EViews.com

Hello!

Recently used optimization algorithms aren't the best. Possiblity to write your own function to optimize is very nice, but there are problems with initial values and convergence. I had lots of cases when eviews didn't find the solution because there was a problem with initial values or with convergence. Other econometric programs had no problems to find the solution (of course parametrization was the same like with eviews).
I strongly believe that there should be an upgrade of your optimization algorithms or at least add some functionality to these algorithms. I think for instance: during optimization if convergence isn't achieved eviews will change initial values and try to solve equation one more time to achieve convergence. I completely realise that problem with finding a global minimum (or maximum) isn't the simplest one, but I believe it is worth searching for better solution.

What do you think ?

Thanks in advance:)

Can you supply some more details as to what part of EViews you are working with?

EViews uses a variety of optimization algorithms and starting value calculations in different areas. It would be helpful if you could provide more information so we can determine what portion of EViews you have experienced problems with.

I especially talk about likelihood functions. For instance to fit Skew Generalized Error Distribution (or other more sophisticated distribution) to data or implement this distribution to GARCH model, it isn't as simple as it should be. I mean that eviews has problem to find proper values of parameters. When I changed likelihood function to another parametrisation (which was of course similar to previous one) than eviews found the solution. It was suprising to me because other econometric programs don't have this problem. After years of using eviews and working with likelihood functions I noticed that eviews is considerably sensitive to initial values of parameters and to parametrisation given to optimize.

Just to be clear, you're talking about the LogL object?

Yes, the LogL object.