Forecasting 101: Going Beyond Automatic Forecasting
Part 2: Improving Your Forecasting with Event Models
This installment of Forecasting 101 presents the second 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. In this article we explore event models—an advanced forecasting method which will often outperform automatic time series approaches for data where special events like promotions, strikes, moveable holidays, etc. have occurred during the historic demand period.
The evolution of a successful forecasting process.
The evolution of a successful forecasting process often involves several steps. It is very common for an organization to start by creating spreadsheets containing the demand history for the items to be forecasted using simple ad-hoc formulas such as “same as last year plus a percentage” to establish the forecasts. The accuracy of the forecasts generated this way is often poor and using Excel to tackle the job usually leads to large unwieldy, complex spreadsheets that are prone to unintentional human error and are difficult to maintain as the business and its forecasting staff changes.
Often, the next step is to purchase a dedicated forecasting package such as Forecast Pro and use its default forecasting procedure. In Forecast Pro the default procedure consists of an automatic algorithm that selects among different time series forecasting methods for each individual item being forecasted. Moving from spreadsheets to this type of approach almost always improves the accuracy of the forecasts, simplifies the mechanics of generating the forecasts and helps to formalize the forecasting process.
Although most organizations find that automated time series approaches work quite well for the vast majority of their items, there is usually a certain fraction of items where customized forecasting approaches can outperform the automated approaches. Thus, the final step is to tweak the forecasting process to incorporate customized approaches where they add value. One common customized approach is event modeling—the subject of this article.
What is an event model?
As we explained in the last installment of Forecasting 101, a time series forecasting method is a forecasting technique that bases the forecast solely on the history of the item being forecasted. When demand for an item is being driven by such factors as sales levels, trends and seasonal patterns, time series methods tend to work quite well. However, business data often contain responses to events that cannot be captured as part of the level, trend and seasonal components. Examples include product promotions, moveable holidays, business interruptions and other irregular occurrences. When a significant amount of demand is being driven by these types of events, time series methods will not work very well.
An event model is a forecasting method that is designed to quantify the impact of events and use this information to improve the forecasts. The input for an event model is both the historic demand for the item to be forecasted and a schedule listing the timing of any events that have occurred historically and (if applicable) the timing of any future events that will occur in the forecast period.
Event models are an extension of exponential smoothing models. They work by generating indexes to adjust for the different types of events. These indexes are estimated and used to model the impact of each event type in almost exactly the same way that the seasonal indexes in a traditional exponential smoothing model are estimated and used to model the impact of each seasonal period.
For a more detailed discussion of how event models work and are implemented, you should consult the Forecast Pro Statistical Reference Manual. This pdf document is accessible via the Forecast Pro Help System and is the primary suggested reference for event models and all forecasting techniques, statistics and algorithms found in Forecast Pro.
The green line above represents monthly demand for a popular consumer packaged product. The product is heavily promoted and the timing of the promotions varies from year to year. A time series method will not work well for a product like this due to the fact that a significant portion of the demand is being driven by the promotions. The event model above successfully models the response to the promotions and yields an accurate forecast (in this example, the timing of future promotions is known).
How hard is it to build event models?
Building an event model is very straightforward and does not require deep statistical skills. It does require that you have a knowledge of the events which have occurred, that you judge whether an event’s impact on your data is significant enough to warrant special treatment and that you decide how many unique event types are needed (e.g., are the promotions that impacted two different periods similar enough to model as one event type or do they need separate event types and therefore separate indexes). You also need to create an event schedule indicating the timing of each event.
Obviously the amount of effort required to build an event model is greater than that required for an automated time series method. However, for data sets where events drive considerable demand, the payoff in improved forecast accuracy is well worth the extra effort.
Coming Next: Improving your forecasts with top-down models.
In the next installment of Forecasting 101 we will explore how top-down models can be used to improve upon automatic time series forecasts. These models leverage statistical forecasts generated for aggregated data to improve lower-level forecasts.
About the author:
Eric Stellwagen is Vice President and co-founder of Business Forecast Systems, Inc. (BFS) and co-author of the Forecast Pro software product line. He consults widely in the area of practical business forecasting—spending 20-30 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 Coca-Cola, Procter & Gamble, Merck, Blue Cross Blue Shield, Nabisco, Owens-Corning and Verizon, and has served on the board of directors of the International Institute of Forecasters (IIF).