Skip to content

NAIL-RAPM v1.3: Expanded Additive Profiles

NAIL-RAPM v1.3 tests four additional additive player-profile features while holding the v1.2 prior, cold-start, gap-returner, and five non-additive lineup terms fixed. It is a documented, non-promoted experiment; production remains NAIL-RAPM v1.2.

Feature Contract

The candidate retains v1.2's eight additive profile components and adds:

Feature Source Shrinkage contract
Defensive rebound percentage Historical player-season panel Existing stat-specific padding
Free-throw attempts per 100 possessions Historical player-season panel Existing published FTA padding
Unassisted rim makes per 100 Assisted shot profiles 480 pseudo-possessions toward the qualified-player median
Unassisted three-point makes per 100 Assisted shot profiles 480 pseudo-possessions toward the possession-weighted mean

The unassisted-shot centers and 480-possession padding constants were selected strictly through 2022-23 by predicting each qualifying player's next-season rate. The frozen 2023-24 through 2025-26 target seasons do not influence that selection.

Additive Weight Audit

The contextual Ridge model standardizes every home-minus-away five-man feature before fitting. A coefficient below is therefore the change in predicted net rating per 100 possessions for a one-standard-deviation change in that lineup differential, conditional on the other profile and non-additive terms. It is a conditional model weight, not a causal player value or a standalone stat ranking.

Each column below is a separate completed source model. For example, the 2022-23 column was used to forecast frozen 2023-24 outcomes; no target-season outcomes enter these weights. This panel is the relevant view for the frozen evaluation.

Additive feature 2022-23 source -> 2023-24 target 2023-24 source -> 2024-25 target 2024-25 source -> 2025-26 target
Assists / 100 +2.447 +2.459 +2.323
Three-point makes / 100 +1.103 +1.975 +2.223
Steals / 100 +1.225 +2.795 +1.724
Offensive-rebound claim total +0.276 +0.764 +1.552
Turnovers / 100 -1.782 -1.040 -1.512
Free-throw attempts / 100 +2.367 +1.455 +1.312
Blocks / 100 +0.642 +0.567 +0.734
Unassisted rim makes / 100 +0.460 +1.908 +0.678
Usage / 100 +0.145 -0.961 -0.342
Three-point attempts / 100 +0.492 +1.121 -0.280
Defensive rebound percentage -0.268 +0.403 +0.262
Unassisted three makes / 100 +0.590 +1.041 +0.246

The four bolded rows are v1.3 additions. FTA rate is consistently strong and positive. Unassisted rim makes are positive in all three source fits but vary in magnitude. Defensive rebound percentage changes sign, while unassisted three-point makes stay positive but have a smaller and declining weight. These patterns motivate ablation work before any future profile release.

Full Historical Trajectories

The same extraction is available for all 29 persisted source models, from 1997-98 through 2025-26. Green traces are inherited terms; orange traces are v1.3 additions. The annotation in each panel reports the dominant-sign share and mean absolute standardized weight over the complete history.

NAIL-RAPM v1.3 additive profile coefficient trajectories

Feature Dominant direction Same-sign seasons Mean absolute weight
Assists / 100 Positive 29 / 29 1.99
Three-point makes / 100 Positive 29 / 29 1.13
Steals / 100 Positive 28 / 29 1.21
Turnovers / 100 Negative 28 / 29 1.03
Blocks / 100 Positive 28 / 29 0.76
Free-throw attempts / 100 Positive 27 / 29 1.33
Offensive-rebound claim total Positive 26 / 29 0.77
Unassisted rim makes / 100 Positive 24 / 29 0.69
Three-point attempts / 100 Positive 23 / 29 0.67
Defensive rebound percentage Positive 23 / 29 0.45
Usage / 100 Positive 19 / 29 0.66
Unassisted three makes / 100 Positive 16 / 29 0.62

This makes the v1.3 diagnosis sharper. FTA rate is a credible persistent addition; unassisted rim finishing is promising but less stable; defensive rebound percentage is modest; and unassisted three-point make volume is too sign-unstable to justify retaining without a more focused experiment.

Frozen Results

Both models replay the same persisted recursive state into 2023-24, 2024-25, and 2025-26 with no target-season refit. v1.3 is effectively tied on pooled regular-season possession error, improves pooled playoff error slightly, but has worse regular full-game RMSE.

Model Regular poss. RMSE Regular full-game RMSE Winner accuracy Team NetRtg RMSE Pythagorean-win RMSE Playoff poss. RMSE Playoff eligible-game RMSE
NAIL-RAPM v1.2 1.197958 14.265989 68.16% 3.2908 7.0899 1.192734 16.6148
NAIL-RAPM v1.3 1.197987 14.298111 68.44% 3.3320 7.0619 1.192630 16.5034

Non-Promotion Gate

The primary metric is full-game margin RMSE. The candidate is eligible only if the upper endpoint of the paired 95% game-block bootstrap interval for

\[ \mathrm{RMSE}_{\mathrm{v1.3}} - \mathrm{RMSE}_{\mathrm{v1.2}} \]

is no greater than 0.5% of v1.2's RMSE in the pooled sample and in each frozen season. Positive values favor v1.2.

Scope v1.3 minus v1.2 Paired 95% interval Gate threshold Gate
Pooled +0.0321 [-0.0114, +0.0740] +0.0713 Fail
2023-24 +0.0171 [-0.0671, +0.1014] +0.0697 Fail
2024-25 +0.1026 [+0.0182, +0.1876] +0.0722 Fail
2025-26 -0.0200 [-0.0685, +0.0285] +0.0720 Pass

The 2024-25 result demonstrates a material loss under the agreed rule. Therefore, v1.3 is not promoted despite its small pooled playoff and Pythagorean-win improvements.

Artifacts

  • Candidate fit: artifacts/models/forward_nail_rapm_v13_additive_profiles/2025-26/forward-nail-rapm-v13-additive-profiles-2025-26-20260821T230733Z-def6e560
  • Frozen replay: artifacts/models/nail_v13_additive_profiles_frozen_backtest/frozen_multiseason_backtest/2023-24_to_2025-26/frozen_multiseason_backtest-2023-24-to-2025-26-20260821T232409Z-81841f20
  • Bootstrap gate: artifacts/models/nail_v13_additive_profiles_bootstrap/2023-24_to_2025-26/nail-v13-additive-profiles-bootstrap-20260821T232731Z-c13f1209
  • Weight audit: artifacts/models/analysis/nail_v13_additive_weight_audit/nail-v13-additive-weight-audit-20260822T002651Z-74535a75

Reproduction commands are in Train NAIL-RAPM v1.3 additive profiles.