Abstract

Customer churn is one of the most consequential predictive problems facing retail banks, where the cost of a departing customer routinely exceeds the cost of retaining one many times over. This article reports an empirical, comparative study of ten machine learning classifiers, four traditional baselines, four bagging- and boosting-based ensembles, and two composite ensembles (Voting and Stacking), applied to a structured retail banking dataset of 10,000 customers built to reflect the statistical properties of established benchmark studies. Following an expanded review of the literature across traditional classifiers, bagging, boosting, stacking and voting ensembles, class imbalance handling, and explainable AI, the study trains and evaluates all ten classifiers under an identical pre-processing, SMOTE-resampling and evaluation protocol. AdaBoost achieved the strongest overall discrimination (ROC-AUC = 0.722) and the best balance between accuracy and churn-class recall, while Gradient Boosting and XGBoost achieved the highest raw accuracy at the cost of substantially lower recall on the minority churn class. The Voting Ensemble delivered the second-best accuracy and a strong, stable ROC-AUC, broadly corroborating the wider literature. Gain-based feature importance identified gender, geography, active-membership status and tenure as the leading predictors of churn. The findings reinforce the argument that model selection for churn prediction should privilege minority-class recall and business cost over headline accuracy alone.