This topic looks at practical issues associated with the selection of the best Statistical Forecasting model for any individual item. For full information on all Trend Models available in IFP, see Trend Models in Statistical Forecasting.
Choice of Forecasting Model
In IFP, the best model to be used for forecasting purposes includes consideration of the following elements:
- Basic shape of the underlying trend
- Seasonal model to be used
- Relevant period of data to be used
- Level of detail to use in the creation of forecast models
- Adjustment of forecasts to reflect special factors for future periods
Automatic Selection of Models
Some other applications claim to select the best forecasting model by automatically searching for the trend model that best fits all available historical data. This is misleading in practice, as the business objective is to produce forecasts that will be close as possible to actual sales for a range of future periods.
A trend model that best fits available historical data will often be a poor choice for forecasting purposes. This is particularly the case for more complex models which use a large number of parameters.
In practice, the best forecasting models are created by combining statistical analysis of available data with the use of expert knowledge of product managers.
Any mathematical analysis of trends and seasonality will be of limited value if the historical data being used has not been previously cleaned to remove the effects of known unusual events. For example, actual sales data are often distorted by out of stock situations. See Data Preparation and Cleaning for details.
Levels of Detail
Practical demand forecasting usually involves the creation of forecasts for groups of similar items, e.g. multiple presentations (SKUs) for a brand. Data may also be available in more detail, e.g. by customer and/or region/country. An important decision in such situations is to select which level(s) of detail of data should be used for analysing trends.
For a particular product group, you may take the view that trends in demand for each individual SKU should be analysed separately and hence should have their own Statistical Forecasting model.
Alternatively, you may take the view that it is better to analyse trends in total combined volumes for the product group. This will often be a good idea if actual demand for each SKU contains large amounts of random variation with substantial switching between SKUs from month to month. It is usually the case that total demand for a product group contains relatively less random fluctuations from month to month, hence it is easier to forecast with reasonable accuracy. Of course, if you take this view then you will need to split total forecast demand for each product group into contributions for each SKU. This is usually best done by using historical average percentage contributions of each SKU to the product group total.
Weighting of Data
You should consider whether data in the selected period should be weighted according to its age when fitting selected trend models. The use of weighting is important when you believe that the parameters of your selected trend model (e.g. level or growth rate) may be changing over time. Clearly, if this is the case then more recent data is more relevant to forecasting future demand than older data.
In IFP, this weighting may be introduced by using a suitable discount factor in conjunction with weighted average or weighted linear trend models. The lower the discount factor the more emphasis is put on recent demand data when estimating the future path of the trend. See Weighted Estimation of Trends in Statistical Forecasting for further details.
Product Life Cycle
The shape of the underlying trend for any product is strongly related to where the product is in its product life cycle. Suitable trend models for each stage are often as follows:
- The Introduction phase often shows an exponential trend.
- The Growth stage will usually show a generally linear trend.
- As products approach Maturity, a Modified Exponential or S-Curve usually describes the underlying trend.
- After that a period of no growth will usually follow, where weighted averages are best for measuring trend values.
- As the product enters decline, an exponential trend will usually be the best model for forecasting purposes.
Best Fit Issues
Clearly, changes in the shape of trend during the product life cycle mean that the trend model that best fits available historical data to date is a poor indicator of the model that will be best for forecasting purposes.
For example, in the introduction phase an exponential trend will be a good fit to historical data available at that time. However, growth will soon start to become more linear and forecasts from an exponential trend will vastly overestimate reality.
Similarly, in the growth phase a linear trend will often be a good fit to historical data available at that time. However, growth will eventually start to flatten and forecasts from a linear trend will again vastly overestimate reality.
You will need to decide how you wish to model Seasonality. There are two main options here:
- Item Specific: Use this option if you believe that the underlying seasonality for this item is unique,
- Group Based: Use this option if you feel that the underlying seasonality is similar to other items in the same product group.
See Seasonal Factor Estimation for more details.
Special Factors in the Forecast Period
It is important to remember that Statistical Forecasting models on their own can only suggest future levels of monthly demand assuming that past trends and seasonality continue into the future. Hence, final demand forecasts must include adjustments to reflect any special factors that may arise in future periods which have not been present in historical demand. For example, adjustments may be required for the following types of special factor:
- Changes in direct selling activity
- Special offers (e.g. buy one, get one free)
- Price changes
- New indications for the use of a product
- New competition
- New products and possible cannibalisation effects
- Customer inventory adjustments
- Effects of supply constraints
To see some examples of appropriate Trend Models being applied in action, please see [video]: Statistical Forecasting Trend Models in IFP.