Widely used methods of automated forecasting for production and inventory control contribute to the severity of recessions. We describe an approach to forecasting that should reduce the damage caused by the use of current statistical packages when encountering a substantial change in business conditions.

Assume that you are a manufacturer. It is early 2008 and you have been reading about a possible recession. How should you go about forecasting your production and inventory? Should you use your knowledge about bad times ahead, or just rely on the statistical models that churn out forecasts for the 5,000 thousand items under your control. If you do want to use your knowledge about the products, how can you do it?

If you are like most managers, your statistical models will not serve you well. They lack crucial information that you have.

One type of information that managers have relates to expectations about trends. Armstrong and Collopy (1993) use the notion of “causal forces” – the influences on demand for a product -- to describe expectations. They identified five types of time-series: growth, decay, opposing, regressing, supporting, and unknown. For example, a growth series is one where managers expect the trend to be upwards no matter what the historical trend has been. Similarly, a decay series is one where they expect the trend to be downward, perhaps because a currently successful product will be replaced by a superior product.

Armstrong and Collopy found that college students with little knowledge about a subject area (say automobile sales) were able to quickly identify the causal forces for a wide variety of time series. Of course, it would probably be better to use people who have expertise in the relevant areas.

Currently, production and inventory forecasting models assume that the causal forces always support the historical trends. In fact, Armstrong and Collopy were unable to find any time-series where the causal forces always support the trend.

Although the assumption of supporting causal forces is unfounded, the world often looks as though the trends are always supported by causal forces. Consequently, when the good times roll, standard forecasting packages perform well. However, when the historical trends are contrary to the expected causal forces, forecast errors become large and contribute to oversupply in recessions and shortages in recoveries.

To deal with this and related problems, Collopy and Armstrong developed and published a set of 99 rules as the basis of a method they dubbed “rule-based forecasting (RBF).” Although a number of RBF programs have been developed privately, there is no commercial package. While applying RBF without software is onerous, one simple and inexpensive rule can achieve much of the benefit of RBF by reducing errors when forecasting contrary series.

Here is the rule: When a time series is identified as “contrary,” do not extrapolate a trend.

Once the causal forces have been identified for each series (or sets of series), a simple rule can be introduced to compare the expected trend with the historical trend. When the expected and historical trends are “contrary” to one another, set the trend forecast to zero. The rule was tested for using a variety of economic data. It was also tested using data on epidemics in China and on U.S. Navy manpower planning. The contrary series rule improved accuracy in every test. The gains were especially large for contrary series with strong causal forces; errors were cut by half.

When a recession is anticipated (or perhaps even in its early stages), a large number of formerly growth series should, on the basis of expectations, be re-coded as decay, thus leading many series to be classified as contrary. Had firms used the contrary series rule, they would have substantially reduced their production and inventories on many items well before the statistical models realized that a downturn was underway. Eventually, the recession will end, and here again, use of the contrary series rule will aid the firms, as they will be able to build inventories to meet demand.

Contrary series also affect prediction intervals because their errors tend to be larger in the direction of the trend expected based on the causal force. This will have an impact on setting the desired inventory levels. A quick solution is to shift the prediction intervals in the direction of the causal forces. This would help to move quickly to reduce inventories in anticipation of a downturn– and to a larger inventory when a recovery is expected. These seem like sensible things to do, but statistical models will not figure that out on their own.

You must ask your statistician to insert a contrary series rule along with your judgments on the causal forces relevant to the types of time-series your are dealing with. Some software providers have told me that they will happy to do this if clients ask.

By now, your time series are heading south, so you missed an opportunity. But if you add this capability to your forecasting package now, you will be in a position to respond quickly when the recession is expected to end – more quickly than those who have not used this approach. It takes a long time for statistical models to realize that a recession is underway.

This structured use of managers’ knowledge almost always leads to better forecasts for production and inventory control. When the economy is in recession or is recovering from one, it is especially useful. Do not let your statisticians tell you otherwise.

Note: For more information on this topic, see J. Scott Armstrong & Fred Collopy (1993), Causal Forces: Structuring Knowledge for Time-series Extrapolation, Journal of Forecasting, 12, 103-115. In addition, adjustments must be made for assessing prediction intervals for contrary series. For this, see J. Scott Armstrong & Fred Collopy (2001), Identification of Asymmetric Prediction Intervals through Causal Forces, Journal of Forecasting, 20, 273-283. Full text of these papers can be obtained at ForPrin.com. They provide all that your statistician (or forecasting software provider) needs to know. We talked to some software providers; they say that they will be happy to make the adjustments if their clients ask. Information about software providers can be obtained from forecastingprinciples.com.