Auctions with Dynamic Scoring

Martino Banchio, Aranyak Mehta, Andrés PerlrotharXiv 2024 · arXiv · OpenReview

Abstract

We study the design of auctions with dynamic scoring, which allocate a single item according to a given scoring rule. We are motivated by online advertising auctions when users interact with a platform over the course of a session. The platform ranks ads based on a combination of bids and quality scores, and updates the quality scores throughout the session based on the user's online activity. The platform must decide when to show an ad during the session. By delaying the auction, the auctioneer acquires information about an ad's quality, improving her chances of selecting a high quality ad. However information is costly, because delay reduces market thickness and in turn revenue. When should the auctioneer allocate the impression to balance these forces? We develop a theoretical model to study the effect of market design on the trade-off between market thickness and information. In particular, we focus on first- and second-price auctions. The auctioneer can commit to the auction format, but not to its timing: her decision can thus be cast as a real options problem. We show that under optimal stopping the first-price auction allocates efficiently but with delay. Instead, the second-price auction generates more revenue by avoiding delay. The auctioneer benefits from introducing reserve prices, more so in a first-price auction.

Introduction

In traditional search advertising, auctions occur instantaneously upon a user's query. However, the rise of conversational AI interfaces introduces a temporal dimension to ad placement. As a user interacts with a chatbot, the platform continuously learns about the user's intent, allowing for more precise ad matching over time. This creates a fundamental trade-off: waiting allows for better information acquisition regarding ad quality, but it risks reducing "market thickness"—the level of competition among advertisers. If the platform waits until only one advertiser is deemed a good match, competition vanishes, potentially leading to a collapse in revenue. This paper investigates the optimal timing for running auctions when the quality of the goods is revealed dynamically.

Methods

The authors develop a theoretical microeconomic model involving a platform (the auctioneer) and two advertisers. Ad quality is modeled as a binary state (good or bad match), initially unknown to all parties. As the conversation progresses, the platform receives stochastic signals and updates its beliefs ($μ_t$) about each ad's quality.

Evolution of beliefs $μ_t^1$ and $μ_t^2$ over time, illustrating how the platform learns about ad quality as the session progresses.

The platform commits to an auction format—either First-Price (FPA) or Second-Price (SPA)—but retains discretion over the timing of the allocation. This decision is framed as a real options problem, solved using martingale theory and Hamilton-Jacobi-Bellman (HJB) equations to determine optimal stopping boundaries in continuous time.

Theoretical Insights: The Timing Tug-of-War

The choice of auction format significantly influences the platform's timing strategy:

  • Second-Price Auctions (SPA) and Information Aversion: In an SPA, the price is determined by the second-highest bidder. If the platform waits and discovers one ad is a poor match, the competitive price drops to zero. To avoid this "thin market" risk, the platform tends to run the auction immediately ($t=0$) while uncertainty—and thus competition—is high.
  • First-Price Auctions (FPA) and Information Seeking: In an FPA, the winner pays their own bid. The platform is less penalized by the loss of a second bidder and prefers to wait until it can identify the most efficient (highest quality) match.
  • The Bid Shading Trap: Under FPAs, advertisers anticipate that the platform will wait until competition is low. Consequently, they strategically lower their bids (bid shading), knowing they will likely face minimal opposition at the moment of allocation.

Key Results and the Role of Reserve Prices

The study reveals a stark contrast in revenue performance based on the presence of reserve prices:

  1. Revenue Dominance without Reserves: In the absence of reserve prices, the SPA generates strictly more revenue than the FPA. The revenue ratio is $\pi_{2P} / \pi_{1P} = 1/\mu_0$, where $\mu_0$ is the prior probability of a match.
  2. Efficiency vs. Revenue: FPAs achieve efficiency by showing the best ad but suffer lower revenue due to bid shading. SPAs prioritize revenue by leveraging market thickness at the cost of potential matching inefficiency.
  3. The Reserve Price Reversal: The introduction of a Myersonian reserve price fundamentally changes the outcome. A reserve price provides a floor that protects the platform from thin markets. With an optimal reserve, the FPA becomes the superior mechanism, achieving both high matching quality (through delay) and high revenue (through the reserve floor).

Conclusion and Future Work

This research highlights that the optimal design of conversational ad markets requires careful consideration of both auction format and timing. While SPAs favor immediate placement, FPAs with reserve prices offer a path to maximizing both efficiency and revenue. Future research could extend this model to include dynamic bidding, where advertisers update their bids in real-time, or explore the impact of user churn if the information acquisition process becomes too lengthy.

Summary updated

Abstract

We study the design of auctions with dynamic scoring, which allocate a single item according to a given scoring rule. We are motivated by online advertising auctions when users interact with a platform over the course of a session. The platform ranks ads based on a combination of bids and quality scores, and updates the quality scores throughout the session based on the user's online activity. The platform must decide when to show an ad during the session. By delaying the auction, the auctioneer acquires information about an ad's quality, improving her chances of selecting a high quality ad. However information is costly, because delay reduces market thickness and in turn revenue. When should the auctioneer allocate the impression to balance these forces? We develop a theoretical model to study the effect of market design on the trade-off between market thickness and information. In particular, we focus on first- and second-price auctions. The auctioneer can commit to the auction format, but not to its timing: her decision can thus be cast as a real options problem. We show that under optimal stopping the first-price auction allocates efficiently but with delay. Instead, the second-price auction generates more revenue by avoiding delay. The auctioneer benefits from introducing reserve prices, more so in a first-price auction.
BibTeX
@misc{banchio2024auctions,
  author = {Martino Banchio and Aranyak Mehta and Andrés Perlroth},
  title = {{Auctions with Dynamic Scoring}},
  year = {2024},
  eprint = {2403.11022},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2403.11022},
}

Loading related content...

Figure