ProQua

Forecasting: The Lost Art

Computers are not enough. Forecasting needs the human touch and intervention at every step of the process.

Like so many operations management skills, forecasting has, in the past several decades, become highly mechanized and computerized. Senior management and high-level planning teams rely on the forecast estimates for a month or quarter ahead and overall volumes for a year or more. Such projections, particularly if reasonably accurate, can smooth inventory use, avoid scheduling difficulties, reduce expediting costs, and ensure reliability of customer delivery.

Most forecasts are developed using spreadsheets such as Excel, stand-alone software, or programs embedded in material requirements planning systems. These software products appear easy to understand, at least superficially, and are often accompanied by glitzy graphics and print format tools. However, a recent review of forecasting practices of several Fortune 50 companies confirmed that few executives understand either the visual displays or the calculations behind them. In fact, computerization appears to have reduced or eliminated the human component (the art) of forecasting and thus increased forecast error.

The apparent simplicity and user-friendliness of computer forecasting may encourage executives and forecasters alike to ignore a major component of forecasting, the forecast error, and, more important, the process of reducing error. To be meaningful, the forecast must be two numbers: first, the projected future value of a specific model and, second, the forecast error. Several forecasting models are then evaluated using one or several error measures. Of course, senior management and teams are directly concerned with the projected forecast value; however, they often disregard how much error is inherent in the method and rarely look toward error reduction.

The projected forecast and the error of that forecast may be viewed as complementary values; that is, for each period, the percent forecast accuracy and the percent error add to one. Differently stated, percent error equals one minus percent accuracy. After the period-by-period error has been calculated, these values are aggregated into various forecast error measures and the diagnosis and reduction of the error component proceeds.

Diagnosis of error values permits the forecaster, for example, to identify and manually or intuitively adjust for a biased (under- or over-forecasting) method, or to select a specific forecast method with a low error. For example, the builder of a new home would likely want a positively biased forecast (understating demand) because overbuilding costs and risk are high. Alternatively, park directors would want a negatively biased forecast of expected swimmers at a pool or beach (overstating demand) because insufficient lifeguards with high numbers of swimmers might risk an accident. Evaluation by other error measures could be similarly conducted.

My sense from numerous corporate visits, as well as a review of recent practitioner literature, is that there is an all-too-rapid effort to get to the forecast without a full diagnosis of the error or attempt to reduce error. The following series of simple activities can help diagnose the error with the expectation that the forecast can be dramatically improved.

Integrate the forecast with corporate strategy. (How much error is acceptable?)
Plot a histogram of monthly/quarterly data for a high-volume product family.
Visually determine the level, trend, and cycle of the plot.
Forecast the data using a baseline (simple) method. Determine overall error.
Intuitively identify where the greatest error occurs and address why.
Reduce error by changing the model, number of periods, exponents, weights, or other factors.
Use improved models to forecast products in this and other families.
Regularly monitor forecast error and continue to reduce it.

This process is highly iterative. Forecasting, like many areas of operations, encourages periodic review of the forecast against demand and the charting of forecast error. Well-managed companies use such processes as a part of the sales and operations planning review to permit a visible, cross-functional assessment of the drivers of forecast error. For example, marketing professionals would be aware of competitor promotional activity and environmental factors, and operations would be aware of potential material or labor problems. But more important, this visibility identifies the forecast as the starting point for mid- and short-range operations planning and emphasizes the impact of forecast error on inventory, scheduling, costs, and customer delivery.

Simply stated, even these basic diagnostic processes can not be done by a computer alone; there must be human diagnosis and intervention at every step of the process. Additionally, there must be a continuous effort to improve the forecast by understanding the underlying movement and vagaries of the data. Careful application of this process can notably improve forecast accuracy.

This, then, is the art of forecasting, and the expected results that can be achieved by proactive companies. Of course, the forecast is always in error; however, the more important questions are: by how much, why, and how can this error be reduced? It is these human and intuitive processes that have been lost over the past several decades of increased dependence on computerized models. Based on my observations, this art may also have been lost in corporate decision-making activities and planning processes.

�Peter W. Stonebraker, Ph.D., CPIM, professor of operations management, College of Business and Management, Northeastern Illinois University, can be reached at (773) 442-6124 or via e-mail at p-stonebraker@neiu.edu.