Theory of Constraints Handbook - James Cox Iii [197]
Why Is It Impossible to Find a Good Forecasting Model?
The advanced forecasting modules existing today try to model the demand and create a good answer to the availability puzzle: What product to hold at which place (where) and when. Notice that this puzzle has three questions: what, where, and when. To be a good forecast of demand, forecasting has to answer each of these questions. The forecasting mechanism, no matter how good it is, cannot really predict what the demand would be like.
FIGURE 11-1 A typical push supply chain.
With respect to forecasting, one must consider some fallacies regarding statistics. These fallacies and a discussion of each are provided in the following sections.
1. The fallacy of disaggregation.
2. The fallacy of the mean.
3. The fallacy of the variance.
4. The fallacy of sudden changes.
The Fallacy of Disaggregation
The first fallacy is that aggregation or disaggregation has no impact on variation. The fact is that the more disaggregated the data is, the higher the variation is of those data elements. In our distribution environment, for the question of “How much demand for this product?” the answer for the M/D location is very accurate with low variability but the answer to this same question for a specific retail location is quite inaccurate with high variability. This phenomenon stems from the fact that fluctuations average out on the aggregated events (assuming they are independent events). If we predict the sales at 100 different locations, we might get an answer that sales in an average location will range from 10 to 25 units a day. If we ask the same question on the overall quantity that we need to manufacture, we will get a much narrower range as an answer—probably something ranging from 1650 to 1850. If we would just take the lows (10) and highs (25) of each of the 100 consumption points and aggregate them, we will get a much worse answer—from 1000 to 2500. This point is demonstrated in Fig. 11-2. Note the high variation at the retailer versus the lower variation at the M/D warehouse. The rule then becomes the higher the aggregation, the better the forecast.
FIGURE 11-2 The mathematical effect of aggregation.
The Fallacy of the Mean
The second phenomenon relates to the wrong interpretation of data—people using statistics must have a good enough understanding of the mathematical logic that stands behind the forecast. Huge mistakes are made daily in almost every organization because of a lack of understanding of statistics. For example, the average demand in the previous example is 17.5 (assuming a normal distribution and a high and low of 10 and 25). Suppose that we stocked 17.5 units at each retail location. Do you think we would sell 1750 units? Never! There are stores that would have demand less than 17.5 units a day and we would have excess inventory (not sold) in these stores. There are other stores where we stocked 17.5 units and the demand was greater than that amount. We can only sell the 17.5 units we have that day. Therefore, overall we would sell far less than the 1750 aggregate demand. A clever man not experienced in statistics might deduce from this example that the consumption will be between 1650 and 1850 for all consumption points, that each consumption point will have a consumption between 16.5 and 18.5, keeping 19 units for each location, and running out of stock in a fairly large number of them, while others will be left with a lot of stock they can’t sell. The fact that we got an aggregated range does not mean that it can be applied to the points that make up this sum. Forecasting algorithms are getting more and more complex