Application of machine learning method under IFRS 9 approach to LGD modeling

Abstract

The aim of our research was to show new methods that can successfully be applied by banks in their internal risk calculations. The methods concern one of the key risk parameters, Loss Given Default (LGD). The proposed approach is admissible under IFRS 9 standard. We have applied gradient boosting algorithm which is a classification algorithm and a transitional generalized linear model to forecast LGD values based on explanatory variables and lagged LGD values. We have introduced a Markov chain structure into our data and built an infinitesimal generator to forecast LGD values based on migration matrices for any period t > 0. Performance of both applied methods was examined by ROC curves. The calculations were done on real data in SAS 9.4.