Machine learning · 2026
ChurnGuard
ROC-AUC 0.84 · PR-AUC 0.66
- Problem
- Flag customers likely to cancel, where missing a churner costs more than a false alarm.
- Approach
- Logistic Regression, Random Forest, and Gradient Boosting against a majority-class baseline. Preprocessing inside a leakage-safe Pipeline, stratified 5-fold CV, threshold tuned on out-of-fold predictions.
- Result
- Held-out test set: ROC-AUC 0.84 (random 0.50), PR-AUC 0.66 (random 0.27). At a 0.36 threshold it catches 70% of churners at 57% precision.
- Data
- Telco Customer Churn · 7,043 customers, 26.5% churn
