Performance of credit risk prediction models via proper loss functions

The performance of predictions models can be assessed using a variety of methods and metrics. Several new measures have recently been proposed that can be seen as refinements of discrimination measures, including variants of the AUC (Area Under the ROC curve), such as the H index. It is widely recognized that AUC suffers from lack of coherency especially when ROC curves cross. On the other hand, the H index requires subjective choices. In our opinion the problem of model comparison should be more adequately handled using a different approach. The main contribution of this paper is to evaluate the performance of prediction models using proper loss function. In order to compare how our approach works with respect to classical measures employed in model comparison, we propose a simulation studies, as well as a real application on credit risk data.