Enhancing Customer Churn Prediction in the Banking Sector through Hybrid Segmented Models with Model-Agnostic Interpretability Techniques

Abstract

The banking industry is experiencing a transformative period due to rapid advancements in big data and artificial intelligence, which present both significant opportunities and challenges. One of the pressing challenges in the domain of customer churn prediction (CCP) is the accurate classification of imbalanced datasets. In this study, we conduct a comprehensive investigation into CCP within the banking sector, utilizing an extensive range of datasets. We integrate robust models capable of capturing complex non-linear relationships to develop hybrid segmented models for CCP. Additionally, we introduce a novel, model-agnostic technique that extends SHAP (SHapley Additive exPlanations) to ensure the interpretability of these segmented hybrid models. The approach rigorously evaluates the performance of various predictive models across 14 customer turnover datasets. The interpretability of the new model-agnostic method is showcased through a detailed case study, providing clear insights into model decision-making processes. The staged comparison trials reveal that the Voting Classifier, XGBoost, CatBoost, and LGBoost achieve accuracies of 0.81, 0.84, 0.82, and 0.83, respectively. Among these, XGBoost demonstrates the highest prediction performance, emerging as the recommended algorithm. This study not only advances the accuracy of CCP models in the banking sector but also enhances their interpretability, facilitating more informed decision-making.