Enhancing Telecom Retention with a Stacking-Based Churn Model

Abstract

Customer churn in the telecommunications sector poses a critical challenge to profitability, particularly due to the difficulty of identifying minority-class churners in highly imbalanced datasets. This study aims to enhance churn detection sensitivity (Recall) and maximize retention profit by proposing a robust Hybrid Stacking Classifier. Utilizing the benchmark Telecom Churn Dataset (3,333 records; Kaggle) with a natural churn prevalence of 14.5%, we implemented a strict Stratified K-Fold Cross-Validation protocol. To preventing data leakage, synthetic oversampling (SMOTE) was applied exclusively within training folds, ensuring that validation and testing remained on naturally distributed data. The proposed ensemble stacks XGBoost, CatBoost, and AdaBoost outputs using a Logistic Regression meta-learner. The Hybrid Stacking model demonstrated superior stability and effectiveness in identifying at-risk customers, achieving a Recall of 77.33% (±2.32%), a Precision-Recall AUC (PR-AUC) of 0.8743 (±2.44%), and a low Brier Score of 0.0366 (±0.0064). While CatBoost offered competitive precision, the Stacking approach provided a balanced improvement in Recall, which is crucial for minimizing false negatives in churn prediction. A cost-benefit analysis reveals that the model yields significant positive ROI even under conservative offer acceptance rates. The study confirms that a strictly validated stacking approach effectively mitigates class imbalance without relying on unrealistic data leakage, providing a deployable, low-latency solution (0.02s inference) for real-time customer retention strategies.