Demand forecasting Mathematics or crystal ball?
One sometimes has the impression that forecasting is the magic of the 21st Century. The ghosts of the past are replaced by exogenous factors and mediums by mathematical models. And as with the magic of days gone by, we know the results are often not very reliable, and we are not too sure where they come from, but it is the only information we have… Unlike magic, the forecast is an information resource without which we would be unable to effectively manage business operations. Fortunately, forecasting is not necessarily as esoteric as magic. With the application of a little method, we can all understand its mechanisms, provided we do not allow ourselves to be blinded by the complex terminology of experts.
Measuring the quality of forecasting
Before turning to the subject of forecast preparation, we should seek to understand how one measures the quality of forecasts. To assess the quality of a forecast, one should examine its reliability, bias and stability. Forecast reliability (often measured as the average of absolute differences) indicates the capacity to forecast the right quantities over the right periods. When the forecast is higher than the actual level over a period and lower over the next, the two errors do not compensate for one another. Forecast reliability is the most rigorous indicator. Forecast bias (often measured as the average of differences) indicates whether the forecast tends to be too optimistic or pessimistic. If the forecast is sometimes too high and sometimes too low, the reliability may be low but the bias is also low. If the forecast is higher for each month compared to reality, the bias may be much greater for the same level of reliability. From a planning point of view, the forecast bias is very important: if the bias is near to zero, supply based on a too high forecast will be compensated for the following month. If the forecast has a positive bias, stock levels will always be higher than planned. If the forecast has a negative bias, frequent disruptions in supply may occur. Stability indicates how the forecast for a certain period changes over time. If one observes large differences between forecasts from one month to the next for the same sales months, it will be more difficult to manage operations correctly than if the forecast converges towards actual figures.
Does the past guarantee the future?
Traditional forecasting methods divide the sales history into 3 parts: the level, trend and seasonality.
- The level is the simplest: by observing the sales history, one can estimate the average demand. The level does not depend on time. Once one has identified it, the level remains the same for all future periods for which one performs a forecast.
- The trend is not much more complicated: by observing the sales history, one can estimate if demand will increase or decrease over time. The rate of increase or decrease of demand will be constant for future periods.
With these two elements, forecasting is a straight line that starts at the level identified and rises or falls over each period according to the trend.
- The final element traditionally added is seasonality. Seasonality is the cyclical fluctuation of demand around the average. The seasonality will make the straight line vary according to variations observed in the past.
When deciding which effects to include in the forecast, one question is crucial: what is the probability that phenomena observed in the past will occur again in the future? Take the example of the trend. Suppose that you observe a rise in demand for a product over the last 6 months; would it be wise to imagine that this increase will continue over the next 3 months? And over the next 6 or 12 months? The same question is even more pertinent as concerns seasonality. If the October 2009 sales for a product represented 80% of average monthly sales in 2009, what is the probability that the October 2010 sales will also be around 80% of average monthly sales? The majority of techniques limit the risk of identifying false seasonalities by calculating seasonality factors over several years. For this reason, one normally requires 36 months of history to define reliable seasonality factors. Even with these precautions however, the use of seasonality factors remains a dangerous technique. If variations observed in the past are not really recurrent, one risks generating false forecasts that will disrupt operations more than they assist with them. The simplest forecasts are sometimes the most effective.
Including external factors
Up to this point, we have considered techniques that are based only on sales history. Clearly however, demand for a product will also depend on certain external events. We should therefore include these in the forecasting model. Firstly, we should include the effects of promotions. This approach is very popular in the world of mass consumption products where companies quite easily achieve half of their sales during promotions. To model promotions, it is important to be able to distinguish reference price sales from promotional sales. The forecasting model will calculate a reference price forecast, to which it will apply a promotion history. Some market software enables promotional data to be manipulated with relative ease, making the forecaster’s task correspondingly easier. In certain sectors, it seems logical to take into account other external factors, such as the weather forecast or econometric parameters, including the price of raw materials or conjunctural indicators. Regression techniques enable the generation of these types of forecasts. During the construction of a regression model, one must answer two questions: what is the relation between the demand and the external factor, and what is the reliability of the external parameter forecast? Firstly, one must assess the correlation between the demand and the external parameter(s). As was the case for seasonality factors, it is important not to assess only the reliability of the model in the reproduction of the past, but also its capacity to forecast future demand. Several sophisticated techniques exist to check this predictive capacity. Based on experience, we can add that the best validation of a regression model remains validation using commercial intuition. When the correlations identified seem logical to market experts, the model is very probably useable. When market experts are surprised by the correlations identified, the model may be generating abnormal forecasts. It is logical that a forecast based on external factors will only provide reliable results if the external factor forecasts themselves are reliable.
The importance of the human factor
Finally, we cannot over emphasise the importance of the forecast process. After all, it is the men and women involved in the forecast calculations who determine its performance. To ensure that all those involved work well together, it is important to ensure that each team member works in the area in which they specialise. Forecasters are more valuable when they make forecasts based on a large number of references or if there are many actors involved in the process. In the first case, forecasters are mathematicians who produce statistical forecasts. In the second case, they require considerable interpersonal skills to co-ordinate all the actors. Sales team members often have a good sense of the development of demand in relation to major products or customers over the coming 3 months. As soon as one requires more long-term forecasts (12 - 18 months for example, to supply an industrial and commercial plan), it is sometimes best to include marketing in the process. The sales management is best placed to provide forecast corrections over future weeks… Sometimes, the process of creating a reliable forecast remains shrouded in mystery. On occasions, it is even impossible. But if one understands the main techniques and if all members of staff in the company work together, simple common sense is all that is required to generate an acceptable forecast, and the crystal ball will become a thing of the past.