Forward Aging Bounded Hierarchical Portable-Matchup Contextual RAPM

This candidate tests whether a strictly forward age-informed transition improves the returning-player prior in the bounded hierarchical portable contextual model. It preserves the exposure-gated draft/replacement branch for first-NBA- season players and the existing portable context contract:

\[ C(A,B)=h(x(A))-h(x(B))+q(x(A),x(B)). \]

For a returning player (i) entering season (t), the prior is the output of an aging ridge model trained only on completed candidate-state transitions:

\[ \mu_{i,t}^{\mathrm{age}} = f_{<t}\left( r_{i,t-1},\; \log(1+p_{i,t-1}),\; a_{i,t},\; e_{i,t},\; d_i,\; b_i \right). \]

Here (r_{i,t-1}) is the prior completed candidate RAPM, (p_{i,t-1}) is its on-court possession exposure, (a_{i,t}) is known age, (e_{i,t}) is NBA experience, (d_i) is draft information, and (b_i) is the physical profile. The aging model uses the existing spline-age ridge specification and performs its own expanding historical regularization selection inside every recursive season.

Frozen 2025-26 Result

The completed through-2025-26 run uses a 0.1 aging-ridge regularization for the 2025-26 prior. It trains on 10,794 prior completed player transitions through 2024-25, then forecasts 462 returning players; the 100 first-year players continue through the exposure-gated cold-start branch.

Cohort	Possession RMSE	Eligible-game margin RMSE	Full-game margin RMSE
Regular season	1.198831	14.3654	14.6265
Playoffs	1.192544	16.5388	-

At the team level, the candidate records a NetRtg RMSE of 3.7737 and a Pythagorean-win RMSE of 7.7289. These are the best currently published frozen values for each of those metrics. The complete comparison is maintained in the Frozen Preseason Leaderboard.

Artifact: forward-aging-bounded-hierarchical-portable-matchup-contextual-rapm-2025-26-20260810T024224Z-a4e478fe.

Use the training guide to reproduce it.