A wavelet-based method for churn detection

Abstract

Customer churn, or customer attrition, is a major concern for many companies because acquiring new customers is much more expensive than retaining existing ones. Customers may stop purchasing for diverse reasons: financial problems, switch to competitors, dissatisfaction of the products, etc. Getting insights on these reasons may help companies to adapt their business strategy. Therefore, an accurate identification of churned and churning customers is desirable. In this master thesis, we develop a method to detect such behaviors for Siemens customers. Most of the methods found in the literature are supervised which means that the data sets on which they are applied have already labels "churned" and "not churned". This is the case for the contract-based businesses in the domain of telecommunications. In non-contractual business, like Siemens, defining if a customer is churned or not is more challenging. Siemens customers also exhibit very different behaviors such as non-constant purchase frequencies over time that results in making state-of-the-art methods like the RFM (Recency-Frequency-Monetary) fail. Wavelets and their associated scaleograms identify purchase decreases across a range of frequencies predefined from past customer purchase behavior. Wavelets allow the estimation of thresholds to identify churned customers. The wavelet-based method developed in this thesis can be applied to other kinds of data sets especially on non-stationary signals of a high frequency variability over time.