思路清楚了不就很简单了吗，一般用AR（1)-GARCH(1,1) 拟合正态部分足够了，再用Hermit偏离，所以下面我下面的代码只调控Kz 和Kx 两个关键参数。其中子函数Hermit，你自己动动脑筋，相信能够解决
function [Loglikelihood,BIC,Density,Lnf]=SNP(data0,data,Kz,Kx,param)
%% Illustraion: This code was written to calculate the 'real' density
%% on the base of hermit polynomial. As Gallant and Tauchen(2002) suggest
%% we only use AR(1)-GARCH(1,1) for nomal part and control the non-nomality
%% through Kz and Kx!
%% Warning:Copyright belongs to LvLong in School of Economics, HUST!
u0=param(1);
u1=param(2);
b0=param(3);
b1=param(4);
g1=param(5);
param_poly=param(6:end);
Num=length(param_poly);
if(Num~=(Kx+1)*(Kz+1)-1)
    error('The length of param-vector is not right');
end
Y=data;
%%%%prepare something%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%
T=length(data);
Lnf=zeros(T,1);
Density=zeros(T,1);
Mu=zeros(T,1);
R=zeros(T,1);
Z=zeros(T,1);
R0=sqrt(var(data));
Mu0=mean(data);
Z0=(data0-Mu0)/R0;
%%%%%%%%%start the major cycle%%%%
for t=1:T
    if(t==1)
        Mu(t)=u0+u1*data0;
        R(t)=b0+b1*abs(data0-Mu0)+g1*R0;
        Z(t)=(Y(t)-Mu(t))/R(t);
        hvalue=Hermit(Z(t),Z0,Kz,Kx,param_poly);
    else
        Mu(t)=u0+u1*Y(t-1);
        R(t)=b0+b1*abs(Y(t-1)-Mu(t-1))+g1*R(t-1);
        Z(t)=(Y(t)-Mu(t))/R(t);
        hvalue=Hermit(Z(t),Z(t-1),Kz,Kx,param_poly);
    end
   Density(t)=hvalue/R(t);
   Lnf(t)=log(Density(t));
end
Loglikelihood=sum(Lnf)/T;
vartotal=length(param);
BIC=-Loglikelihood+0.5*log(T)*vartotal/(T);

如果程序运行成功，可追加论坛币~

有大神会么

啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊

不能自己写吗，实现起来不难的，EMM关键弄懂什么是SNP(Semi-nonparametric) density , SNP 关键是写出hermit多项式展开，hermit多项式展开最难的是底部的标准化常数的计算，你就按这个思路来计算SNP，然后其余的就是对AR和GARCH模型定阶，矩条件做模拟大样本的GMM。纯个人体会，希望对你有帮助