Shocks contribute to FEVD but little to shock decomposition

[Jesse]

Dear All,
I have estimated my DSGE model, I have generated both in-sample shock decomposition and out-of-sample forecast error variance decomposition, i find some shocks contribute a lot to out-of-sample forecast error variance decomposition but contribute very little to in-sample shock decomposition, i am wondering that is it normal or does it indicate model misspecification?
Thank you very much and look forward to hearing from you.
Kind regards,
Jesse

[jpfeifer]

That is not completely unheard of, but quite unusual. Usually the two should quite coincide (at horizon infinity for the FEVD). This might indeed be a sign of misspecification

[Jesse]

Dear Johannes,
Thank you very much! Does this mean my DSGE model will be useless for forecasting?
thank you and look forward to hearing from you.
Kind regards,
Jesse

[jpfeifer]

I would tend to say that is the case. Of course, you can always argue that the in-sample realizations of the shocks were special and that the FEVD, which cares about theoretical/asymptotic properties is a more realistic guidance. But most people would probably reject this argument, saying that if it does not work in-sample, it will not work asymptotically.

[Jesse]

Dear Johannes,
Thank you very much for your helpful guidance. in my model, there are 20 structural shocks and 20 observed variables, only 2 of them encounter such a problem,
in shock decomposition the shock X accounts for 5% of observed variable Y, but in forecast error variance decomposition, shock X accounts for 20% of observed variable Y's variation.
Since the number of shocks are relatively large-20, and only 2 shocks encounter such problems, does this mean the misspecification problem is less severe?
Thank you very much and look forward to hearing from you.
Kind regards,
Jesse

[jpfeifer]

That still sounds reasonable. The last factor to consider is how long your sample is. If it's rather short, then you should be fine.

[Jesse]

Dear Johannes,
Thank you very much.
I just have one last question. My in-sample size is 90 quarters, and my forecast horizon is 8 quarters. is 90 quarters a relatively short in sample or not?
Thank you again and look forward to hearing from you.
Kind regards,
Jesse

[jpfeifer]

Yes, that is short. And forecast horizons smaller than infinity in the FEVD cannot easily be compared to the historical decomposition results. Horizon 8 is rather small. Or were you talking about the actual forecasts you are trying to make? In that case, forget my last comment.