Skip to content

Latest commit

 

History

History
63 lines (49 loc) · 2.36 KB

File metadata and controls

63 lines (49 loc) · 2.36 KB
last_updated 2026-08-11

Controlled No-Context Value-Conditioned RAPM

This model is the direct player-only control for Value-Conditioned Aging HPM. It preserves the same historical regular-season data, lambda schedule, possession-weighted prior centering, value-conditioned aging prior, and exposure-gated cold-start prior. It removes only the recursive contextual state:

[ C_{t-1}(U,V)=0. ]

Therefore it answers a narrow question: what does the HPM player-prior system predict when no lineup-composition or opponent-matchup adjustment can be allocated separately from player coefficients? Every season is retrained in sequence, so this is a genuine counterfactual state rather than a translation of an older uncentered RAPM artifact.

Frozen 2025-26 Result

Metric No-context control Value-Conditioned Aging HPM
Regular possession RMSE 1.198800 1.198763
Regular game-margin RMSE 14.4991 14.3469
Team NetRtg RMSE 4.0654 3.7568
Pythagorean-win RMSE 8.6225 8.0346
Playoff possession RMSE 1.192332 1.192460
Playoff game-margin RMSE 16.2971 16.4605

Context improves the primary regular-season game and team outcomes, while the control is marginally stronger on the small target-season playoff cohort. The result supports keeping context in HPM for preseason regular-season prediction, but makes its effect on player attribution directly inspectable.

Curry Diagnostic

The control materially changes Stephen Curry's 2015--19 player state:

Season No-context RAPM Value-Conditioned Aging HPM
2014-15 +7.49 +5.06
2015-16 +10.00 +6.43
2016-17 +9.54 +5.35
2017-18 +8.06 +4.01
2018-19 +7.91 +4.30

That difference is consistent with a substantial part of the Warriors-era unit environment being represented by the contextual state in HPM. It does not establish a unique redistribution of credit among teammates: lineup context is shared exposure, not a player-level causal attribution.

Artifact

artifacts/models/forward_centered_value_conditioned_aging_no_context_rapm/2025-26/

metadata.json records context_enabled: false, and all frozen-evaluation context corrections are exactly zero. The training guide contains the command and boundary contract.