Skip to content

NAIL-RAPM Additive-Only Context Ablation

Last updated: 2026-08-17

This is the direct all-or-none test of NAIL-RAPM v1.0's six non-additive lineup coordinates. It keeps the recursive value-conditioned aging prior, exposure-gated cold starts, historical data, annual RAPM penalty schedule, and linear Ridge estimator fixed. The only difference is the context design matrix.

Contract

The ablation retains the eight player-additive basketball profile totals:

  • three-point attempts and makes per 100 possessions;
  • assists, turnovers, and usage events per 100 possessions;
  • offensive-rebound claim percentage; and
  • steals and blocks per 100 possessions.

It removes the non-additive lineup bundle: bottom-two shooting, credible shooter count, top-two assists, usage concentration, shooting-by-usage, and shooter-by-passing. Consequently, the contextual contribution is exactly

\[ C(H,A)=\sum_k \gamma_k\,[x_k(H)-x_k(A)], \]

where each \(x_k(U)\) is a sum of the five players' lagged profiles. This is not the earlier Additive Profile-Prior RAPM experiment: the eight coordinates remain in the lineup-level Ridge fit rather than being moved into next season's player prior.

Frozen Comparison

The completed replay compares this artifact directly to NAIL-RAPM v1.0 on 2023-24 through 2025-26. The player prior, annual lambda schedule, target cohorts, and scoring procedure are identical; the six removed context coordinates are the sole model difference.

Cohort Metric NAIL-RAPM v1.0 Additive-only context Additive-only minus NAIL
Regular season Possession RMSE 1.197977 1.197989 +0.000012
Regular season Possession MAE 1.141387 1.141430 +0.000043
Regular season Eligible game RMSE 14.1119 14.1286 +0.0166
Regular season Full-game RMSE 14.3517 14.3754 +0.0237
Regular season Winner accuracy 68.33% 67.93% -0.40 pp
Regular season Team NetRtg RMSE 3.4307 3.4783 +0.0477
Regular season Pythagorean-win RMSE 7.3460 7.4728 +0.1267
Playoffs Possession RMSE 1.192752 1.192713 -0.000039
Playoffs Possession MAE 1.137640 1.137590 -0.000050
Playoffs Eligible game RMSE 16.6636 16.5724 -0.0912

For regular-season uncertainty, a 10,000-draw paired bootstrap stratified by season compares additive-only minus NAIL. Full-game RMSE is directionally worse at +0.0237, but its 95% interval \([-0.0105, +0.0594]\) crosses zero. Possession MAE is reliably worse: +0.000042, 95% interval \([+0.000022, +0.000063]\), with only 0.01% of draws favoring the additive-only ablation.

The result is therefore nuanced: the bundle's aggregate regular-season signal is real but small at the game level. It is not evidence that every one of the six coordinates is independently identified or necessary; that requires a within-bundle ablation. The playoff point estimates favor additive-only, which is useful caution against treating the regular-season result as universal.

Artifact:

artifacts/models/analysis/nail_additive_only_context_frozen/
frozen_multiseason_backtest/2023-24_to_2025-26/
frozen_multiseason_backtest-2023-24-to-2025-26-20260818T012932Z-40442bd8