Results.mat size

Estimation
1

Dear Dynare Team,
When I run one stoch_simul or one estimation, each one including a lot of replic/periods or chains/replications, respectively, the results.mat generated is incredibly heavy compared with the current Dynare version to, let say, 4.4.3 version. Of course I speak about similare options for each stoch_simul or estimation blocks.

  1. Is there a way to substantially increase the size without removing essential results?
  2. Does such optimization could improve simulation/estimation time?

Best,
Jonathan

2
  1. Which options are you using? The largest new arrays that are stored are related to variance decompositions and forecast error variances in smoothers.
  2. As a general rule, you should only request the options that you really need. Deleting variables after they have been computed should not speed up things by much.
3

Hi Johannes,
It seems I badly formulated my questions.
Let’s assume I have one estimation with Dynare 4.4.3 with the estimation command line A. The same estimation (ie the same mod file with the same command line A) with Dynare 4.5.6

I obtain the result.mat generated from Dynare 4.5.6 is heavier than from Dynare 4.4.3.

  1. Is there a way to substantially decrease this size (ie from Dynare 4.4.3) without removing essential results?
  2. Does the Dynare Team optimizes/aware about this issue?
  3. Is there an optimal way to decrease this size and then improve estimation computation time ?

Best,
Jonathan

4

Could you provide me with the two files to check what is going on?

5

Of course Johannes. The attached files are slightly modified versions of SW2007.
I run this mod file with this data file with Dynare 4.4.3 and got a results.mat file of about 10Mo
I run this mod file with this data file with Dynare 4.5.6 and got a results.mat file of about 126Mo

How can you, or someone else, explain this substantial difference ?

Best,
Jonathan

image

usdata1.xlsx (16.2 KB)
us11.mod (9.0 KB)

6

Hi,
starting in 4.5, the filter_covariance option works with the MCMC. See https://archives.dynare.org/DynareWiki/NewFeatures
That will create huge matrices when you specify filter_step_ahead. This should explain the difference.

7

Hi Johannes,
I am sorry to continue this conversation but I run the same model over the 2 Dynare versions again but without the filter_step_ahead command in the estimation function.

I still have a difference of almost the same order between the Dynare versions.

image

Do you have an idea why? And how I can reduce again this file size?

Best,
Jonathan