Automatic Model Selection

Purpose

You can use this process to automatically select a suitable forecast model. If the forecast run for a location product had a bad forecast result during the last forecast period, that is the results varied greatly from the reality, it is advisable to use a different forecast model. During automatic model selection, the system checks which model would best create an exact as possible forecast for a product.

Note Note

The system only considers the declining demand forecast model during automatic model selection if you have selected one of the following periodicities in Customizing:

Month
Posting period with exactly 12 posting periods per year

End of the note.

Prerequisites

You have executed the following IMG activities in Customizing for Advanced Planning and Optimization , under Supply Chain Planning Service Parts Planning (SPP) Forecasting:
- Define Forecast Service Profile
- Define Smoothing Factor Record
- Define Profile for Automatic Model Selection
In the forecast profile, on the Model Selection tab page, you have entered one of the profiles in the Profile for Automatic Model Selection field that you created in Customizing in the IMG activity Define Profile for Automatic Model Selection . You get to the forecast profile on theSAP Easy Accessscreen underAdvanced Planning and Optimization Service Parts Planning (SPP) Planning Forecasting Forecast Profile.

Process

Planning Services for Automatic Model Selection

In the Planning Service Manager (PSM) the following services trigger automatic model selection:

Execute combined forecast

If you schedule this planning service, the system checks if the current forecast model is still valid ( Trigg’s tracking signal ). The system only carries out automatic model selection if the model shows a bad result.

If you have recently changed forecast model, the system can prevent another model change for reasons of stability. The system then simply reconciles the smoothing factors of the existing model. You define the length of time that each model must be in use before it can be changed with the parameters for stability periods.To do so, on the SAP Easy Access screen, choose Start of the navigation path Advanced Planning and Optimization Next navigation step Service Parts Planning (SPP) Planning Forecast Forecast Profile End of the navigation path . You define the parameters for each forecast model on the Model Selection tab page in the following fields:

Forecast Model	Field Name
First-order exponential smoothing	Stability Periods for FOES
Second-order exponential smoothing	Stability Periods for SOES
Moving average	Stability Periods for Moving Average Model
Linear regression model	Stability Periods for LR Model
Seasonal trend model	Stability Periods for Seasonal Trend Model
Seasonal trend model with FPG	Stability Periods for Seasonal Trend Model with FPG
Intermittent forecast model	Stability Periods for Intermittent Forecast Model
Dynamic moving average	Stability Periods for DMA Model
Declining demand forecast (DDF)	Stability Periods for Declining Demand Forecast Model

For more information about the composite forecast, see Planning Services for Forecasting

Execute automatic model selection

If you choose this service, the system executes automatic model selection without first checking whether it is necessary or useful to choose a new model.

Start Automatic Model Selection Automatically

You can start automatic model selection for a location product manually on the SAP Easy Access screen under Start of the navigation path Service Parts Planning (SPP) Next navigation step Planning Forecasting Interactive Forecasting End of the navigation path .

To do so, choose the Change pushbutton and the Model Selection pushbutton. At runtime, the system then calculates values for the Demand: AMS Forecast and Demand: AMS RMSE key figures as of the current period.

You can thus compare forecast values calculated by the system during automatic model selection with forecast values that already exist. You can also compare the root of the mean square error (RMSE) calculated by the system during automatic model selection with the standard deviation that already exists.

Selection Procedure

The system executes statistical tests in the sequence that you defined in the profile for automatic model selection. If a test is successful, the test series is completed, that is the system cancels the test series at this point and executes no further tests.

The following tests are available:
The system chooses all forecast strategies that are assigned to the successful test. You have made these assignments in the profile for automatic model selection.

For more information aboutforecast strategies, see Forecast Models, Strategies, and Profiles .
For all forecast models that are assigned to the successful test, the system calculates a forecast error from the ex-post forecast and the past actual demand. The system calculates the forecast error that you have chosen in the forecast profile on the Model Selection tab page in the Forecast Error in Automatic Model Selection field For more information about forecast errors that are available for automatic model selection, see Forecast Errors .

Note
If you work with a customer-defined forecast strategy, and assign it to one or more forecast tests in the profile for automatic model selection, you must also make an ex-post forecast available for the customer-defined forecast strategy. Otherwise the system cannot calculate a forecast error.

End of the note.
The system weights the forecast error calculated for each forecast model using the forecast model weighting factor that you have defined for each model in in the profile for automatic model selection. A high value leads to a high error weighting, that is the higher you weight the error, the less likely it is that the system calculates the relevant forecast model as optimal.
The system selects the forecast model with the smallest forecast error as the valid model. The system then updates this model as well as the alpha and beta smoothing factors in the forecast profile, if necessary.
Note
The forecast error is always zero for the following forecast models:
- Seasonal trend model with fixed period groupings (FPG)
- Dynamic moving average
- Declining demand forecast (DDF)
If a test to which you have assigned one of these forecast models is successful, the system always selects this forecast model. If you assign one or more of these forecast models to one or more forecast tests, it is advisable (for performance reasons) to always assign the sequence number one to these forecast models.

End of the note.