Let us start by simplifying the task. We can throw a third of it away from the start. Textbooks traditionally divide forecasting techniques into three: qualitative techniques, time series techniques and causal models. A little research reveals that "qualitative techniques" is academicspeak for guesswork. The books talk of Delphi methods, where a lot of people make guesses and you average out the answers. Or "scenario methods", where a few people make the guesses instead. These techniques are not of much use to the average business. They are long-range guesstimates employed to "forecast" things like technological developments in a particular industry.
That leaves time series analysis techniques and causal modelling. Another glance at the contents page shows that these, too, divide into impressive-sounding individual techniques, with names like "exponential smoothing" and "multiple regression". Then, as page after page of equations swim into view, confusion sets in: not just because of the mathematics but also because it is difficult to tell which technique is suited to which circumstances.
However, the choice is not really so difficult. It revolves largely around two very simple questions. First, can what actually causes demand be identified? Second, can any quantified data on it be obtained? These questions lie behind any choice of forecasting technique. Forecasting is rather like selecting a gear in a car: no one technique is always right - it all depends on the circumstances. The forecaster chooses a technique appropriate to the data available and its relevance to the underlying causal factors.
The most readily available data often comes in the form of month-by-month sales histories. But these simply say what happened to sales, not why. The fact that sales have been increasing by 5% a month for the past year might mean that they will go up 5% again next month. Then again, they might not. There is no causal link between turning the page of the calendar and an automatic increase in sale.
It is customers, not calendars, that make buying decisions. The key to causal forecasting is understanding the factors behind these decisions. One of the most important is simply the raw requirement: the level of demand which customers are experiencing for the products and services that they sell. But other factors impinge too: the general economic climate, price, competitiveness, and so on. Aggregated together, these individual buying decisions build up to the graph on the sales manager's wall.
Theoretically, therefore, it is better to try and get a handle on the causal factors themselves, rather than to treat time as a proxy for them. This is the principle behind the causal modelling approach. Causal modelling involves building a "model" (consisting of one or more equations) which embodies the relationship between one or more "independent" variables and the "dependent" variable that they influence - namely demand. The equations can be very simple indeed. A typical multiple regression might say "so much of the movement in month-on-month sales is explained by factor A, so much by factor B and so much by factor C".
Provided that the model is well constructed - at least conceptually - then causal techniques offer high levels of accuracy. But there is a snag: they do need data - stretching into the future - on the factors that have been incorporated. It is pointless to consider causal techniques if there are not any numbers to plug into them. And while industrial giants might consider making the investment necessary to compile such data, this is simply not an option for most companies.