Model fit


I have a question concerning the evalutation of the model fit. I perform Bayesian estimation of the model parameters with four observables. Then I compare unconditional moments from the model with those from the data. The standard deviations of observable variables in the model are all higher than in the data. Some of the correlations from the model are quite close to those from the data, but some even have an opposite sign.

How should I proceed if my model fit looks quite poor. Can I still use the model for policy Analysis or should I re-estimate the model?


Or alternatively - if I try to provide an Explanation which features the model lacks (I use a quite stylized Setup) and give an Intuition why it doesn’t fit the data well, could it be a solution to the problem?

I am going to assume your estimation is correct and found the mode. In this case, your model seems to be misspecified along some dimensions you seem to care a lot. In this case, if you can point to a particular parameter that you think seems off, you can tighten the prior on this parameter. But more generally, there is a problem with the model. If you know why/which features are missing, it is recommended to add these features rather than relying on an obviously wrong model for policy analysis. The principle is: garbage in, garbage out.

Thank you for your reply!

Yes, the estimation is correct - mode was found and convergence is also good. I compared the theoretical moments that I got after stoch_simul with those from the data that I used for estimation.
I will probably also try to use the option moments_varendo, but I suppose this should not change a lot?
I suppose, the problem could arise also because of calibrated parameters (those that I fix before estimation)?

My focus is here on policy analysis, I want to show how the policy implications of the model change when I change some assumptions. Then, the parameterization is probably not crucial on itself?

Of course the parameters matter a lot. They are the objects you estimate to fit the model to the data (the degrees of freedom you have)