Churn Prediction in the Mobile Telecommunications Industry An application of Survival Analysis in Data Mining

Recently, the mobile telecommunication market in the Netherlands has changed from a rapidly growing market, into a state of saturation and fierce competition. The focus of telecommunication companies has therefore shifted from building a large customer base into keeping customers ‘in house’. Customers who switch to a competitor are so called churned customers. Churn prevention, through churn prediction, is one way to keep customers ‘in house’. In this study we focus solely on prepaid customers. In contrast to postpaid customers, prepaid customers are not bound by a contract. The central problem concerning prepaid customers is that the actual churn date in most cases is difficult to assess. This is a direct consequence of the difficulty in providing a unequivocal definition of churning and a lack of understanding in churn behavior. To overcome this problem, here a custom and flexible churn definition is proposed. The predictive churn model presented in this study is based on the theory of survival analysis. Survival analysis is predominantly used in medical sciences to examine the influence of variables on the length of survival of patients. In survival analysis, the time until the occurance of a well-defined event is modelled. In the present case, the event of interest is churn. In this research the focus is on the extended Cox model. This is a variant of the original proportional hazards model, that is used for churn modelling. Since survival models are not designed to act as predictive models, some adjustments had to be made. To be able to compare the performance of the extended Cox model with the established predictive models, a decision tree is also considered. Both models performed approximately similar with a sensitivity ranging from 93% to 99% and a specificity ranging from 92% to 97%, depending on the model and the churn definition. The extended Cox model can be considered as a perfect alternative to the established predictive models and offers some unique qualities.