Analysis of the performance and explainability of different model architectures for Loss Given Default

Estimating loss given default is a fundamental requirement for credit institutions to comply with international banking regulations. This thesis presents different model architectures for estimating loss given default, using traditional statistical techniques and machine learning. Their performance was evaluated on a sample of exposures secured by mortgages on immovable property and compared with each other to identify the best architecture using evaluation metrics such as the accuracy ratio AR and Somers' D. An analysis was also conducted into the explainability of each model, a fundamental requirement for models that are then used by credit institutions to carry out their activities.