A Generalized Count Model on Customers’ Purchases in O2O Market

In the Online-to-Offline (O2O) ecommerce model, one challenge facing the online business is to predict customers' future purchases towards each product or subcategory of products, and consequently, coordinate the large amount of offline businesses involved. The main obstacle in doing that originates from the highly diversified services and thus the customer base which offline businesses bring in. The heterogeneity of customers, geographic or demographic, needs to be accurately accounted for. However, although the previous transactions for each customer are well documented, his/her demographic data is difficult or costly to acquire. Traditional wisdom relies on fitting customers into some specific statistical distribution to arrive at a satisfactory stochastic model, which may be accurate, to some extent, at a higher level. This is the case for the classic Beta-Binomial/Negative Binomial Distribution (BB/NBD) model on customers' repeat purchasing in offline context. Nevertheless, to deal with the complex level in customers' heterogeneity at an O2O business, using specific distribution is inadequate, let alonethe mathematical challenges.We propose a new model to deal with the diversity of customers. Using BB/NBD as a starting point, we relax the Beta assumption in the model to include a generalized distribution. The generalization is made possible through using the Gaussian quadrature. The results retain the elegance of stochastic model while at the same time it captures customers' heterogeneity at a better, granular level. We use a dataset from Ctrip.com, a leading O2O provider in China, to show that the proposed method outperforms the BB/NBD model in both in-sample and out of- sample predictive performance. Our approach provides a practical solution for O2O practitioners to forecast their future demands