Data-driven approaches to predicting customer churn in a non-contractual car-sharing company

Customer churn is a commonly found problem in most businesses. Yet, it is not well studied in sharing economy businesses, due largely to difficulty in observing customer attrition across different customer segments. To better address customer churn—and so the enhancement of sustainable urban mobility under diverse user behavior and service engagement patterns—six data-driven approaches, with and without data balancing techniques (Synthetic Minority Oversampling Technique, SMOTE), have been herein adopted and applied to a dataset from a car-sharing operator in Thailand. Our results indicate that, within specific user groups, certain algorithms excel without the need for a data balancing technique. In particular, the Transformer model without SMOTE performs best in predicting churn for one-time user groups, whereas the Artificial Neural Network (ANN) model without SMOTE and the Extreme Gradient Boosting (XGBoost) model exhibit the highest prediction performance for frequent and infrequent users, respectively. We also find that important features influencing churn tend to vary greatly across different customer segments, underscoring the necessity for churn retention strategies tailored to specific segments. In this regard, financial and service engagements are highly correlated with churn, implying that customers with better engagement are less likely to churn, which is expected in a sharing economy business.