WEEKLY PROJECTION SCORECARD · FULL PPR
32,664 forecasts. 6 seasons. No peeking.
Every season stayed hidden until its model was trained. Every score below uses the same eligible players as the trailing-points baseline.
2020–2025 · QB, RB, WR, TEBY POSITION
The model beat the baseline everywhere.
MAE is the average absolute miss in PPR points. Lower is better. Startable overlap asks how often projected starters matched the week's actual top finishers.
| Position | Forecasts | Model MAE | Trailing MAE | MAE improvement | Rank correlation | Startable overlap | 80% coverage |
|---|---|---|---|---|---|---|---|
| QB | 3,470 | 6.57 | 7.19 | +8.6% | 0.54 | 53.2% | 78.4% |
| RB | 8,534 | 4.70 | 5.05 | +7.0% | 0.62 | 58.9% | 80.3% |
| WR | 13,711 | 4.65 | 4.96 | +6.3% | 0.60 | 55.7% | 80.6% |
| TE | 6,949 | 3.62 | 3.98 | +9.1% | 0.55 | 44.9% | 80.5% |
BY HELD-OUT SEASON
6 chances to fail. 6 wins.
The model never trains on its test season. This prevents a good result from being manufactured by mixing future outcomes into the inputs.
| Season | Forecasts | Model MAE | Trailing MAE | MAE improvement | Rank correlation | 80% coverage |
|---|---|---|---|---|---|---|
| 2020 | 5,162 | 4.93 | 5.35 | +8.0% | 0.56 | 78.7% |
| 2021 | 5,511 | 4.83 | 5.26 | +8.1% | 0.55 | 78.9% |
| 2022 | 5,446 | 4.60 | 5.00 | +7.9% | 0.58 | 81.0% |
| 2023 | 5,430 | 4.50 | 4.82 | +6.7% | 0.60 | 81.7% |
| 2024 | 5,451 | 4.60 | 4.90 | +6.0% | 0.59 | 79.8% |
| 2025 | 5,664 | 4.44 | 4.78 | +7.0% | 0.59 | 81.3% |
RESEARCH FILE
What this result does—and doesn't—prove.
Design. Rolling future-season evaluation. Each position model trains only on seasons before its held-out season; every player-week feature uses completed prior weeks.
Comparison. Trailing four-appearance mean PPR points on the identical eligible player-weeks.
Reproducibility. Two complete executions produced the same fa913128ea2a… forecast hash.
Limits and source files
- Historical injury, depth-chart, and news states are not available in this public-data benchmark.
- The player pool is reconstructed from prior appearances and may lag transactions or role changes.
- Playing probability is a simple trailing-appearance estimate, not an injury model.
- Intervals are position-specific training-residual bands rather than player-specific simulations.
8 nflverse weekly files, 2018–2025. CC BY 4.0 (nflverse compilation; underlying NFL data retain owner terms).