With the increasing popularity of machine learning techniques, it has become
common to see prediction algorithms operating within some larger process.
However, the criteria by which we train these algorithms often differ from the
ultimate criteria on which we evaluate them. This paper