极端VaR风险测度的新方法:QRNN+POT  

A new method for extreme value at risk measure: QRNN+POT

作　　者：许启发[1,2] 陈士俊[1] 蒋翠侠[1,2] 刘曦[1] 

机构地区：[1]合肥工业大学管理学院,安徽合肥230009 [2]合肥工业大学过程优化与智能决策教育部重点实验室,安徽合肥230009

出　　处：《系统工程学报》2016年第1期33-44,共12页Journal of Systems Engineering

基　　金：国家自然科学基金资助项目(71071087;70901048);教育部人文社科规划基金资助项目(14YJA790015);安徽省哲学社会科学规划基金资助项目(AHSKY2014D103);合肥工业大学产业转移研究中心资助项目(SK2014A073)

摘　　要：由于金融时间序列极端尾部数据的稀疏性,一方面非线性分位数回归存在非线性函数形式选择困难;另一方面非线性分位数回归的极端VaR风险测度精度一直不高.为此,提出了使用神经网络分位数回归(QRNN)模拟金融系统的非线性结构,并使用极值理论的POT方法弥补非线性分位数回归对极端尾部数据信息处理能力的不足,得到了一个新的金融风险测度方法:QRNN+POT,给出了其基本算法,并将其应用于极端VaR风险测度.选取了世界范围内代表性国家股票市场为研究对象,从样本内与样本外两个方面实证比较了QRNN+POT方法与已有的非线性分位数回归模型在VaR风险测度中的表现,结果表明:第一,直接使用非线性分位数回归模型能够准确地得到正常VaR风险测度,而极端VaR风险测度效果却差强人意;第二,使用QRNN+POT方法,极大地改善了极端VaR风险测度效果,能够有效地描述金融危机期间出现的极端风险.It is difficult to choose an appropriate nonlinear functional form in nonlinear quantile regression and give an accurate measure of extreme VaR due to the data sparsity in the upper or lower tail of financial time series distribution. To this end, we proposed to describe the nonlinear structure in financial system through the quantile regression neural network（QRNN）. Furthermore, we improved the ability to handle tail data information for nonlinear quantile regression using the POT method from extreme value theory, and advanced a new financial risk measure method： QRNN ＋ POT. We studied in detail how to implement the new method and provided the procedure of the new method to estimate extreme VaR. For empirical illustration, we selected four worldwide stock markets to test the performance of the new method and classical nonlinear quantile regression models both in-sample and out-of-sample. The empirical results show that the classical nonlinear quantile regression models perform bad in extreme VaR measure while they have good performance in normal VaR measure. Our new method turns out to be superior to those classical models in terms of the accuracy of VaR measure and is able to describe the extreme risk effectively during the financial crisis.

关 键 词：极端VaR 非线性分位数回归 神经网络分位数回归 POT方法 QRNN+POT方法 

分 类 号：F224.0[经济管理—国民经济]