Quantifying uncertainty of machine learning methods for loss given default

Machine learning has increasingly found its way into the credit risk literature. When applied to forecasting parameters, approaches have been outperform standard statistical models. The quantification of prediction uncertainty is typically not analyzed in machine setting. However, this vital interests managers and regulators alike as increases transparency stability management reporting tasks. We fill gap by applying novel approach deep evidential regression loss given defaults (LGDs). evaluate aleatoric epistemic for LGD estimation techniques apply explainable artificial intelligence (XAI) methods analyze main drivers. find that considerably larger than uncertainty. Hence, majority estimates appears be irreducible it stems from data itself.