Methodology
Walk-forward backtest of Model C. We train on 2010–2020 (the earliest year nflverse's rich features become reliable) and hide recent seasons to grade the model. The numbers below are the model arguing with reality.
Why 2021, 2022, and 2023? Why not 2024 or 2025?
Model C predicts average PPR per game over the next 2 NFL seasons. To grade a prediction made from season S, we need to know what happened in seasons S+1 and S+2. The 2026 season hasn't been played yet, so 2024 and 2025 predictions can't be graded. 2023 is the most recent year for which the answer key exists.
| Test year | Target uses | Available? |
|---|---|---|
| 2021 | 2022 + 2023 | yes |
| 2022 | 2023 + 2024 | yes |
| 2023 | 2024 + 2025 | yes |
| 2024 | 2025 + 2026 | not yet |
| 2025 | 2026 + 2027 | not yet |
Strict no-leak: when grading test year T, the trainer never sees rows from S > T − 3 because their forward target depends on data from year T or later. So 2021 grades a model that only saw 2010 to 2018.