Prompt file imported from isaacsilver13/nba (
.github/prompts/nba-model-review.prompt.md). Copyright stays with the author.
Review the user's named NBA model, output, or experiment slice using current files and artifacts.
- Locate the producer script, selected input files, version or modification date, and any join-audit or manifest output. Do not infer the latest artifact from a filename alone.
- Check required columns and row counts before calculating diagnostics. Identify unmatched or fallback odds rows, duplicate game keys, missing dates, and side/sign conventions.
- Separate the conclusions for walk-forward RMSE, calibration or probability quality, market-match quality, ROI, drawdown, bet coverage, Kelly sizing, and daily exposure. Explain when metrics disagree.
- Check for leakage risks: postgame fields, outcome-derived columns, future rows in rolling features, global preprocessing, and odds or market data unavailable at prediction time.
- Compare against a reproducible baseline using the same folds and candidate rows. Do not promote a change because of one favorable fold or one aggregate number.
- Recommend at most one next experiment or code change. Preserve frozen experiment boundaries and write down the exact configuration and expected discriminating metric.
- If implementation is requested, edit only the owning slice and run its narrowest test or diagnostic before broader runs.
