PREVISÃO DE SÉRIES TEMPORAIS EM UMA INDÚSTRIA METAL MECÂNICA UTILIZANDO MÉTODO CLÁSSICO DE BOX-JENKINS E REDES NEURAIS ARTIFICIAIS MLP.

Abstract

The demand forecasting is of essential importance for business environments, in a way to serve as a decision making supporting tool during the development of companies strategic planning. This work strived to compare statistics with artificial intelligence methods applied to provisioning on demand issues using temporal series through Box-Jenkins and Artificial Neural Networks Multilayer Perceptron (MLP) methods. Studies were performed to identify and define the main demand forecasting methods. Subsequently, the selected prediction methods for the analysis of the three most relevant products of a metalworking industry were applied in the period 2012 to 2014. The four last periods were used only for performance validation of both methods, through the analysis of forecast errors. Softwares R, Matlab and SPSS supported the data deployment, modeling and analysis. From those models, a step ahead provisioning of sales of a metal mechanic industry was performed, followed by the comparison of the errors of each method based on root mean squared error, RMSE, and mean absolute percentage error, MAPE, to identify the most satisfactory and adequate provisioning method. The results indicated that the performance of the forecasts using the statistical method of Box-Jenkins in Products 1 and 3 were higher than the application of the MLP neural network models. While, for Product 2 the method of neural networks achieved better results. In the statistics analysis, one could verify that the series present some behavior patterns associated to seasonality and oscillations, being possible to observe that both methods show satisfactory results for each data characteristics of the temporal series.