Train Age-Informed Prior RAPM

This command fits possession-level prior-centered RAPM using the frozen output of a forward RAPM aging model. It does not refit the aging model and rejects a prior table containing target-season outcomes.

Prerequisites

Build the multi-season player panel and train the target-season aging model:

uv run nba-train-aging-model --holdout-season 2025-26

The aging run must publish a label-free player_priors.parquet for the same target season. Pin its immutable run ID explicitly:

uv run nba-train-aging-prior-rapm \
  --season 2025-26 \
  --aging-run-id aging-2025-26-20260801T220356Z-4de5f001

Without --aging-run-id, the command uses artifacts/models/aging/{season}/latest.json.

Contract

For each player in the target RAPM design, the aging forecast is the prior mean. Players absent from the aging table receive the explicit zero cold-start prior. Lambda is selected only on chronological folds within the target regular season. The final regular holdout and playoff outputs use the model fitted on the first 1,044 2025-26 regular-season games; playoff outcomes never enter fitting or selection.

Outputs

Each immutable run is written under:

artifacts/models/aging_prior_rapm/{season}/{run_id}/

File	Purpose
`aging_player_priors.parquet`	Pinned label-free aging-model input
`holdout_metrics.parquet`	Stint-level regular-holdout results
`holdout_predictions.parquet`	Regular-holdout stint predictions
`player_rankings.parquet`	Frozen final-training coefficients, priors, and adjustments
`final_training_player_coefficients.parquet`	Coefficients fit on the first 1,044 games
`final_training_state.json`	Frozen intercept and selected lambda
`frozen_playoff_metrics.parquet`	Common eligible-possession playoff metrics without refitting
`frozen_playoff_predictions.parquet`	Frozen playoff predictions
`metadata.json` / `manifest.json`	Aging-run provenance, hashes, split identity, and artifact integrity

The completed run is indexed in MLflow under the aging_prior_rapm kind.