Author(s): Kimura, T.
Since acquiring new customers is often more costly than retaining existing ones, customer retention management is critical for many business organizations. Identifying potential churners can lead to effective retention management. However, predicting customer churn is difficult because there are diverse predictors of customer churn, and their effect sizes are not evident. The technical advancement of data storage and data analytics has enabled us to implement customer churn prediction using machine learning techniques. Therefore, as one of the keys to retaining customers, customer churn prediction has drawn the growing interest of both academic researchers and marketing practitioners. Researchers have applied supervised machine learning algorithms to customer churn prediction, regarding it as a binary classification problem. Among those algorithms used in previous studies, the most popular ones are logistic regression, K-Nearest Neighbor, and Decision Tree. Recent studies have shown that advanced ensemble learning models such as XGBoost, LightGBM, and CatBoost achieve high prediction performance in classification problems. However, only a few studies applied them to customer churn prediction. In many cases, the datasets used in customer churn prediction are imbalanced: with only a few churn cases and many non-churn cases. Therefore, previous studies have mainly applied Synthetic Minority Over-sampling Technique (SMOTE) to balance the data. Recently, researchers have proposed hybrid resampling such as SMOTE-ENN and SMOTE Tomek-Links as novel and effective resampling methods. However, few studies applied these hybrid methods to customer churn prediction. Therefore, by developing a prediction model combining ensemble learning algorithms and hybrid resampling methods and comparing the model’s prediction performance with traditional methods and previous studies, this study aims to make a unique contribution to research in customer churn prediction.