Dear Sandeep
That depends on your use case.
I think it is more useful to first run substitute missing values before I let the outlier correction run - if I would want to use both.
Promotion sales lift elimination is originally designed for Demand Sensing, in order to "identify positive outliers (sales lifts) associated with promotions, and to remove them from the sales history" - like outlier correction but only in one direction. It is not compatible with monthly buckets. It does not seem to be wise to combine this with the general outlier correction as you would have twice the outlier correction. Well, there may be some very exceptional cases I cannot think about right now that prove otherwise
I guess you will just have white noise if you run all the three together...
And if I can I first identify if a combination is seasonal, before I decide on the preprocessing steps. If you run too much outlier correction on a seasonal product, the result of the forecast algorithm may be dissatisfying because in the identification of the outlier correction model there is no check on seasonality. So if you can define a group of products that are seasonal, for those I would potentially NOT have the outlier correction, but only substitute missing values.
It is planned to enhance the possibilities to identify seasonal behavior in future releases, so then the choice may be different.
If you have products that are not seasonal but have a strong trend, you need to check if you want to run the substitute missing values at all, or only fill in with constant value 0 but potentially not use mean or median.
So, the question is not only on the correct order, but more on which to choose and how
Irmi
Add comment