Manual versus built in Maximum Likelihood: Different results

[strypste]

Hi

I would like to estimate an AR(1)ARCH(1) model. The program I attached does this in two ways. One way is with the built in command and the other is by creating an LogL object. I set the options of these procedures so that they are exactly the same. Unfortunately I do get slightly different results and I do not see why. Does somebody have an idea?

Best
s

[strypste]

Oh I forgot to mention the following.

One difference I can see in the output is the derivs option. For the built in command it says "accurate mixed (linear)" and for the manual method it says "accurate numeric". For the built in command I have set the derivs option to accurate numeric derivatives and always use numeric. For the manual built in I use the default.

Hope this helps..

Best
s

[EViews Gareth]

I think the most likely cause is presample variance issues. If you change your estimation command to this:

Code: Select all

equation eq1.arch(1,0,showopts,m=1000,c=1e-8,s=0,z,derivs="an", backcast=0.000001) y c ar(1)

you get much closer to the logl.