Skip to content

Train Era-Conditioned Aging HPM

This candidate retains the centered, value-conditioned aging HIPSTER PM prior and allows its smooth age function to evolve across NBA eras. For target season \(t\), the known season index is

\[ e_t = \frac{\operatorname{startYear}(t)-2010}{10}. \]

The aging regression includes an age-spline basis \(B(a)\) and the penalized interaction \(e_t B(a)\):

\[ \mu_{i,t} = B(a_{i,t})\gamma + e_tB(a_{i,t})\delta + \beta_r r_{i,t-1} + \beta_{ar}(a_{i,t}-27)r_{i,t-1} + \cdots. \]

All interaction coefficients are ridge-regularized. The era index is known before a season starts, so this adds no target-season leakage. The model does not introduce a three-way age-by-era-by-value interaction in this first pass.

uv run nba-train-forward-centered-era-conditioned-aging-bounded-hierarchical-portable-matchup-contextual-rapm \
  --through-season 2025-26

Outputs

artifacts/models/forward_centered_era_conditioned_aging_bounded_hierarchical_portable_matchup_contextual_rapm/2025-26/

In addition to the standard frozen evaluation tables, the run publishes the annual player rating, seasonal fit metadata, serialized aging pipelines, and population aging curve grid described in Forward model artifacts.