Selling an item among a strategic bidder and a profiled agent

Ioannis Caragiannis, Georgios Kalantzis · 2025

Read Selling an item among a strategic bidder and a profiled agent by Ioannis Caragiannis, Georgios Kalantzis (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-12313.txt.

Introduction

Selling one item gets tricky when one buyer bids openly and the other buyer's worth is only guessed. Ioannis Caragiannis and Georgios Kalantzis study that setup in work from twenty twenty-five. Both buyers draw their worth from the same pool of chance, but the seller sees just one bid plus a forecast for the other, and the forecast may be wrong.

Two sale rules are tested against the best take from an honest pair. The first rule earns the top take when the forecast is sure, and a solid share when it errs. The second rule skips the forecast and still earns a solid share. The team also shows the deal breaks down for some chance pools, so the shape of the pool is key.

Catalog record

CREATORIoannis Caragiannis, Georgios Kalantzis
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE41 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-12313

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Selling an item among a strategic bidder and a profiled agent Ioannis Caragiannis, Georgios Kalantzis Abstract We consider the fundamental scenario where a single item is to be sold to one of two agents. Both agents draw their valuation for the item from the same probability distribution. However, only one of them submits a bid to the mechanism. The other agent is profiled , i.e., the mechanism receives a prediction for her valuation, which can be true or false. Our goal is to design mechanisms for selling the item that make as much revenue as possible in cases of a correct or incorrect prediction. As a benchmark for proving our revenue-approximation guarantees, we use the maximum expected revenue that can be obtained by a strategic and an honest bidder. We study two mechanisms. The first one yields optimal revenue when the prediction is guaranteed to be correct and a constant revenue approximation when the prediction is incorrect, assuming that the agent valuations are drawn from a monotone hazard rate (MHR) distribution. The second mechanism ignores the prediction for the second agent and simulates the revenue-optimal mechanism when no bid information for the bidders is available. We prove, again assuming that valuations are drawn from MHR distributions, that this mechanism achieves a constant revenue approximation guarantee compared to our revenue benchmark. The MHR assumption is necessary; we show that there are non-MHR but regular probability distributions for which no constant approximation of our revenue benchmark is possible. 1 Introduction Traditionally, selling an item is approached in one of two ways. The first is to use an auction: potential buyers submit their bids to an auctioneer, who decides who should get the item and at which price. Typically, the price depends on the bid of all potential buyers. This approach leverages competition to maximize revenue. The second is to use a fixed price, often determined using statistical information about the potential buyers. Personalized information may differentiate the price (e.g., via discounts to specific buyers). Here, achieving high revenue due to the simplicity of pricing is the goal. In this paper, we study a hybrid setting that combines elements of the two approaches. We consider two potential buyers of a single available item: one agent who submits a bid and another agent who responds to a fixed price. The first acts strategically, aiming to buy the item at the lowest price possible and maximizing her utility. The second agent has no expressive power herself; a profiling service provides information about her preference. How should we design high-revenue mechanisms for selling an item in such an environment? More specifically, our model assumes that both potential buyers draw their valuation for the item independently from a common probability distribution. The first is a strategic bidder and mechanisms need to ensure her truthful behavior (i.e., so that her bid is equal to her true valuation for the item). The second is a profiled agent . This means that the mechanism is just given a prediction of her valuation for the item. This prediction can be correct, i.e., equal to the true valuation of the profiled agent, or incorrect. Our goal is to design mechanisms that make as high a revenue as possible in this Bayesian setting, regardless of whether the prediction is correct or not. What does high revenue mean in our setting? One approach could be to assume that both agents are strategic and use classical auction theory to compute the optimal revenue as the maximum virtual social welfare Myerson (1981) . However, this estimate can be very conservative. As mentioned above, the profiled agent has no expressive power and, hence, cannot affect the outcome of the mechanism. Furthermore, in the case the prediction of her valuation is correct, it is like having a strategic and an honest bidder. The maximum expected revenue that can be extracted by a strategic and an honest bidder will be our revenue benchmark in this work. We would like to design mechanisms which achieve as high a fraction of the revenue benchmark as possible, regardless of whether the prediction is correct or not. This is an ambitious goal that goes beyond traditional auction design. Borrowing our terminology from the recent trend of learning-augmented algorithms (or algorithms with predictions; see, e.g., the survey by Mitzenmacher and Vassilvitskii 2022 ), we will use the terms consistency and robustness to evaluate our mechanisms. Both terms quantify how far from the revenue benchmark the revenue extracted by a mechanism is. Consistency is the ratio of the revenue benchmark over the revenue extracted by the mechanism when given a correct prediction of the profiled agent valuation. Robustness is the ratio of the revenue benchmark over the worst-possible revenue that is extracted by the mechanism which utilizes an incorrect prediction. We specifically study two mechanisms. The first one is the optimal-consistency mechanism M_{oc} . Here, we can assume that the mechanism receives as input a possibly strategic bid by the strategic bidder and an honest bid by the profiled one. Myerson’s theory tells us that the maximum expected revenue we can extract from the strategic bidder is her expected virtual valuation. However, from the honest bidder, we can extract the entire valuation as revenue simply by asking the agent to pay her honest bid. Hence, the mechanism M_{oc} is defined as follows. If the virtual bid of the strategic bidder is higher than the prediction of the valuation for the profiled agent, the strategic bidder gets the item paying the inverse virtual value of the prediction. Otherwise, the profiled agent gets the item, paying her predicted (true) value. Notice that in case of a true prediction, the item is always sold. However, if the true valuation of the profiled agent is smaller than the prediction, buying the item would result in negative utility. To bound the robustness of mechanism M_{oc} , we pessimistically assume that no revenue is ever extracted from the profiled agent when the prediction is incorrect. Our second mechanism M_{pi} is designed with this pessimistic scenario in mind and completely ignores the prediction. Instead, to decide the price offered to the profiled agent, the information about the probability distribution of valuations (and predictions) is used. Hence, the mechanism M_{pi} is designed using a backward induction argument. If the item is to be sold to the profiled agent, the price will be set so that its virtual value is 0 ; by definition, this maximizes the expected revenue of the profiled agent if she is offered the item. Then, to maximize expected revenue, the price offered to the first agent should be such that she gets the item when her virtual valuation is higher than the price offered to the profiled agent. Since this mechanism ignores the prediction, it has the same consistency and robustness. Even in the very simple setting we consider, the revenue analysis is highly non-trivial. Even though we get closed-form expressions for the revenue benchmark, computing the worst-case robustness bound among a large class of distributions is very challenging. Specifically for mechanism M_{pi} , where a close-form expression of its expected revenue is available as well, we cannot compare its expected revenue directly to the revenue benchmark. Instead, since

Read the complete text