Accurate prediction of customer churn is essential for telecommunications companies to enhance retention and maintain revenue stability. This study proposes an optimized stacking model, integrating base learners Extra Trees Classifier, Random Forest Classifier, XGB Classifier, and LGBM Classifier trained using K-fold cross validation, with a Logistic Regression meta-model leveraging base model predictions. Class imbalance is addressed through SMOTE, and a feedback loop is employed to refine hyperparameters and features. The model achieves an Accuracy of 97.46%, a Precision of 97.26%, an F1 Score of 97.37%, and an AUC Score of 97.46%, demonstrating a favorable performance compared to ensemble voting and showing modest improvements over prior studies on the same dataset. These results suggest an enhanced capability for effective churn prediction.