 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
Overview
Forecast Models
Model Selection
Manual Model Selection
Automatic Model Selection
Manual Model Selection with Additional System Check
Model Initialization
Ex-Post Forecast
Parameter Optimization
Maintaining Forecast Parameters
Forecast Parameters: Independent of the Forecast Model
Forecast Parameters: Dependent on the Forecast Model
Creating Historical Data
Creating Forecast Values
Carrying out the Forecast with Reference to Another Material
In order to carry out the forecast for a material, you must first of all maintain the forecast parameters for this material in the material master record.
In the following section, a description is given of,
When a series of consumption values is analysed, certain patterns can usually be detected. From these patterns it is then possible to differentiate between various forecast models: A constant consumption flow applies if consumption values vary very little from a stable mean value. With a trend model, consumption values fall or rise constantly over a long period of time with only occasional deviations. If periodically recurring peak or low values which differ significantly from a stable mean value are observed, it is a case of a seasonal consumption flow. A seasonal trend consumption model is characterized by a continual increase or decrease of the mean value. If none of the above patterns can be detected in a series of past consumption values, then we have an irregular consumption flow.
Before the first forecast can be carried out, you must specify which forecast model the system is to use to calculate the forecast values.
There are three basic possibilities:
If you want to select a model manually, then you must first of all analyse past consumption data to determine whether a distinct pattern or trend exists according to which you can manually select a model for the system.
Constant requirements pattern
If your past data represents a constant consumption flow, you can then select either the constant model or the constant model with adaption of the smoothing factors. In both cases, the forecast is carried out by first-order exponential smoothing. When adapting the smoothing parameters, the system calculates different parameter combinations and then selects the optimum parameter combination which is the one which results in the lowest mean absolute deviation.
You have another two possibilities if your past consumption pattern is constant; either the moving average model or the weighted moving average model.
In the weighted moving average model, you can weight individual past consumption values, which means that the system will not give equal value to past data when calculating the forecast values. In so doing, you can influence the calculation so that the most recent consumption values play a more important role in the forecast than the previous periods - as is also the case in exponential smoothing.
Trend requirements pattern
If your past consumption data represents a trend, it makes sense to select either the trend model or the second-order exponential smoothing model. In the trend model, the system calculates the forecast values by means of the first-order exponential smoothing procedure.
In the second-order exponential smoothing models, you can choose between a model with or without parameter optimization.
Seasonal requirements pattern
If your past consumption data represents a seasonal pattern, you can specify the seasonal model you want. The system calculates the forecast values for the seasonal model by means of first-order exponential smoothing.
Seasonal trend requirements pattern
If your past consumption data represents a seasonal trend pattern, you can select the seasonal trend model you want. The system calculates the forecast values by means of first-order exponential smoothing.
Irregular requirements pattern
None of the patterns or trends mentioned in the above examples can be recognized in an irregular consumption flow. If the system is to carry out a forecast for an irregular pattern, then it is usually advisable to select either the moving average model or the weighted moving average model.
Forecast Models for Different Requirements Patterns
Requirements pattern | Forecast model |
Constant | constant model |
| constant model with smoothing factor adaption | |
| moving average model | |
| weighted moving average model | |
Trend | trend model (first-order exponent smoothing) |
Irregular | no forecast |
| moving average model | |
| weighted moving average model | |
Extended forecasting component used: | |
Trend | second-order exponential smoothing model (with and without parameter optimization) |
Seasonal | seasonal model |
Seasonal trend | seasonal trend model |
If you do not want to specify a forecast model manually, you must instruct the system to make an automatic selection. The system then analyzes the historical data for different patterns and selects the appropriate model. The following models are possible:
If the system does not detect any regular pattern in the past consumption data, it automatically selects the constant model.
The system requires a different amount of historical values for the individual tests. For more information on the exact number that it requires, see Model Initialization.
Model Selection Procedures
If the system is to make an automatic model selection, you have the choice between two model selection procedures:
Procedure 1
The system carries out statistical tests and checks whether a trend or a seasonal requirements pattern applies.
In the trend test, the system subjects the historical values to a regression analysis and checks to see whether there is a significant trend pattern.
In the seasonal test, the system clears the historical values of any possible trends and carries out an autocorrelation test.
Procedure 2
The system calculates the models to be tested using various combinations for alpha, beta and gamma. The smoothing factors are also varied between 0.2 to 0.8 in intervals of 0.2.
The model which is then chosen is the the model which displays the lowest mean absolute deviation (MAD).
Procedure 2 is more precise than procedure 1 but takes considerably longer.
In the third case, you can have a combination of manual and automatic model selection: you specify a model and set the system so that it additionally checks historical values for a seasonal pattern or trend.
Possible combinations that you can choose are as follows:
A reference is given to the forecast formulae used by the system for the individual models in Forecast Formulae.
Model initialization is the process of specifying the necessary model parameters, such as the basic value, the trend value and the seasonal indices for the selected forecast model. In each case, initialization occurs during the first forecast for a material. It must also be carried out in cases of structural interruption, that is, if the existing forecast becomes invalid. Which model parameters are necessary for which forecast model is shown in the following Table: Model Parameters
Model | Model parameters |
Constant model | basic value |
Trend model | basic value, trend value |
Seasonal model | basic value, seasonal indices |
Seasonal trend model | basic value, trend value, seasonal indices |
The forecast model is usually automatically initialized. In order to do this, the system requires a certain amount of historical values. The number required varies depending on the forecast model as shown in the table below.
Number of Historical Values Required for Model Initialization
Model | No. of historical values |
Constant model | 1 |
Trend model | 3 |
Seasonal model | 1 season |
Seasonal trend model | 1 season + 3 |
2nd-order exp. smoothing | 3 |
Moving average | 1 |
Weighted moving average | 1 |
The system calculates the basic value on the basis of an average value, and the trend using the results of the regression analysis. The seasonal indices result from the quotient of the actual past value and the basic value which has been adjusted for the trend value.
These calculation methods are used for the constant, trend, seasonal and seasonal trend models, depending on which parameters are to be determined.
For the second-order exponential smoothing model, a regression analysis is carried out.
For the moving average and the weighted moving average models, the system calculates an average value.
If more past values are available than the system requires or is to use for model initialization (refer to "Number of historical values" and "Initialization periods" in Forecast Parameters: Independent of the Forecast Model), then the system will carry out an ex-post forecast. In order to do this, the system divides the time series of past values into two groups. The first group with the older values is used for initialization. The system carries out an ex-post forecast using the second group as it is best to adapt the parameters to the most recent developments.
Furthermore, you can monitor the forecast model during the initialization phase by comparing the forecast values from the ex-post forecast with the actual consumption values.
The ex-post forecast does not only play an important role in the initial forecast. It is also significant for subsequent forecasts. It is, therefore, possible to omit periods during the forecast calculation. This means that you can determine, for example, daily forecast values for a material even if you only execute a weekly forecast run for this particular material. The system forecasts the missing periods in retrospect. The ex-post forecast is only possible for models with the exponential smoothing procedure.
For forecast models with exponential smoothing, you can instruct the system to optimize the smoothing factors (see Forecast Parameters: Dependent on the Forecast Model).
If the system is instructed to optimize the smoothing factors, it calculates several parameter combinations and then selects the combination with the lowest mean absolute deviation (MAD). The finer the degree of optimization, the more exact, but also the more time-consuming, the parameter optimization. You can vary the increment (also known as the degree of optimization) from 0.1 (low), 0.2 (medium) and 0.3 (high).
If you have set the system correctly, it will carry out parameter optimization during initialization for the initial forecast as well as for all other forecasts. The most effective combination of the smoothing factors is determined by means of the ex-post forecast.
Please refer to Appendix B in this manual for additional information on the calculation of the MAD as well as the significance of the smoothing factors.
Optimizing the smoothing factor is only possible for models with the exponential smoothing procedure.
You maintain forecast parameters in the material master record. In order to do this, proceed as is described below:
The pop-up window Select View(s) appears on your screen.
The Organizational Levels/Profiles pop-up window now appears.
The "Forecasting" data screen now appears.
A more detailed description is given of the individual forecast parameters in the follwoing sections:
Forecast Parameters: Independent of the Forecast Model
Forecast Parameters: Dependent on the Forecast Model
For more information on maintaining the material master record, please refer to MM Managing Material Master Data.
You can/must maintain the following parameters independent of the method you use for model selection and which model you choose.
Period indicator
The period indicator determines the time interval the system is to use to store consumption or forecast values. You can define the period indicator when maintaining the MRP parameters as well as when maintaining the forecast parameters.
Fiscal year variant
You must maintain the fiscal year variant if you want a flexible period length for the material equal to that of the accounting period.
Historical values
You specify how many historical values the system should take into account for the forecast by filling in this field. The default value in this case is the maximum number possible (60). If fewer values are available, then they are used for initialization.
Initialization periods
You specify the number of periods the system is to use for model initialization here. If the number of historical values is greater than the number of initialization periods, then the system will carry out an ex-post forecast for those periods which were not used for initialization. The number of periods that the system requires for initialization depends on the forecast model.
Forecast periods
Via the number of forecast periods, you specify for how many periods the system determines forecast values.
Fixed periods
Via the number of fixed periods, you determine how many periods are fixed in the future and are no longer automatically changed by the next forecast run.
Initialization
This field activates re-initialization during the forecast. This is necessary for the first forecast and also when structural changes (that is, if the forecast model which has been used until now has lost its validity) appear in the time series.
Model initialization can either be carried out automatically (X) or manually (M). In the case of manual initialization (which should be the exception), you must determine the model parameters manually. To do this, select Details ® Forecast details ® Model parameters. Depending on the forecast model, you can manually maintain such parameters as the basic value and the trend in the pop-up window. If you do not set these parameters here, you will be requested to set them during the forecast. No historical data has to be stored in the system in order to carry out manual initialization.
Tracking limit
During the forecast calculation, the system compares the tracking limit with the quotient of the error total and the mean absolute deviation (MAD). This quotient is known as the tracking signal. If the value calculated by the system lies below that of the tracking limit, then the forecast remains valid. If not, the system generates an exception message which indicates that a structural interruption has taken place. This means that the forecast model should be revised and possibly re-initialized.
Autom. reset
Causes the system to reset the forecast model automatically if the tracking limit is exceeded. Furthermore, the system carries out a new model selection during the next forecast.
Correction factors
If you fill in this field, historical values and forecast values are weighted with the appropriate period factors which you determined via customizing.
Reference material for consumption
If you define a reference material, the system accesses the historical data of the reference material up to the specified validity date. This makes sense, if no historical data exists yet for the specified material.
Reference plant for consumption
If you enter a reference material for consumption, you must also define a reference plant where the historical data to be used comes from.
Validity date
During the forecast, the system accesses the historical data of the reference material up to the specified validity date. From this date, it uses the material’s own historical data.
Multiplier
By entering a multiplier, you can define that only a certain percentage of the consumption quantity of the reference material is used in the forecast.
You can/must maintain the following parameters depending on how you carry out model selection and which model you choose. The relationship between parameters and model selection is shown in the following two tables. Parameters Dependent on Model Selection
Model selection | Parameters |
Manual model | forecast model selection |
Automatic | model selection, model selection procedure model selection |
Parameters Dependent on the Forecast Model
Specified model/ model to be tested | Possible/required parameters |
Constant model | parameter optimization |
| optimization level | |
| alpha and delta factors | |
Constant model with optimization of the smoothing factors | - |
Trend model | parameter optimization |
| optimization level | |
| alpha, beta and delta factors | |
Moving average model | initialization periods |
Weighted moving average | weighting group model |
Extended forecast component used: | |
Seasonal model | parameter optimization |
| optimization level | |
| alpha, beta and delta factors | |
Seasonal trend model | parameter optimization |
| optimization level | |
| alpha, beta, gamma and | |
| delta factors | |
2nd-order exponential smoothing model | parameter optimization |
| alpha and delta factors | |
2nd-order exponential smoothing model with smoothing factor optimization | alpha and delta factors |
Forecast model
You determine via the forecast model which model the system uses as a basis when calculating the forecast values. If you do not know the forecast model, you can have it determined by the system via automatic model selection.
Model selection
This indicator specifies for which model the system is to examine the historical values. You can specify whether the system searches the historical values
Please note that depending on the model test, a minimum number of historical values must be available. This field is significant if you do not know the model and you want the system to determine it automatically. Furthermore, you also have the possibility of pre-selecting a trend model, but at the same time instruct the system to search for a seasonal pattern and vice versa.
Selection procedure
This indicator specifies how the system is to carry out the model selection. Here, you can choose between two procedures:
Parameter optimization
Via this indicator, you can specify that the system is to optimize the necessary smoothing factors for the appropriate model. The system calculates several parameter combinations and selects the one that displays the lowest MAD. Parameter optimization is carried for the initial forecast as well as for the subsequent forecasts.
Periods per seasonal cycle
You must enter the number of periods that constitute a season here if you have selected a seasonal model or if the system is to carry out a seasonal test.
Optimization level
By determining the optimization level, you are specifying the increment with which the system is to carry out parameter optimization. The lower the increment, the more exact but also the more time consuming the optimization process will be.
Weighting group
You only have to maintain this field if you selected the forecast model, "weighted moving average". This key specifies how many historical values are taken into account for the forecast and how these values are weighted in the forecast calculation.
The following factors are used by the system, depending on the model, for exponential smoothing. Thus, for example, only the alpha and the delta factors are required for the constant model whereas all of the smoothing factors are required for the seasonal trend model.
Alpha factor
The system uses the alpha factor for smoothing the basic value. If you do not specify an alpha factor, the system will automatically use the alpha factor 0.2.
Beta factor
The system uses the beta factor for smoothing the trend value. If you do not specify a beta factor, the system will autimatically use the beta factor 0.1.
Gamma factor
The system uses the gamma factor for smoothing the seasonal index. If you do not specify a gamma factor, the system will automatically use the gamma factor 0.3.
Delta factor
The system uses the delta factor for smoothing the mean absolute deviation and the error total. If you do not specify a delta factor, the system will automatically use the delta factor 0.3.
If you set parameter optimization, the system will overwrite the originally set smoothing factors with those which have been newly calculated by the optimization process.
