Who Leaves and When? A Survival and Machine-Learning Analysis of Customer Churn in the IBM Telco Dataset.

Customer churn erodes the recurring revenue of subscription businesses, yet analyses often report classification accuracy without asking when customers leave or what a prediction is worth. Using the public IBM Telco Customer Churn sample (7,032 customers; 1,869 churners, 26.6%), this study combines (i) Kaplan–Meier and Cox proportional-hazards analysis of time to churn, (ii) three classifiers (regularised logistic regression, random forest, gradient boosting) tuned on training data only and evaluated on a stratified hold-out set for discrimination, calibration and subgroup performance, and (iii) an explicit asymmetric-cost analysis of the retention-offer decision. Churn is concentrated among month-to-month customers (42.7%, against 2.8% on two-year contracts); estimated 12-month retention is 70.3% for month-to-month and 99.9% for two-year customers. In the Cox model, one-year and two-year contracts carry hazard ratios of 0.20 and 0.04 relative to month-to-month, and electronic-check payment a hazard ratio of 1.80 relative to bank transfer; the proportional-hazards assumption is rejected for contract type, so a contract-stratified model is reported as a sensitivity analysis. The three classifiers reach almost identical hold-out AUC (0.835– 0.840), paired bootstrap differences between them include zero, and all are reasonably well calibrated; discrimination within the month-to-month segment, where most churners are found, is clearly lower than the pooled AUC suggests. Under illustrative cost ratios, contacting only the customers flagged by the model lowers expected cost relative to contacting everyone. All results are associational and rest on a vendorpublished sample whose collection process is not documented, so the study is best read as a transparent, fully reproducible methodological case study rather than evidence about any real operator.