There are many different approaches to creating a Statistical Forecast but the most common forms are:
- Moving Average - Smooths data to create a less volatile line. Smoothed past data can be used to create smoother forecasts.
- Exponential Smoothing - Similar to Moving Average but uses the forecast error to drive forecast direction.
- Regression - Uses at least one variable to predict another.
The advantage of the various Solutions that are available is that they can automate the analysis of hundreds of forecast types and their variations and then select the most appropriate. Two typical approaches are:
Best Fit: Apply the most appropriate from a selection (ODP)
Bayesian: Combine the best of many in one forecast (Demantra)
The results of a Statistical Forecast will differ depending on the Level that it has been created at. Generally lower is better but issues like short lifecycles and unique customer products can force grouping levels. If you set levels too low the variations in data might be large causing decreased accuracy. Statistical Forecasts generated from higher grouping levels can produce better results than using the lowest levels but inaccuracies can be magnified at aggregate levels! If you set levels to high, the forecast might pick up localised trends and allocate them incorrectly.
The Dimension Levels that you choose should relate to your company data and processes and the needs of the business.
The accuracy of the Statistical Forecast will depend on the length of History analysed and the horizon of Forecast that is created. 3 years of past data is enough to generate season trends but this requies product lifecycle stability or the creation of suitable levels to group like-items.