Skip to content

NAIL-RAPM v1.2.1: Pruned Non-Additive Context

NAIL-RAPM v1.2.1 is a controlled ablation of the production v1.2 gap-returner model. It retains all eight additive profile totals, the value-conditioned aging prior, exposure-gated cold starts, gap-returner bridge, stat-specific profile padding, and Ridge regularization. It removes four of v1.2's six non-additive lineup-context terms.

Feature Contract

v1.2 non-additive term v1.2.1
Usage concentration Retained
Top-two assists Retained
Bottom-two three-point makes Removed
Credible shooter count Removed
Shooting-by-usage Removed
Shooter-by-passing Removed

The selection is based on the v1.2 coefficient audit: usage concentration has the same positive sign in every completed source season, while top-two assists is positive in 89.7%. The four removed terms are directionally unresolved over the historical panel. This is a feature-selection hypothesis, not a claim that the removed concepts are unimportant in basketball.

Retained-Term Trajectories

Each panel below is a standardized Ridge coefficient from one completed source season. It estimates the conditional change in a home-minus-away stint rating for a one-standard-deviation increase in that five-man feature, holding the eight additive profile totals and the other retained non-additive term fixed. These are lineup-context coefficients, not player ratings or causal effects.

NAIL-RAPM v1.2.1 retained non-additive coefficient trajectories

There are 29 fitted source-season states because the first historical season has no prior context state. usage_concentration is positive in all 29, with a median standardized coefficient of +0.91 and mean absolute coefficient of 0.89. top_two_assists is positive in 26 of 29 (89.7%), with median +0.59 and mean absolute coefficient 0.69. The chart also makes the remaining uncertainty visible: top-two assists has a wider historical range (-0.67 to +2.32), so its retention rests on strong directional consistency rather than a claim of a time-invariant effect size.

Additive Profile Trajectories

The eight additive player-profile totals are unchanged from v1.2 and remain eligible for exact player-level attribution. Their fitted weights are not, however, mechanically identical: the Ridge model refits all ten retained coordinates jointly after the four non-additive terms are removed. The chart therefore shows the v1.2.1 weights themselves rather than reusing v1.2's fourteen-term fit.

Across the 232 matched feature-season coefficients, the v1.2 and v1.2.1 additive weights have Pearson correlation 0.967 and mean absolute change 0.126 standardized units. The largest changes are in three-point attempts and makes, which is expected because the removed terms explicitly involved shooting. The remaining additive contract is therefore broadly stable, but not identical.

NAIL-RAPM v1.2.1 additive profile coefficient trajectories

Frozen Results

The recursive fit uses all 30 completed seasons. The strict three-season replay forecasts 2023-24, 2024-25, and 2025-26 from each season's preseason information set. v1.2 is the direct incumbent.

Model Poss. RMSE Eligible game RMSE Full-game RMSE Winner accuracy Team NetRtg RMSE Pythagorean-win RMSE Playoff eligible-game RMSE
NAIL-RAPM v1.2 1.197958 14.0414 14.2660 68.16% 3.2908 7.0899 16.6148
NAIL-RAPM v1.2.1 1.197952 14.0246 14.2521 68.24% 3.2706 7.0351 16.5942

The candidate lowers the pooled full-game RMSE by 0.0139 points. It also has slightly lower playoff possession RMSE (1.192709 versus 1.192734) and playoff eligible-game RMSE (16.5942 versus 16.6148). These are small changes, so the decision uses the predeclared paired game-block bootstrap rather than the point estimates alone.

Differences below are v1.2.1 minus v1.2: negative values favor the pruned candidate. The predefined non-promotion gate permits an upper 95% bound no larger than 0.5% of the incumbent full-game RMSE in every scope.

Scope Full-game RMSE difference Paired 95% interval Gate threshold Gate
Pooled -0.0139 [-0.0304, +0.0028] +0.0713 Pass
2023-24 +0.0149 [-0.0031, +0.0325] +0.0697 Pass
2024-25 -0.0593 [-0.1022, -0.0142] +0.0722 Pass
2025-26 +0.0024 [-0.0146, +0.0209] +0.0720 Pass

Decision

v1.2.1 clears the agreed non-promotion gate in the pooled replay and each target season. It is the new regular-season leader in the three-season frozen table while using four fewer non-additive terms. The website bundle remains on v1.2 until this candidate is explicitly promoted for deployment.

Artifacts

  • Recursive candidate: artifacts/models/forward_nail_rapm_v121_pruned_nonadditive/2025-26/forward-nail-rapm-v121-pruned-nonadditive-2025-26-20260822T171824Z-bd4da5c4
  • Frozen replay: artifacts/models/nail_v121_pruned_nonadditive_frozen_backtest/frozen_multiseason_backtest/2023-24_to_2025-26/frozen_multiseason_backtest-2023-24-to-2025-26-20260822T202752Z-d470f6d9
  • Paired bootstrap: artifacts/models/nail_v121_pruned_nonadditive_bootstrap/2023-24_to_2025-26/nail-v121-pruned-nonadditive-bootstrap-20260822T202828Z-d42bb340
  • Coefficient audit: artifacts/models/analysis/nail_v121_pruned_nonadditive_weight_audit/nail-v121-pruned-nonadditive-weight-audit-20260822T204159Z-0c9436cd

Reproduction

See Train NAIL-RAPM v1.2.1 Pruned Non-Additive Context.