Brief
Compare grouped vs time CV on the housing split inside the pipeline skeleton and recommend one scheme with evidence.
Dataset
Housing prices train/holdout. Group column (neighborhood blocks) + date column provided; holdout hidden.
Starter guardrails (88)
- Skeleton: ColumnTransformer + model with 3 labeled slots (grouped-KFold run, time-split run, recommendation).
- Preprocessing inside CV folds only; random_state fixed and recorded.
Deterministic checks
Both CV schemes execute; reported means/stds recompute; holdout RMSE within tolerance of the notebook's estimate for the recommended scheme.
100-pt rubric
- Correct grouped + time implementations: 30
- Leakage-free pipeline discipline: 20
- Comparison reasoning (bias/variance per scheme): 20
- Recommendation with evidence: 15
- Holdout agreement: 10
- Reproducibility (fresh run): 5