Designing Auctions when Algorithms Learn to Bid

A new research paper by economist Pranjal Rawat tackles a critical question in digital advertising and online markets: how should auctions be designed when the bidders are AI algorithms? The study, 'Designing Auctions when Algorithms Learn to Bid,' develops a novel computational laboratory framework. This framework uses factorial experimental designs and large-scale Monte Carlo simulations to treat each auction simulation as a black-box observation, systematically varying inputs to rank factors by their association with outcomes like seller revenue and bid suppression. It tests across multiple modern algorithm classes, including Q-learning and contextual bandits, moving beyond the simplified models of prior work.

The central, counterintuitive finding is that the structural parameters of the market itself dominate the impact of algorithmic design choices. In unconstrained settings, the level of competitive pressure is the strongest predictor of seller revenue. When algorithms operate under budget constraints, the tightness of those budgets becomes the primary driver. The study also reveals that the optimal auction format—first-price versus second-price—is highly context-dependent and can reverse depending on the bidding technology in use. For example, second-price auctions tend to perform better under learning algorithms, but first-price can be superior under budget-constrained pacing. This means there is no universally superior auction format when bidders are algorithms, and applying recommendations from one algorithmic context to another can be counterproductive for platform designers and regulators.