Problems with forecasting from VAR-Models

[MiWe]

Hello, I am using EViews 12, the latest version. There are problems with the prediction of VAR models. When I estimate the same model several times, I get different results: When saving the estimation results in an Excel spreadsheet or internally in a matrix, different columns give different results. There are always two different results that are split between the columns. One variant is (with e.g. 9 identical estimates) always in the columns 1,3,7,8. The other variant is in the columns 2,4,5,6,9. This is very strange and I cannot use these calculations (and selection calculations with the RMSE). Do you have any idea what could be the reason for this? I would be very grateful for any help! Many greetings, Michael

The estimates look like:
smpl 1995:1 2022:3
var var1.var 1 5 y x
var var2.var 1 5 y x
smpl 2022:4 2023:4
var1.makemodel(mod1)
solve mod1
series y1 = y_0
var2.makemodel(mod2)
solve mod2
series y2 = y_0
write M:\test.xls y1 y2

The ouput look like in the attachment.

[EViews Gareth]

Could you provide your workfile?

A simple test indicates the results are identical:

Code: Select all

wfcreate q 1995 2023

series y=nrnd
series x=nrnd


smpl 1995:1 2022:3
var var1.ls 1 5 y x
var var2.ls 1 5 y x
smpl 2022:4 2023:4
var1.makemodel(mod1)
solve mod1
series y1 = y_0
var2.makemodel(mod2)
solve mod2
series y2 = y_0

show y1 y2


[MiWe]

Hallo Gareth,
Thanks for your reply! Your example works without problems... Attached you will find my workfile (sheet "quarterly" is relevant) and a simple program-example where I get different results.
Thanks a lot, Michael

[MiWe]

Hello Gareth,
I have the guess that the problem is that when estimating with a program, the existing models/settings in the workfile are not overwritten. This problem is new to me - it exists since I got a new version installed with a new notebook. If I delete all models from the workfile in advance, the results are identical. Could this be a general problem?

[EViews Gareth]

Yes, you are correct. If you re-estimate a the BVAR model, it appears to be keeping some if its original settings (not specification, but settings)