Customer churn remains a critical challenge in the telecommunications industry, with annual churn rates that can be high, causing significant revenue loss. This study develops a machine learning solution to predict customer churn using a public telecom dataset of 5634 preprocessed records. Four classification models (decision tree, random forest, XGBoost, and logistic regression) were trained and evaluated using accuracy, precision, recall, F1 score, and ROC‐AUC. After systematic hyperparameter tuning with cross‐validation, the three best‐performing models (tuned random forest, tuned XGBoost, and tuned logistic regression) were combined into a soft‐voting ensemble model. This ensemble averages the predicted probabilities from each model to make a final, more reliable prediction. The ensemble achieved 85.44% accuracy, 85.48% F1 score, and 93.63% ROC‐AUC. Class‐specific performance showed 85.79% recall and 85.19% precision for churners, and 85.23% recall and 85.79% precision for nonchurners. The soft‐voting ensemble demonstrated a modest but consistent improvement over the best individual tuned model. This enhanced predictive performance and trustworthy probability outputs increase the practical reliability of churn predictions, enabling telecom operators to make more effective, data‐driven retention decisions. It enables telecom companies to identify at‐risk customers early and apply targeted retention strategies such as personalized offers, improved service, or loyalty rewards, ultimately reducing churn and increasing profitability.
Mettle et al. (2026) studied this question.