Loss Given Default Modeling and Variable Importance Analysis of Credit Loans

The loss given default (LGD) is one of the most important credit risk parameters in Basel Capital Accord risk management. We used eight interpretable models to model and predict LGD for a credit loan dataset with more than 40,000 contracts. We find that the tree based non-parametric model performs significantly better than the parametric model. Among the parametric models, the two-stage model does not perform as well as the single-stage model, and among the non-parametric models, XGBoost has the best prediction ability. We also analyzed the importance of the variables in the LGD prediction model, and found that whether the customer is a defaulted executee and loan balance at default were the most important in predicting LGD. By dividing the sample by disposal methods, we further analyzed the differences in the importance of variables in the models constructed from samples with different disposal methods. We find that information on lenders and loan contracts play a more important role in LGD prediction models when adopting external agency disposition models such as delegated collection and litigation. We also find that lenders' pre-default repayment behavior plays a more important role in the prediction of LGD when a decentralized management model is used.