Mining Rare Events Data for Assessing Customer Attrition Risk

Abstract

Customer attrition refers to the phenomenon whereby a customer leaves a service provider. As competition intensifies, preventing customers from leaving is a major challenge to many businesses such as telecom service providers. Research has shown that retaining existing customers is more profitable than acquiring new customers due primarily to savings on acquisition costs, the higher volume of service consumption, and customer referrals. For a large enterprise, its customer base consists of tens of millions service subscribers, more often the events, such as switching to competitors or canceling services are large in absolute number, but rare in percentage, far less than 5%. Based on a simple random sample, popular statistical procedures, such as logistic regression, tree-based method and neural network, can sharply underestimate the probability of rare events, and often result a null model (no significant predictors). To improve efficiency and accuracy for event probability estimation, a case-based data collection technique is then considered. A case-based sample is formed by taking all available events and a small, but representative fraction of nonevents from a dataset of interest. In this article we showed a consistent prior correction method for events probability estimation and demonstrated the performance of the above data collection techniques in predicting customer attrition with actual telecommunications data.