To choose the highest-valued bidder, Yahoo develops predictors of how many clicks and sales result from each impression. For example, if one click occurs for every ten impressions, an advertiser would have to bid more than 10 times as high for a click as for an impression in order to win the auction.
Yahoo was very proud of its predictors, but was puzzled that they systematically over-predicted the actual number of clicks or sales after the auctions closed. A well-trained economist would recognize this as an example of the winners' curse:
In a standard auction context, the winner’s curse states that the bidder who over-estimates the value of an item is more likely to win the bidding, and thus that the winner will typically be a bidder who over-estimated the value of the item, even if every bidder estimates in an unbiased fashion. The winner’s curse arises because the auction selects in a biased manner, favoring high estimates. In the advertising setting, however, it is not the bidders who are over-estimating the value. Instead, the auction will tend to favor the bidder whose click probability is overestimated, even if the click probability was estimated in an unbiased fashion.
As with the winner's curse, there is a simple fix--bid as if your estimate is the highest among all the bidders. This requires shading your predictors downwards, based on the variance of the prediction.
The paper also details two other features that Yahoo uses to make their auctions more efficient: randomized bidding, used when a less-informed bidder bids against a more-informed one; and sometimes the auction is not won by the highest bidder, due to the value of learning. When a bidder has a novel or unusual use for the display, sometimes Yahoo lets the novel user win so that Yahoo can learn more about the value of the use. If Yahoo learns that the novel use of the display is more valuable than they thought, then they can earn enough in the future to more than compensate them for giving up some revenue earlier on.