A novel profit maximizing metric for measuring classification performance of customer churn prediction models
Authors: Thomas Verbraken, Stefan Lessmann, and Bart Baesens
Presented at: WEB2012: The Eleventh Workshop on e-Business, Orlando (US), 15 December 2012
Keywords: Decision Support Systems, Customer Relationship Management, Data Mining, Knowledge Management and Business Intelligence
Predictive analytics (PA) is concerned with methods to build and assess data-driven forecasting models. It is routinely used to aid decision making in various applications. We argue that current practices in PA-based decision support are not well aligned with managersí requirements. This issue is illustrated for churn management, where prediction models are used to estimate the likelihood of customer churn. Whereas some progress has been made to assess churn models on the basis of profitability and thus in a way consistent with business goals, these goals are ignored in the actual model building step. Through testing four hypotheses, we show that concentrating on model assessment is insufficient. We also show that a profit-based model building approach yields significantly higher profits than current churn modeling practices. The main implication following from our results is that prediction models should be built in awareness of the decision task they are meant to support.