In S-PLUS the default p-values used by summary.mgarch() are based on the
Student's t distribution. Here is the relevant code where the p-value is
computed

pv <- 2 * (1 - pt(abs(tv), xdf))

and pt() is the function for computing the cdf of the Student's t
distribution with xdf degrees of freedom.

What is important is not the p-value distribution but the formula for the
asymptotic variance of the garch estimates. In splus the default is to use
the inverse of the numerical Hessian of the garch likelihood. Typically this
is the Gaussian distribution. However, S-plus has the option of computing
the Quasi maximum likelihood asymptotic variance based on the so-called
sandwich formula.

In my experience garch estimates vary considerably across software packages
(see the review articles by Chris Brooks for more details). This is mainly
due to different optimizers, different convergence parameters etc. As a
result, one would not expect the p-values (or estimates) to match exactly
across different software.

与一般的回归模型的显著性检验一样，可参考任何计量经济学书。

2# 耕耘使者
那为什么 看matlab的程序 对于garch模型的系数的统计量的构造怎么会与hessian矩阵有关呢？（是通过hessian矩阵来计算）

模型的估计到是比较简单~
关键是这个系数的显著性检验的统计量与p值如何求出来！~

还希望各位大哥大姐帮帮忙啊！~