Skip to content

NAIL Teammate-Continuity Replacement Candidate

Question

The first prior teammate-continuity candidate added continuity on top of production's two non-additive terms. Its continuity coefficient was much more directionally stable than the top_two_assists coefficient, while the fitted top-two-assists weight fell by about 0.33 standardized points on average. This experiment asks the narrower question:

Can prior teammate continuity replace top-two assists rather than compete with it?

Controlled Contract

Everything from NAIL-RAPM v1.2.1.3 remains fixed:

  • eight Medvedovsky-padded additive player-profile coordinates;
  • usage concentration as a non-additive lineup term;
  • back-to-back and home-court schedule controls;
  • residualized-target player-lambda cross-validation;
  • the same recursive aging, cold-start, centering, playoff-training, and regularization contracts.

The only substitution is

\[ \{\text{usage concentration},\ \text{top-two assists}\} \longrightarrow \{\text{usage concentration},\ \text{prior teammate continuity}\}. \]

For five-player unit \(U\), let \(c_{ij,t-1}\) be prior-regular-season shared possessions for teammate pair \((i,j)\). The continuity coordinate is

\[ \phi_{\text{continuity}}(U,t) =\frac{1}{10}\sum_{i<j}\log\left(1+c_{ij,t-1}\right). \]

Every one of the ten player pairs contributes. An unseen pair contributes zero. This is a prediction-only relationship feature; it is not an intrinsic, portable property of a hypothetical lineup.

Evaluation

The candidate is fit recursively through the full historical sequence, then evaluated against production on the standard frozen 2023-24, 2024-25, and 2025-26 regular seasons and their pooled playoffs. The final decision also uses a paired 10,000-draw game-block bootstrap and the complete historical coefficient trajectories.

Results

Frozen Comparison

Metric Production v1.2.1.3 Replacement Change
Regular possession RMSE 1.198147 1.198149 +0.000002
Regular possession MAE 1.141455 1.141537 +0.000082
Eligible game-margin RMSE 14.0051 14.0031 -0.0021
Full-game margin RMSE 14.2166 14.2119 -0.0047
Winner accuracy 68.30% 68.41% +0.11 pp
Team NetRtg RMSE 3.2351 3.2853 +0.0502
Pythagorean win RMSE 6.9423 6.9815 +0.0393
Playoff possession RMSE 1.192728 1.192657 -0.000071
Playoff game-margin RMSE 16.5782 16.4638 -0.1144

The paired 10,000-draw game-block bootstrap estimates a pooled full-game RMSE change of -0.0047, with a 95% interval of [-0.0364, +0.0272] and a 61.7% probability that the replacement is better. The interval includes zero by a wide margin. Possession MAE is worse in every bootstrap draw, although the absolute difference is very small.

Coefficient Stability

Continuity-replacement standardized weights

Feature Median standardized weight Positive seasons Positive mass SD
Usage concentration +1.06 29 / 29 100.000% 0.43
Prior teammate continuity +1.12 28 / 29 99.999% 0.52

Usage concentration remains close to its production trajectory (r = 0.963), with mean absolute movement of 0.097 standardized points. The replacement therefore behaves as intended: continuity occupies the second context coordinate without destabilizing the primary non-additive term.

Decision

Do not promote this candidate. It demonstrates that the historically weak top_two_assists term can be removed without materially damaging game-margin prediction. It does not demonstrate that continuity is a better production coordinate: possession MAE and team NetRtg error worsen, the primary bootstrap interval includes zero, and prior teammate continuity cannot score new or hypothetical teammate combinations as an intrinsic lineup property.

The result is still useful. It confirms that the stable continuity signal is largely interchangeable with existing recursive state and shared-playmaking context rather than a missing independent source of forecast value.

Artifacts

  • Recursive fit: artifacts/models/forward_nail_rapm_teammate_continuity_replacement/2025-26/forward-nail-rapm-teammate-continuity-replacement-2025-26-20260830T152259Z-2e6dd7d8
  • Frozen replay: artifacts/models/nail_teammate_continuity_replacement_frozen_backtest/frozen_multiseason_backtest/2023-24_to_2025-26/frozen_multiseason_backtest-2023-24-to-2025-26-20260830T153016Z-24f97568
  • Coefficient audit: artifacts/models/analysis/nail_teammate_continuity_replacement_weight_audit/nail-teammate-continuity-replacement-weight-audit-20260830T153056Z-bfa0f6d2
  • Paired bootstrap: artifacts/models/nail_teammate_continuity_replacement_bootstrap/2023-24_to_2025-26/nail-teammate-continuity-replacement-bootstrap-20260830T153059Z-1db6feb0

Reproduction

See Train the NAIL Teammate-Continuity Replacement Candidate.