Top-Down or Bottom-Up: Which Is the Best Approach to Forecasting?

Abstract

Bottom-up approach to forecasting yields better forecasts than the when forecasts are statistically generated ... top-down approach works well when SKUs within a family are of equal size ... carefully specified statistical models give better forecasts than those judgmentally developed. Frms often market families of items that are closely related in function, styling, and design. For example, a manufacturer of irrigation equipment may offer several families of irrigation equipment including impact sprinklers, rotator sprinklers, and gun sprinklers. The impact sprinkler family might then consist of two individual stock keeping units (SKUs) such as low angle and high angle impact sprinklers. Two general approaches have been suggested for developing forecasts for individual items in a family. One approach might be referred to as topdown (TD). In this approach, the family data are used to develop a family forecast which is then disaggregated into individual items based on their historical fraction of sales. The alternative employs a bottomup (BU) approach where a separate forecasting model is developed for each item in the family. The forecasting methods used by planners can be either objective or subjective in nature. Objective methods are statistical or other well-specified processes through which data of any type (subjective or objective) can be translated into forecasts. With these formal methods, the forecast could be replicated by anyone using the same data and methodology. Subjective methods are based on informal, experiential and intuitive processes and are often referred to as judgmental forecasting. Like formal methods, judgmental forecasting can also be based on either objective or subjective data. In some situations, formal analysis may also be involved. The important distinction is that the forecast is derived intuitively and is unlikely to be identical to another person's forecast developed from the same information. In most cases, the decision maker would be better off using an objective forecasting technique. However, despite the existence of formal forecasting methodologies and the ready availability of computer resources, most forecasts are still made the old fashion way-- intuitively. This study examines how accurately professional planners can forecast items in a family when they are presented with either aggregate (family) time series or item time series and asked to make judgmental forecasts. In addition, the planners ' forecasts are compared with those generated by computer using exponential smoothing models. Computer forecasts using both the TD and BU approaches are also compared. There is a growing debate concerning which approach -- TD or BU -- yields more accurate forecasts. Several authors have pointed out the weaknesses of TD which include the introduction of bias, the loss of information due to aggregation, and measurement error. Three studies have empirically compared the relative performance of the two approaches. Dunn, William, and Spiney (1971) found that forecasts aggregated from lower-level modeling worked best in forecasting demand for telephones. Dangerfield and Morris (1988) found that forecasts for individual items were more accurate when separate exponential smoothing models for each item (bottom-up) were used. Dangerfield and Morris ( 1992) then used a subset of the M-competition data to examine this issue further. They used Winters' triple smoothing exponential models with mean absolute percentage error (MAPE) as the primary measure of accuracy. They found that BU forecasting resulted in more accurate forecasts for nearly three out of four series tested. BU was even more successful when item series were highly correlated or when one item dominated the aggregate series. Similar results have been obtained using econometric methods to forecast financial information: Kinney (1971), Collins (1976), and Pacter (1993). METHODOLOGY We used a series of forecasting case problems to examine the forecasting performance of individuals as well as those produced by a computer-based exponential smoothing model using both the BU and TD methods. …