Basket data-driven approach for omnichannel demand forecasting

Abstract

Omnichannel retailing has changed the purchasing behavior of customers in recent years, especially in online shopping, which has led to higher complexity in supply chain demand forecasting. Nowadays customers buy a variety of products in baskets that do not share similar characteristics and across various channels. In this article, we propose a new approach to forecasting demand, driven by data on customers shopping baskets. Drawing on network graph theory and findings from the marketing literature, we identify for a given product four attributes to promote the connectivity with other products sold together in a basket: degree and strength for cross-categories connection, substitutability and complementarity for within-categories connection. These attributes are used as predictor variables within four proposed methods: an autoregressive integrated moving average model with exogeneous variables (ARIMAX), a linear and a polynomial regression with one lag of sales and a machine learning method. We conduct an empirical investigation using online and physical sales related to an assortment of 24,000 products of a major cosmetics retailer in France. We provide empirical evidence that using the shopping basket data with the proposed forecasting methods improves the forecasting accuracy and the stock control performance in omnichannel retailing. We also show that there is a benefit from joint forecasting of the online and store channels, and a benefit of shared inventory between both channels in terms of shortage reduction.