A Rolling-Window Framework for Churn Prediction and Behavioral Driver Identification

In three linesChurn prediction framework using 30-day rolling behavioral windows. On real non-contractual data: feature-based model achieves 87.6% accuracy and ROC-AUC 0.94; sequence-based model reaches 96.1% recall. Robustness confirmed on future unseen data (accuracy >83%, ROC-AUC >0.91) without retraining.

Summary generated by Claude — human-verified