Monitoring the Selected Forecast Model
In the course of time, the characteristics of a consumption series can change. In order to be able to react to this in time, the system calculates a so-called tracking signal during every forecast. The tracking signal links the error total (FS) and the mean absolute devialtion (MAD) in the following manner:
Tracking signal = │ FS/MAD │
The forecast error is the difference between actual consumption values and the forecast values from the same period whereas the error total is the sum of all the forecast errors in a consumption series.
The error total is used to check the validity of the forecast model in operation.
If a model is still valid, that is, if the consumption series has not changed, then you can assume that the error total is distributed normally and has an average of zero.
If the consumption pattern has changed, however, the error total will no longer be equal to zero. You must standardize the calculated forecast error in order to set boundaries. Therefore, in addition to the error total, the system also calculates the mean absolute deviation (MAD) as a second value. The system then adds the errors (irrespective of the plus or minus sign) and divides them by the number of consumption values. The first-order exponential smoothing procedure is used to calculate the MAD. The smoothing factor used for this is the delta factor.
Using the quotient from the error total and the MAD, the system can now define a warning limit, the tracking signal which it then compares with the tracking limit specified in the article master. When the tracking signal is greater than the tracking limit, the RP controller receives a message stating that the forecast model should be checked.
After a model change or a forecast model initialization, the error total is automatically reset to zero and the MAD to its initial value.
The system automatically sets the tracking limit (the default value is 4.00). You can change it, however, when maintaining the article master record.