Time varying SVAR

[dakila]

No. Could you post the workfile?

[asif_1210]

hi, i'm trying to restrict Gibbs sampling iterations but program is still running for 1000 iterations, how do i control for iterations?

[dakila]

There is a bug. You can use the dialog interface instead of the command interface. I will fix it soon.

[sakisgw3]

No. Could you post the workfile?

Thank you for your response.

I have attached the workfile. I experimented a bit with the example and this error does not occur when i include three time series in a var. However, after i tried to run my var(3) at the end i get a message saying "gm01 is not the same length as the sample" and no figures are presented.

EDIT: I forgot to include the date vector. You can use this: svector date=@wsplit("2009:03 2013:07")

[dakila]

You are right. Try to put at least 3 variables.
If there are missing observations with NA then it will give you the error ("gm01 is not the same length as the sample").
So adjust your sample. For example, use pagestruct command or click on the range.
I will fix this bug soon.

[dakila]

The tvsvar add-in is updated. It includes some fixes.

[asif_1210]

Hi, Thanks for updation, but it is still giving some msg like 'GM01 is not the same lenght as sample lenght'.. What does this mean can anyone help please?

[asif_1210]

Ok now its working. I figured out the problem. But could you please guide me about IRs. Are they accumulated or not? How can I have data series of impulse responses generated through TV SVAR.

[nadybe]

Hello Davaa,
I have some problems with my data, i mainly use stationary variables and your program returns a near singular matrix error, i also tried using data in level (and transformed in logarithm), reducing the number of endogenous variables, changing the range but the problem remain the same. Please see my workfile below. For your information the impulse variable is the global liquidity indicator.

Thank you davaa.

EDIT: I made a mistake concerning the Global liquidity indicator so i fixed it. I add also the growth rate version and HP filtered version of my data, which return the near singular matrix error too.

tvsvar_first_country.wf1

(74.22 KiB) Downloaded 1161 times

[nadybe]

I forgot to specify something, The forex and reserves are basically the same variables so you should include just one of them in your specification.

Thank you davaa.

edit: I also transformed the data into growth rate but i get the same error (near singular matrix)

EDIT 2: I managed to launch the program with my data (stationary and in level) but i have new error after the iteration ends : "size do not match in matrix function" WHAT SHOULD I DO TO FIX THIS???

Best regards

[nadybe]

Hello Davaa
I managed to get results, but i have one last question, What is the maximum endogenous variables you can use with your program, i have out of memory and sometimes near singular matrix errors when i put more than 4 endogenous variables in the program. Also the date selection vector cannot support more than approx 7 dates for me, at least considering my data.

Best regards

[dakila]

Actually there is no restriction on maximum endogenous variables. But It is very computationally demanding since it uses MCMC method. That's why you got out of memory and sometimes near singular matrix. So use a computer with core i7 and 8 gb ram.

You can use the command interface instead of dialog interface. There is no restriction on the date selection vector. For example, use the following command:

Code: Select all

svector dtsel=@wsplit("1975q1 1981q3 1996q1") tvsvar(comp=1) 2 40 dtsel interest @ inflation unempl interest

[asif_1210]

Hi, I'm also facing the same issues, but my computer have 8 GB ram and its i5. Does it still too sensitive for computations since I'm also getting near singular matrix problem with six endogenous variables.

[dakila]

I can not solve the near singularity problem. It is related your research, more specifically your choice of variables and the prior parameters.

[nadybe]

Thank you davaa. Actually i was already running the program using the Line of command. I am using à virtual computer with Intel core i5 and 16 giga of ram, but i think that using windows as à virtual machine maybe limits its capabilities. I'll try on different computer to see if that changes someting. I got 2 more question, do you have to change the sample size on the Line command? How do you know what are the Best priors suitable with your data? As i am working on the global liquidité topic i know that nobody ever used à tvpvar before so i cannot based my priors on previous research.
Best regards