Frozen No-Prior Window RAPM
These are deliberately simple preseason controls. They fit ordinary zero-centered ridge RAPM on completed regular-season stints, with no lagged player prior, age, draft, box-score, exposure gate, or other player-profile input. They test whether the forward model’s gains come merely from retaining more historical possessions.
Specification
For each training stint, the model estimates one coefficient per player:
by minimizing possession-weighted squared error plus \(\lambda\sum_i\beta_i^2\). The player-prior mean is zero for every player. Each candidate selects \(\lambda\) through chronological folds that end before the 2025-26 target season, then refits on all of its completed training window.
| Candidate | Regular-season training window | Selected lambda | Cold starts |
|---|---|---|---|
| One-year no-prior RAPM | 2024-25 | 0.03 | 0.0 |
| Pooled three-year no-prior RAPM | 2022-23 through 2024-25 | 0.01 | 0.0 |
The three-year model does not have a formal player prior. Its stability comes from constraining a player to one coefficient across the entire three-season window. That reduces variance but can lag genuine improvement or decline.
Frozen 2025-26 Results
The realized 2025-26 lineup allocation is supplied as an oracle, while every player coefficient, lambda, home-court intercept, and scoring mean is fixed before opening night. The full comparison is maintained in the Frozen Preseason Leaderboard.
| Candidate | Regular possession RMSE | Regular game-margin RMSE | Team NetRtg RMSE | Pythagorean-wins RMSE |
|---|---|---|---|---|
| One-year no-prior RAPM | 1.199061 | 15.0694 | 4.9792 | 11.5828 |
| Pooled three-year no-prior RAPM | 1.198919 | 14.7820 | 4.6618 | 10.7999 |
| Forward lagged RAPM | 1.199000 | 14.8894 | 4.8538 | 10.7006 |
| Forward contextual RAPM | 1.199008 | 14.6525 | 4.1572 | 9.3153 |
The pooled model is clearly more stable than the one-year model. It improves on the lagged state in possession RMSE, game-margin RMSE, and team NetRtg, while the lagged state retains a small Pythagorean-wins advantage. Forward contextual RAPM remains the strongest game, team-NetRtg, and win forecast in this cohort.
Artifacts
The one-year coefficient/CV artifact is
frozen-one_year-rapm-2025-26-20260807T224659Z-420d7857 under
artifacts/models/frozen_one_year_rapm/2025-26/; its immutable evaluation is
frozen-one-year-rapm-2025-26-20260807T224708Z-b39c8bd3 under
artifacts/models/frozen_prior_evaluation/2025-26/.
The three-year coefficient/CV artifact is
frozen-three_year-rapm-2025-26-20260807T224747Z-bb4eb660 under
artifacts/models/frozen_three_year_rapm/2025-26/; its immutable evaluation is
frozen-three-year-rapm-2025-26-20260807T224756Z-e8e90b17 under
artifacts/models/frozen_prior_evaluation/2025-26/.
Each coefficient artifact includes the frozen player vector, chronological CV results, selected lambda, fitted home-court intercept, and a season-level training summary. Each evaluation artifact includes regular-season and playoff possession predictions, eligible-possession game margins, team NetRtg, and Pythagorean-wins tables.