 5.47 | 3.90 | 4.72 | 46.6% | 63.6% | 554 |
Stored training metadata
- fantasy_points_predictor (GradientBoostingRegressor): MAE 3.41, RMSE 5.19, R2 0.570
Where the model is useful, and where it is not
The model is strongest when a player's role is stable: snaps, targets, and carries repeat, so a season PPG and a short rolling xFP both beat a one-week memory. It is weak on the Friday news the box score does not contain yet. An Out or Questionable tag, and a Vegas implied team total, are the supplements. A low implied total or a dome-versus-outdoors spot will not show up inside the MAE table. Read those on the player page and the projections board.