Khalid Mohammed Jibreel Abdalgader

 Gafar Zen Alabdeen Salh Hassan

Abstract :

Customer churn remains a critical challenge for the global banking sector, where the cost of acquiring new clients significantly exceeds the cost of retaining existing ones. While machine learning has been widely adopted for churn prediction, existing models often rely on static demographic data, failing to capture the dynamic, “silent” behavioral shifts that precede account abandonment. This study aims to enhance churn prediction accuracy by introducing novel behavioral features and advanced feature engineering techniques. Utilizing a comprehensive retail banking dataset which downloaded from Kaggle, we engineered four unique behavioral markers: the Balance-to-Salary Ratio (BSR), Product Utilization Index (PUI), Tenure-to-Age Ratio (TAR), and Credit Score Stability (CSS). A comparative analysis was conducted between traditional ensemble methods (XGBoost, Random Forest) and a deep learning framework using a TensorFlow-based Artificial Neural Network (ANN). To address class imbalance, the SMOTE-Tomek integration technique was employed during preprocessing. Experimental results demonstrate that the inclusion of novel behavioral features significantly improved model performance. The TensorFlow ANN achieved the highest predictive power with an accuracy of 89.7% and an F1-Score of 0.77, representing a 15% improvement over baseline models. SHAP analysis confirmed that engineered behavioral ratios were more influential than traditional demographic variables. This research highlights that the “input space” (feature engineering) is as vital as the “model space” in financial analytics. The proposed framework provides banks with a proactive tool to detect subtle attrition signals, enabling more effective customer retention strategies.