subtopic

Target leakage check

Look for features that are 'too good' and confirm the watch-list from E01. Concepts: Leakage; post-outcome features; benchmark columns Project A: No obvious leakage. checking_account has IV ≈ 0.63 (> 0.5 'suspicious' rule): investigate and conclude it is legitimate (account status is known at application). The rule of thumb is a prompt to check, not an automatic drop. Project B: TSO 'total load forecast' correlates ~0.99 with actual load: use it only as the benchmark to beat, never as a feature. Same-hour 'actual' generation columns are also unavailable at forecast time.