skipped/missing data

[wendyyuan]

Hi, dear friends,
Anyone is familiar with Quadratic ARCH model?
I am running bi-variate Quadratic ARCH model, and i use Maximise command to run it with the variance formulas:

h11=c0+c1*u{1}**2+c2*h11{1}+c3*u{1}
h12=d0+d1*u{1}*v{1}+d2*h12{1}
h22=e0+e1*v{1}**2+e2*h22{1}+e3*v{1}

i can get convergence by using either BHHH or BFGS method, but i always have more than half skipped/missing data. I have tried to change different initial guess, tighter cvcrit, higher sub iterations, but it still did not work.
Is there anything wrong with programming or some other mistakes?

Any help please!

[TomDoan]

That will be very sensitive to guess values since the unsquared residuals in the two variance equations might cause those to go negative if the coefficients are too large. My suggestion would be that you first estimate this with c3 and e3 pegged to zero, then release them only when you have (reasonably) well converged estimates of the rest. That is something like:

Code: Select all

nonlin(parmset=garchparms) c1 c1 c2 c3 d1 d2 d2 e0 e1 e2 e3
nonlin(parmset=pegs) c3=0 e3=0
maximize(parmset=garchparms+pegs+meanparms,iters=20,other options) logl
maximize(parmset=garchparms+meanparms,other options) logl

[wendyyuan]

Thanks a lot. I put c3=e3=0 and readjust other initial guesses, and it turns out that the skipped/missing data is only 1. Appreciate your help