Quantifying uncertainty of machine learning methods for loss given default

Abstract

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