Learning from Clickstream Data in Online Retail

E-tailers now have the resources to record the sequence of clicks their online customers make, also known as their clickstream. Such detailed browsing information reveals customer preferences and, as such, can be used to improve product assortment and display decisions over time. In this paper, we consider a Bayesian framework to study how an e-tailer can draw inferences about customer preferences using clickstream and sales data, for utilitarian and hedonic product categories. For a given product category, the e-tailer dynamically determines the set of products to display on his website, via clickable pictures, and which products among the displayed ones to make available for purchase. We show that it may be optimal for the e-tailer to display "phantom" products, which are clickable, but later shown as unavailable for purchase. Under some conditions on the customers' clicking behavior on the website, we prove the optimality of an "offer less" strategy which is reminiscent of the well-known "stock more" result in multi-period newsvendor problem with demand learning. We also discuss the implications of our results on when the e-tailer should communicate product availability information with customers. Finally, we quantify the benefits which accrue from learning from clickstream data for the e-tailer. These benefits are substantial, especially for product categories for which customers do not have well-established preferences and go through an exploratory browsing process before making a purchase.