On Thu, 6 Jul 2017, Periklis Gogas wrote:
On Fri, Jun 30, 2017 at 6:17 PM, Allin Cottrell
> On Fri, 30 Jun 2017, Periklis Gogas wrote:
>> I run an AR(10)-GARCH(2,2) model just for an example using the included
>> data file djclose.gdt
>> I run the following:
>> *Model 1:*
>> Model>Time Series>GARCH Variants and got this:
>> [image: Inline image 1]
>> *Model 2:*
>> Model>Time Series>GARCH and got this:
>> [image: Inline image 2]
>> Why do I get so different results on the same data and model? The
>> results are very different in both the mean equation and the GARCH
>> part. They are both an AR(10)-GARCH(2,2) in the logs.
> I wouldn't say the results are very different: they're qualitatively
> similar and both sets suggest an over-parameterized/misspecified model.
First of all thank you very much for the response!
I selected these models jut to show this difference they were not the
product of any model selection procedure.
> Gig finds a slightly higher log-likelihood;
What is "gig"?
"gig" is "Garch in gretl", the addon package which supplies the
"GARCH variants" menu item.
> the built-in garch command warns that the norm of the gradient
> "convergence" is too big.
Where can I see this?
With current gretl (the last release is 2017b, from May of this
year), the warning is printed under the GARCH estimation results.
Ah, but I see the message is not shown in the GUI model window, only
when the garch command is executed via script or in the gretl
console -- that's something we should fix. This script will show the
garch 2 2 ; l_djclose 0 l_djclose(-1 to -10)
shows: "Warning: norm of gradient = 4.84663". The norm of the
gradient should be much smaller than that when convergence on the
MLE has truly been achieved.
> Apparently there is not a well-defined MLE.
Thank you very much and sorry for the possibly stupid questions.
They're not stupid questions, but complex nonlinear estimators are
not guaranteed to work well when the model is misspecified and the
MLE is either non-existent or hard to find.