基于广义Beta回归的不良贷款回收率模型  

Recovery Rate Models Based on General Beta Regression for Non-performing Loan

作　　者：陈暮紫[1,2] 陈浩[3] 马宇超[3,4] 王博[3] 唐跃[3] 黄意球[3,4] 陈敏[3,4] 杨晓光[3] 

机构地区：[1]中国科学技术大学统计与金融系,安徽合肥230026 [2]中央财经大学投资系,北京100081 [3]中国科学院数学与系统科学研究院,北京100190 [4]中国科学院随机复杂结构与数据科学重点实验室,北京100190

出　　处：《数理统计与管理》2011年第5期810-823,共14页Journal of Applied Statistics and Management

基　　金：国家重点基础研究发展规划项目(973计划)(No:2007CB814902);国家基金委海外杰出青年基金(No:10628104);国家基金委创新研究群体(No:10721101);国家水利部公益性项目(No:200801027)的资助;中央财经大学"211工程"三期资助项目

摘　　要：依据国内最大的违约损失率数据库-LossMet rics^(TAI),利用广义beta回归模型对时间跨度为2001-2008年的不良贷款回收率分布进行了研究。采用极大似然和OLS估计方法,针对全样本和回收率为(0.1)开区间的非极端回收样本给出包含宏观、贷款和债务人异质因子的多变量广义beta回归模型的拟合参数和各个因素对回收率均值、方差的影响分析,并在多种变换方式下研究了只含宏观因子的单变量模型。结果表明,宏观因子的影响在单、多变量模型中都十分显著;不良贷款的有效抵质押和债务人的经营状况对模型的拟合也有明显影响;各因素对全样本和非极端回收样本的拟合结果差异性显著,同一因素对回收率均值、方差的影响大为不同。应用所得广义beta回归模型可进一步讨论回收率的区间估计、在险价值(VaR)等问题,给不良贷款风险管理提供极大帮助。Based on LossMetrics^TM, we analyze recovery rate distribution of 2001-2008 non-performing loan by general beta regression. Using maximum likelihood and OLS estimation, focusing the whole sample and the recovery rate of （0,1）, we give the fitting parameters of multi-variable general beta regression models with macro, debt and debtor factors and analyze how the various factors effect mean and variance of recovery rate distribution. Finally, we discuss single variable models with macro factor through different transform. The results show that the impacts of macro factor in the single and multivariable models are very significant, the debts＇ credit guarantees and the debtors＇ financial situations are also significant, the fitting parameters are quite different between the whole sample models and the non-extreme sample models, even impacts of the same factors are different between mean and variance. Using the final fitting parameters of general beta regression, we can analyze interval estimation, value at risk and so on.

关 键 词：广义Beta回归 宏观因子 极大似然估计 OLS估计 单因素模型 多因素模型 

分 类 号：F830[经济管理—金融学] O212[理学—概率论与数理统计]