The Role of Prescreening in Auctions with Predictions

Yanwei Sun, Fupeng Sun, Chiwei Yan, Jiahua WuarXiv 2025 · DOI · arXiv

Abstract

In modern auction environments, sellers often face physical or operational constraints that prevent them from admitting every potential bidder. With the advent of machine learning, sellers can now use predictive analytics to identify and invite only the most promising participants. This paper investigates the strategic impact of such "prescreening" on auction revenue. We demonstrate that while traditional theory favors broad competition, the introduction of predictive signals changes the game. Specifically, in All-Pay Auctions (APA), limiting participation to a small, elite group can be revenue-optimal, while in standard First-Price and Second-Price auctions, high-quality predictors significantly mitigate the revenue loss associated with restricted entry.

Introduction

Classic auction theory, most notably the Bulow-Klemperer result, suggests that a seller is generally better off inviting more bidders than trying to extract more value from a smaller pool through optimal mechanism design. However, in practice, sellers often face "hard" constraints—such as limited time slots in online advertising, physical capacity at a venue, or the administrative costs of vetting participants—that necessitate limiting the number of bidders.

As sellers increasingly turn to AI-driven valuation predictors to manage these constraints, a fundamental question arises: Does selectively admitting bidders based on predicted values help or hurt revenue? This research reconciles traditional auction theory with the modern reality of predictive analytics, identifying the conditions under which "less is more."

Theoretical Framework

To analyze this problem, the authors develop a Bayesian model involving $m$ potential bidders with independent and identically distributed (i.i.d.) private valuations. While the seller cannot observe these valuations directly, they have access to a "valuation predictor" that provides noisy signals. The seller commits to a policy of admitting the top $n$ bidders based on these signals.

The Equivalent Game

A primary technical contribution of this work is the concept of the "Equivalent Game." The authors prove that an auction with prescreening is mathematically equivalent to a standard auction (without prescreening) where bidders have correlated valuations. Even if bidders' initial valuations are independent, the act of selecting the "top" candidates based on a signal induces a statistical correlation. This endogenous correlation fundamentally alters bidding behavior, as participants know they have been selected from a larger pool and are likely facing other high-value competitors.

Modeling Accuracy with Copulas

To handle varying levels of prediction accuracy, the researchers employ copula modeling. This allows for a flexible definition of the dependence between true valuations and predicted signals, ranging from a "Null Predictor" (random selection) to a "Perfect Predictor" (where signals perfectly match valuations).

Impact on Auction Formats

The study characterizes Symmetric and Strictly Monotonic (SSM) bidding strategies across three major auction types: Second-Price (SPA), First-Price (FPA), and All-Pay (APA).

Second-Price and First-Price Auctions

In standard SPA and FPA formats, the results largely align with classical intuition: admitting all potential bidders ($n=m$) remains the revenue-maximizing strategy. However, the research highlights a critical nuance: as the accuracy of the AI predictor increases, the revenue gap between the restricted auction ($n < m$) and the full auction ($n = m$) shrinks. With a perfect predictor, the revenue loss from excluding lower-valued bidders disappears entirely, as the top $n$ bidders already represent the most significant competitive pressure.

All-Pay Auctions: The "Less is More" Result

The most striking findings occur in All-Pay Auctions, where every participant pays their bid regardless of whether they win. This format is common in crowdsourcing contests and R&D races. Counter-intuitively, the authors find that admitting more bidders can actually decrease revenue by discouraging aggressive bidding.

Key insights for APA include:

  • Optimal Admission: Under a Perfect Predictor, the optimal number of bidders is exactly $n^* = 2$, regardless of how many potential bidders ($m$) are available. This creates an intense "head-to-head" competition among the elite.
  • The Inflated Type: Prescreening creates an "inflated" perceived valuation. Bidders, knowing they were selected from a large pool, realize they and their opponents are likely high-value, which drives up the effort or bids in a small, select group.
  • Non-Monotonicity: Unlike SPA/FPA, revenue in APA is often non-monotonic with respect to the number of bidders. While a Null Predictor suggests admitting everyone, a high-quality predictor suggests a very restrictive entry policy.

Revenue Ranking and Conclusions

The research establishes a new revenue ranking for sellers using predictive analytics: APA with optimal prescreening ≥ FPA ≥ SPA. Notably, an All-Pay Auction with just two optimally selected bidders can generate more revenue than a standard SPA or FPA with an additional potential bidder.

Limitations and Future Directions

While this model provides a robust framework for understanding prescreening, it assumes risk-neutral bidders. In high-stakes contests, risk aversion might alter the optimal number of participants. Additionally, the current model assumes the seller keeps the predictive signals private. Future work could explore "information disclosure" policies, where sharing signal data with bidders might further influence their beliefs and increase revenue. Finally, the joint optimization of auction rules and admission policies remains a complex frontier for mechanism design in the age of AI.

Summary updated

BibTeX
@misc{sun2025role,
  author = {Yanwei Sun and Fupeng Sun and Chiwei Yan and Jiahua Wu},
  title = {{The Role of Prescreening in Auctions with Predictions}},
  year = {2025},
  eprint = {2502.12117},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2502.12117},
}
Figure