Machine Learning Algorithms for Customer Churn Prediction in the Banking Sector: A Comparative Study

Abstract

Predicting customer churn in retail banking is essential for sustaining profitability. This study compares four supervised machine-learning models—Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), and Artificial Neural Network (ANN)—using the publicly available bank customer churn dataset from Kaggle (10,000 records, 18 attributes; publicly available at Kaggle repository. Data preprocessing included one-hot encoding for categorical variables, label encoding for gender, and feature selection via an ExtraTreesClassifier retaining nine informative predictors (e.g., age, credit score, balance). To address class imbalance (≈80% nonchurners vs. 20% churners), models were trained and evaluated with and without the Synthetic Minority Oversampling Technique (SMOTE), which was applied only to the training folds under stratified 5-fold cross-validation. Evaluation metrics comprised accuracy, precision, recall (for churn class), F1-score, ROC–AUC, and PR–AUC. RF achieved the best balance between recall (0.484 imbalanced; 0.619 balanced) and accuracy (0.867 imbalanced), while LR with SMOTE attained the highest recall (0.715) at the cost of reduced accuracy (0.718). Overall, the results highlight RF as the most robust model across both distributions and emphasize the importance of imbalance-aware evaluation in bank churn prediction.