Strange Estimation IRFs

Dear Dynare Forum Members

I cannot understand why the IRFs for the Bayesian Estimation of my code gives such strange results (please see attached example). I have run dozens of estimations and get good acceptance rates, good prior-posterior distributions, and good diagnostic check results. There are some steady-states that are non-zero, but these numbers are very small (approx. e^-16); does this make any difference to the starting point of the IRFs?

I have double and triple-checked the data and code (although there’s always the element of human error); and the estimated parameters are quite consistent with the literature. Thank you for your time.

Regards,
Hylton
MPshock.pdf (11.4 KB)

Why do you think they look strange?

Hi jpfeifer

Thanks for the reply.

Well, one: IRFs fluctuating around the steady-state, and two: not starting from the steady-state equal to zero. In addition to having good diagnostic results, prior-posterior distributions and literature/theoretically consistent paramter values, my question being two parts: does this most likely imply something wrong with the model setup, or could it be that I struggled with a non-positive definite matrix for some time before being able to run the estimation through mode_compute=4 (i.e. I used option 6 to get mode values then run through option 4) and therefore have poorly specified priors or calibrated variables that I am not estimating?

I know your answer cannot be too specific, but any advice from your experience would be greatly appreciated.

Regards,
Hylton

P.S.

When I run through dynare 4.3.1 I get this warning sign which I did not get before: Warning: File 'mh_scale_fname' not found.

You can ignore the warning. I still don’t understand the problem. IRFs fluctuating around the steady state, i.e. overshooting and returning is relatively common, even in the basic RBC model. And why should IRFs start at the steady state if there is a shock at t=0 that drives the system out of steady state?

Hi jpfeifer

You are most certainly correct. I think the fact that I am getting such counterintuitive results (i.e. productivity shock generally good but MP shock always “incorrect” with increased interest rates and increased inflation, although the setup of the model is New-Keynesian) thus making it hard to interpret - or make sense of - the results.

Thanks for the reply.

Regards,
Hylton

Does your model with a sensible baseline calibration yield good results? If not, the problem is not with estimation, but with the model.

Hi jpfeifer

I managed to sort the issue out. I specified two equations slightly differently and then was able to run with mode_compute=4. Results are good now.

I have one quick question on the output of figures in dynare 4.3.1: is there any way that I can generate .fig IRF figures instead of (or in addition to) .eps figures after I run an estimation? The previous version of dynare 4.3.0 gave .fig files.

Regards,
Hylton

See the manual on graph_format

Awesome, thanks. Appreciate your help.