If you do not want to enter smoothing factors for exponential smoothing yourself, or you are unable to, you can let the system determine the best smoothing factors by means of optimization.
The optimization of smoothing factors is a mathematical process to determine the best possible smoothing factors. The aim is to fit the underlying forecast model to the historic data as well as possible by varying the smoothing factors. In this way you can generally expect to find the best quality forecast for the forecast model.
The variation of smoothing factors takes place within certain limits. These are defined by the search space.
Optimization consists of two steps:
· Performing a grid search which variegates the smoothing factors within the respective limits with the respective step sizes. The error is determined for each combination of smoothing factors (in accordance with the optimization variable). The result of the grid search delivers the starting value for the second step.
· Executing the direct search according to the Hooke-Jeeves algorithm.
The system cannot guarantee that this optimization algorithm finds the absolute optimum as the direct search can converge to a local minimum for the error measure.
Choose this option to obtain the highest quality forecast.
As with automatic model selection, the system proposes the optimization of smoothing factors supplied by the system with exponential smoothing forecast models (see Automatic Model Selection). This can be costly in terms of time when using trend-seasonal exponential smoothing. In this case SAP recommends:
· Speeding up the search by reducing the search area for the smoothing factors or increasing the step size.
· Entering suitable smoothing factors yourself. This is only advisable if, for example, you have recognized suitable values for the smoothing factors from past experience.
Optimization variables
The optimization variable is an error measure that is used for optimizing smoothing factors. Error measures allow a judgment to be made on the quality of the forecast. Optimization variables determine when one combination of smoothing factors is to be preferred to another combination.
You can select one of the following optimization variables:
· Mean absolute error
· Mean absolute percentage error
· Mean squared error (MSE)
· Exp. smoothed absolute error
Search area
You can restrict the search space for smoothing factors by defining an upper and a lower limit for each smoothing factor.
Furthermore, you can set a step sizefor each smoothing factor with which the system should begin to search, within the appropriate search space.