Método bagging para aprimoramento de previsões de séries temporais

Abstract

Different methodologies are proposed and explored aiming to reduce time series forecasting error. A promising approach consists in combining different forecasts from different models in order to get a better accuracy, i.e., a smaller forecast error. This work aims to review and apply the bootstrap aggregating method, also known as bagging, in order to improve time series forecasting. First, each time series is divided into training and testing time series, and then the moving block bootstrap methodology is applied to the training series to generate different resampled time series, and then forecasting for each one of the series is performed and combined, thus obtaining the final combined forecast. The test data set is used to calculate the accuracy of the models, individual and combined. A simulation study of time series and application to a real time series data sets are presented. The chosen and fitted model for each of the time series was an autoregressive integrated moving average (ARIMA). The accuracy measurements used were the mean square error and its root, mean arctangent absolute percentage error and the symmetric mean absolute percentage error. Finally, the impact on the forecasts of the combined model by varying the resampling method parameters was explored and comparisons between the combined and individual forecasting methods were also carried out.