How to create a statistical forecast model in IBP
December 2017
Statistical forecast models should be used for medium-term and long-term demand planning. Without a forecast there is a limited basis on which to decide on what to manufacture, what to ship and what to stock. Forecasting in IBP is used for the short term, the medium term and the long term – known as demand sensing. New product introductions can be handled via phase-in, and old products can be handled via phase-out curves and decisions, which are all important means to best estimate your sales. This blog will focus on the medium-term and long-term forecasting by using statistical forecast models.
When you create a forecast model, you define what algorithm(s) to be used and what the input and output key figures should be. This means that you decide how the forecast should be calculated, from where it should read the historical data and to where it should store the forecast results. However, the forecast model is only the container of a functionality. In the forecast model itself, you cannot define when it should be run or at what aggregation level it should run on.
The forecast model consists of three steps: a preprocessing step, a forecasting step and a postprocessing step. You decide on which steps are necessary for the forecast model.
It is possible to clean your historical data by using the key figure logic, which is described in blog post 1. Then you can define how, for example, negative values should be treated or have a manual overwrite key figure. But you can also use the algorithm in the forecast model to do outlier correction and substitute missing values.
For the outlier correction, you define what outlier detection method should be used and how the outliers should be substituted by defining the outlier correction method. Our recommendation is that it is important for the demand planners to understand the output of the forecast model, and therefore the demand planner must also understand the consequences of the different outlier correction methods, if it is decided to use automated cleaning of historical data.
Plenty of forecasting algorithms are available in IBP. The most important thing to bear in mind when deciding on what forecast algorithm to choose is that just because a forecast model fits very well with historical data there is no guarantee that it will fit future demand. Instead of drowning in complexity, analysing and maintaining lots of forecast models, we believe that spending the time on getting the right input from the right people adds more value than changing the forecast models every fortnight.
A single exponential smoothing with a low alpha coefficient creates a constant future forecast with the alpha value controlling the reactiveness of the model. Seasonality cannot be ignored neither, so if you expect seasonality, this should be applied in your model as well. Finally, for step changes and long-term trends, it is required to get input from both Sales and Marketing and build a logic in key figures to accommodate for these changes.
In the postprocessing step, you can choose which error measures should be used to evaluate your forecast. The total forecast error is composed of three different elements: demand variation, bias and forecast variation. Separation of these elements and a structured approach are key to reduce the forecast error.
When a forecast model is created, you can assign planning combinations to the model or run the model using planning filters. This is how you combine your planning combinations with your forecast models. A forecast model can both be executed from the Excel front-end or setup as a recurring job in the Web UI.
When you execute the model, you define the aggregation level. Should the forecast model be run on a detailed level for product, location and customer combination, or should it be run on a product family and region level? If the aggregation level is more aggregated than the base planning level of the input key figure, it will use the key figures’ aggregation mode to aggregate the data before the calculation is done.
When the job has finished, you can see the forecast in your selected output key figure for the model. After the run, you have created the baseline for your forecast, and further information from Sales and Marketing can be added in other key figures to finalise your demand planning.