Composite Forecasting in Demand Planning
The majority of Demand Planning installations use Univariate Forecasting. Even when Multiple Linear Regression is used, its equation may not be complete enough to fully model the nature of the demand. This blog discusses how Composite Forecasting can be used to blend Univariate and MLR forecasts to get a more accurate overall forecast.
Definitions of Forecast Profiles available in SCM APO Demand Planning (definitions from help.sap.com)
Definition of Univariate ForecastingUnivariate forecasting provides methods that allow you to forecast the following time series patterns. You use the history of a key figure, for example: historical sales volumes, to detect a pattern and then use this pattern to predict the future values of this pattern: – Constant – demand varies very little from a stable mean value – Trend – demand falls or rises constantly over a long period of time with only occasional deviations – Seasonal demand – periodically recurring peaks and troughs differ significantly from a stable mean value – Seasonal trend – periodically recurring peaks and troughs, but with a continual increase or decrease in the mean value – Intermittent demand – demand is sporadic – No change from previous year – no forecast is carried out; instead, the system copies the actual data from the previous year Definition of Causal Forecasting (using MLR)One of three categories of models used for forecasting. The other two are time series models and judgmental models. The basic premise of a causal model is that the future sales of a particular product or service are closely associated with changes in some other variable(s). Therefore, once the nature of that association or relationship is quantified, information about that other variable(s) can be used to develop a demand forecast. For example, you can gauge what price point you need to hit in order to reach a particular sales volume. That is because when the price of an item goes up, generally the demand for it goes down. The causal model used in Demand Planning is multiple linear regression (MLR). Definition of composite forecastingThis function combines forecasts from alternative forecasting methods (such as times series, causal, and/or judgmental) for a particular brand, product family or product. Each forecast is based on the same historical data but uses a different technique. The underlying objective is to take advantage of the strengths of each method to create a single “one number” forecast. Either you can average the forecasts giving each one equal weight, or you can weight each one differently, or you can vary the weightings of each forecast over time. By combining the forecasts, the business analyst’s objective is to develop the best forecast possible. The composite forecasts of several mathematical and/or judgmental methods have been proven to out-perform the individual forecasts of any of those methods used to generate the composite.
Notes about the Composite Profile
– Management by exception using the Alert Monitor is not possible with composite forecasting. You can still create your own macro alerts however. – The Standard mode of Composite forecasting can be used to blend different univariate and MLR forecasts. The other modes are selection of model that has the lowest error (MAD, MAPE, etc.). The example in this blog only addresses the use of the standard mode of composite forecasting.
When to use univariate, MLR, and composite profiles
You can use univariate models when historical patterns are a good indicator of the future. Use MLR when events or factors that influenced sales volumes of the past can be modeled, are recorded, and these factors are known or can be estimated in the future Use composite when you need a blend of these methods. Composite works especially well when there are historical factors that cannot be modeled neatly by causal variables.
Real world example
A certain consumer goods producer had been performing univariate forecasting with another tool prior to implementing Demand Planning in APO. They therefore had the parameters (alpha, beta, gamma, etc.) that could be used to create univariate profiles. It was apparent, however, that their previous models and those to be created in APO would not adequately explain the nature of demand for some products. Some products were extremely sensitive to two casual factors: increased volume of package (temporary promotions) and price changes. When the absolute value (actual price and total volume) of these variables were used as causal factors, more of the behavior of the demand was explained. However, the two variables were not enough to explain all variation in sales volume. For many products a blend of MLR and univariate models were found to be a better indicator than either one alone.
How it was modeled
Initially, using intrinsic knowledge of the planners, composite models were developed using for certain products using their univariate models and an MLR model that used the variables mentioned above. It was found that some manual tweaking to set the percentage of each model to use yielded the best results.
How to measure accuracy/goodness of composite profiles
The system does not have an automatic way to measure the accuracy of these derived composite models. However, with macros one could calculate errors and monitor forecast accuracy. Alternatively, BW can be used to develop reports that can compare forecast accuracy of the composite forecast to other calculated forecast; the consensus forecast and actual volumes after sales become history.
Composite forecasting is one more tool in the toolset of the demand planner. The demand planner must have a good understanding of the nature of demand of the company’s products to make the best use of it. The system initially will not “know” which combinations and ratios of univariate and MLR results will yield the best overall results. After empirical data is collected, the planner can begin to refine the forecast profiles and use the parameters that use the best result. However, there is some art to knowing how to apply the science. When planners understand their products and the different forecasting models that are being used, composite forecasting can produce a very favorable result.