Show TOC

Function documentationForecast Strategies

 

The forecasting strategy defines how forecast values will be computed.

All forecast strategies are based on statistical forecast procedures and forecast models that represent the time series mathematically.

The exponential smoothing methods are currently the most widely used time series patterns (see Exponential Smoothing).

If you expect historic values to continue to develop as they have in the past, choose a forecast model that fits the time series pattern.

Note Note

The Automatic Model Selection strategy allows you to let the system select the forecast model that best fits the trend of historic data (see Automatic Model Selection).

End of the note.

Features

The following forecasting strategies can be used:

Average

The forecast value is calculated from the mean of the historic values.

Floating Average

The forecast value is calculated according to the order.

  • Obligatory forecast parameter: Order of Floating Average

    The order of the floating average is a number N that determines the length of the time interval for calculating the average. This is the number of chronologically sequential historic values. The forecast value is calculated simply as the average of the last N historic values.

    Enter a positive number for the order.

  • Optional Forecast Parameters Outlier Correction

    Logging Statistical Key Figures

    Ignoring Initial Zero Values

Weighted Floating Average

When the system calculates the floating average, each historic value is given a particular weight.

  • Obligatory forecast parameter: Order of Floating Average

    The order of the floating average is a number N that determines the length of the time interval for calculating the average. This is the number of chronologically sequential historic values.

    Enter a positive number for the order.

  • Mandatory forecast parameter: Weighting Factors.

    The weighting factors specify the relationship between the individual historic values and the average calculation. The sequence is important: Weighting factor 1 refers to the previous periods, while weighting factor 2 refers to the periods before that and so on.

    Example Example

    You want to create a forecast based on monthly values and choose a weighted moving average with an order that has the value 6. In this case, you want to place more weight on the most recent monthly values than on the less recent monthly values. The historic data is taken from months 5 to 10. The six weighting factors and the relevant months are as follows:

    No.

    Weighting Factor

    Month

    1

    3,00

    10

    2

    2,00

    9

    3

    2,00

    8

    4

    1,00

    7

    5

    1,00

    6

    6

    1,00

    5
    End of the example.
Linear Regression

Simple linear regression (method of smallest quadrats).

Seasonal linear regression

Seasonal linear regression is based on the same statistical procedures as used in demand planning.

Note Note

For more information, see http://help.sap.com/ Start of the navigation path SAP Business Suite Next navigation step SAP Supply Chain Management Next navigation step SAP APO 3.1 Next navigation step Application Help Next navigation step Demand Planning Next navigation step Demand Planning Process Next navigation step Definition/Redefinition of Forecast Models Next navigation step Creating a Master Forecast Profile Next navigation step Univariate Forecasting Next navigation step Forecast Strategies Next navigation step Seasonal Linear Regression End of the navigation path

End of the note.
Simple Exponential Smoothing (Constant Model)

Simple exponential smoothing is suitable if the historic data shows a constant trend.

Simple exponential smoothing with alpha optimization (constant model)

This procedure is the same as the “simple exponential smoothing“ described above with just one difference: The system also calculates the Alpha smoothing factor. The Alpha value is variegated in the interval using the defined step size and a forecast calculation (for the historic time frame) is performed in each case. The optimum value for Alpha is the value that produces the smallest error in the forecast results.

  • Smoothing factor settings: Optimization Variable, Alpha from, Alpha tos, Alpha Step Size

Linear exponential smoothing (trend model)

The forecast is calculated according to Holt’s method and is suitable if historic values display an upward or downward trend.

Seasonal Exponential Smoothing (Season Model)

Choose this strategy if your historic values show seasonal fluctuations (for example, annual fluctuations) from a constant base value.

Seasonal Trend Exponential Smoothing

Forecasting is based on Winter/Holt's multiplicative method and is appropriate if historical values fluctuate on a seasonal basis from an increasing or decreasing trend. The fluctuation increases with an upward trend.

Example Example

Ice cream sales in summer: Assume that ice cream sales rise by a trend of 10% annually. A seasonal increase of 30% each summer then leads to ever greater absolute fluctuations.

End of the example.
Croston Method

The Croston method was developed specifically for sporadic trends. This procedure uses exponential smoothing to calculate a mean time interval between the values in the time series that are not equal to zero.

Note Note

For more information, see http://help.sap.com/ Start of the navigation path SAP Business Suite Next navigation step SAP Supply Chain Management Next navigation step SAP APO 3.1 Next navigation step Application Help Next navigation step Demand Planning Next navigation step Demand Planning Process Next navigation step Definition/Redefinition of Forecast Models Next navigation step Creating a Master Forecast Profile Next navigation step Univariate Forecasting Next navigation step Forecast Strategies Next navigation step Croston Method End of the navigation path

End of the note.

Check whether it might be possible to aggregate the data in order to remove the gaps in the time series. This would make it possible to use procedures that take account of trend or seasonal time series patterns. You can aggregate data in this way by choosing a rough time characteristic (month instead of day) or by forecasting values for product groups instead of individual products.