L20-NAIL-MSP v0.2

This slice repeats v0.1 with the exact NAIL forecast that a preseason product can display for the target season.

For target season \(t\), each player receives:

\[ R^{\mathrm{pre}}_{i,t} = \mu^{\mathrm{forward}}_{i,t} + \Delta^{\mathrm{additive}}_{i,t-1}, \]

where the first term is the value-conditioned forward prior and the second is the frozen additive profile computed only from information through \(t-1\). There is no target-season RAPM update or non-additive lineup observation.

The forecast applies the same rating tilt and expanding-window parameter selection as v0.1. It is the only rating-aware allocation version eligible to replace the Win Projections baseline, because it matches the public preseason NAIL contract rather than using a completed rating as a proxy.

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

Initial Frozen Result

The expanding-window replay covered ten frozen holdouts, 2016-17 through 2025-26. For each holdout, both the rating-aware candidate and the prior-only control selected their own parameters from earlier seasons. The candidate won seven of ten seasons on mean team allocation total variation:

Metric	Independently tuned prior-minute control	Exact preseason NAIL	Difference
Mean team allocation total variation	0.22205	0.21980	-0.00226
Mean Brier score	0.02110	0.01977	-0.00133
Player-share MAE	0.02704	0.02677	-0.00027

The full completed history (2015-16 through 2025-26) selects \(\epsilon=0.32\) and \(\beta=0.10\) for the next preseason. A one-standard- deviation NAIL advantage therefore multiplies a player's relative allocation score by \(\exp(0.10) \approx 1.105\) before the team total is normalized.

This is still an allocation model, not a claim that coaches mechanically play the highest-rated roster. Prior minutes remain the dominant signal, while the preseason NAIL forecast provides a measured within-roster adjustment. This version is eligible for local Win Projections integration because the replay uses the same forecast contract available in the product.

Local Product Contract

Win Projections now uses this allocation for its local preseason baseline. The website retains its practical active-15 roster gate: prior-season minutes choose the initial 15 players, then the exact preseason NAIL forecast rebalances only those 15 shares. This keeps the user-facing manual-only designation for players outside the initial rotation. The gate itself was not part of the opening-roster replay and remains a product constraint rather than a validated allocation term.