Vector Autoregression model without using past values of predicted value I have 16 time series: A, B, C ... P.
And I am trying to use Vector Autoregression model to capture the relationship between A and other series.
But when I use VAR, I found that the model relies too much on the past values of A.
So I want to remove past values of A from the model.
I want to describe A only using series B, C ... P.
Is there any way to implement this or enforce such constraint in VAR?
I am using statsmodels.tsa.api in Python.
Thank you for the help.
 A: A VAR model is efficiently estimated with equation-by-equation OLS. If you restrict certain coefficients to be zero (which is what dropping lags of the dependent variable amounts to), you may gain some efficiency with a GLS type approach, also known as seemingly unrelated regressions. You would still get consistent estimates when simply running an OLS regression of A on the lags of B to P, though (provided your restriction is correct, of course - also, you may or may not care about such consistency if, for example, your ultimate goal is, e.g., forecasting).
It may be worth asking, though, what "relying too much" is to mean. In my experience, it is - intuitively, I would say - not uncommon that own lags are more important than cross lags, so maybe your VAR captures that strong own-dependence just right instead of "relying too much"?
I cannot help you with Python, but in R, such an approach can for example be implemented with the systemfit package. Alternatively, you could use the vars package using the restrict command.
