Entering content frame

Function documentation Automatic Model Selection Procedure 2 Locate the document in its SAP Library structure

Use

This procedure is applied in the forecast strategy 56. It is also the basis of the Structure linkAdaptive Forecasting technique, in which in background processing the system automatically runs this procedure if any error value specified in the diagnosis group is exceeded.

Prerequisites

You need at least two seasonal cycles and three periods as historical values to initiate the model. However if less are available the procedure will run, but models that require more initialization periods, such as seasonal trend, are not used.

Features

The procedure conducts a series of tests used to determine which type of forecast model (constant, trend, seasonal, and so on) to use. The system then varies the relevant forecast parameters (alpha, beta, and gamma) in the intervals and with the increments you specified in the forecast profile. If you do not make any entries, the system uses default values, in all cases 0.1. It uses these parameters to execute a forecast. It then chooses the parameters that lead to the lowest error of measure that defined in the forecast profile – default is MAD.

The system can use all measures of error and allows you to use a measure of error that you have defined yourself using the Define Measure of Forecast Error Business Add-In (BAdI) (see Customizing for Demand Planning).

Note

For procedure 2, you must bear in mind that when you use the outlier correction, the results are not comparable with the results of the individual processes, since another procedure can be selected for the outlier correction than for the final forecast.

Activities

...

       1.      The system first tests for intermittent historical data by determining the number of periods that do not contain any data in the historical key figure. If this is larger than 66% of the total number of periods, the system automatically stops model selection and uses the Croston method. You can change this value by using the BAdI /SAPAPO/SCM_FCSTPARA – Method PARAMETER_SET.

       2.      The system then checks for white noise. If it finds white noise, it automatically uses the constant method.

       3.      If both tests are negative, the system proceeds to test for seasonal and trend effects

                            a.      The system first eliminates any trend that it finds. To test for seasonal effects, the system determines the autocorrelation coefficient for all possible number of periods (from Number of PeriodsLength Variation to Number of Periods + Length Variation ). If the largest value is larger than 0.3 or another value for the limit as specified in the /SAPAPO/SCM_FCSTPARA BAdI – Method PARAMETER_SET, the test is positive.

For more information about the autocorrelation coefficient including formula, see Automatic Model Selection Procedure 1.

                            b.      To test for trend effects the system determines a trend significance parameter. As above it first eliminates any seasonal effects if the seasonal test is positive. If no seasonal effects have been found, it executes this test for the number of historical periods (as determined in the forecast profile) minus 2. If seasonal effects have been found, the system executes the test for the number of periods in a season plus 1.

Note

Since the results of these two tests determine which models the system checks in the next stage, the Periods per Season value in the forecast profile is very important. For instance, if your historical data has a season of seven periods and you enter a Periods per Season value of 3, the seasonal test will probably be negative. No seasonal models are then tried; only trend and constant models.

       4.      The system then runs forecasts with the models selected (see table below), calculating all the measures of errors. For models that use forecast parameters (alpha, beta, gamma) these parameters are varied in the ranges and with the step size specified in the forecast profile.

 

White Noise Test

Sporadic Data Test

Seasonal Test

Trend Test

Croston Model

 

X

 

 

Trend model

 

 

 

X

Seasonal model

 

 

X

 

Seasonal trend

 

 

A

A

Linear regression

 

 

o

X

Seasonal linear regression

 

 

A

A

X – The model is used if the test is positive

A – The model is used if all tests are positive

o- - The model is used if this testis negative

The constant model runs always, the one exception to this is when the sporadic data test is positive. In this case only the Croston model is used (which is a special type of constant model).

The system then chooses the model with the parameters that result in the lowest measure of error as chosen in the Error measure field of the forecast profile.

 

Leaving content frame