Forward Bounded Hierarchical P-spline Contextual RAPM

This forward contextual RAPM candidate retains the original relative context function, \(C(A,B)=g(x(A)-x(B))\), without portable-unit or matchup decomposition. For every completed season it clips each relative feature to its possession-weighted 5th--95th percentile interval before fitting and prediction. It then applies Ridge level shrinkage, a second-difference P-spline penalty, and a projected completed-prior-season function prior.

The completed 2024-25 bounded state is the only contextual information used for the frozen 2025-26 forecast. The 2025-26 state is retained only for a future forecast.

Frozen Result

The model records a regular eligible-game margin RMSE of 14.5485, a full-game margin RMSE of 14.8706, and a Pythagorean-win RMSE of 9.2772. The full comparison is maintained in the Frozen Preseason Leaderboard.

Artifact: forward-bounded-hierarchical-pspline-contextual-rapm-2025-26-20260809T152421Z-658f4071.