رامین حیدرزاده آذر

This study tackles the challenge of customer churn in life insurance, which leads to substantial financial losses. It introduces a tra  arent, reproducible, and leakage-free machine learning framework designed to identify at-risk policyholders accurately and efficiently. Using ۲۰,۰۰۰ anonymized Iranian life insurance policies with a churn rate of ۲۶%, the study develops a complete modeling pipeline that includes imputation, standardization, and encoding steps performed strictly within the training process to prevent data leakage. Three model types—logistic regression, random forest, and XGBoost—were trained and evaluated using stratified cross-validation, F۱-optimized thresholds, and performance metrics such as ROC-AUC, F۱, and precision. Results show that all models perform similarly (ROC-AUC ? ۰.۷۰), with logistic regression achieving the highest F۱ score (۰.۶۶) and XGBoost offering the best precision (۰.۶۸). The top-ranked predictions captured about ۶۹% of churners, demonstrating strong operational potential. Policyholders with major payment delays were identified as the most churn-prone and easily detectable group. Overall, the findings confirm that tra  arent and interpretable machine learning models can effectively balance accuracy, simplicity, and practicality—supporting data-driven, value-focused customer retention strategies in the life insurance industry.