Are there any suggested approaches for using non-stationary series in a VAR model? As per otexts.org:

If the series are non-stationary we take differences to make them stationary and then we fit a VAR model (known as a “VAR in differences”).

Are there any other approaches for creating a forecasting model non-stationary series in a multivariate series?

Any leads on this would be helpful. I'm looking for implementing this model in R.

  • $\begingroup$ Either a VAR in first differences or a vector error correction model (VECM) depending on whether your series are cointegrated. See e.g. this which is a more general case, but easy to simplify. See also this for a general cookbook approach to VAR modelling. $\endgroup$ – Richard Hardy Feb 14 '17 at 15:18
  • $\begingroup$ Thanks a lot for the information provided. The only way to identify stationarity of var is to check the stationarity of constituent variables? $\endgroup$ – Lal Prasad R Feb 15 '17 at 4:48
  • $\begingroup$ Yes, that's right. $\endgroup$ – Richard Hardy Feb 16 '17 at 17:54
  • $\begingroup$ @LalPrasadR you DON'T need to first difference data to forecast with VAR (sorry this is one of my pet pervs)!!! Read my post here: stats.stackexchange.com/questions/191851/… $\endgroup$ – Jacob H Jan 7 '19 at 19:07

To use VAR model for non-stationary series, you have to test the cointegration If there is cointegration you use the model VECM Otherwise a VAR on the first differences of the variables

| cite | improve this answer | |
  • $\begingroup$ Thank you for the information provide. So if we need to predict from the VECM mode, we need to go for Rolling forecasts? $\endgroup$ – Lal Prasad R Feb 15 '17 at 16:11

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.