Train Student-t Forward RAPM

Train the robust-likelihood counterpart to the recursive forward exposure-gated RAPM state through the latest completed season:

uv run nba-train-student-t-forward-rapm --through-season 2025-26

The default has five Student-t degrees of freedom. It uses iteratively reweighted least squares (IRLS), preserving the Gaussian player-coefficient prior and the exact per-season ridge regularization chosen by the completed Gaussian forward run. This isolates the effect of robust stint errors.

The trainer checkpoints after every completed season at artifacts/models/student_t_forward_rapm/<season>/.checkpoint.joblib. Re-run the same command after interruption; it resumes from the next season. The checkpoint is removed only after the immutable artifact run is published.

For an intentional short batch, add --max-seasons 3. To test a different fixed tail thickness, pass --degrees-of-freedom 8; that produces a distinct artifact but is not directly comparable to the default until evaluated under the same contract.

See Student-t Forward RAPM for the statistical contract and current result.