Precision-Oriented Churn Prediction with a Fine-Tuned Meta-Learner Stack Model and SHAP: A Case Study on IBM Telco

Abstract

Customer churn prediction is essential in the telecommunications industry, where maintaining existing customers is significantly more cost-effective than acquiring new ones. This study introduces a precision-oriented stacked ensemble model to predict churn using the IBM Telco Customer Churn dataset. Emphasis is placed on maximizing precision to reduce false positives, thereby minimizing unnecessary and costly intervention efforts. The proposed architecture employs LightGBM, CatBoost, and Logistic Regression as base learners, with a fine-tuned ElasticNet serving as the meta-learner. Evaluation results show that the stacking model achieves strong overall performance, attaining an AUC of 0.917 and the highest precision among all models tested. To ensure interpretability, SHapley Additive exPlanations (SHAP) are applied to identify key drivers of churn such as number of referrals, contract type, monthly charges, and tenure. These findings demonstrate that a precision-focused approach can balance business efficiency and predictive power, offering a robust framework for proactive and cost-sensitive churn management.