subtopic

Error, rare event or anomaly? Decide and treat

For each flagged group decide: correct, cap, transform, keep or remove — and log the reason. Concepts: Treatment options: keep, correct, cap (winsorise), transform, exclude Project A: Outliers are informative: long-duration outliers bad rate ≈ 57% vs 28% normal; large-amount outliers ≈ 54% vs 28% → keep them; winsorise at 1st/99th percentile with caps learned on the TRAIN split (E13). Project B: Fix pressure errors to NaN then interpolate; keep peak-load hours.