Explaining Stacking Ensemble Predictions: An Application of SHAP and LIME to Bank Telemarketing Data

Machine learning (ML) is changing the way banks market their products, making it easier to target customers for things like term deposits. However, the complex “black box” models often used—especially ensemble methods—can be difficult for business leaders to trust and adopt. This study addresses that challenge by building a strong stacking ensemble model to predict which clients might subscribe to products and by comparing two popular tools for explaining model predictions: SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME). Our model, trained on the UCI Bank Marketing dataset, achieved impressive results, with an accuracy of 0.91 and an F1-score of 0.89 on the test set. Both SHAP and LIME helped clarify how the model made decisions. SHAP provided more consistent insights into which factors mattered most overall, while LIME excelled at offering clear explanations for individual predictions. We found that combining both methods—using SHAP for a big-picture view and LIME for specific cases—makes complex model predictions much easier to understand and use in a business setting.