Export Reviews, Discussions, Author Feedback and Meta-Reviews

We thank the reviewers for their insightful comments, and will address the issues raised and questions asked as much as possible in this author response. Several reviewers made comments about the figures, and in particularly Fig.5. In the final version we will (i) make sure all figures are clear in B/W; (ii) provide separate captions for the left and right figures; (iii) move Fig.5 to the main paper; (iv) clearly explain all models used in the experiment. In his paper Hand set out to relate AUC to expected misclassification loss under varying cost proportions, but the derived relation was model-dependent as it assumed optimal decision thresholds only. It has since been shown that if thresholds are aggregated in a different way the model dependence vanishes and AUC can be translated into expected loss, see citation [7] in the paper.