Using the Unix version of dynare++ 1.3.5, I find that estimated the IRFs change slightly every time I run a model file and use a nonlinear approximation. This occurs even though I am using the same seed and same # of simulations each time.
Is there a way to avoid this?
yes, that happens. It is described on page 16 of the tutorial, look for ‘–seed’.
I have already started working on this in my working copy, so hopefully, it will appear in the version 1.3.6. Each thread then will get its own random generator instance. Until that time, if you despearatelly need your simulations being identical, say ‘–threads 1’.