A Seasonal Model with Dropout to Improve Forecasts of Purchase Levels

Abstract

Predicting future purchase levels is an important and constant challenge for marketing professionals, as purchase patterns often vary over time and across customers. Moreover, purchases often follow individual and cross-sectional seasonal patterns, which affect forecasts of purchase propensity and customer dropout. The authors develop the hierarchical Bayesian seasonal model with dropout (HSMDO), which captures the interrelation between individual and cross-sectional seasonality, purchase, and dropout rates, with the aim of improving forecast accuracy at specific points in time. They perform (1) a parameter recovery analysis with synthetic data; (2) an empirical validation on three noncontractual retail data sets; (3) an analysis of different model variants to isolate the effects of dropout, seasonality, and hierarchical seasonality; and (4) a comparison with several probabilistic models from the marketing literature. The results demonstrate that the HSMDO provides increased forecast accuracy and that tracking errors decrease further with data exhibiting strong seasonality and high customer retention. The HSMDO yields a measure of individual seasonality that has high discriminative power even with sparse data sets and is useful to customer relationship management analysts for customer segmentation, portfolio management, and improvement in the timing of marketing actions.

Editors’ Parting Thoughts: Retrospective and Prospective

“…It is also important to carefully select the benchmark models to avoid cherry-picking, and to compare the different methods both conceptually and empirically to articulate the conditions under which the proposed method is “superior” (e.g., Sarkar and De Bruyn 2021). Implementing a well-designed simulation study can enhance the contribution of a methods paper significantly (e.g., Wünderlich, Wünderlich, and Wangenheim 2022).…”