| Summary: | In this study, we examine the problem of predicting customer defection in a noncontractual setting. Motivated by recent work on machine learning using multiple time slices, we develop a novel training and testing framework, the sliding multi-time slicing (SMTS) method. We apply this method to data from the largest marketplace in Greece, namely, Skroutz, considering the standard features that account for the important characteristics of customer activity and custom performance metrics aimed at capturing business-related goals established by the company. The dataset comprises customers over a relatively short period, since April 2018, the number of which has also exhibited a significant increase in recent months. Despite these difficulties and the inherent seasonality of customer defection, our results demonstrate that, with SMTS, developing models that outperform previous approaches and optimize decision-making is possible. We validate the approach to a benchmark dataset from the commerce sector and discuss the practical considerations and requirements of the proposed method.
|
|---|