Churn prediction is the canonical ML business application. End-to-end implementation: 1) Feature engineering: usage signals (DAU, session frequency, feature adoption), engagement quality (depth of use, collaborative features), health signals (support tickets, NPS responses), contractual signals (contract renewal date proximity, pricing changes). Time series aggregations: rolling 7d, 30d, 90d for each metric. 2) Model: gradient boosting (LightGBM) with binary target (churned within 30 days). Features: 50-100 engineered features. Performance: typical AUC 0.80-0.92 for B2B SaaS. 3) Survival analysis: alternative to binary prediction. Cox Proportional Hazards model estimates time-to-churn probability curve for each customer. More informative for understanding when customers will churn. 4) Uplift modeling: do not just predict who will churn - model who will respond positively to intervention. Customers who will churn regardless = "lost causes". Customers who will stay regardless = "sure things". Target "persuadables" who will stay if intervened. 5) Intervention optimization: different interventions (discount, executive call, feature training) for different risk tiers. 6) ROI measurement: control group (no intervention) vs treatment, measure 90-day retention rates. Attribution challenge: customers who would have stayed anyway inflate apparent ROI. Use uplift model to isolate true treatment effect. Tools: scikit-survival for survival analysis, causalml for uplift modeling.