Customer Churn Prediction In The Banking Sector: A Machine Learning And Deep Learning-based Hybrid Approach

Customer churn poses a significant challenge to businesses, necessitating robust predictive solutions. We propose a novel hybrid stacking framework that integrates four diverse base classifiers—logistic regression (LR), random forest (RF), artificial neural network (ANN), and XGBoost—with a meta-learner to enhance churn prediction performance. In the first stage (Level 0), the base models independently learn from preprocessed customer behaviour and demographic features, capturing both linear and non-linear patterns. Their predicted class probabilities subsequently serve as input features to a deep feedforward neural network at Level 1, which functions as the meta-learner. This architecture is trained using categorical cross-entropy loss with the Adam optimiser, incorporating dropout to mitigate overfitting. The stacking ensemble leverages the complementary strengths of the base models (e.g., interpretability from LR, decision-boundary flexibility from RF, complex pattern recognition from ANN, and from XGBoost to achieve superior predictive accuracy and generalisation compared to any individual classifier. Experimental results on a real-world churn dataset demonstrate that the hybrid model consistently outperforms traditional baselines, achieving statistically significant improvements in AUC and F1-score. The findings suggest that stacking heterogeneous learners with a deep meta-model provides a powerful methodology for addressing the complexities of churn prediction.