It has been shown that feedforward multilayer perceptron networks can learn to approximate the time series of sales in supermarkets. For a special group of articles neural networks have been trained to forecast future demands on the basis of the past data. To improve the prediction quality we use additional price and advertising information. Thereby the prediction accuracy is sufficient.
The time consuming back-propagation learning has been parallelized and so accelerated significantly. The necessary training for prediction has been reduced to an acceptable value.
Another approach in order to reduce training time is to minimize the number of input neurons. By correlation analysis we want to find out only the relevant time series that have to be taken into consideration.
For the future the modelling of the input vectors should be improved in order to minimize the prediction error: especially season and holiday information have to be given to the net; the value of changing prices can be modelled quantitatively.
The aim of our research is to develop a forecasting system for supermarkets. This system will reduce stock-keeping costs by flexible adaptability to changing circumstances.