Skip to content

State-Precision NAIL: No Forgetting

Last updated: 2026-08-24

This candidate tests a narrow question: does NAIL improve when its existing forward player prior is regularized according to each player's estimated uncertainty?

It is the first non-uniform State-Precision NAIL candidate. It received the same strict three-season frozen evaluation as production NAIL and is not promoted.

Contract

The player-prior mean remains the current NAIL v1.2.1.2 forward prior. Only the ridge precision changes by player.

After season \(t\), the RAPM fit supplies diagonal posterior variance \(P^+_{i,t}\). For the first candidate, no offseason variance is added:

\[ P^-_{i,t+1}=P^+_{i,t}. \]

The next fit rescales each player's prior penalty relative to the active season's median variance:

\[ r_{i,t+1}=\frac{\operatorname{median}_j(P^-_{j,t+1})}{P^-_{i,t+1}}. \]

Higher-confidence estimates receive \(r_i>1\), while uncertain estimates receive \(r_i<1\). The published season lambda remains the global penalty scale.

This differs from production NAIL, which uses \(r_i=1\) for every player. It also differs from a decay experiment: this candidate is deliberately equivalent to \(\rho=1\), so it isolates uncertainty-aware regularization without offseason forgetting.

Promotion Gate

Run the existing three-season frozen evaluation. Compare it with NAIL v1.2.1.2 using the bootstrap confidence intervals already used for promotion decisions. Only if it clears that gate should we tune \(\rho<1\).

Frozen Result

The completed through-2025-26 artifact was replayed directly for 2023-24, 2024-25, and 2025-26. The replay uses each stored target prior vector and its matching prior-season context and schedule model; it does not retrain any seasons.

Cohort Model Poss. RMSE Eligible game RMSE Full-game RMSE
Regular v1.2.1.2 production 1.197946 14.0107 14.2330
Regular State-Precision, no forgetting 1.197970 14.0255 14.2783
Playoffs v1.2.1.2 production 1.192710 16.6032 --
Playoffs State-Precision, no forgetting 1.192641 16.5191 --

The small playoff improvement does not offset weaker regular-season possession, game, and team-level metrics. The candidate remains an evaluated, non-promoted row on the Three-Season Frozen Leaderboard.

Artifact: artifacts/models/nail_state_precision_no_forgetting_frozen_backtest/frozen_multiseason_backtest/2023-24_to_2025-26/frozen_multiseason_backtest-2023-24-to-2025-26-20260825T043905Z-da55e4b8.