GARCH-BEKK

[EDIT: 20110617 09:06 CDT - reformat - WDR]

Hi,

I want to evaluate the volatility spill over between bonds, cds and equity using company data.

However, I have a problem with my GARCH BEKK model. I used UCSD toolbox, and followed the following steps for the estimation of the model. Built a ARMA model and obtained the residuals, then demeaned the residuals and run the GARCH BEKK model. Everything is fine so far, but the problem is that I get insignificant results for the coefficients that reflect the volatility spillover.

This is my code:

%ARMA fiting
[parameters, errors, LLF , SEregression, stderrors, robustSE, scores, likelihoods]=armaxfilter(rstock,1,1,1)

%Ljung Box test for serial correlation
[H,pValue,Qstat,CriticalValue] = lbqtest(errors,[5 ,10],0.01)
resstock=errors
resstock=resstock-mean(resstock)

%ARMA fiting
[parameters, errors, LLF , SEregression, stderrors, robustSE, scores, likelihoods]=armaxfilter(rbond,1,1,1)

%Ljung Box test for serial correlation
[H,pValue,Qstat,CriticalValue] = lbqtest(errors,[5 ,10],0.01)
resbond=errors
resbond=resbond-mean(resbond)

%ARMA fiting
[parameters, errors, LLF , SEregression, stderrors, robustSE, scores, likelihoods]=armaxfilter(rcds,1,1,1)

%Ljung Box test for serial correlation
[H,pValue,Qstat,CriticalValue] = lbqtest(errors,[5 ,10],0.01)
rescds=errors
rescds=rescds-mean(rescds)

x=[resbond rescds resstock]

%GARCH BEKK estimation

[parameters, loglikelihood,likelihoods, stdresid, stderrors, A, B, scores]  = full_bekk_mvgarch(x,1,1)

%%%%%Test for adequacy

stdresid2=stdresid.^2   %square of the std resid
%Ljunk Box Test
[h,pValue,stat,cValue] = lbqtest(stdresid2,20) 
qqplot(stdresid2)

%%%%Extracting std errors from A
st=diag(A,0)
st=sqrt(st)

Could anybody help me?

Answer 1

[Junjun]

Why does this coefficient ought to be significant?

[Javier]

Normally you get the error of the coefficient estimated with the sqrt(diag(hessian(something))). Don not what is A.

Answer 2

[Javier]

Hello Kaloyan

I just start checking UCSD GARCH toolbox. Once you get the H value for the lbqtest of the square residuals equal to 0, it means that the model is ok (UNIVARIATE). To get the Significance of the parameters I use this formula:

parameters/sqrt(diag(A)).

Now you have to test significance in the multivariate sense. There is no function in Matlab or MFE toolbox for that purpose. In this book you can find a multivariate test (New introduction to multiple time series analysis (Luetkepohl 2005) Hope this helps

Answer 3

[Javier]

Hello Kaloyan

I just talk with professor Sheppard. There is a new toolbox MFE with the same functionalities. Erase UCSD Toolbox and prove with the new one.

Best regards

Mauricio