Application of the kNN-Based Method and Survival Approach in Estimating Loss Given Default for Unresolved Cases

Abstract

A vast majority of Loss Given Default (LGD) models are currently in use. Over all the years since the new Capital Accord was published in June 2004, there has been increasing interest in the modelling of the LGD parameter on the part of both academics and practitioners. The main purpose of this paper is to propose new LGD estimation approaches that provide more effective results and include the unresolved cases in the estimation procedure. The motivation for the proposed project was the fact that many LGD models discussed in the literature are based on complete cases and mainly based on the estimation of LGD distribution or regression techniques. This paper presents two different approaches. The first is the KNN non-parametric model, and the other is based on the Cox survival model. The results suggest that the KNN model has higher performance. The Cox model was used to assign observations to LGD pools, and the LGD estimator was proposed as the average of realized values in the pools. These two approaches are quite a new idea for estimating LGD, as the results become more promising. The main advantage of the proposed approaches, especially kNN-based approaches, is that they can be applied to the unresolved cases. In our paper we focus on how to treat the unresolved cases when estimating the LGD parameter. We examined a kNN-based method for estimating LGD that outperforms the traditional Cox model. Furthermore, we also proposed a novel algorithm for selecting the risk drivers.

Keywords

• Loss Given Default
• survival analysis
• Cox proportional hazard model
• unresolved cases
• non-parametric model