It is easy to understand why most Forecast Pro users rely heavily on the software’s expert selection mode to generate their forecasts–expert selection automatically analyzes their data, selects the appropriate forecasting techniques and generates the forecasts. This article discusses how the length of a data set can influence your decision to choose expert selection or an alternative approach for forecasting your data.
Models Included in Expert Selection Mode
Forecast Pro’s expert selection mode only considers time series methods which are forecasting techniques that generate a forecast solely based upon the item’s demand history. Time series methods considered in expert selection mode include:
- different forms of exponential smoothing;
- Box-Jenkins models;
- the Croston’s model;
- discrete data models; and
- moving averages.
Methods not considered in expert selection mode include:
- event models (which provide adjustments for promotions, business interruptions or other irregular occurrences);
- dynamic regression models (which allow you to incorporate explanatory variables):
- top-down approaches (which allow you to use aggregate-level forecasts to improve lower-level forecasts);
- new product models (forecast by analogy and the Bass diffusion model); and
- curve fitting.
If you are forecasting a data set that calls for one of these methods, expert selection should not be used regardless of the length of the data set.
Data Considerations with Expert Selection Mode
The amount of historical data available often has a direct impact on the forecasting model expert selection chooses. In this installment of Tips & Tricks, we will examine different data length scenarios and consider their impact on expert selection’s forecasting approach as well as suggest alternative approaches you may wish to consider. We then discuss cases where you might consider discarding some of the historical data.
Forecasting with up to four months of demand history
In general, expert selection is unlikely to yield an accurate forecast when you have fewer than five data points. There is simply not enough history to determine the statistical properties of the data and answer questions such as, “are the data trended?” and “are the data seasonal?”
When there are no data points, expert selection will set the forecast to zero. If there are between one and four data points, expert selection will use a moving average.
In these cases you should consider alternative approaches. If you are forecasting replacement products or product line extensions, you might consider mapping product histories or using top-down approaches. If you are forecasting new-to-company products, you might consider using Forecast by Analogy or Bass diffusion models.
Forecasting with more than four months but less than two years of demand history
Generating an accurate forecast for an item with less than two years’ worth of history can be challenging. If you have between five and 23 months of data, expert selection will consider a variety of both trended and non-trended models; however, it will not consider seasonal models. Until you have at least two observations for each month of the year, expert selection cannot automatically determine whether the data are seasonal. Thus, in cases where you have less than two years’ worth of data and you know the data to be seasonal, you should not use expert selection.
In these cases, appropriate alternatives include using top-down approaches, applying user-defined seasonal patterns via the weighting transform and judgmentally adjusting the forecasts generated by expert selection. If you have 17 or more months’ worth of data (one year plus five periods) you could also dictate that a seasonal exponential smoothing model be used. All of these approaches are fully supported in Forecast Pro.
Forecasting with two years or more demand history
Expert selection generally works quite well for items with two years or more of demand history. Having stated that, if you have between two and three years’ worth of seasonal data and the seasonality is not very pronounced, you will want to keep an eye on expert selection. If expert selection opts for a non-seasonal model for data that you know to be seasonal, you will want to override the selection and dictate that a seasonal method be used, such as Winters exponential smoothing. Remember, your knowledge of the data is important—to Forecast Pro it is simply a stream of numbers.
When to consider discarding data
Many forecasters are much too quick to throw away historical data. Most of the forecasting methods in Forecast Pro are adaptive and give the current data more emphasis than the more distant past. Retaining the older data is important to help establish the rate of change in the data.
On the other hand, there are times where discarding the earliest data is appropriate. These times include:
- When you have more than seven years of history. Expert selection uses time series methods only, and using more than seven years of history does not usually improve time series forecasts.
- When there are distinct start-up behaviors. Some product launches have very distinct start-up sales behaviors. For example, pharmaceutical drugs often have very high sales immediately after a drug is approved as pharmacies order initial inventories. After this initial “pipeline fill,” the demand pattern is driven by the pharmacies reordering the drug as customers fill individual prescriptions. Discarding the initial pipeline fill data makes sense if you are trying to forecast ongoing demand to fill prescriptions.
- When there is a radical change in the data. Some industries experience radical changes that render the older demand history irrelevant in terms of forecasting the future. For instance, you may want to discard data prior to the breakup of a monopoly or the advent of new technology which causes certain products to become obsolete.