Predictive Analytics for Customer Churn in Digital Banking Using Integrated Ensemble ML Approaches

Abstract

Customer churn prediction in digital banking has become a predictive analytics problem because retaining customers is less expensive than acquiring them. The increasing competition of digital financial services heightens the necessity of smart systems that can extract the customers who are likely to depart, but conventional statistical solutions can be inadequate in identifying the behavioral patterns that are not easily identified. The project aims to develop a reliable and robust machine-learning (ML) model that uses demographic data to predict client attrition, financial, and behavioral attributes based on a 10,000-record Kaggle digital banking customer churn dataset. Its methodology involves thorough preprocessing - removal of irrelevant identifiers, categorical encoding, feature standardization, and class balancing using SMOTE - followed by training four optimized models: Decision Tree, LightGBM, XGBoost, and an Ensemble Voting Classifier. The models that performed the best as ensemble classifiers (accuracy of 90% and ROC-AUC of 0.95%) were assessed using accuracy and precision, recall and F1-score, and ROC-AUC. When compared to the current models (Random Forest (77% accuracy), Logistic Regression (70.98% accuracy), and Naive Bayes (47.63% accuracy), the efficacy of the suggested ensemble technique demonstrates its superior capacity for generalization. It was found that the combination of multiple learners greatly minimizes false negatives, which are critical in avoiding loss of customers, and provides maximum discriminative powers relative to all other models that were tested. The work enhances a scaled and data-based churn prediction pipeline specific to the digital banking industry, with better accuracy, reduced error, and realistic applicability to real-time customer retention systems.