Numerical noise in the likelihood

This forum is closed. You can read the posts but cannot write. We have migrated the forum to a new location where you will have to reset your password.
Forum rules
This forum is closed. You can read the posts but cannot write. We have migrated the forum to a new location (https://forum.dynare.org) where you will have to reset your password.

Numerical noise in the likelihood

Postby andwes » Mon Sep 26, 2011 12:23 pm

When running th estimation command for a medium-size DSGE model, I get the message "Numerical noise in the likelihood". What kind of problem is that (from a numerical point of view), and it is an indication that there is some type of problem in the model I am using ?

Best

Andreas Westermark
andwes
 
Posts: 20
Joined: Wed Mar 17, 2010 2:17 pm

Re: Numerical noise in the likelihood

Postby rattoma » Wed Oct 19, 2011 8:11 am

Hi,

I guess this is displayed with mode_compute = 5. In that optimizer, some numerical checks are performed when computing the gradient. In particular we try to tune the perturbation of each parameter (say dx(i)) such that the change in the log-likelihood (df) is in the order of 1.e-5. Typically df versus dx(i) is monotonic, so it is quite easy to tune dx(i), up to reasonable numerical precision. When this is not possible, usually the estimation is near to some unit root or prior boundary, then the precision for df is relaxed to higher values (1.e-4 1.e-3 and so on) and the warning is displayed.

This warning may appear in some iterations, then if the optimizer manages to escape the bad region, it disappear, otherwise, in most cases, your optimization will end-up in some boundary.
rattoma
 
Posts: 51
Joined: Tue Jan 04, 2005 10:48 am

Re: Numerical noise in the likelihood

Postby GilBurt » Fri Oct 28, 2011 11:06 am

Thanks for this good discussion forum...
GilBurt
 
Posts: 4
Joined: Fri Oct 21, 2011 11:44 am


Return to Dynare User Guide

Who is online

Users browsing this forum: No registered users and 0 guests