Dynamic behavior based churn prediction in mobile telecom

Abstract Customer churn is one of the most challenging problems that affects revenue and customer base in mobile telecom operators. The success of retention campaigns depends not only on the accuracy of predicting potential churners, but with equal importance, it depends on the timing when the prediction is done. Previous works related to churn prediction presented models to predict churn monthly with a focus on the static behavior of customers, and even the studies that considered the dynamic behavior of the customer, looked mainly at the monthly level behavior. However, customer behavior is susceptible to changes over days of month, and during the time leading up to a customer decision to churn, he/she starts behaving differently. Therefore, considering monthly behavioral features negatively affects the predictive performance, because it ignores changes in behavior over days of month. Moreover, predicting churners on monthly basis will be late for customers who decided to leave at the beginning of the month because they will not be detected as churners until the next month. To address these issues, in this paper, we propose daily churn prediction instead of monthly based on the daily dynamic behavior of customer instead of his monthly one. More precisely, we represent customer’s daily behavior as multivariate time series and propose four models to predict churn daily based on this representation. Two models depend on features extracted from the multivariate time series, namely RFM-based model and statistics-based model. While the other models, exploit deep learning techniques for automatic feature extraction, namely LSTM-based model and CNN-based model. The predictive performance of the proposed models were investigated by evaluating them using a 150-day-long dataset collected from MTN operator in the country. The results showed that the daily models significantly outperform the monthly models in terms of predicting churners earlier and more accurately. Furthermore, the LSTM-based model significantly outperforms the CNN-based model. However, the prediction performances of the LSTM-based and the CNN-based models are equal to the prediction performance of the RFM-based model. Moreover, all of these three models significantly outperform the Statistics-based model.