An application of learning machines to sales forecasting under promotions

This paper deals with sales forecasting in retail stores of large distribution. For several years statistical methods such as ARIMA and Exponential Smoothing have been used to this aim. However the statistical methods could fail if high irregularity of sales are present, as happens in case of promotions, because they are not well suited to model the nonlinear behaviors of the sales process. In the last years new methods based on Learning Machines are being employed for forecasting problems. These methods realize universal approximators of non linear functions, thus resulting more able to model complex nonlinear phenomena. The paper proposes an assessment of the use of Learning Machines for sales forecasting under promotions, and a comparison with the statistical methods, making reference to two real world cases. The learning machines have been trained using several configuration of input attributes, to point out the importance of a suitable inputs selection.