针对非对称厚尾GARCH模型参数的预选分布很难确定的问题。对模型参数空间进行数据扩张,把模型中的厚尾残差分布表示成正态分布和逆伽玛分布的混合分布,然后通过对参数的后验条件分布进行变换获得参数的预选分布,从而利用M-H抽样实现了非对称厚尾GARCH模型的贝叶斯分析。中国原油收益率波动的实证研究发现中国原油收益率的波动具有高峰厚尾性但不存在"杠杆效应",样本内的预测评价发现基于M-H抽样的贝叶斯方法优于极大似然方法,说明了M-H抽样方案设计的有效性。
The proposal densities for the parameters in the GARCH models with heavy-tailed errors are unobtainable,especially if we aim to sample the degrees of freedom as well.To overcome this problem, we adopt some suitable data augmentation of the parameter space of the GARCH model,by a mixture of normal distribution and inverse Gamma distribution to represent the Student t distribution,such that the resulting MCMC algorithm is based on Metropolis-Hastings sampling steps.An empirical application of the method modeling the China crude oil price dynamic illustrates that there exists conditional leptokurtosis but not"leverage effect",the model estimated Metropolis-Hastings sampler provides better out-of-sample forecasting power than maximum likelihood estimation,and to prove the validity of the devise of Metropolis-Hastings sampler.