Abstract:

The paper presents results of studies aimed at the determination of the possibility of demand prediction based on its sales variability in the past. For a group of selected products, the coefficient of variation in subsequent weeks was calculated and introduced as an input to a Nonlinear Autoregressive Neural Network model. The neural network model predicted future demand on the basis of what it has learned guided by historical data input to it. The quality of the match was determined using the coefficient of the residual sum, calculated as the sum of residuals is the distance between the actual and the predicted value and by means of the mean squared error value. In the analyses, the impact of the type of learning function and the size of the delay vector, which are parameters of the neural network, were investigated. The type of function which guarantees the best quality of the predictive model was determined.