HPM x2: Raw OREB/100 Context

Last updated: 2026-08-14

HPM x2 is a narrow rebounding context experiment. It retains the same centered, value-conditioned aging prior and exposure-gated cold-start procedure as HPM x1, but substitutes the original raw offensive-rebounding-rate ingredient:

\[ \operatorname{OREB100}(U) = \sum_{i \in U}\operatorname{OREB100}_i. \]

This is a sum of forward-safe individual player offensive rebounds per 100 possessions. It is deliberately not transformed, capped, or interacted with usage. Therefore x2 tests whether raw summed OREB/100 itself is a useful lone contextual feature, relative to x1's summed ORB% player claims.

Run

uv run nba-train-hpm-x2 --through-season 2025-26

The recursive artifact is written under artifacts/models/forward_hpm_x2_orb_per_100_total/, then replayed on the frozen 2023-24 through 2025-26 regular-season and playoff targets.

Frozen Result

The completed run is artifacts/models/forward_hpm_x2_orb_per_100_total/2025-26/forward-hpm-x2-orb-per-100-total-2025-26-20260815T041121Z-55b9ac34. Across the three frozen regular seasons, x2 recorded a 14.4641 full-game RMSE, 1.198048 possession RMSE, 67.79% winner accuracy, and 3.6558 team NetRtg RMSE. Its pooled playoff checks were 1.192760 possession RMSE and 16.6382 eligible-game RMSE.

The appropriate control is the recovered-coverage Complete Player-Prior RAPM Baseline, which preserves the same value-conditioned aging and exposure-gated cold-start prior while disabling all lineup context and box-score terms.

Metric	No-context control	HPM x2	x2 minus control	Paired 95% interval	P(x2 better)
Full-game RMSE	14.475623	14.464065	-0.011558	[-0.029468, +0.006483]	89.68%
Winner accuracy	67.64%	67.79%	+0.14 pp	[-0.37 pp, +0.66 pp]	68.46%
Possession RMSE	1.198061	1.198048	-0.000013	[-0.000023, -0.000003]	99.35%
Possession MAE	1.141675	1.141660	-0.000015	[-0.000026, -0.000005]	99.80%

Thus raw summed OREB/100 improves the possession-level forecast relative to the proper no-context control, but its full-game improvement is not yet precise enough to satisfy the promotion rule. Against HPM x1, the two single-feature variants remain effectively tied on full-game RMSE: x2's -0.0024 difference has a paired 95% interval of [-0.0084, +0.0036].