Leveraging purchase regularity for predicting customer behavior the easy way

The valuation of future customer activity is a mainstay of any organization seeking to efficiently manage its customer portfolio. In the area of customer-base analytics, the ongoing race for predictive power has yielded a large corpus of research to assist managers in this respect. Approaches in the tradition of stochastic models have been particularly successful because they rely only on easy-to-compute key metrics and integrate them within a parsimonious probability-modeling framework. Recent advances in this field have demonstrated that incorporating the timing regularity of past purchases can improve predictive accuracy relative to purely recency/frequency-based approaches. This paper expands that idea and introduces generalizations of a well-established probability model, the BG/NBD (Fader et al., 2005a), by replacing the exponential with a more flexible Erlang-k interarrival timing process. The resulting model variants are capable of leveraging regularity while retaining almost the same level of data requirements and algorithmic efficiency. Using extensive simulation studies and six data sets covering a wide range of empirical settings the authors demonstrate substantial improvements in predictive accuracy against the baseline models and performance gains close to or on par with a more complex model alternative. The availability of efficient and easily accessible implementations of the new model variants in the R-package BTYDplus allows marketing analysts to apply them in large-scale scenarios of data-rich environments on a continuous basis.