基于分位数回归森林的VaR估计及风险因素分析

VaR estimation based on quantile regression forest and risk factors analysis

构建非参数、集成性的分位数回归森林算法,对上证综指和标普500指数的VaR进行了估计;同时构建了其他一些主流的方法,包括历史模拟、GARCH族方法、弹性网络、门限分位数回归、CAViaR等,进行检验和对比.通过对不同置信水平下的VaR估计进行多种检验,验证了该方法的有效性和稳健性.进一步,基于分位数回归森林模型定义了一种特征重要性度量方法,评估了各个因素对于风险值的影响权重大小,发现过去一日收益率对上证综指的风险值影响较大,而波动率对标普500指数的风险值影响较大,整体来看两国股市间的风险传导性较弱;并引入偏相依关系,动态地分析了各个因素在不同水平下对于风险值的作用方向,一定程度上弥补了机器学习算法在金融应用中一直存在的“黑箱性”问题.

Quantile regression forests as a nonparametric and ensemble method were built to estimate the VaR of Shanghai Composite Index and the S&P 500 Index at different confidence levels.Meanwhile,other methods were built for comparison,including historic simulation,GARCH,elastic net,threshold quantile regression model and CAViaR,and the superiority of the proposed method was verified.Further,a new measurement method of variable importance based on the quantile regression forest was defined to judge the importance of various factors on the risk value,and it was discovered that the past one day yield has the greatest influence on the risk value of the Shanghai Composite Index,and that the volatility has the greatest influence on S&P 500 Index risk value.At the same time,the risk conduction between China and US is weak.Further,by dynamically analyzing the partial dependence between the factors and risk value,the“black box”problem of machine learning used in financial applications has been remedied to some extent.