Train Value-Conditioned Aging HPM

This HIPSTER PM candidate retains the centered aging prior and adds a learned, ridge-penalized interaction between target age and the player's completed prior-season RAPM:

\[ z_{i,t}=(\operatorname{age}_{i,t}-27)\operatorname{RAPM}_{i,t-1}. \]

The interaction permits the strictly forward aging model to learn whether players with different prior values age differently. It does not encode a special rule for stars: the coefficient is selected jointly with the existing aging features under expanding-season validation.

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

The standard possession-weighted prior centering remains in place, so the new feature changes relative player forecasts rather than only translating the coefficient origin.

Outputs

artifacts/models/forward_centered_value_conditioned_aging_bounded_hierarchical_portable_matchup_contextual_rapm/2025-26/

season_player_prior_metadata.parquet records the selected aging regularization and the enabled age_by_prior_rapm feature for every season.