on 03-08-2011 1:36 PM
Hi,
Would like to highlight a few issues with Seasonal Models Selected through Auto Model 2:
1. Though Seasonality Test is Positive Constant Model is selected: When you run Auto model 2 in the interactive screen for Data with seasonality, The messages appear that seasonal Test is Positive but Auto model Ultimately chooses a Constant Model
2. Same Error for Different Gamma Factors: When you run Seasonal Models with Multiple Gamma Values the statistical forecast generated is different for each value of Gamma but the Errors remain the same for all the Gamma Values
3. Inconsistency for 3 Years Past History: Auto model 2 is picking the parameters with least MAD if the history period is 2 Years (for Seasonal Pattern in Past). But if the history period considered is 3 Years the model does not pick the parameters with least MAD
Kindly suggest.
Thanks & Regards,
Bharath
Hi Bharath,
Forgot to mention one more thing. Make sure that BAdI implementation /SAPAPO/VMIFCST_SCMB is "inactive" in your system according to the note 976935. It is valid even for 5.1 Release.
Thanks and Regards,
Nithin.
You must be a registered user to add a comment. If you've already registered, sign in. Otherwise, register and sign in.
Hi Bharath,
After further investigation all these days i have found out the following information :
The tests performed by Automodel Selection II are elimination tests, not selection tests.
To be precise, "Seasonal Test is positive" means that season model is NOT eliminated, but it does not mean that Season model will be chosen. The message "Trend test is negative" implies that the Trend model is eliminated. All the non-eliminated models which will go through the Forecast Error Optimisation process. This Forecast Error Optimisation process will execute a series of forecasts for each remainiing models, varying the alpha, beta and gamma values. The forecast model which gives the smallest error measure will be the chosen model.
I kindly request you to test this
Now with the same selection which you are using now, just execute forecast once. Goto Parameters tab. Change Error measure from MAD to say "MAPE"(you can also change to other errors, try to for all once). Then again execute forecast by clicking on Univariate forecast button(Shift + F1) at the top or in the Model tab. See if seasonal model is selected for any.
Are you from HUL?
Thanks and Regards,
Nithin.
Hi Nithin,
Sorry for the delayed response. Even i observed the same behavior of automodel, assume have to live with it.
For instance automodel should be suggesting seasonal model for historical data having week seasonality even if the error is slightly more than that of a constant model (since the seasonal test is positive).
Ruling out a model purely on the basis of error may not always be right.
Regards,
Bharath
Hi Bharath,
The problem may be because of the implementation of any of the user exit's or BAdI's. I have handled such a situation where the season test was positive but the auto model 2 used to select constant model. Some parameters may be hard coded because of which auto model 2 will be forced to select constant model.
Check if you have any implementation of the BAdI "/SAPAPO/SCM_FCSTPARA". If yes then check inside all of the methods whether some parameter is hard coded to 'X'. This definitely affects the model which is chosen during forecast by auto model 2.
If the error which needs to be considered is only MAD(which you would have mentioned in forecast profile) then definitely least MAD should be chosen. If there are other errors also and if it is less then i think those with less error will be chosen. But kindly check about the BAdI implementation which i have mentioned above.
Thanks and Regards,
Nithin.
You must be a registered user to add a comment. If you've already registered, sign in. Otherwise, register and sign in.
User | Count |
---|---|
15 | |
4 | |
3 | |
2 | |
1 | |
1 | |
1 | |
1 | |
1 |
You must be a registered user to add a comment. If you've already registered, sign in. Otherwise, register and sign in.