REVENUE AND RETURN MANAGEMENT IN E-COMMERCE

Abstract

In this dissertation, we explore the intersection of return management and dynamic pricing strategies in online retailing. The dissertation consists of five chapters. In Chapter 1, we present the research motivations and provide an overview of the problems studied. Chapter 2 investigates the \Returnless Refund" policy, a novel return strategy in which retailers offer a full refund without requiring customers to return the product. We show that the optimal returnless refund policy is in the form of a threshold policy, offering a returnless refund when the salvage value of the returned product is below a positive threshold. This method allows retailers to decide between granting a refund and reselling the product efficiently. It is also shown that, for items with a high expected salvage value, this threshold-based policy is advantageous for both retailers and customers. We also show that in the early stages of policy implementation, when customers are unaware of returnless refunds, a naive policy is optimal, but the threshold rises as customer awareness increases. Furthermore, our findings show that dishonest customers, who may fake request a return to exploit this policy, can enhance the retailer profits when the price exceeds a certain level. In Chapter 3, we study a conservative dynamic pricing problem with demand learning in the presence of covariates, where the demand function follows a generalized linear model. We address managers’ concerns about transitioning from a legacy pricing system to a learning-based approach, focusing on risks of revenue loss. We propose two dynamic pricing models. The first, a stage-wise safe model, ensures that the instantaneous expected revenue from algorithmic pricing matches or exceeds a fraction of the baseline policy’s revenue in each period. Using a modified UCB algorithm, we show that the regret of this model is composed of two parts: the regret from the learning process and the regret from applying perturbed baseline prices. The second, a cumulative revenue safe, model extends this by ensuring the algorithm’s cumulative revenue meets a target compared to the baseline. Our analysis shows that the algorithm uses the baseline prices a finite number of times, even when the expected revenue of the baseline prices must be learned, offering a balance between exploration and revenue safety constraints. Chapter 4 addresses a dynamic pricing problem where customers can return products within a specified grace window, and purchasing and returning probabilities are unknown. We propose two approaches: in the first approach, the retailer learns the probabilities separately, leading to a higher regret due to censored data from return decisions. The second approach focuses on joint learning, where the final demand |calculated as the product of purchasing and keeping probabilities |is learned directly, resulting in lower overall regret. For both approaches, we extend the analysis to scenarios where return delays are dominated by a Pareto distribution. Finally, Chapter 5 summarizes the contributions and suggests directions for future research.