Today, online retailers (e-tailers) have the resources to analyze detailed consumer response data, such as their clicks on the website (also known as clickstream), in addition to sales data. Such detailed browsing and purchase information may reveal consumer preferences and can be used to improve product assortments over time. We study how an e-tailer can dynamically make assortment decisions by learning about consumer preferences using clickstream and sales data. Clickstream data provides rich (yet, noisy) information about the consumers' preferences. Practitioners and academics alike, have noted the relevance of using clickstream data in decision models, yet, to the best of our knowledge, our paper is the first one to consider a dynamic assortment planning problem with learning based on clickstream data. We formulate the e-tailer's problem as a dynamic optimization problem in a Bayesian learning framework where the e-tailer determines the set of clickable and the set of purchasable products. We model consumer click and purchase process based on the Random Consideration Set (RCS) model. When consumer have a common preference ranking over the products but possibly, different clicking behaviors, we show how assortments can be partially ordered in terms of their informativeness based the concept of Blackwell ordering. Further, when profit margins are aligned with the preference ranking, we show that the optimal set of clickable and purchasable products are popular sets. Interestingly, we find that it may be optimal for the e-tailer to offer phantom products, which are clickable on the website, but unavailable for purchase and, and show that these are among the least preferred products by the consumers. Learning from clickstream data incentivizes the e-tailer to offer a greater variety of products for purchase and as phantoms. Of the maximum incremental profit (between 1.25%-3%) accrued from learning about consumer preferences (compared to not learning), we find that learning from sales-only data contributes, on average, 38.85% and learning from clickstream data contributes, on average, an additional 7.70%
