Optimized CatBoost with Class-Weight Balancing and Threshold Tuning for Bank Marketing Term Deposit Prediction

Abstract

Direct marketing campaigns in the banking sector face significant challenges due to highly imbalanced customer response distributions, where the majority of clients decline term deposit subscriptions. Class imbalance poses a major challenge in financial predictive analytics, particularly in bank marketing campaigns where subscriber conversion rates are inherently low. This study presents an optimised machine learning framework that combines CatBoost classification, class-weight balancing, and decision-threshold optimisation. Evaluated on the public UCI Bank Marketing dataset using a strictly leakage-free nested cross-validation scheme, the proposed Optimized CatBoost (OCB) pipeline showed strong performance across composite and rank-based metrics, with an F1-Score of 0.6256, ROC-AUC of 0.9353, PR-AUC of 0.6277, and MCC of 0.5747. While baseline models such as Random Forest and XGBoost achieved higher raw precision or recall in specific uncalibrated configurations, OCB provided a more balanced overall decision profile. These results demonstrate the effectiveness of combining algorithmic weighting with post-hoc threshold tuning for tabular financial data. Overall, the findings indicate that the integration of ordered boosting, automatic imbalance handling, and threshold optimisation provides a robust solution for real-world bank marketing decision support systems.