04-10-2013 10:45 AM
I am running a UCM model logsales=logbaseprice dummy1 dummy2 promo1 promo2
(promo1 promo2 as dummies also) as well as requesting the unobserved components.
My problem is the following : my level seems to be getting the most of the contribution for the prediction while my promo variables are not contributing as i had expected even in the weeks that the indication of promotion is obvious.
Am i doing something wrong when it comes to the estimates that i get from the run? should i convert them in someway to get the contribution? i am a bit sceptical since using logs but i revert back the logs as should, is there something similar that needs to be done for the promo and dummy coeffs?
Thanks in advance
04-11-2013 04:27 PM
Would it be possible to share your code and some example data?
As a wild guess:
If you are using a local linear model, you could try to set the LEVEL variance to 0, but leave the SLOPE variance to be determined by the procedure.
This will result in a smoother trend estimation, which might be helpful.
04-11-2013 04:55 PM
In addition to Udo's reponse I would also ask what your particular objective is? If it is prediction, then perhaps not finding a huge response is acceptable. If, on the other hand, your objective is to find the marginal effect of the program then you might either re-examine the tool (PROC AUTOREG) or the functional form (lags: the effect might not be contemporaneous). My two cents.