Build The Draft-Prior Study
This diagnostic uses the validated player-season panel to estimate a simple, interpretable first-NBA-season RAPM prior from draft profile fields. It is not a lineup model and does not modify the frozen preseason Leaderboard.
uv run nba-build-draft-prior-study --season 2025-26
The default run trains on first-NBA-season players through 2024-25, chooses the ridge regularization through expanding season-level folds, and scores the 2025-26 first-NBA-season cohort using profile fields only. It never passes 2025-26 RAPM or possession labels to the target profile table.
Outputs
Each run is immutable:
artifacts/models/draft_prior/<season>/<run_id>/
| File | Purpose |
|---|---|
training_first_nba_season_players.parquet |
Historical labels and model inputs through the cutoff |
target_first_nba_season_players.parquet |
Target-season profile inputs with RAPM and exposure columns removed |
cross_validation.parquet |
Expanding-fold ridge selection results |
regularization_stability.parquet |
Expanding-cutoff history of selected ridge regularization |
empirical_draft_curve.parquet |
Possession-weighted observed RAPM by draft tier |
adjusted_draft_curve.parquet |
Pick 1-60 partial curve with season-block bootstrap bands |
rookie_rankings.parquet |
Scores for the target first-NBA-season class |
model.joblib |
Fitted standardized ridge pipeline |
draft-prior-curve.svg |
Published diagnostic graphic |
metadata.json / manifest.json |
Cutoff, feature, hash, and artifact-integrity contract |
The current curve SVG is copied to
docs/assets/images/draft-prior/draft-prior-curve.svg; rebuilding the study
updates that documentation asset while retaining prior immutable runs.
See Draft-Informed Cold Starts for the model, conditioning caveat, and published rankings.