The Need for Market Segmentation in Buy-Till-You-Defect Models

Buy-till-you-defect [BTYD] models are built for companies operating in a non- contractual setting to predict customers’ transaction frequency, amount and timing as well as customer lifetime. These models tend to perform well, although they often predict unrealistically long lifetimes for a substantial fraction of the customer base. This obvious lack of face validity limits the adoption of these models by practitioners. Moreover, it highlights a flaw in these models. Based on a simulation study and an empirical analysis of different datasets, we argue that such long lifetime predictions can result from the existence of multiple segments in the customer base. In most cases there are at least two segments: one consisting of customers who purchase the service or product only a few times and the other of those who are frequent purchasers. Customer heterogeneity modeling in the current BTYD models is insufficient to account for such segments, thereby producing unrealistic lifetime predictions. We present an extension over the current BTYD models to address the extreme lifetime prediction issue where we allow for segments within the customer base. More specifically, we consider a mixture of log-normals distribution to capture the heterogeneity across customers. Our model can be seen as a variant of the hierarchical Bayes [HB] Pareto/NBD model. In addition, the proposed model allows us to relate segment membership as well as within segment customer heterogeneity to selected customer characteristics. Our model, therefore, also increases the explanatory power of BTYD models to a great extent. We are now able to evaluate the impact of customers’ characteristics on the membership probabilities of different segments. This allows, for example, one to a-priori predict which customers are likely to become frequent purchasers. The proposed model is compared against the benchmark Pareto/NBD model (Schmittlein, Morrison, and Colombo 1987) and its HB extension (Abe 2009) on simulated datasets as well as on a real dataset from a large grocery e-retailer in a Western European country. Our BTYD model indeed provides a useful customer segmentation that allows managers to draw conclusions on how customers’ purchase and defection behavior are associated with their shopping characteristics such as basket size and the delivery fee paid.