0
$\begingroup$

I am setting a VAR model with two variables. One of the variables looks like this:

timeseries

Do you think it's still appropriate to setup the VAR as a time series with "drift" or do you think it must be "drift and trend"? The data is already log-transformed.

I read a few papers and my impression is that every author is handling it differently. The "trend" graph shows that there is a trend but when looking at the graph, it does not seem very clear.

$\endgroup$
  • $\begingroup$ Whether it is drift or "trend", you want to supply (V)AR model with stationary or pseudo-stationary data. What if you take: a) first differences, b) first log differences of your vector and try to investigate the existence of: a) autoregressive part, b) seasonality? $\endgroup$ – Alexey Burnakov Oct 13 '17 at 9:34
  • $\begingroup$ The data is log-level and ADF indicates stationarity at 5% level when accounting for a "drift" trend. $\endgroup$ – user179929 Oct 13 '17 at 11:06
  • $\begingroup$ Are data log scale, or differences are on log scale? $\endgroup$ – Alexey Burnakov Oct 13 '17 at 11:12
  • $\begingroup$ Data is log scale, not differenced $\endgroup$ – user179929 Oct 13 '17 at 16:58
  • $\begingroup$ OK. If you want to make forecast using your decomposition, you are about to know how to forecast all the components. I assume you can model trend as a linear function of time. You can model seasonality as either autoregressive model or seasonal difference model. Besides you can model residuals by taking AR component of them. Does it make sense? $\endgroup$ – Alexey Burnakov Oct 13 '17 at 17:07

Your Answer

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