Automatic Model Selection
Use
Automatic model selection allows you to let the system determine which forecast model best fits your historic data.
We recommend that you use automatic model selection if you do not know the trend of your historic data, if you cannot estimate how your data will develop, or if you do not want to specify a model.
Features
The system performs a number of tests and uses the results to determine the model to be used (see Forecast Strategies). If the model chosen is exponential smoothing, the system optimizes the relevant smoothing factors (Alpha, Beta, Gamma).
Note that automatic model selection requires a high calculation effort. This is particularly true of the seasonal trend model. The calculation effort also depends on the scope of the search space and the precision of the step sizes that you have set.
Activities
...
1. First the system tests for sporadic historic data by determining the number of periods that do not contain any data for the history key figure. If this number accounts for more than 66% of the total number of periods, the system automatically uses the Croston method.
2. Then the system tests for white noise. If there is white noise, the system automatically uses the constant method.
3. If both tests are negative, the system tests for seasonal and trend effects.
a. First the system deletes existing trends. To test for seasonal effects, the system determines the auto-correlation coefficients. If the auto-correlation coefficient is greater than 0.3, the test is positive.
b. To test for trend effects, the system determines the trend significance parameters. If the seasonal test is positive, the system removes possible seasonal effects. If no seasonal tests are determined, the system runs the test using the number of past periods minus 2. If seasonal effects are determined, it runs the test using the number of periods in a season plus 1.
Since the results of these tests determine the model that the system checks in the next step, the Periods per Season parameter is particularly significant. If your historic data contains, for example, a season with seven periods and you enter the value “3“ for Periods per Season, the seasonal test will probably be negative. In this case, the system does not check the seasonal model but checks the trend and constant models only.
4. The system uses the selected model (see the table below) and performs the forecast. It calculates all error measures. In models that use forecast parameters (Alpha, Beta, Gamma), these parameters vary to reflect the areas and step sizes specified in the forecast profile.
Test results and model selection
|
Test for White Noise |
Test for Sporadic Data |
Seasonal Test |
Trend Test |
Croston model |
|
X |
|
|
Trend model |
|
|
|
X |
Seasonal model |
|
|
X |
|
Trend season |
|
|
A |
A |
Linear regression |
|
|
o |
X |
Seasonal linear regression |
|
|
A |
A |
Legend for the characters in the table:
X – model is used if the test is positive.
A – model is used if all tests are positive.
o – model is used if the test is negative.
The constant model always runs, unless the test for sporadic data is positive. In this case, the Croston model is used exclusively (as a special variant of the constant model).
The system chooses the model with the parameters that produce the lowest error measure. The error measure is specified by the selection made in the Error Measure field in the forecast profile.