Skip to content

NAIL-RAPM v1.2.1.1: Standard USG%

NAIL-RAPM v1.2.1.1 is a convention-correction patch to v1.2.1. It retains the same eight additive profile coordinates, two non-additive lineup terms, value-conditioned aging prior, exposure-gated cold starts, gap-returner bridge, profile padding, and Ridge estimators. It changes only the definition of usage.

Usage Contract

The prior release used a player action-event rate per RAPM on-court possession:

\[ 100 \cdot \frac{FGA_i + 0.44 FTA_i + TOV_i}{\text{on-court possessions}_i}. \]

v1.2.1.1 uses the conventional box-score usage percentage, computed at the game level and aggregated across a player-season:

\[ USG_i = 100 \cdot \frac{\sum_g (FGA_{ig} + 0.44 FTA_{ig} + TOV_{ig})(TM_g/5)} {\sum_g MIN_{ig}(FGA_{tg} + 0.44 FTA_{tg} + TOV_{tg})}. \]

Here, \(TM_g\) is the team’s game minutes, \(MIN_{ig}\) is the player’s game minutes, and the subscript \(t\) denotes the player’s team. This makes the denominator an estimate of the team’s offensive opportunities while the player was on the court. It puts familiar player values, such as a 31.8% season for Anthony Edwards in 2025-26, on the same scale used in standard NBA analysis.

The feature receives the existing 300 pseudo-opportunity profile shrinkage and replaces the legacy usage-event coordinate in both the additive profile and the non-additive usage_concentration term. No target-season information is used in frozen evaluation: each target season uses only its pre-season player profile state and prior-year fitted context state.

Frozen Results

The strict frozen replay forecasts 2023-24 through 2025-26 from each target season’s pre-season information set. Values below are pooled over the three regular seasons; lower is better except winner accuracy.

Model Poss. RMSE Poss. MAE Eligible game RMSE Full-game RMSE Winner accuracy Team NetRtg RMSE Pythagorean-win RMSE
v1.2.1 legacy usage rate 1.197952 1.141355 14.0246 14.2521 68.24% 3.2706 7.0351
v1.2.1.1 standard USG% 1.197951 1.141332 14.0264 14.2516 68.10% 3.2658 7.0279

The changes are deliberately small. The candidate improves full-game RMSE by 0.00055 and possession MAE by 0.000023, while eligible-game RMSE and winner accuracy are slightly worse. The predeclared paired game-block bootstrap is therefore the decision criterion rather than point-estimate cherry-picking.

Scope Full-game RMSE difference (v1.2.1.1 - v1.2.1) Paired 95% interval Gate threshold Gate
Pooled -0.00055 [-0.00743, +0.00608] +0.07126 Pass
2023-24 -0.01101 [-0.02278, +0.00067] +0.06976 Pass
2024-25 +0.00799 [-0.00558, +0.02171] +0.07189 Pass
2025-26 +0.00096 [-0.00834, +0.01040] +0.07205 Pass

Decision

v1.2.1.1 passes the non-promotion gate, but is not promoted. Conventional USG% is cleaner as a display convention, yet it does not materially change the Minnesota-style concentration interpretation. The candidate also gives up small point-estimate advantages in eligible-game RMSE and winner accuracy. v1.2.1 therefore remains the product model and source for the website bundle.

Artifacts

  • Recursive fit: artifacts/models/forward_nail_rapm_v1211_standard_usage/2025-26/forward-nail-rapm-v1211-standard-usage-2025-26-20260823T141226Z-ddb46a76
  • Frozen replay: artifacts/models/nail_v1211_standard_usage_frozen_backtest/frozen_multiseason_backtest/2023-24_to_2025-26/frozen_multiseason_backtest-2023-24-to-2025-26-20260823T142112Z-d5cea78f
  • Paired bootstrap: artifacts/models/nail_v1211_standard_usage_bootstrap/2023-24_to_2025-26/nail-v1211-standard-usage-bootstrap-20260823T142343Z-07ea4b70

Reproduction

See Train NAIL-RAPM v1.2.1.1 Standard USG%.