Predictive analytics for customer repurchase: Interdisciplinary integration of buy till you die modeling and machine learning

Abstract

Predicting customer repurchase propensity/frequency has received broad research interests from marketing, operations research, statistics, and computer science. In the field of marketing, Buy till You Die (BTYD) models are perhaps the most representative techniques for customer repurchase prediction. Those probabilistic models are parsimonious and typically involve only recency and frequency of customer activities. Contrary to BTYD models, a distinctly different class of predictive models for customer repurchase is machine learning. This class of models include a wide variety of computational and statistical learning algorithms. Unlike BTYD models built on low-dimensional inputs and behavioral assumptions, machine learning is more data-driven and excels at fitting predictive models to a large array of features from customer transactions. Using a large online retailing data, we empirically assess the prediction performance of BTYD modeling and machine learning. More importantly, we investigate how the two approaches can complement each other for repurchase prediction. We use the BG/BB model given the discrete and non-contractual problem setting and incorporate BG/BB estimates into high-dimensional Lasso regression. In addition to showing significant improvement over BG/BB and Lasso without BG/BB, the integrated Lasso-BG/BB provides interpretability and identifies BG/BB predictions as the most influential feature among ∼100 predictors. The lately developed CART-artificial neural networks exhibit similar patterns. Robustness checks further show the proposed Lasso-BG/BB outperforms two sophisticated recurrent neural networks, validating the complementarity of machine learning and BTYD modeling. We conclude by articulating how our interdisciplinary integration of the two modeling paradigms contributes to the theory and practice of predictive analytics.