The use of control card

Another aspect of the error measurement is the use of a control card that can also be done in the forecasting software. However, in this case the errors are shown as individual occurrences in the control card, which shows the development of the forecast accuracy. Based on that, it can be, for example, estimated if the accuracy is getting better or worse over time. The use of control card was briefly discussed in the theory section of this study. While the theory section introduced briefly some of the reasons why the control card should be used, this chapter focuses on explaining the different examples it cannot be used for.

The examples that are presented here have been chosen based on typical error situation mentioned in subchapter 5.3 of this study. The specific control card presented here is the one available in the forecasting software. However, one of the most important things to remember when using the control card is that only one error measure should be chosen at once, because of the different scales and measures. For example, if percentage errors of tens and absolute errors of hundreds are shown simultaneously in the control card the scale is obviously determined by the larger errors, which means that it can be difficult to see the changes in the smaller (percentage) error values. This can deteriorate the overall image of the forecast and its errors.

The first example presents the situation where the forecast is regularly larger than the actual demand (AC-segment, New Parts). In this situation, most of the error values are below the zero line in the lower graph. The same can be seen in the upper graph, which shows the forecasted and the actual amounts, whereas the lower graph shows only errors. Figure 5.2 depicts the first situation.

Picture 5.2. Forecast is regularly larger than demand (adapted from the Forecasting Software).

The different error measures on the right column are the ones that can be seen in the control card and in the forecast graph as well. The first situation can best be seen when using error and mean error (also abs. percentage error). If the mean error was used in this situation it would be used the mean error (ME) line would be constantly below the zero, which implies that the forecast is uncontrollable (subchapter 3.2.3). In general the

use of either of these measures, error or mean error, can help to find a possible bias in the forecasts.

The second situation, seen already in figure 5.1, shows how the control card can be used to find demand peaks or lows with the help of MAD. In this case, MAD is decreasing or relatively constant but after a demand peak or low, it increases drastically in one period and then starts to decline again or stay at a new constant relatively constant level. As figure 5.1 also shows, peaks or lows are easily detectable from the forecast graph as well and control card is not always needed to find them. However, a strong increase in MAD is a consequence of a relatively strong peak or low: if the peak or low is only strong by absolute value, the change in MAD is not always so apparent, as figure 5.3 shows.

Figure 5.3. Demand peaks (adapted from the Forecasting Software).

As figure 5.3, in this case a proper measure to use is either error or absolute error. The aforementioned brings up an important with the use of control card and forecast graph, which is to be mindful of the scale of the graph and control card: sometimes the changes are relatively small but can be very significant in the absolute value. This is also one of

the reasons why the use of both absolute values and control card complement each other.

In the third situation the forecast has been in control at first, with (absolute) error values close to zero, but after a while the errors have been increasing for some reason.

Figure 5.4. Forecast that is out of control (adapted from the Forecasting Software). In the third situation the measure absolute error was used, because it shows the overall development of the accuracy. The same could have been noticed when using simple error: in that case the errors would have been close, in both sides of the zero in the beginning but after period 01/2011 they would have started to properly deteriorate on both sides. The error movement would have been similar to the one in the upper graph where the actual demand deteriorates to both sides of the forecast. After period 01/2011, the situation resembles a case where the random variation of demand for certain product is strong. The errors are distributed on the both sides of the zero line (when used the measure “error”) and are usually large on percentage value.

As previous examples showed, absolute error (or absolute percentage error) can be used when wanting to find out the progress of the forecasts; is it getting better (decreasing errors) or worse (increasing errors) over time. MAD and MAPE are also suitable for this since they show the average after every period; decreasing MAD or MAPE means forecasts are improving and vice versa. However, sometimes these measures can be slightly misleading, as was the case in figure 5.3. Additionally, MAD or MAPE should not be used alone, when estimating the progress of the forecast; especially in the case of strong random variation. Figure 5.5 shows why.

Figure 5.5. Misleading MAPE (adapted from the Forecasting Software).

As seen above, the development of MAPE would imply that the forecast is getting more and more accurate with the decreasing MAPE after period 09/2009. However, when looking at the individual errors it is quite difficult to suggest that the forecasting accuracy is actually improving. Surely there is some increase in accuracy in the beginning but the overall development is not quite as reliable.

The use of the aforementioned examples as feedback mechanisms are discussed further in chapter 6, which discusses the aspects presented in chapter 5 in more critical detail in relation to prior studies and their possible effect on the original research problem of this particular study.