Train Forward Hierarchical P-spline Contextual RAPM
MPLCONFIGDIR=/private/tmp uv run nba-train-forward-hierarchical-pspline-contextual-rapm
This experiment uses the same 1996-97 through 2025-26 forward player-prior recursion and portable-matchup context contract as the published contextual model. Its contextual state adds a P-spline curvature penalty and a prior season function prior. The frozen 2025-26 evaluation is still scored only with the completed 2024-25 state.
The default penalties are:
| Penalty | Value | Role |
|---|---|---|
| Ridge level | 10,000 | Shrinks every spline coefficient toward zero. |
| P-spline curvature | 1,000 | Penalizes second differences of each feature's unscaled B-spline coefficients. |
| Temporal hierarchy | 10,000 | Pulls each completed season toward the prior completed season after response-function projection onto the current basis. |
Override an individual value when running a sensitivity experiment:
MPLCONFIGDIR=/private/tmp uv run nba-train-forward-hierarchical-pspline-contextual-rapm \
--context-curvature-alpha 1000 \
--context-temporal-alpha 10000
The immutable artifact is written below
artifacts/models/forward_hierarchical_pspline_contextual_rapm/2025-26/.
After a successful run, refresh the shared full-game table:
MPLCONFIGDIR=/private/tmp uv run nba-report-frozen-game-outcomes --season 2025-26
See Forward Hierarchical P-spline Contextual RAPM for the statistical construction and frozen boundary.