Student-t Talent-Prior Contextual RAPM
This controlled ablation combines the current Forward Contextual RAPM state with the Student-t talent prior. It keeps Gaussian stint errors, the recursive exposure-gated cold-start policy, the published per-season lambda schedule, and the lagged contextual spline function unchanged.
Seasonal Transition
For season \(t\), the completed context function \(g_{t-1}\) is subtracted from the raw stint target before fitting player coefficients:
The player adjustment around the forward prior \(\mu_{i,t}\) receives a Student-t penalty instead of the Gaussian ridge penalty used by the base contextual model:
IRLS turns that into a player-specific local shrinkage multiplier. The fitted additive state then defines the raw residual used to fit \(g_t\), which is carried into season \(t+1\). Thus the model changes only the player-prior departure distribution; it does not alter cold starts, context features, or the information boundary.
Frozen 2025-26 Result
The run
artifacts/models/student_t_talent_contextual_rapm/2025-26/student-t-talent-contextual-rapm-2025-26-20260808T012017Z-7bbeadf3/
fits 1996-97 through 2025-26. Its frozen 2025-26 forecast uses the completed
2024-25 player state and \(g_{2024-25}\), then scores realized target-season
lineup exposure without a player refit.
| Cohort | Possession RMSE | Possession MAE | Eligible-game RMSE | Team NetRtg RMSE | Pythagorean win RMSE |
|---|---|---|---|---|---|
| Regular season | 1.199026 | 1.141556 | 14.6770 | 4.1771 | 9.3594 |
| Playoffs | 1.192959 | 1.135643 | 17.6146 | - | - |
It narrowly improves the base contextual model's regular-season possession MAE (1.141571 to 1.141556), but is worse on its primary full-game margin metric (15.0190 to 15.0487), team NetRtg RMSE, and Pythagorean win RMSE. It is therefore retained as a transparent combination test, not promoted over Forward Contextual RAPM.
Run it with Train Student-t Talent-Prior Contextual RAPM.
Outputs
The immutable artifact includes the standard frozen possession, eligible-game, full regular-game, team NetRtg, and Pythagorean-win outputs. It also retains the per-season player priors, exposure-gated replacement tokens, cold-start metadata, contextual models, and contextual fit metadata needed to reproduce the state transition.