Adverse selection and auction design for internet display advertising

Nick Arnosti, Marissa Beck, Paul Milgrom

Research output: Contribution to journalArticlepeer-review


We model an online display advertising environment in which "performance" advertisers can measure the value of individual impressions, whereas "brand" advertisers cannot. If advertiser values for ad opportunities are positively correlated, second-price auctions for impressions can be inefficient and expose brand advertisers to adverse selection. Bayesian-optimal auctions have other drawbacks: they are complex, introduce incentives for false-name bidding, and do not resolve adverse selection. We introduce "modified second bid" auctions as the unique auctions that overcome these disadvantages. When advertiser match values are drawn independently from heavy-tailed distributions, a modified second bid auction captures at least 94.8 percent of the first-best expected value. In that setting and similar ones, the benefits of switching from an ordinary second-price auction to the modified second bid auction may be large, and the cost of defending against shill bidding and adverse selection may be low.

Original languageEnglish (US)
Pages (from-to)2852-2866
Number of pages15
JournalAmerican Economic Review
Issue number10
StatePublished - Oct 2016
Externally publishedYes


Dive into the research topics of 'Adverse selection and auction design for internet display advertising'. Together they form a unique fingerprint.

Cite this