Marketing analytics in banking 4.0: A two-stage explainable AI framework for high-accuracy and well-calibrated predictions

Forecasting customer conversion in bank marketing is challenged by imbalanced class distributions, where scarce minority responses lead to underfitting of true patterns or overfitting of limited instances. Sampling techniques are commonly applied to address this class imbalance problem; however, synthetic data generation may introduce distributional distortion, inflate apparent performance while degrade calibration, fairness, and explainability. Finding and selecting suitable model-sampling combinations that transparently balance performance against distortion remains a challenging task. To address this, we propose a two-stage quadrilateral evaluation framework that assesses models with sampling techniques across discrimination, calibration, computational cost, and explainability to select the suitable. In the first stage, we test ensemble learning and deep learning models on datasets with imbalanced class distributions using k-fold cross-validation and hyper-parameter tuning. In the second stage same models were re-evaluated under five sampling strategies, including Borderline2Smote, to quantify the trade-offs introduced by synthetic sampling. Results show that the combination of XGBoost with Borderline2SMOTE sampling demonstrates improved adaptability with reduced synthetic distortion. Shapley Additive exPlanations (SHAP) further support stable interpretability by identifying key drivers of customer responses for predictive inference and marketing strategies. This framework enables context-aware, transparent, and reproducible model selection strategy for responsible predictive analytics in Banking 4.0. This study also provides a foundation for further research into fairness-aware and explainable predictive analytics to support data-driven decision-making.