Research On Loss Given Default Of Corporates’ Debt

Loss given default(LGD) is the core indicator to measure the credit risk ofbanks, it is also an important variable for the calculation of regulatory capital andeconomic capital. It plays a decisive role in the risk management of bank. Withthe Basel New Capital Accord is implemented in China, the development of LGDmodel becomes the research focus of banking industry. Due to the lack of defaultloan and loss data, many and complex influence factors, the research of Lossgiven default (LGD) is late in China, it is still the qualitative description, maturemethod of measurement and predicting model are less. Therefore, this paperattempts to makes a tentative inquiry into measuring enterprises’ debt loss givendefault of our country, in order to provide a new idea for the research on themeasurement model of loss given default in our country and promote the bank toimprove the risk mangement.This paper adopts the method of regression after discrimination on themeasurement of loss given default. First of all, tease out the influence factors ofLGD with the domestic and foreign literature and the actual situation of ourcountry. Secondly, according to the analysis of the influence factors, preilminarydetemine the alternative variables of LGD, and use stepwise discriminantanalysis and stepwise regression to reduce the dimensionality respectively, getexplanatory variables of distinguishing model and predicting model. Then, loanswill be divided into completely without recovery loans, partial recovery loansand fully recovery loans three categories, try to build distinguishing model byusing support vector machine, and use PSO algorithm to optimize parameter, theestablished three classification PSO_SVM model can be used to classify LGD.Finally, aimed at the loans of partial recovery, combine the PSO algorithm andsupport vector regression method, establish the predicting model of LGD.Through empirical research, the validity and robustness of the distinguishing model and the predicting model is good.