Using Machine Learning for Modeling Human Behavior and Analyzing Friction in Generalized Second Price Auctions

Abstract

Recent advances in technology have reduced frictions in various markets. In this research, we specifically investigate the role of frictions in determining the efficiency and bidding behavior in a Generalized Second Price Auction (GSP) — the most preferred mechanism for sponsored search advertisements. First, we simulate computational agents in the GSP auction setting and obtain predictions for the metrics of interest. Second, we test these predictions by conducting a human subject experiment. We find that, contrary to the theoretical prediction, the lower valued advertisers (who do not win the auction) substantially overbid. Moreover, we find that the presence of market frictions moderates this phenomenon and results in higher allocative efficiency. These results have implications for policymakers and auction platform managers in designing incentives for more efficient auctions. Finally, after establishing the validity of our computational model, we simulate counterfactuals that provide additional insights into the role that frictions play in the markets that are not feasible (or practical) to investigate with human-subject experiments.