NEURAL NETWORKS AS A MEANS OF PRECISE DAY -TO-DAY SALES PREDICTION FOR MILK-PRODUCTS

Within the framework of tactical planning in dairy enterprises precise day-to-day sales prediction is an important means. Precision of forec asting of the models on the basis of (linear) regression available so far is unsatisfactory. One reason might be the constraint immanent in the system to have to commit oneself to the functional form of the rel ationship between potential factors of influence and forecasting value s without economic justification. This problem can be avoided by using models on the basis of neural networks. It was, therefore, investigat ed whether use of Multi-Layer-Feed Forward networks lead to better for ecasting results. In the test these networks with two hidden layers an d weight calculation on the basis of the back propagation algorithm ga ve forecasting errors which were lower by 4 to 36 per cent compared wi th the results obtained using regression. Although the computational e xpenditure in terms of time has still to be considerably reduced by ha rd-and software improvements until use on a practical scale will be po ssible and in the case of forecasting errors of 20 % of the mean sales quantities of individual products in special shops further improvemen ts are required, the latter have, however, mainly to be achieved via t he determination of more suited parameters of influence on sales quant ities. From the methodical viewpoint neural networks are at present th e best available tool of sales prediction.