Work in Progress
Auctions, Adverse Selection, and Internet Display Advertising
Joint work with Marissa Beck and Paul Milgrom.
Online advertisements can be sold through pre-negotiated contracts that promise the delivery of a specified number of impressions, or through single-impression spot auctions. Typically, content providers use both methods simultaneously, which creates the question of which impressions should be used to fill pre-existing contracts and which should be sold by auction. The most common approach is to first fill existing contracts, and auction the remaining ad slots as remnants. This paper demonstrates that this is highly inefficient. We propose an alternative that performs better, both in theory and (we believe) in practice.
One of our contributions is to introduce a tractable analytical framework, which captures the idea that an impression's value has both a common component and an advertiser-specific match value. Within our model, we define several desirable properties (truthful, impervious to shill bidders, provides "representative" impressions to pre-existing contracts), and characterize the set of mechanisms with these properties.
Much of the paper focuses on the "modified second bid" auction, which solicits bids for each impression, but only allocates it to the highest bidder if the ratio of their bid to the second-highest bid is sufficiently large. We demonstrate that this auction is a Pareto improvement over current practice, and that the gains from adopting it may be quite large. We also show that no mechanism can significantly improve upon the efficiency of the modified second bid auction without assigning pre-existing contracts disproportionately low-quality impressions.
Availability in Dynamic Matching Markets
Joint work with Ramesh Johari and Yash Kanoria.
This research project concerns the operation of online marketplaces, such as oDesk, Airbnb, and OkCupid. One common theme from such markets is that attention is a scarce resource. On oDesk, for example, job posters often receive a flurry of applications, and do not have the time or energy to evaluate and respond to each of them. This congestion causes job seekers to apply to a many positions, which in turn means that potential employers receive hastily written messages from applicants whose interest and qualifications are tenuous at best. My work seeks to model and understand this phenomenon.
These marketplaces pose several modeling challenges. In particular, individuals generally arrive, are matched, and depart asynchronously. This creates the possibility that a job applicant may no longer be available by the time their potential employer makes an offer. These types of timing-related frictions are difficult to capture in static models, and served as one motivation for my work.
This paper considers a dynamic model of a two-sided market without wage negotiation.
Agents derive benefits from being matched, but must pay application costs to contact those on the other side, and screening costs to evaluate the suitability of potential match partners. We prove the existence of a unique equilibrium in large markets, and demonstrate that this equilibrium generically involves more applications and higher total screening costs than would be socially optimal. We quantify this loss, and discuss when it is most acute.
Having shown that the unregulated market may perform poorly, we study the effect of several simple marketplace interventions. We show that introducing barriers to search and application may cause job seekers to apply more selectively, which can in fact boost welfare on both sides of the market. This finding is worth careful consideration, because one of the primary purposes of these online marketplaces is to reduce search frictions in order to allowing agents on opposite sides to find each other.
Unraveling and Uncertainty in Two-Sided Matching Markets
Joint work with Nicole Immorlica and Brendan Lucier.
In a number of two-sided matching markets, matches are formed using a centralized clearinghouse (prominent examples include the National Residency Matching Program and high school admissions in major metropolitan areas such as New York and Boston). The clearinghouse first asks participants to rank options on the other side, and then applies an algorithm to the reported preferences to produce a recommended matching.
When preferences are known, the deferred-acceptance algorithm is known to produce a core outcome; that is, a matching such that no group of agents could jointly benefit by contracting independently. We show that this result does not hold when agents are uncertain about the preferences of others. Indeed, we prove that when agents can sign contracts before learning the preferences of others, every matching algorithm is at risk of being circumvented by a group of would-be participants. This impossibility result becomes event stronger when, additionally, agents are uncertain about their own preferences.