Loss Functions for Predicted Click-through Rates in Auctions for Online Advertising

We consider the problem of the optimal loss functions for predicted clickthrough rates in auctions for online advertising. While standard loss functions such as mean squared error or the log likelihood loss function severely penalize large mispredictions while imposing little penalty on smaller mistakes, we nd that a loss function re ecting the true underlying economic loss resulting from mispredictions would impose signi cant penalties for small mispredictions while only imposing slightly larger penalties on large mispredictions. We illustrate that using such a loss function can signi cantly improve economic e ciency and revenue from online auctions if one is trying to t a model that is misspeci ed even when one has an arbitrarily large amount of training data.