| title | HPM x1 |
|---|
Last updated: 2026-08-14
HPM x1 is a deliberately narrow context experiment. It retains the complete value-conditioned aging RAPM prior and exposure-gated cold-start procedure, but its contextual model receives exactly one side feature:
[ \operatorname{ORBClaims}(U) = \sum_{i \in U}\operatorname{ORB%}_i. ]
The individual ORB% values are forward-safe player profile estimates from
the prior season. The sum is measured in percentage points and is not a
predicted team offensive-rebound percentage: overlapping player claims mean it
is a claim total, not a realized rate.
No shooting, passing, usage, defensive-event, defensive-rebounding, square-root-rebound, or rebound-by-usage context feature enters HPM x1. The bounded hierarchical P-spline fit learns the response to the relative home-minus-away ORB claim total with the same forward seasonal hierarchy as HPM v1.
uv run nba-train-hpm-x1 --through-season 2025-26The recursive artifact is written under
artifacts/models/forward_hpm_x1_orb_claim_total/. It will be evaluated on
the frozen 2023-24, 2024-25, and 2025-26 regular-season holdouts before any
promotion decision.
The relevant ablation comparison is the Controlled No-Context Value-Conditioned RAPM, not the broader Complete Player-Prior RAPM baseline. That control is being rebuilt from the corrected stint data and will be replayed on the same frozen 2023-24 through 2025-26 targets before a standalone-context conclusion is drawn.
Original HPM v1 remains materially stronger on the primary full-game and team net-rating metrics: 14.3802 versus 14.4665 full-game RMSE and 3.4290 versus 3.6653 team NetRtg RMSE. See the Frozen Model Tournament for the broader model comparison.