Predicting customer churn for a telecom operator using data mining techniques

Abstract

Churn prediction is an important requirement for customer retention and customer relationship management (CRM). It is because, lost customers must be replaced by new customers who are not only expensive to acquire but also generate less revenue in near term than established customers. This is more evident in a mature industry such as the telecom sector1.Consequently, retention campaigns are rolled out that could be effective in containing the churn but at the same time, they could be very expensive. Therefore, it is important to find out who is most at risk for attrition so that retention offers could be made to those appropriate customers who might leave without the additional incentives. In this report, we attempt to predict the voluntary customer churn for a telecom operator in the US. A data set of 71,407 records is used to model churn prediction as a binary outcome for a specific period of 31-60 days. The data set consists of both, Metric and Nominal variables of behavioral and demographic data of users. We evaluate various combinations of possible Data Mining techniques (Artificial Neural Network, Decision Tree, Discriminant analysis and Logistic Regression) and arrive at building a hybrid model using a combination of these techniques, in order to increase the overall accuracy and precision of the model.