Prediction of bank credit customers churn based on machine learning and interpretability analysis

Nowadays, traditional machine learning methods for building predictive models of credit card customer churn are no longer sufficient for effective customer management. Additionally, interpreting these models has become essential. This study aims to balance the data using sampling techniques to forecast whether a customer will churn, combine machine learning methods to build a comprehensive customer churn prediction model, and select the model with the best performance. The optimal model is then interpreted using the Shapley Additive exPlanations (SHAP) values method to analyze the correlation between each independent variable and customer churn. Finally, the causal impacts of these variables on customer churn are explored using the R-learner causal inference method. The results show that the complete customer churn prediction model using Extreme Gradient Boosting (XGBoost) achieved significant performance, with accuracy, precision, recall, F1 score, and area under the curve (AUC) all reaching 97%. The SHAP values method and causal inference method demonstrate that several variables, such as the customer's total number of transactions, the total transaction amount, the total number of bank products, and the changes in both the amount and the number of transactions from the fourth quarter to the first quarter, have an impact on customer churn, providing a theoretical foundation for customer management.