Dear all,

I have one questions regarding the implementation of the new block in dynare for SMM/GMM. In fact, I find it quite convenient simulation tool for moment comparison. However, I have one question regarding the interpretation of the ‘black box’. Usually, the new block could always search for the best optimized results within the given parameterized boundary, however, since the given boundaries are kind of subjective and it could always rule out the continuous optimization process like SMM, which contributes to no standard errors calculated, I am wondering how to interpret the given results when the boundary does limit the optimization. Can I still interpret it as SMM results since it’s optimized within a given region ?

Best

The interpretation is a bit tricky. It’s probably best to think about it as a kind of Bayesian approach where your prior is flat but involves parameter bounds. But in strictly frequentist terms, the bounds may indeed be a problem.

Professor, is there a way to match with data correlation not covariance in match_moment block with matlab ? It seems that the SMM block could only match with covariance but not correlation since there is only demean option for data but no normalize option for the simulated data ? For instance, if I want to match a series autocorrelation ACF(1) not cov(y,y(-1)). Is there any way to do with it ?

No, that is not possible as the standard errors cannot easily be computed. But the autocorrelation is just a function of the autocovariance and the variance.

Thanks, professor! Just one more question. Do you know why my dynare crash sometimes when it’s running to the end of printing and suddenly quit the system ? It happens lots of time. My code should be no problem. This just happens by occasion. Sometimes, it doesn’t happen but sometimes it does.

Without any more specifics/details about the problem it’s impossible to tell, particularly if it’s not reproducible.

Professor, can I ask you one question regarding the new block of SMM estimation ? I am using it for estimation and obtain some result from it, but when I tried to use the estimator in my own stoch_simul. I find that it cannot replicate the data and model comparison as the estimation shows. Where is my problem ? The log shows the estimation result and the New5.mod is the file I am trying to replicate the log result. By adjusting the log parameters, I find the SMM reports result and the New5.mod result have large difference.

ClimateDSGE_EZ_LRR_SMM5.log (48.9 KB)

ClimateDSGE_EZ_LRR_SMM5.mod (12.2 KB)

ClimateMacroData_SMM_EZ_LRR1990.mat (332.8 KB)

New5.mod (11.2 KB)

None of the two mod-file can be run due to missing files and macro-processor settings. What exactly is the problem?

Sorry，professor. I tried to upload the inc file but it doesn’t work. The problem is that I do Smm with the estimation of the new block and it generates the results as the end of the log file. However when I give the estimator into the simulation code in New.mod,I find that the estimation matched moments cannot be generated and there’s large difference. Very confused where is the problem. Is this because of incorrect parameter settings? Or something else?

You should now be able to upload the `.inc`

-file.

Great. It’s uploaded professor. So, basically, New5.mod should be the same model as the inc. I just input the parameter estimation result from log file but find that I cannot replicate the log file comparison between model and data.

Climate_EZ_LRR_common4.inc (10.2 KB)

find_climate_steady_state_LRR_EZ.m (985 Bytes)

find_ss_kd_habit_LRR_EZ.m (1.2 KB)

find_ss_theta_habit_LRR_EZ.m (801 Bytes)

Sorry, forget these three for steady state codes.

Best

Ethan

What do you mean with

?

How did you compare the moments? The ones reported by `stoch_simul`

are centered, the ones by `method_of_moments`

are not.

1 Like

Thanks for your reply, professor. Please correct me if I make mistakes. I do have non-stationary component in the neoclassical model so I normalize it by the long run non-stationary TFP. When I compare it, I just compare the manually centered data with normalized stationary variable to make sure that the centered second moments in estimation could be matched. Maybe, there are some unobserved mistakes here, but this is the logic I am using. Please correct me if my thought is wrong.

Then please provide the code you use for the comparison.