I have a couple of time series, say, T1 and T2. I have established (using the grangercausalitytest library of Statsmodels in Python) that T1 Granger-causes T2 at 10% significance for lags 4 and 5. I now want to know whether I can validate the result by using VAR (unrestricted model).

Going by the definition of Granger causality mentioned in the book "Introduction to Modern Time Series Analysis", for 2 time series, "x" and "y",

...if future values of y can be better predicted, i.e. with a smaller forecast error variance, if current and past values of x are used,

I should get less RMSE for the VAR model as compared to the Autoregression model with only T2. (I'm using ARIMA(p,0,0) for autoregression with p = 4,5 as the Granger - cause occurs at these lags)

My first question is my interpretation correct?

If not, why not? And if that is the case, the RMSE of what should I be comparing? I ask this question because T1 Granger-causes T2 at lags 4 and 5. Should I take the RMSE (for both VAR and Autoregression models) after lags of 4?

Also, should there be any other caveats?


1 Answer 1


Let me denote the first time series $\{Y_1\}$ and the second one $\{Y_2\}$.

Your interpretation is correct.

RMSE of what?
You may compare the RMSE of one-step ahead forecasts from the AR(p) model for $\{Y_2\}$ with one of the ARDL(p,p) model for $\{Y_2\}$ for either $p=4$ or $p=5$. The ARDL(p,p) model corresponds to the equation for $\{Y_2\}$ from the VAR(p) model that includes both $\{Y_1\}$ and $\{Y_2\}$. I think RMSE of $h$-step ahead forecast for $h>1$ would also work but may be a less common choice.

For more details on the logic of comparing AR and ARDL models in the context of Granger causality, see this.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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