Explainable Machine Learning for Bank Marketing Campaign Success Prediction and Decision-Support Recommendations

Objective: Inappropriately targeted direct bank-marketing campaigns may involve high contact costs and low conversion rates. This study aimed to develop an explainable machine-learning framework to predict the probability of a customer subscribing to a term deposit and to translate the predictions into practical recommendations for customer prioritization. Methods: The Bank Marketing dataset, comprising ۴۱,۱۸۸ records from direct marketing campaigns of a Portuguese bank, was used. After removing ۱۲ duplicate records, ۴۱,۱۷۶ observations remained for analysis. To prevent information leakage, the call-duration variable (duration), which is only known after a call has been completed, was excluded from the primary pre-contact model. Following feature engineering and a stratified random split of the data into training and test sets, Logistic Regression, Random Forest, and CatBoost models were compared. The final model was calibrated using isotonic regression and interpreted using SHAP. Evaluation was conducted using accuracy, precision, sensitivity, specificity, F۱-score, ROC-AUC, PR-AUC, and Brier score. Results: The campaign success rate was ۱۱.۲۷%. The calibrated Random Forest was selected as the final model and achieved an accuracy of ۰.۸۷۷۴, sensitivity of ۰.۵۹۹۱, specificity of ۰.۹۱۲۷, F۱-score of ۰.۵۲۴۰, ROC-AUC of ۰.۸۰۴۷, PR-AUC of ۰.۴۶۰۰, and Brier score of ۰.۰۷۵۸ on the test set. According to the SHAP analysis, the most influential predictors were employment variation rate, number of employees, three-month Euribor rate, contact type, contact month, age, job, and previous campaign outcome. Under the contact-capacity scenario, selecting the top ۱۵% of customers by predicted probability captured ۶۰.۶۷% of all successful outcomes while maintaining an observed success rate of ۴۵.۵۵%. Conclusion: Combining leakage-aware modelling, probability calibration, and explainability enables machine-learning predictions to be used as a decision-support tool for customer prioritization. Nevertheless, the identified relationships are observational and should be assessed through continuous monitoring, recalibration, and prospective evaluation before operational deployment.

کلیدواژه ها:

bank marketing ، explainable machine learning ، Random Forest ، probability calibration ، SHAP ، decision support ، term deposit