Skip to content

NAIL-RAPM v1.2.1.3 Residualized-Target Lambda CV

This release changes exactly one modeling decision from v1.2.1.2: the player-RAPM Ridge penalty is selected from the completed source season's residualized training target rather than imported from a separate forward RAPM run.

The rest of the contract is unchanged: reconstructed regular-season stints, forward value-conditioned aging and gap-returner priors, exposure-gated cold starts, Medvedovsky-style additive-profile padding, the retained two-feature non-additive context bundle, linear antisymmetric context Ridge with alpha 10,000, and the forward B2B control with alpha 10,000.

Lambda Selection

For a forecast target season \(t\), let \(s=t-1\) be its completed source season. In season \(s\), the context and B2B forecasts from \(s-1\) are first removed from the source-season training target. Three chronological folds of that residualized \(s\) data then select the player penalty:

\[ \lambda_s \in \{10^{-5}, 3\!\times\!10^{-5}, 10^{-4}, 3\!\times\!10^{-4}, 10^{-3}, 3\!\times\!10^{-3}, 10^{-2}, 0.03, 0.1, 0.3, 1, 3, 10\}. \]

For example, the frozen 2025-26 forecast uses \(\lambda_{2024\text{-}25}\), selected only from chronological folds within 2024-25. The fit performed on 2025-26 after that season ends is retained for the next state update and is not used to score the 2025-26 frozen cohort.

No fitted lambda schedule, state, or hyperparameter is imported from another model artifact. The only fixed profile-shrinkage constants are the accepted Medvedovsky-style padding contract shared with v1.2.1.2.

Frozen Result

The immutable comparison artifact is artifacts/models/nail_v1212_residualized_lambda_frozen_backtest/frozen_multiseason_backtest/2023-24_to_2025-26/frozen_multiseason_backtest-2023-24-to-2025-26-20260827T221820Z-1a0cc90d.

Metric v1.2.1.2 imported schedule v1.2.1.3 residualized CV
Regular full-game RMSE 14.2330 14.2166
Regular winner accuracy 67.96% 68.30%
Team net-rating RMSE 3.2847 3.2351
Pythagorean-win RMSE 7.0551 6.9423
Playoff eligible game-margin RMSE 16.6032 16.5784

It cleared the frozen metric comparison against v1.2.1.2 and passed review of the player rankings and coefficient histories. It is the promoted production release. The frozen leaderboard contains the full metric table and ranks.

Promotion Review

The following review is calculated from the completed 2025-26 candidate fit. Its player rating is the fitted player RAPM plus the centered, exactly compilable additive-profile adjustment. It is not a frozen 2026-27 forecast.

Rank Player Team Pos. NAIL Prior Season update Additive profile Poss.
1 Victor Wembanyama SAS F-C +10.26 +2.50 +4.79 +2.96 3,896
2 Nikola Jokic DEN C +9.70 +4.36 +1.39 +3.95 4,786
3 Shai Gilgeous-Alexander OKC G +9.18 +5.18 +1.28 +2.72 4,730
4 Kawhi Leonard LAC F +7.27 +2.89 +3.27 +1.11 4,167
5 Derrick White BOS G +6.68 +2.04 +3.30 +1.34 5,180
6 Bam Adebayo MIA C-F +6.28 +2.61 +2.84 +0.82 5,031
7 Alex Caruso OKC G +6.07 +1.01 +1.83 +3.24 2,125
8 Dyson Daniels ATL G +6.01 +0.29 +2.78 +2.93 5,317
9 Jimmy Butler III GSW F +6.00 +1.80 +2.41 +1.79 2,449
10 Chet Holmgren OKC C-F +5.96 +3.03 +2.63 +0.30 4,129
11 Giannis Antetokounmpo MIL F +5.78 +4.03 +1.15 +0.61 2,129
12 Jayson Tatum BOS F-G +5.30 +4.39 -0.20 +1.10 1,046
13 Lauri Markkanen UTA F-C +5.08 +2.90 +2.07 +0.10 3,080
14 Donovan Mitchell CLE G +4.93 +1.57 +0.94 +2.42 4,925
15 LaMelo Ball CHA G +4.78 +0.66 +1.41 +2.71 4,101
16 Ausar Thompson DET G-F +4.76 +0.93 +1.52 +2.31 3,897
17 OG Anunoby NYK F-G +4.76 +2.56 +1.63 +0.57 4,531
18 Luka Doncic LAL F-G +4.69 +2.46 -0.71 +2.93 4,759
19 Devin Booker PHX G +4.57 +2.19 +2.11 +0.28 4,442
20 Franz Wagner ORL F +4.49 +4.26 -0.81 +1.04 2,190
21 Steven Adams HOU C +4.35 +1.34 +1.43 +1.58 1,455
22 Collin Gillespie PHX G +4.19 -0.80 +2.85 +2.13 4,626
23 Jalen Smith CHI F-C +4.15 +0.65 +2.20 +1.31 2,316
24 Day'Ron Sharpe BKN C +4.12 +0.84 +0.22 +3.05 2,330
25 Karl-Anthony Towns NYK C-F +4.04 +3.04 +0.29 +0.71 4,716

Home Court And Back-To-Back

Home-court advantage is a season-specific home-net-rating intercept. The B2B series is the raw coefficient for home back-to-back - away back-to-back; a negative value means a home B2B lowers the predicted home edge.

NAIL-RAPM v1.2.1.3 home-court and B2B coefficient trajectories

Context Coefficients

Context weights are standardized Ridge coefficients on each home-minus-away five-man feature. Blue panels are player-attributable additive profile terms; orange panels are the two retained non-additive lineup terms.

NAIL-RAPM v1.2.1.3 additive profile coefficient trajectories

NAIL-RAPM v1.2.1.3 non-additive coefficient trajectories

The immutable review artifact is artifacts/models/analysis/nail_v1213_promotion_review/nail-v1213-promotion-review-20260828T002159Z-62888daf.

Reproduction

uv run python -m nba_lineup_model.modeling.forward_nail_v1212_residualized_lambda \
  --log-path artifacts/logs/nail-v1212-residualized-lambda.log

uv run python -m nba_lineup_model.modeling.nail_v1212_residualized_lambda_frozen_backtest \
  --log-path artifacts/logs/nail-v1212-residualized-lambda-frozen.log

uv run python -m nba_lineup_model.modeling.nail_v1213_promotion_review

See the training guide and NAIL model contract for the complete provenance and remaining validation priorities. Proposed lineup-context additions and their resolved tests are tracked in the Non-Additive Context Feature Registry.