Bidder Behavior in Complex Trading Environments: Modeling, Simulations, and Agent-Enabled Experiments

Abstract

Combinatorial auctions represent sophisticated market mechanisms that are becoming increasingly important in various business applications due to their ability to improve economic efficiency and auction revenue, especially in settings where participants tend to exhibit more complex user preferences and valuations. While recent studies on such auctions have found heterogeneity in bidder behavior and its varying effect on auction outcomes, the area of bidder behavior and its impact on economic outcomes in combinatorial auctions is still largely underexplored. One of the main reasons is that it is nearly impossible to control for the type of bidder behavior in real world or experimental auction setups. In my dissertation I propose two data-driven approaches (heuristic-based in the first part and machine-learning-based in the second part) to design and develop software agents that replicate several canonical types of human behavior observed in this complex trading mechanism. Leveraging these agents in an agent-based simulation framework, I examine the effect of different bidder compositions (i.e., competing against bidders with different bidding strategies) on auction outcomes and bidder behavior. I use the case of continuous combinatorial auctions to demonstrate both approaches and provide insights that facilitate the implementation of this combinatorial design for online marketplaces. In the third part of my thesis, I conduct human vs. machine style experiments by integrating the bidding agents into an experimental combinatorial auction platform, where participants play against (human-like) agents with certain pre-determined bidding strategies. This part investigates the impact of different competitive environments on bidder behavior and auction outcomes, the underlying reasons for different behaviors, and how bidders learn under different competitive environments.