Forecasting 101: Going Beyond Automatic Forecasting
Part 3: Improving Your Forecasting with Topdown Models
This installment of Forecasting 101 presents the third of three articles about going beyond automatic forecasting. The first article presented an overview of automatic time series approaches, examining how they work, the pros and cons of using them, and situations where they should not be utilized. The second article described the evolution of a successful forecasting process, noting that it often involves a progression from ad hoc spreadsheets to automatic time series approaches and finally to customized approaches being applied to the subset of the items where they add value. The second article described one such customized approach—event models—an advanced forecasting method which will often outperform automatic time series approaches for data where special events such as promotions, strikes, moveable holidays, etc. have occurred during the historic demand period.
In this article we will look at another common customized approach—topdown modeling. Topdown approaches leverage structure that exists in higherlevel aggregate data to improve forecasts at lower levels of the forecasting hierarchy.
What is a topdown forecast?
Most organizations deal with multiple levels of aggregation and require consistent forecasts at all levels. For instance a beverage company might need a forecast for total sales, as well as a forecast for each brand, each customer segment, each container type and each SKU.
When preparing forecasts for hierarchical data, you must decide upon a reconciliation strategy (i.e., you must decide how to enforce that the forecasts are consistent across levels). One approach is to apply statistical forecasting methods directly to the lowestlevel demand histories and construct all grouplevel forecasts by summing the lowerlevel forecasts—this is known as a bottomup forecast. An alternative approach is to use statistical forecasting methods on more aggregated data and then to apply an allocation scheme to generate the lowerlevel forecasts—this is known as a topdown forecast.
Let’s illustrate these approaches with a very simple example.


ModelBased


BottomUp


TopDown

6 packs


70


70


84

12 packs


30


30


36

Cans


120


100


120

The column labeled ModelBased contains the forecast that would be created by applying a statistical forecasting method directly to the given data set. Thus in our example, if you forecasted the demand for 6 packs directly the forecast would equal 70, if you forecasted the demand for 12 packs directly the forecast would equal 30 and if you forecasted the demand for total cans directly the forecast would equal 120. Notice that forecast for Cans does not equal the sum of the forecast for 6 packs and the forecast for 12 packs. When the three data sets are forecasted independently using their own histories, there is no statistical mechanism that forces them to reconcile and they are extremely unlikely to do so. At times, the modelbased forecast for the group can be quite different than the sum of the component series’ modelbased forecasts.
Notice that in the bottomup approach the modelbased forecasts are used for the itemlevel data (6 packs and 12 packs) and the grouplevel forecast (Cans) is calculated as their sum. In the topdown approach the modelbased forecast is used for the grouplevel and the itemlevel forecasts are calculated by adjusting their modelbased forecasts proportionally so that they sum to the grouplevel forecast.
An alternative topdown approach is to not forecast the itemlevel data at all and just disaggregate the grouplevel forecast by applying proportionality factors. This method would be appropriate when the proportions are constant and known (e.g., disaggregating shoe sales using a size chart or a finished good using a bill of materials).
When do topdown approaches improve the forecasts?
The decision to use a topdown or a bottomup approach often hinges upon two important issues.
1. Are the lower level units likely to require distinct statistical models?
This would be the case if the market forces that shape sales at the lower level are different. Different markets, different advertising and promotion, and different distribution all favor creation of distinct modelbased forecasts. If apples and oranges have distinctly different markets, then you will probably do better to forecast them separately.
If not, then there is often a distinct advantage to forecasting top down from the aggregated “fruit.” If the lowerlevel data are statistically similar, forecasting at a group level will generally result in a more accurate forecast because:
A. a higher volume of data is available.
B. there is less “noise” (random variation) in the aggregated data that could skew the forecast.
C. the aggregated data will often exhibit a more pronounced structure, making patterns easier to recognize and forecast.
2. Is there sufficient statistical information in the lower level historic sales to construct a model based only on those sales?
Many organizations that need to generate lowlevel forecasts discover that at the lowest levels there simply is not enough structure to generate meaningful statistical forecasts directly from the lowlevel data. In these cases there is little choice but to generate the lowestlevel forecasts not with statistical models, but rather by using some type of topdown allocation scheme.
Let’s illustrate this with an example.
Figure 1
Figure 2
Figure 1 shows monthly sales for a brand of cough syrup. Figure 2 shows monthly sales for a specific SKU. The company assigns a unique SKU number to each flavorbybottlesize combination that it produces.
Consider the two graphs. Notice that at the brand level, there is more structure to the data. The seasonal pattern is readily apparent and there is less noise. More than three years of demand history is available at the brand level, while only 10 months of history exists for the recently introduced SKU.
In this example, the lack of history at the SKU level doesn’t allow you to build a seasonal forecasting model directly from the data. Thus, since cough syrup is clearly a seasonal product, a bottomup approach will yield very poor forecasts. On the other hand, a topdown approach allows you to capture the seasonal structure that exists at the group level and introduce it to the SKUlevel forecasts via the topdown adjustments.
Summary
Most organizations find that using automated time series approaches—such as those implemented in Forecast Pro—work quite well for the vast majority of their items and offer substantial advantages over ad hoc forecasting with spreadsheets. In this series of articles, we have discussed the pros and cons of automatic time series approaches and also explored event modeling and topdown forecasting—two alternative forecasting approaches that often improve forecast accuracy for the subset of your items where automatic time series models do not perform well.
About the author:
Eric Stellwagen is Vice President and cofounder of Business Forecast Systems, Inc. (BFS) and coauthor of the Forecast Pro software product line. He consults widely in the area of practical business forecasting—spending 2030 days a year presenting workshops on the subject—and frequently addresses professional groups such as the University of Tennessee’s Sales Forecasting Management Forum, APICS and the Institute for Business Forecasting. Recognized as a leading expert in the field, he has worked with numerous firms including CocaCola, Procter & Gamble, Merck, Blue Cross Blue Shield, Nabisco, OwensCorning and Verizon. He is also currently serving on the board of directors of the International Institute of Forecasters (IIF).
