Yanwei Sun (Imperial College London)- The Role of Prescreening in Auctions with Predictions

Abstract: Motivated by the rise of generative AI—which can simulate bidders’ behaviors—and by machine-learning algorithms capable of predicting bidders’ valuations, we consider an auction environment with IID privately known valuations, where the auction designer employs a noisy predictor. Specifically, the designer receives a coarse signal about each bidder’s valuation, where the signal is fully informative with a given probability. Based on the posterior expectation of the valuations, the designer selects only the top players to admit—a procedure we call \emph{prescreening}. We show that this prescreening game is equivalent to a standard auction (without prescreening) but with \emph{correlated} valuations. Notably, the commonly assumed concept of \emph{affiliated} valuations does not generally hold in our context, except in certain special cases. We characterize conditions for the existence of a symmetric and strictly monotone equilibrium strategy in classical auctions. Our results reveal that prescreening can significantly improve the designer’s revenue in all-pay auctions; in fact, when the prediction accuracy is one, admitting only two players is optimal. In contrast, prescreening is unnecessary in both first-price and second-price auctions. We also discuss the joint decision on the admitted number and the auction format.