A Question about Estimation Result and Impulse Responses

Hi Everyone,

Recently, when I estimate a model similar to Smets and Wouters (2007) using my own codes, the posterior mode I get seems a little strange to me. Specifically, when there is a positive technology shock, the level of output falls below its stochastic trend immediately and persistently increases to a new trend that is higher than the original trend. In other words, output falls given a positive TFP shock. This seems to be inconsistent with the standard results in the literature. But it is not always like this. When I change to another vector of parameter values which is not necessarily the mode, I get the standard impulse responses. This is obviously related to the parameter values. And I think imposing some kind of sign restrictions during the estimation would rule out those unreasonable parameters. And this is just what people do in the sign restriction literature of SVAR. But I am not sure if imposing these sign restrictions has been routinely done in the DSGE estimation, especially behind Dynare.

Thank you!

No, this is not routinely done and not straightforward to implement in Dynare. There may be parameter combinations where the inefficiency compared to the flex-price economy leads to an output drop. The aim of full information estimation is to minimize the forecast errors. You are suggesting to augment this by an additional check based on conditional IRFs, which kind of contradicts the FIML you are otherwise doing. On the other hand it would be similar to imposing a unique stable saddle-path solution so it could be justified. But as in the case of sVAR sign restrictions, you must be sure that the sign restriction you are imposing is uncontroversial.
Moreover, are you sure the mode and the model are correct?

Thanks for your reply, jpfeifer! I think maybe it is a good idea to double check my model and estimation results by doing it again in Dynare and see if the problem still exists. Let me postpone discussing this with you until finishing that.

Thank you again for your prompt responses!