Skip to content

NAIL-RAPM v1.2.1.2: Back-to-Back Schedule Control

Last updated: 2026-08-24

This controlled candidate starts from the standard-USG% NAIL v1.2.1.1 profile contract and adds a single schedule adjustment. It does not alter the player prior, additive profile, or the two retained non-additive lineup features.

The Production NAIL Model Contract is the complete assumption register for this release. It distinguishes components with direct historical evidence from inherited or still-unvalidated numerical choices.

For each game \(g\), the schedule feature is

\[ x_g = \mathbb{1}[\text{home played the previous calendar day}] - \mathbb{1}[\text{away played the previous calendar day}]. \]

The completed source-season weighted Ridge coefficient is carried forward:

\[ \widehat{y}_{g,t+1} = \text{player edge} + \text{lineup edge} + \widehat{\beta}_{t}^{\mathrm{B2B}}x_g. \]

Before the player RAPM refit, the same source-season schedule term is removed from the current stint target. This prevents a repeatable calendar effect from being assigned to the players who happened to play under it.

The full catalog is used to derive the flag, so a failed or excluded possession file cannot make the following game appear more rested than it was. Target season outcomes remain outside the fit boundary; its calendar is allowed because it is available before tipoff.

Frozen Results

The strict three-season replay forecasts 2023-24 through 2025-26 using only the state available before each target season. The candidate is compared to its direct parent, v1.2.1.1, which keeps the same conventional-USG% profile contract but omits the schedule control.

Split Model Poss. RMSE Poss. MAE Eligible game RMSE Full-game RMSE Winner accuracy Team NetRtg RMSE
Regular v1.2.1.1 standard USG% 1.197951 1.141332 14.0264 14.2516 68.10% 3.2658
Regular v1.2.1.2 B2B control 1.197946 1.141313 14.0107 14.2330 67.96% 3.2847
Playoffs v1.2.1.1 standard USG% 1.192708 1.137593 16.5989 -- -- --
Playoffs v1.2.1.2 B2B control 1.192710 1.137604 16.6032 -- -- --

The B2B control wins the possession and regular-season margin metrics, with a small 0.0186-point improvement in pooled full-game RMSE. It loses the team net-rating, winner-accuracy, and pooled playoff comparisons. The comprehensive Three-Season Frozen Leaderboard remains the source of record for ranks across all candidates.

Bootstrap Gate

The paired bootstrap resamples games within each frozen season and compares full-game margin RMSE to the direct parent. The promotion guardrail allows at most 0.5% practical harm. A lower difference favors the candidate.

Scope Candidate minus parent 95% CI Better-draw probability Gate
Pooled -0.0186 [-0.0559, +0.0194] 83.33% Pass
2023-24 -0.0178 [-0.0608, +0.0255] -- Pass
2024-25 -0.0662 [-0.1126, -0.0209] -- Pass
2025-26 +0.0246 [-0.0661, +0.1141] -- Fail

The pooled estimate favors the B2B control but its confidence interval crosses zero, and the final frozen season fails the gate. It has weaker team and playoff metrics, which remain visible in the leaderboard. It is nevertheless promoted for the website because it has the best median regular-season rank and the best pooled full-game RMSE, while passing the direct pooled no-material-harm gate.

Annual Schedule Effect

The plotted coefficient is the source-season Ridge coefficient in home net-rating points for home back-to-back - away back-to-back. A value of -3.03 in 2024-25 means a home team on a B2B versus a rested visitor receives a -3.03 expected-margin adjustment before player and lineup terms; reversing the situation changes the sign. Every completed source season is negative.

Annual raw B2B schedule-control coefficient

The sign stability supports the basketball interpretation. The magnitude varies by season, including the compressed 2020-21 schedule, so the model retains the rolling one-season source-state rather than imposing a pooled constant.

Artifacts

  • Recursive fit: artifacts/models/forward_nail_rapm_v1212_back_to_back/2025-26/forward-nail-rapm-v1212-back-to-back-2025-26-20260824T140929Z-e99e9646
  • Frozen replay: artifacts/models/nail_v1212_back_to_back_frozen_backtest/frozen_multiseason_backtest/2023-24_to_2025-26/frozen_multiseason_backtest-2023-24-to-2025-26-20260824T141701Z-1b7f1e15
  • Bootstrap: artifacts/models/nail_v1212_back_to_back_bootstrap/2023-24_to_2025-26/nail-v1212-back-to-back-bootstrap-20260824T142245Z-1f66ebb7
  • Weight audit: artifacts/models/analysis/nail_v1212_back_to_back_weight_audit/nail-v1212-back-to-back-weight-audit-20260824T142248Z-cba6d361

See Schedule Controls and the training guide.