Value-Conditioned Aging HPM

HIPSTER PM with Value-Conditioned Aging extends the centered aging portable contextual model with a regularized interaction between age and the completed prior-season player rating:

\[ \mu_{i,t}=f(\operatorname{age}_{i,t}, \ldots) + \beta_{\mathrm{value-age}} (\operatorname{age}_{i,t}-27)\operatorname{RAPM}_{i,t-1}. \]

The previous aging branch modeled the effect of age and prior RAPM additively. This specification can learn a different aging trajectory for a player with a high prior rating, but retains ridge shrinkage and strictly forward expanding validation. It remains a population estimate, not a hand-tuned exception for individual stars.

The full prior vector is then centered with prior-season possession weights and combined with the same exposure-gated cold-start branch and bounded hierarchical portable-matchup contextual term:

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

Rebuilt Frozen 2025-26 Evaluation

The historical raw-to-stint recovery rebuilt the one-season RAPM layer and the player-season panel before this strictly forward rerun: forward-centered-value-conditioned-aging-bounded-hierarchical-portable-matchup-contextual-rapm-2025-26-20260813T050955Z-c9b2e37d. The final-season aging branch again selected regularization 0.01 and recorded the value-conditioned feature in its season metadata.

Metric	Rebuilt Value-Conditioned Aging HPM
Regular possession RMSE	1.198736
Regular eligible game-margin RMSE	14.3214
Full-game margin RMSE	14.5898
Full-game winner accuracy	68.46%
Playoff possession RMSE	1.192465
Playoff eligible game-margin RMSE	16.4069
Team NetRtg RMSE	3.5858
Pythagorean wins RMSE	7.6300

This is the current primary lineup-prediction reference: it has the best regular possession and eligible game-margin performance among the models rerun on the recovered history. The frozen leaderboard will be regenerated only after its full candidate set is rebuilt on the same data contract.

See the training guide for the reproducible command and artifact contract.

Deterministic Reproducibility

Value-Conditioned Aging HPM has no random initialization, sampling, or split selection. A second seed is therefore not a meaningful perturbation. This audit compares a full rerun against the previous immutable artifact with absolute tolerance 1e-10.

Output	Reference rows	Rerun rows	Shared rows	Max absolute difference	All close
`cohort_metrics.parquet`	2	2	2	0.000e+00	yes
`game_predictions.parquet`	1,315	1,315	1,315	0.000e+00	yes
`historical_player_coefficients.parquet`	14,560	14,560	14,560	0.000e+00	yes
`team_net_rating_predictions.parquet`	30	30	30	0.000e+00	yes
`team_win_predictions.parquet`	30	30	30	0.000e+00	yes

Audit artifact: artifacts/analysis/value_conditioned_aging_reproducibility/2025-26/value-conditioned-aging-reproducibility-2025-26-20260810T174447Z-c3454967. It retains per-column numerical differences.