基于神经网络分位数回归的多期CVaR风险测度

Multi-period CVaR Measure via Neural Network Quantile Regression

与VaR金融风险测度相比,CVaR具有更好的数理性质,其计算方法成为关注的焦点。相对于单期CVaR而言,多期CVaR风险测度具有较强的非线性特征,其建模过程更加复杂。在神经网络分位数回归基础上,建立了一种新的多期CVaR风险测度方法;基于似然比检验,建立了多期CVaR风险测度返回测试评价准则。将该新方法应用于沪深300指数的多期CVaR风险测度,并将其与传统的测度方法进行了对比,返回测试结果表明:第一,该新方法具有较强的稳健性,各期平均绝对误差大小基本不变,特别适合于多期CVaR风险测度;第二,基于神经网络分位数回归的多期CVaR风险测度效果优于传统测度方法,表现为似然比检验拒绝次数最少和平均绝对误差最小。

Compared with VaR, CVaR is a coherent risk measure, which was paid more attention. Multi-period CVaR measure is more complicated than one-period CVaR in that it requires nonlinear modeling. We propose a new method for estimating multi-period CVaR via neural network quantile regression neural network. To evaluate the performance of the new method, we develop a new program to backtesting CVaR by likelihood ratio test. The Hushen 300 index is used to compare the performance of our new method with those of traditional methods. The empirical results show that, firstly, the new method is robustness and more suitable for the multi-period CVaR risk measure as the value of the mean absolute error is almost the same at each period, secondly, the new method outperforms several traditional methods in terms of the small number of rejected by the LR test and the small value of the mean absolute error