A systematic critique of how the IBM Telco churn dataset is analyzed across 36 public notebooks and papers, revealing that ~90% report F1/accuracy while fewer than 15% include a profit curve and none use survival analysis for LTV. The author demonstrates that using a default 0.5 classification threshold implicitly assumes equal cost for false positives and false negatives — a 13:1 asymmetry in reality. Using Kaplan-Meier survival analysis to compute LTV and a brute-force threshold sweep instead of the Bayes-optimal formula (which breaks on SMOTE-balanced data due to probability miscalibration), the optimal threshold drops to ~0.03, recovering roughly $86 per customer or $8.6M on a 100K-subscriber book. The piece concludes with concrete recommendations: report profit curves, anchor LTV in survival analysis, disclose calibration assumptions, and segment interventions by customer value.
