Demand forecasting is one of the key processes for enterprises that manufacture, distribute and sell material products. In the face of strong market competition, the demand for goods needs to be covered immediately. Therefore, is necessary to always maintain a certain inventory level. However, storage entails different kinds of costs – from logistic-related to depreciation due to physical and moral aging of products. The question is: How many goods a company should have to meet clients’ needs and at the same time not to freeze too much funds in inventories?
As demand forecasting is determined based on the analysis of time series, it is essential for big companies to operate a well-designed and efficient ERP system. Statistical models may be grouped in terms of their purposes as follows:
- Short-term forecasting without any trends (e.g. the naïve model, the arithmetic mean, the weighted moving average, Brown’s exponential smoothing)
- Short-term forecasting including trends (Holt’s exponential smoothing)
- Mid-term forecasting without seasonality (the classical method of least squares)
- Mid-term forecasting including seasonality (Winters’ exponential smoothing).
The manner of grouping presented above is only a suggestion. The selection of a final model is much more complex and involves many different factors such as: a standard (weekends) or non-standard (holidays) plan of non-working days, promotions, sales cannibalism as a result of similar products entering the market, innovations, demand shocks, sales values (rounding small fractions to integers, etc).
Under laboratory conditions, the effectiveness of the model is determined using the sum of root-mean-square errors (RMSE). In practice, RMSE minimisation may or may not be an optimum method. Depending on its profile and competition, an enterprise has to decide which of the threats is greater: the potential loss of a client or maintaining an extensive inventory.
Therefore, the selection of the model always needs to be decided individually for each enterprise.