When learning about forecasting algorithms, it is easy to get lost in technicalities and forget a few preliminary points.
Forecasting is not about a single number. We are already familiar with inferential statistics and confidence intervals. Hence, we should keep in mind that a single number, i.e., a point forecast, may be of quite little use without some measure of uncertainty. As far as possible, forecasting should be about building a whole probability distribution, not just a number to bet on.
Choose the right time bucket. Imagine that you order raw materials at the end of every week; should you bother about daily forecasting? Doing so, you add unnecessary complexity to your task, as weekly forecasts are what you really need. Indeed, it is tempting to use large time buckets in order to aggregate demand with respect to time and reduce forecast errors; this is not advisable if your business process requires forecasts with small time buckets.
What should we forecast? If this sounds like a dumb question, think again. Imagine that you are a producer of T-shirts, available in customary sizes and a wide array of colors. Forecasting sales of each single item may be a daunting task. As a general rule, forecasts are more reliable if we can aggregate items. Rather than forecasting demand for each combination of size and color, we could aggregate sizes and consider only forecasting for a set of colors. In fact, demand for a specific combination of color and size may be rather volatile, but the fraction of population corresponding to each size is much more stable. We may forecast aggregate sales for each T-shirt model, and then use common factors across models to disaggregate and obtain forecasts for each single combination. Of course, in the end, we want to forecast sales for each individual item, but in the process of building a forecast we use suitable aggregation and disaggregation strategies. This approach is also quite powerful in pooling demand data across items, thus improving the quality of estimates of common factors, but it must be applied to compatible product families. We cannot apply size factors that are standard for adults to T-shirts depicting cartoon characters (maybe).
Forecasts are a necessary evil. Common sense says that forecasts are always wrong, but we cannot do without them. This is pretty true, but sometimes we can do something to make our life easier and/or mitigate the effect of forecast errors. If our suppliers have a long delivery lead time, we must forecast material requirements well in advance. Since the larger the forecast horizon, the larger the uncertainty in forecasting, it is wise to try whatever we can to shorten lead time (possibly by choosing geographically closer suppliers). The same applies to manufacturing lead times. Sometimes, seemingly absurd approaches may improve forecasting. Consider the production process for sweaters. Just like with T-shirts, forecasting a specific combination of color, model, or size is quite hard. Two key steps in producing a sweater are dying the fabric with whatever color we like, and knitting. The commonsense approach is to dye first and knit later. However, the time needed to knit is much longer than the time needed to dye. This means that one has to forecast sales of a specific combination much in advance of sales, with a corresponding criticality. By swapping the two steps, knitting first, one may postpone the final decisions, and rely on more accurate forecasts. This is a typical postponement decision,3 whereby the impact on cost and quality must be traded off against the payoff from better matching of supply and demand.
The general point is that the forecasting process should support the surrounding business process and should help in making decisions. Forecast errors may be numerically large or small, but what really matters is their economic consequence.
As a last observation, let us consider another seemingly odd question: Can we observe what we are forecasting? To be concrete, let us consider demand again. Can we observe demand at a retail store? In an age of bar codes and point-of-sale data acquisition, the answer could be “yes.” Now imagine that you wish a pot of your favorite cherry yoghurt, but you find the shelf empty; what are you going to do? Depending on how picky you are, you might settle for a different packaging, a different flavor, or a different brand, or you could just go home quite angry. But unless you are very angry and start yelling at any clerk around, no one will know that potential demand has been lost. What we may easily measure are sales, not demand. Hence, we use sales as a proxy of demand, but this may result in underforecasting. If this looks like a peculiar case, consider a business-to-business setting. A potential customer phones and needs 10 electric motors now; your inventory level is down, but you will produce a new batch in 5 days. Unfortunately, the customer really needs those motors right now; so, he hangs up and phones a competitor. Chances are, this potential demand is never recorded anywhere in the information system. Again, we risk underestimating demand. Statistical techniques have been devised to correct for these effects, but sometimes it is more a matter of organization than sophisticated math.
Leave a Reply