Probability Of Default Based On Generalized Partial Linear Model

In the credit risk management of the commercial bank, probability of de-fault refers to the possibility that the borrowers who are not able to repay the principal and interest of the bank loan or fulfill the related obligations in a future certain time according to the contract requirement. In the credit risk management system of modern commercial bank, the evaluation and measure-ment of probability is an important asspect and has already been one of the core tools to compute the expected loss, Var and economic capital.Probability of default model(PD) has always been a hotspot of finance and scientific research fields. Traditional PD models are easy to be used in practice for their good interpretation and computational simpleness. But it has some defects, such as, strict statistical hypothesis, being easy to produce model bias and loss of systemic comprehension about credit risks.and this affects models'forecasting accuracy. So this paper use a new PD model based on theories of generalized partial linear models: P(Y=1|U,T)= E(Y|U,T)= G(UTβ+m(T)) where G(-) is a given link function,βis the unknown parametric vector, m is the unknown nonparametric function. This model has parametric and non-parametric vectors simultaneously. So it has good interpretation and adaption.This paper discusses estimate theories and computational process of GPLM model based on various literatures and put two-classify PD model to the or-dered multiple-classify situation. In the end, through dealing with German credit data from machine learning database of University of California at Irvine, I compare the fitness and accuracy between GPLM PD model and Logistic PD model in two-classify and ordered multiple-classify situations. I also appraise two models'effects through ROC curves, CAP curves and K-S testing curves. Research shows GPLM PD model has high accuracy and better manifestation.The new idea in this paper is to make application in credit risks manage-ment fields with GPLM model and do some study about GPLM PD model in two-classify and ordered multiple-classify situations. This paper provides a new way to evaluate probability of default. Feasibility and effectiveness of GPLM PD model are illustrated through practical analysises.