Combinação de previsões por métodos lineares e não-lineares: uma aplicação em séries industriais

Abstract

Forecasting is an important activity in planning, process control and decision making, especially in the industrial context, and the combination of forecasts is an attractive approach in order to obtain accurate results. The objective of the present research was to investigate the application and combination of prediction methods aiming to verify the accuracy and the gains in terms of error reduction for time series of the industrial sector. The research is structured in two articles: article I discusses a systematic review of the literature on the development and applications of forecasting models in industrial processes. In this study, the Web of Science, Scopus and IEE databases were searched, composing a portfolio of 354 articles published in scientific journals in the last 10 years. The analysis of the literature was carried out in three stages: (i) initially an analysis of the frequencies of the publications was carried out regarding the periodicals, years of publications, authors, countries and number of citations; (ii) analysis of cocitation, bibliographic coupling and similarity analysis of the co-occurrence of the terms in the studies; (iii) finally, a unified framework was developed to classify the applications of forecast methods in industrial processes. Article II addressed the application of the combination of forecasts in a case study conducted in a large mining and logistics company with industrial production series from an integrated port system, in this study the general objective was to fit a model of accurate combined forecasting capable of capturing the serial time behavior of the system. The models of Exponential Smoothing, ARIMA modeling from the Box-Jenkins methodology and the Artificial Neural Networks models were used as individual predictors. The combination of the forecasts was carried out by three different approaches: the combination by arithmetic mean, by the Minimum Variance method, consisting of a linear combination from the variance of the prediction errors, and from Copula models, being a nonlinear approach based on the degree of dependence on individual forecasts. As evaluation measures of the proposed models, the RMSE and U-Theil criteria were used. The results showed that the ARIMA and ANN models were superior in terms of accuracy in relation to the individual predictions, and the methods of combination by model of copula produced more accurate predictions in relation to the other approaches of combinations adopted.