Skip to content

NAIL Critical-Spacing Candidate

Status: experimental. This is a controlled one-feature extension of NAIL-RAPM v1.2.1, not a released model.

Hypothesis

A lineup can contain a large aggregate shooting total while still having multiple players whom a defense can help off. critical_spacing tests the discrete version of that idea: a unit may become materially harder to space once it has at least two low-threat shooters. This is structurally different from summing player shooting rates, which is additive and already belongs in the player-attributable profile layer.

For season-state profile pool \(P_t\), define its lower-tercile shrunk shooting cutoff as

\[ q_t = Q_{1/3}\left(\{\mathrm{3PM100}_{i,t}:i\in P_t\}\right). \]

For a unit \(U\), the feature is

\[ \mathrm{CriticalSpacing}_t(U)= \mathbb{1}\!\left[\sum_{i\in U} \mathbb{1}[\mathrm{3PM100}_{i,t}<q_t]\ge2\right]. \]

The strict inequality prevents players exactly at the seasonal cutoff from being classified as low threat. 3PM100 is the existing shrunk player-profile rate, and both the profile and threshold come only from information available before the target season. The lineup-context Ridge fit then estimates the coefficient jointly with the eight additive profile totals, top_two_assists, and usage_concentration.

Why This Differs From the Retired Shooter Count

The earlier credible_shooter_count supplied a broad linear count with a fixed absolute cutoff. It was not directionally stable enough to retain in v1.2.1. Critical Spacing instead tests one narrowly specified threshold event: two or more players below an era-specific, shrunk shooting baseline. It may still fail. The separation is intentional so the frozen comparison answers a single, interpretable question.

Evaluation Plan

The candidate is fit recursively across the historical panel and evaluated on the frozen 2023-24, 2024-25, and 2025-26 regular seasons against v1.2.1. The same paired game-block bootstrap non-promotion gate applies: the upper 95% bound for full-game RMSE harm must not exceed 0.5% of the incumbent RMSE in the pooled result or any target season.

Frozen Results

The recursive fit completed through 2025-26 and the strict frozen replay forecast 2023-24 through 2025-26 using each target season's preseason information set. The candidate is evaluated directly against v1.2.1.

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 1.197952 14.024550 14.252137 68.24% 3.270613 7.035098 16.594200
Critical Spacing candidate 1.197954 14.024052 14.253192 68.47% 3.280515 7.060052 16.589610

The candidate changes pooled full-game RMSE by +0.0011 (candidate minus v1.2.1), effectively no aggregate movement. Its point estimates improve in 2023-24 and 2024-25 but worsen by +0.0176 in 2025-26. The pooled paired 10,000-draw game-block bootstrap interval is [-0.0059, +0.0080]; it passes the predeclared no-material-harm threshold of +0.0713, but does not establish a meaningful advantage.

Scope Full-game RMSE difference Paired 95% interval Gate
Pooled +0.0011 [-0.0059, +0.0080] Pass
2023-24 -0.0078 [-0.0168, +0.0014] Pass
2024-25 -0.0083 [-0.0157, -0.0008] Pass
2025-26 +0.0176 [+0.0003, +0.0348] Pass

Non-Additive Coefficient Audit

The chart reports the full three-term non-additive contract from the candidate fit: the two retained v1.2.1 terms plus Critical Spacing. This is required to inspect possible coefficient displacement rather than assessing the new term in isolation.

Across 29 completed source-season states, the standardized critical-spacing coefficient is negative in 19 seasons and positive in 10. The agreed stability measure is total one-sided mass, not binary sign count: the negative side accounts for 9.38 coefficient-mass versus 2.36 on the positive side, or a 79.9% negative directional-mass share. Its median is -0.20, mean absolute magnitude is 0.40, and range is -1.55 to +0.59. The chart shows a broadly downward historical path, with the large-magnitude observations concentrated on the expected negative side. The small positive crossings are real but do not overturn the directional-mass evidence.

Critical Spacing non-additive coefficient trajectories by source season

The candidate does not materially displace either incumbent term. Matched source-season weights for usage_concentration have Pearson correlation 0.9885 with v1.2.1 and mean absolute change 0.0449 standardized units; top_two_assists has correlation 0.9868 and mean absolute change 0.0850. Critical Spacing therefore adds a mostly independent third non-additive signal rather than substituting for either retained feature.

Decision

Do not yet promote the candidate. It clears the non-promotion gate and the directional-mass audit supports a negative Critical Spacing effect. The reason to hold the production model at v1.2.1 is narrower: there is no meaningful pooled full-game improvement (+0.0011 RMSE). This candidate remains a directionally credible feature rather than a rejected one.

Artifacts

  • Recursive candidate: artifacts/models/forward_nail_rapm_critical_spacing/2025-26/forward-nail-critical-spacing-2025-26-20260823T223333Z-f9430d81
  • Frozen replay: artifacts/models/nail_critical_spacing_frozen_backtest/frozen_multiseason_backtest/2023-24_to_2025-26/frozen_multiseason_backtest-2023-24-to-2025-26-20260823T224105Z-a201166a
  • Paired bootstrap: artifacts/models/nail_critical_spacing_bootstrap/2023-24_to_2025-26/nail-critical-spacing-bootstrap-20260823T224136Z-c1b037d5
  • Coefficient audit: artifacts/models/analysis/nail_critical_spacing_weight_audit/

Reproduction

See Train the NAIL Critical-Spacing Candidate.