irf create run time

[Jiansheng Zhong]

I am running Stata/SE 12.0 on a 2013 iMac.

After running "svar" on 12 variables with 6 lags (there are a total of 105 observations), I used the command "irf create model1, set(myirf, replace) step (30)" to generate the irf results. I waited for about 15 hours and then gave up.

I then reduced the lags from 6 to 4 for the "svar" procedure, and reduced the steps from 30 to 20 in the "irf create" command. The program has been running for 7 hours and is still running.

Is such a long time for "irf create" normal? Do I just need to be patient and wait longer, or did I do something wrong?

Many thanks!

[Daniel Schneider]

Hi Jiansheng,

A 12 variables system is very large for a SVAR model, much larger than what you typically find in the literature. Be sure that the model is identified and converges reliably.

Regarding irf, the computation time depends on the number of equations, the length of the forecast horizon, and the number of lags. In particular, the computation of the asymptotic error bands may take long. You could try the following: Use the nose option of irf to avoid calculating the asymptotic error bands, and set the forecast horizon to a small value, say, 5. If this converges quickly, increase the forecast horizon up to whatever you require, 20 or so. If this is still reasonable quick, say, under a minute, get the error bands by means of a bootstrap (options bs or bsp) which should then be possible within 1-3 hours.

Best,
Daniel

[Jiansheng Zhong]

Hi Daniel,

Thanks so much for your reply! Your information and suggestions are very helpful.

I have tried the "nose" option and the results came out almost immediately.

I have also reduced the number of variables from 12 to 8 and used the "bs" option in generating the impulse response functions. I got the results in about 3 hours. However, without specifying "bs", the program did not yield anything and looked like it actually crashed (every time).

Although I am still not sure why the irf create command crashes on this particular dataset (it works on a different dataset, which has 9 variables and 2 lags), I can at least get the IRFs using the "bs" option, which is good enough for me.

Thanks again!

[Daniel Schneider]

Hi Jiansheng, this is great, I am glad you could get things to work.