Skip to content

Study and Train NAIL Profile Padding

NAIL-RAPM v1.0 shrinks every possession-rate profile with the same 300 pseudo-possessions. This controlled study changes only that padding contract; the player prior, RAPM lambda schedule, context features, context penalty, and three frozen evaluation seasons remain fixed.

1. Estimate Cross-Season Constants

uv run nba-build-profile-padding-study --through-target-season 2022-23

For each primitive statistic, the command minimizes next-season weighted squared error over adjacent player-season transitions ending by 2022-23. The 2023-24, 2024-25, and 2025-26 leaderboard seasons are excluded from selection. It writes an immutable artifact containing the selected constants, transition predictions, metadata, and checksums.

2. Train the Controlled Candidates

uv run nba-train-nail-profile-padding --contract uniform-season
uv run nba-train-nail-profile-padding --contract published
uv run nba-train-nail-profile-padding --contract cross-season

The uniform-season candidate isolates the effect of changing the league anchor while keeping 300 pseudo-possessions. The published candidate uses the statistic-specific stabilization constants reported by Krishna Medvedovsky. The cross-season candidate uses this project's forward-safe estimates. Three-point makes are reconstructed from a separately padded 3PA rate and 3P percentage. Usage is reconstructed from separately padded FGA, FTA, and turnover rates.

3. Run the Frozen Comparison

uv run nba-evaluate-nail-profile-padding
uv run nba-bootstrap-nail-profile-padding

The first command replays canonical uniform-300 NAIL, published-rate NAIL, and cross-season padding NAIL over the fixed 2023-24 through 2025-26 regular seasons and pooled playoffs. No target-season state is refit during evaluation.

The second command runs a paired 10,000-draw game-block bootstrap, stratified by season, for full-game margin RMSE, winner accuracy, possession RMSE, and possession MAE.

The external starting values come from NBA Stabilization Rates and the Padding Approach.