Comparison of Traditional Modelling Techniques and Machine Learning for Prediction of LGD by FCG

The main purpose of this paper is, using a pooled data set of default data ( GCD LGD ) , to evaluate if ML can increase the accuracy of LGD prediction compared to traditional pooling and regression techniques. The main question of the study therefore is whether ML is worth the model risk it entails. In the development of ML “challenger” models, we also explore if ML can help in discovering additional risk drivers apart from those commonly used when estimating LGD. We also briefly address modeling the cure definition.