Demand forecasting models in companies are often a mix of quantitative models and qualitative methods. As there are so many existing forecasting approaches, many forecasters have difficulty in deciding on which model to select as they may perform “best” in a specific error measure, and not in another. Currently, there is no approach that evaluates different model classes and several interdependent error measures simultaneously, making forecasting model selection particularly difficult when error measures yield conflicting results.
This paper proposes a novel procedure of multi-criteria evaluation of demand forecasting models, simultaneously considering several error measures and their interdependencies based on a two-stage multi-criteria decision-making approach. Analytical Network Process combined with the Technique for Order of Preference by Similarity to Ideal Solution (ANP-TOPSIS) is developed, evaluated and validated through an implementation case of a plastic bag manufacturer.
The results show that the approach identifies the best forecasting model when considering many error measures, even in the presence of conflicting error measures. Furthermore, considering the interdependence between error measures is essential to determine their relative importance for the final ranking calculation.
The paper's contribution is a novel multi-criteria approach to evaluate multiclass demand forecasting models and select the best model, considering several interdependent error measures simultaneously, which is lacking in the literature. The work helps structuring decision making in forecasting and avoiding the selection of inappropriate or “worse” forecasting model.
