Artificial Intelligence, Algorithmic Bidding and Collusion in Online Advertising

Algorithms are becoming the standard tool for bidding in auctions through which digital advertising is sold. To explore how algorithmic bidding might affect functioning of these auctions, this study undertakes a series of simulated experiments where bidders employ Artificial Intelligence algorithms (Q-learning and Neural Network) to bid in online advertising auctions. We consider both the generalized second-price (GSP) auction and the Vickrey-Clarke-Groves (VCG) auction. We find that the more detailed information is available to the algorithms, the better it is for the efficiency of the allocations and the advertisers profit. Conversely, the auctioneer revenues tend to decline as more complete information is available to the advertiser bidding algorithms. We also compare the outcomes of algorithmic bidding to those of equilibrium behavior in a range of different specifications and find that algorithmic bidding has a tendency to sustain low bids both under the GSP and VCG relative to competitive benchmarks. Moreover, the auctioneer revenues under the VCG setting are either close to or lower than those under the GSP setting. In addition, we consider three extensions commonly observed in the data: introduction of a non-stategic player, bidding through a common intermediary, and asymmetry of the information across bidders. Consistent with the theory, the non-strategic player presence leads to increased efficiency, whereas bidding through a common intermediary leads to lower auctioneer revenue compared to the case of individual bidding. Moreover, in experiments with information asymmetry, more informed players earn higher rewards.