Empirical validation and comparison of Pareto/NBD and BG/NBD models using a grocery retailer customer database [Working Paper]

Understanding the benefits of retaining customers, companies have started to search for means to profile their customers individually and track their retention and defection behaviors. Capability to predict changes in individual customer purchase patterns is an indispensable advantage in identifying and targeting the right customers for retention investments. The major issues addressed in customer base analysis literature are identifying and counting the customers, who are active at any time, and predicting their future purchase levels and remaining lifetimes. Fader et al, 2004 have recently proposed a BG/NBD model as an easier alternative to the highly regarded Pareto/NBD model proposed by Schmittlein et al., 1987 for customer base analysis. In this study, we validate and compare the empirical performance of these two models at both aggregate and individual customer levels, using grocery retail transaction data. We measure the predicting performance of the models on repeat purchase levels, active status and lifetime of customers. In the individual level evaluations, both models capture the major characteristics of the purchase and dropout process and produce similar forecasts of the future purchase patterns. At the aggregate level, there are differences in forecasts of two models, which may be due to the way seasonality in the actual data is handled.