Patrick Hummel and R. Preston McAfee, "Loss Functions for Predicted
Click-Through Rates in Auctions for Online Advertising", Journal of
Applied Econometrics, Vol. 32, No. 7, 2017, pp. 1314-1328.

There are three different sources of data used in this paper.  The
first two were generated using the simulation procedures described in
Section 5.  Each of these data sources consists of 60 rows of data
corresponding to each of the 60 simulations used in creating Tables 1
and 2.  These data sources also contain five columns corresponding to
the values of lambda, the number of missing features, the
misspecification errors, the changes in logit(pCTR), and the
percentage efficiency increases.

The last data source is confidential Google data that are used in
Section 6.  These data are only available to full-time Google employees
and are sufficiently sensitive that the first author of the paper
(Hummel) was not allowed to share the data with his co-author (McAfee)
because McAfee does not work at Google.  Any individual who wishes to
use these data will need to first become a full-time Google employee by
applying for employment at Google through any of Google's public job
listings.

The generated data are contained in the ASCII files data1.csv and
data2.csv, which are zipped in the file hm-data.zip. Unix/Linux users
should use "unzip -a".

Please address questions to the corresponding author:

Patrick Hummel
Google Inc.
1600 Amphitheatre Parkway
Mountain View, CA 94043 USA