We are in consumer goods industry and planning to implement sap APO-DP for forecasting process. What we are planning is to , to run the forecast , just select automatic model selection assuming that the this model will give us a good forecast accuracy . And also some new items are in the MTF items list, without any historical sales details, for the same how sap will generate forecast ?
In consumer goods environment i assume number of CVCs will be very high and also there would be seasonality and promotion involved with the products. I don't think auto model as an appropriate solution for forecasting in consumer goods environment. Also since number of CVCs would be very high and as you must be aware auto model does a lot of calculation to find the pattern and then deciding the appropriate forecast model. This is not recommended for performance reason as well. Rather i would suggest to classify the products in terms of seasonality , promotion and trend and then select corresponding forecast model . Also we can think of classifying the products in ABC categories based on profitability and then using auto model for C class products.
First, for history, we copy the history from similar product.
For this we can use the Life cycle planning or realignment option(to copy the key figure data to other key figure).
Secondly, for Model selection, if we use automatic model selection, then system proposes the best model fit for the historical data it has.
also, in automatic model selection, you can have system propose the Alpha, Beta and Gamma values.
But, if you want to test different models, then I suggest you create different Forecast profiles and assign them to the CVCs/selection profile and test it.
The best model fit is selected based on couple of criteria, like the one which gives least error total, or the one which is able to mimic the historical pattern.
if you are using the expost method in forecast profile, then system can analyze and tell the best fit alpha, beta, gamma and error totals and use them for generating future values or forecast.
Also, if you know the alpha, beta gamma values, then a value close to 1 for these parameters will mimic the recent pattern and a value like 0.1 will not.
so, decide on from which product to copy the historical data, then maintain couple of Forecast profile with different combinations of these parameters and test for the best fit.
Also, you can use the help of other key figure, like corrected history, and corrected forecast and macros, which aid in testing these models.
Finally, all the details I have mentioned are related to uni variate models, you might want to choose MLR or a composite.
You must be a registered user to add a comment. If you've already registered, sign in. Otherwise, register and sign in.