I have weekly sales for around 6000 products, for which I would like to obtain forecasts for n periods. I believe that estimate forecasts for all 6000 products is too much, so i thought about aggregating the products in certain categories/clusters (based on some rules).

I end up with around 50 products (or more correct groups/clusters of products), for which again I want forecasts for n periods.

My question is, are 50 products again too many for a VAR model ? I am planning on using the vars package in R

Or maybe it would be a better idea to estimate 50 VAR models (because 50 is the groups/clusters of products), with the logic that since the products in each group are similar, then the sales of a product will affect the sales of a product (in the same category)

  • $\begingroup$ How many observations are you planning on using? Do your weekly observations go back for several years? $\endgroup$ – ERT Aug 9 '18 at 22:40
  • $\begingroup$ the total number of observations for each group of products is around 150, so it is a bit more than 2 years $\endgroup$ – quant Aug 10 '18 at 6:43
  • $\begingroup$ Why do you believe 50 factors would be too many? Are you concerned about computation constraints or viability of results? $\endgroup$ – ERT Aug 10 '18 at 10:41
  • $\begingroup$ i never used VAR before so i dont really know. Moreover if you have 50 groups and you add lag order of 1, then multicollinearity can be an issue, right ? $\endgroup$ – quant Aug 10 '18 at 11:34

Your Answer

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

Browse other questions tagged or ask your own question.