L20-NAIL-MSP v0.1

Rating-Aware First-20-Game Preseason Minute-Share Persistence tests whether completed NAIL-RAPM adds usable information to preseason role allocation.

Contract

For a target season \(t\), the model uses only:

transaction-derived opening-roster candidates;
prior-season total regular-season minutes; and
completed NAIL-RAPM from \(t-1\).

It never uses a completed rating, box score, or minutes from the target season. Players missing from the completed NAIL cache receive a neutral rating of zero.

Forecast

L20-MSP first creates its smoothed prior-minute share \(p_i\). Let \(z_i\) be player \(i\)'s completed NAIL-RAPM, divided by the cross-league standard deviation among target-season opening candidates. L20-NAIL-MSP applies a within-team rating tilt:

\[ \hat y_i = \frac{p_i\exp(\beta z_i)} {\sum_{j \in C_T}p_j\exp(\beta z_j)}. \]

The cold-start mass \(\epsilon\) and NAIL weight \(\beta\) are selected by mean team allocation total variation. \(\beta=0\) exactly recovers the L20 prior-minute allocation for the same \(\epsilon\), so the experiment is a nested comparison rather than a new heuristic.

Evaluation

The runner performs expanding-window frozen tests across 2015-16 through 2025-26. To forecast a holdout season, it chooses \(\epsilon,\beta\) from earlier first-20-game seasons only, then scores the holdout against realized cumulative team-minute shares.

uv run nba-evaluate-l20-nail-minute-share

Artifacts include every parameter grid, selected parameters, team-level metrics, player allocations, and an independently tuned beta = 0 prior-minute control for each holdout. The production parameter grid uses all eleven completed source seasons only after frozen scoring is complete.

Initial Frozen Result

The first rolling evaluation covered ten holdouts, 2016-17 through 2025-26. The control selected its own \(\epsilon\) from the same expanding source window; it selected \(\epsilon=0.32\) in every holdout. The rating-aware candidate won seven of ten seasons on the primary total-variation metric:

Metric	Independently tuned prior-minute control	Rating-aware L20	Difference
Mean team allocation total variation	0.22205	0.22076	-0.00129
Mean Brier score	0.02110	0.02039	-0.00071
Player-share MAE	0.02704	0.02689	-0.00015

The frozen selections kept the NAIL tilt deliberately small: \(\beta\) was 0.10 in five holdouts and 0.05 in five. Fitting all eleven completed source seasons for the current preseason selects \(\epsilon=0.32\) and \(\beta=0.05\). At that value, a one-standard-deviation NAIL advantage multiplies a player's relative minute-allocation score by only \(\exp(0.05) \approx 1.05\) before the team total is renormalized.

This is evidence that player value helps break role ties, not evidence that it should override established minutes. It is not yet wired into the Win Projections page: this initial test uses the prior season's completed NAIL rating. A promotion candidate must repeat the same frozen contract using the exact preseason NAIL forecast that would be shown for the target season.