A hybrid econometric—neural network modeling approach for sales forecasting

Abstract Business sales forecasting is an example of management decision making in an ill-structured, uncertain problem domain. Due to the dynamic complexities of both internal and external corporate environments, many firms resort to qualitative forecasting techniques. However, these qualitative techniques lack the structure and extrapolation capability of quantitative forecasting models, and forecasting inaccuracies typically lead to dramatic disturbances in production planning. This paper presents the development of a hybrid econometric-neural network model for forecasting total monthly sales. This model attempts to integrate the structural characteristics of econometric models with the non-linear pattern recognition features of neural networks to create a “hybrid” modeling approach. A three-stage model is created that attempts to sequentially “filter” forecasts where the output from one stage becomes part of the input to the next stage. The forecasts from each of the individual sub-models are then “averaged” to compute the hybrid forecast. Model development is discussed in the content of an actual sales forecasting problem from a Danish company that produces consumer goods. Actual model performance is reported for a six-month time period. Knowledge gained from the modeling approach is placed in the context of organizational learning about the nature of sales forecasting for this particular company.