Gauss-Newton method in ARIMA
Posted: Tue Oct 11, 2011 10:18 pm
by kiber_master
Hello!
In User's Guide it is said that I can choose Gauss-Newton method for estimation NLS model. How exactly may I do that? I can't get an option.
And if your default method called "Marquardt" is it the same as Wiki said (
http://en.wikipedia.org/wiki/Levenberg% ... _algorithm)? Or you have more consistent modification?
Thank you!
Re: Gauss-Newton method in ARIMA
Posted: Mon Oct 17, 2011 11:14 pm
by kiber_master
Please, help me! I want to set Gauss-Newton method for ARIMA estimation!
Re: Gauss-Newton method in ARIMA
Posted: Tue Oct 18, 2011 10:39 am
by EViews Chris
EViews currently always uses a Marquardt style step for ARMA (NLS) problems.
As a broad generalization, this tends to work better than a pure Gauss-Newton with linesearch approach, so we currently don't provide an option to switch.
Note that the optimization method used does not affect the final results (unless the method is failing to converge or you are finding one of several local optima - both of which can also happen when you use different starting values).
The wikipedia entry you reference doesn't really contain enough details to describe an implementation. It's the right general idea, but you really need to specify how you're going to adjust the size of lambda in the diagonal adjustment before you could implement things.
For what it's worth, EViews v7 follows a more ad hoc style adjustment of lambda, more like the code in Numerical Recipes rather than a formal trust region method as described, for example, in the More & Sorensen paper listed in the references.
If you could provide some more information as to what you're trying to acheive, I might be able to give you some more details.