Churn prediction on imbalanced data is a challenging task. Ensemble solutions exhibit good performance in dealing with class imbalance but fail to improve the profit-oriented goal in churn prediction. This paper attempts to develop a new bagging-based selective ensemble paradigm for profit-oriented churn prediction in class imbalance scenarios. The proposed approach exploits an over-produce and choose strategy, which uses a cost-weighted negative binomial distribution to generate training subsets and a cost-sensitive logistic regression with a lasso penalty to combine base classifiers selectively. Extensive experiments were carried out on ten real-world data sets exhibiting a high level of imbalance from the telecommunication industry. The experimental results show that our proposed method obtains better performance than the other twelve state-of-the-art ensemble solutions for class imbalance in both accuracy-based and profit-based measures. Our research provides a new ensemble tool for imbalanced churn prediction for both academicians and practitioners.