Customer Churn Prediction: Does Technique Matter?

• Activity Number: 615
• Type: Topic Contributed
• Date/Time: Thursday, August 5, 2010 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics and Marketing
• Author(s): Wouter Verbeke*+ and Karel Dejaeger and David Martens and Bart Baesens
• Companies: Katholieke Universiteit Leuven and Katholieke Universiteit Leuven and Katholieke Universiteit Leuven and Katholieke Universiteit Leuven
• Keywords: Customer churn prediction ; Data mining ; Benchmarking study ; Classification techniques

Abstract

A myriad of data mining techniques has been tested to predict customer churn, but the literature reports contradictory results. This study presents the results of an extensive benchmarking experiment, including various state-of-the-art classification algorithms which are applied on twelve real-life churn prediction datasets from wireless telecom operators around the world. The experimental design of the benchmarking study consists of a full factorial experimental setup, in order to assess the effects of three different factors on the performance of a churn prediction model: sampling, input selection, and classification technique. The significance of differences in performance is rigorously tested using multiple performance measures such as for instance the AUC and the novel H-measure. Finally, a ranking is provided, indicating the best performing techniques to predict customer churn.