CMStatistics 2021: Start Registration
View Submission - CFE
Title: Evaluation of classifiers through bilateral projections of the ROC curve Authors:  Andres Romeu - Universidad de Murcia (Spain) [presenting]
Maximo Camacho - Universidad de Murcia (Spain)
Salvador Ramallo - Universidad de Murcia (Spain)
Abstract: Being a popular measure in other disciplines, the area under the ROC curve (AUC) has been recently proposed as a method to evaluate the classification ability of different indices in the business cycle literature. We show that the standard formulation of this measure poses some conceptual problems when there are anomalies such as either very large or very small differences in the scale of the signal along the time span. We propose an alternative measure based on adding a new dimension to the ROC, the threshold levels themselves. In this context, the AUROC is simply the projection on the true positive/false positive rate plane. Analogously, we consider the areas under the projections of the curve on the true positive vs. threshold plane and the false positive vs. threshold plane. We show that the difference between these two gives a measure of classifiers that is robust and able to discriminate abnormal signals that make little economic sense and that the standard AUROC fails to detect.