Price dynamics on Amazon Marketplace : a multivariate random forest variable selection approach

Amazon Marketplace is an ecommerce platform where Amazon is a seller itself and also the platform host to third-party (3P) sellers. Business press has reported that Amazon uses complex pricing strategies for items it sells directly. However, there is limited academic and business press understanding of price dynamics in the marketplace. The modeling challenge is the high dimensionality of the covariate space and sparsity of price changes for some sellers. We develop a novel multivariate random forest (MVRF) based variable selection method to identify the key predictors of price change. This variable selection method has potentially wider application to similar high dimensional modeling situations, e.g., in selecting variables for the design of field and laboratory experiments, and alternative statistical and structural models. We test the robustness of the proposed variable selection method against extant benchmarks such as univariate random forest (RF) and LASSO. We provide descriptive insights on the price change mechanism among the sellers. Our results reveal several unexpected patterns including that price and feature changes of even the smallest and new entrant 3P sellers are significant predictors of Amazon's price change. These findings have academic, managerial and regulatory relevance. We demonstrate additional managerial relevance via a stylized price forecasting example