Skip to content

hpm-x3.md

Latest commit

 

History

History
59 lines (45 loc) · 2.27 KB

hpm-x3.md

File metadata and controls

59 lines (45 loc) · 2.27 KB
Raw
title HPM x3

HPM x3: ORB Claim Rebound Replacement

Last updated: 2026-08-14

HPM x3 is a conditional replacement experiment. It keeps the complete HPM v1 context contract, but replaces its entire rebounding bundle:

  • offensive_rebounds_per_100
  • defensive_rebounds_per_100
  • sqrt_offensive_rebounds
  • sqrt_defensive_rebounds
  • rebounding_usage_interaction

with one forward-safe player-profile feature:

[ \operatorname{ORBClaims}(U)=\sum_{i \in U}\operatorname{ORB%}_i. ]

All shooting, creation, usage, turnover, defensive-event, uncertainty, and other HPM v1 context features remain. This directly tests whether summed ORB% claims are a better rebounding representation when the rest of the incumbent context model is held intact.

Run

uv run nba-train-hpm-x3 --through-season 2025-26

The recursive artifact is written under artifacts/models/forward_hpm_x3_v1_orb_claim_replacement/, then replayed on the frozen 2023-24 through 2025-26 targets against HPM v1.

Frozen Result

Artifact: artifacts/models/forward_hpm_x3_v1_orb_claim_replacement/2025-26/forward-hpm-x3-v1-orb-claim-replacement-2025-26-20260815T055927Z-d20c526b.

Metric HPM v1 HPM x3 x3 minus v1 Paired 95% interval P(x3 better)
Full-game RMSE 14.380249 14.377365 -0.002884 [-0.016831, +0.011267] 65.17%
Winner accuracy 67.84% 67.73% -0.11 pp [-0.54 pp, +0.31 pp] 28.37%
Possession RMSE 1.198010 1.198000 -0.000011 [-0.000020, -0.000001] 98.75%
Possession MAE 1.141402 1.141400 -0.000002 [-0.000012, +0.000008] 66.12%
Team NetRtg RMSE 3.429015 3.422924 -0.006091 -- --
Pythagorean-win RMSE 7.353235 7.326897 -0.026338 -- --

x3 is the best point estimate on regular possession RMSE, eligible-game RMSE, full-game RMSE, team NetRtg RMSE, and Pythagorean-win RMSE among the current HPM candidates. Its full-game confidence interval still crosses zero, so it is not statistically established as superior to HPM v1. It is nevertheless the published production reference under the documented parsimony non-inferiority rule: it removes the coherent five-term rebounding bundle while its +0.0113 upper full-game-RMSE interval remains below the +0.015 practical margin.