Discrete choice models (DMs) have been used to understand customers' behavior and predict demand for a long history. In recent years, the rapid development of e-commerce brings big challenges to traditional DCMs. First, firms can get access to much more data and have stronger data-processing power. The prediction accuracy of parametric models improves marginally with big data because of misspecification errors. How to strengthen the predictive power and use the data more efficiently? Second, traditional DCMs assume a customer selects exactly one alternative, but it is convenient to purchase multiple products in online shopping. How to model and estimate customers' multiple purchase behaviors? How to optimize price and assortment? Is there a Nash equilibrium under competition? Third, mos...[ Read more ]

Discrete choice models (DMs) have been used to understand customers' behavior and predict demand for a long history. In recent years, the rapid development of e-commerce brings big challenges to traditional DCMs. First, firms can get access to much more data and have stronger data-processing power. The prediction accuracy of parametric models improves marginally with big data because of misspecification errors. How to strengthen the predictive power and use the data more efficiently? Second, traditional DCMs assume a customer selects exactly one alternative, but it is convenient to purchase multiple products in online shopping. How to model and estimate customers' multiple purchase behaviors? How to optimize price and assortment? Is there a Nash equilibrium under competition? Third, most existing literature focuses on product-level optimization, but now firms have greater flexibility when designing a product's features. How should they design and price products to maximize profits?

This thesis answers the above three questions in online retailing. The first part proposes a binary choice forest that can capture any customer's behavior. We apply the random forest algorithm with strong predictive power. We prove theoretical results including consistency and error bound. The numerical studies on real data show our framework can outperform the best parametric model. The second part studies price, assortment, and competition problems under the Threshold Utility Model (TUM). We show the advantages of TUM compared to traditional models. We solve the price and assortment optimization problem and show the existence of Nash equilibrium under competition. The third part solves the product line design problem that a firm jointly optimizes the price and feature configurations of products under the Basic Attraction Model (BAM). We develop algorithms based on the K-shortest path and an FPTAS. We also provide performance guarantees under a special case of mixed BAM.