System Proc Make Loglikelihoods returns zeros for CCC-GARCH

[JimS]

Eviews 8

I have a 3 x 3 garch(1,1) model (no threshold term), with four exogenous variables in the variance equation, and some dummy variables and AR terms in the mean equation. I can estimate it fine (Marquardt / Normal) using both the in-built DVECH (rank 1) and CCC system ARCH routines. For the DVECH I can get the loglikelihood contributions, when I access Proc / Make Loglikelihoods, but for the CCC model, I get a sequence of zeros. When I drop all but the constant in the mean equation and the garch(1,1) terms in the variance /covariances, and estimate both specifications, I get loglikelihoods produced in both cases. Is this a bug, or is there something about the production of the loglikelihoods in the CCC model that can't be done when there's exogenous variables in the variance, or more than a constant in the mean equation? Thanks.

[EViews Gareth]

Could you post the workfile?

[JimS]

Thanks for the prompt response. Workfile attached, but it doesn't have the system spec saved within it, so I have included the two "stats" output files (.csv) that were saved out from the results of the two ARCH system estimations.

In both cases, I used the default Marquardt algorithm with Normal errors and all other estimation "options" at the default values.

The 3 dependent variables are rsh, rlg and rft. The mean equation has c, Tues, Wed, Thur, Fri, Indexchg and 5 lags of each of rsh, rlh and rft, so 21 coefficients in each mean equation.

The variance is GARCH(1,1), with nrock, d1, d2 and d3 as "individual" added regressors.

The two correlation specs are CCC (no variance targeting), and DVECH (rank one restriction on all matrices).

Thanks.

[EViews Gareth]

Could you save a copy of the workfile with the system in it? I want to be 100% certain I match your specification.

[JimS]

Workfile with the two systems included, as you can see there's just zeros for the loglikelihoods in the CCC case.

Thanks.

[EViews Gareth]

Thanks. We'll take a look. I'm reasonably confident it is a bug. Note that if you change the order of the variance regressors to be D1 D2 D3 NROCK, it appears to work properly.

[JimS]

Thanks so much, that's a simple fix for me, I hope that you are able to find the "bug".

[JimS]

Actually the loglikelihoods for the CCC model (when the variable ordering is changed) appear rather too variable to be believable. Changing the variable ordering for the DVECH does not change the loglikelihoods to any real noticeable degree. For the CCC model, there are massive negative outlying values on somedays. I got nothing like this instability, when the model did not include all the extra variables in the mean and variance specifications. So, while there are now loglikelihoods produced for the CCC model when the variable ordering in the variance spec is changed, I'm not sure they're going to be useful for comparison to the DVECH. I will try other orderings to see if they are more stable.

[EViews Gareth]

The latest patch has a fix.

[JimS]

Thanks. Patch installed; problem solved. Apologies if this patch has been available a while and I had you revisiting a problem you'd solved some time ago.

[EViews Gareth]

Nope, we found the big and fixed it just because of your post - thanks.